Paper deep dive

GUIDE: Computationally Augmented Consensus for Trustworthy Clinical Practice Guidelines

Yuanfan Lin, Jie Di, Xujin He, Zhenzhe Lin, Jialing Chen, Ruiyang Li, Changyu Zheng, Yuanrui Shang, Weiqin Peng, Jingxi Qin, Xiong Wang, Dongxin Ye, Yining Huang, Guoqiang Li, Andi Xu, Sijia Yu, Ruixin Wang, Yahan Yang, Xiaohang Wu, Duoru Lin, Jingjing Chen, Zhenzhen Liu, Wenben Chen, Haotian Lin

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 175

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Maintaining up-to-date clinical practice guidelines (CPGs) is essential for evidence-based medicine, but rapid evidence growth and complex expert consensus formation limit timely updates. We developed GUIDE, a multi-agent copilot system for human-AI collaborative guideline generation and updating. GUIDE delegates labor-intensive tasks to specialized AI agents while preserving expert judgment and final arbitration. In 12 NICE ophthalmology datasets, GUIDE achieved 88.85% literature recall, 86.06% alignment in certainty-of-evidence ratings, and 91.67% recommendation agreement, requiring 162.17 minutes per dataset on average (28.68% human intervention). Performance remained stable across 20 additional datasets from cardiology, endocrinology, gastroenterology, and respiratory medicine. Ablation analysis showed that removing expert intervention reduced performance in judgment-intensive steps such as Risk-of-Bias assessment. In six post-cutoff datasets, GUIDE maintained robust end-to-end performance on unseen evidence. Prospective real-world validation showed that GUIDE reduced expert workload and produced more clinically acceptable recommendations than the original decade-old guideline (mean score, 3.73 vs. 1.53; agreement, 66.7% vs. 13.3%), with outcomes comparable to human-only updating (P = 0.588) but substantially less time and effort. GUIDE offers a practical framework for timely, trustworthy, expert-centric guideline development and validates computationally augmented consensus.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

171,576 characters extracted from source content.

Expand or collapse full text

1 GUIDE:computationallyaugmentedconsensusfortrustworthy clinicalpracticeguidelines Coauthors:YuanfanLin 1† ;JieDi 1† ;XujinHe 1† ;ZhenzheLin 1† ;JialingChen 1 ;RuiyangLi 1 ; ChangyuZheng 2 ;YuanruiShang 1 ;WeiqinPeng 1 ;JingxiQin 2 ;XiongWang 2 ;DongxinYe 2 ; YiningHuang 1 ;GuoqiangLi 1 ;AndiXu 1 ;SijiaYu 2 ;RuixinWang 1 ;YahanYang 1 ;XiaohangWu 1 ; DuoruLin 1 ;JingjingChen 1 ;ZhenzhenLiu 1 ;WenbenChen 1* ;HaotianLin 1,3,4,5* Affiliations: 1 ZhongshanOphthalmicCenter,SunYat-senUniversity,WHOCollaboratingCentreforEyeCare andVision,StateKeyLaboratoryofOphthalmology,GuangdongProvincialKeyLaboratoryof OphthalmologyandVisualScience,GuangdongProvincialClinicalResearchCenterforOcular Diseases,Guangzhou,China; 2 ZhongshanSchoolofMedicine,SunYat-senUniversity,Guangzhou,China; 3 HainanEyeHospitalandKeyLaboratoryofOphthalmology,ZhongshanOphthalmicCenter,Sun Yat-senUniversity,Haikou,Hainan,China; 4 DepartmentofGeneticsandBiomedicalInformatics,ZhongshanSchoolofMedicine,Institute forFrontierInterdisciplinaryResearchinHealthSciencesandTechnology,SunYat-sen University,Guangzhou,China; 5 Leadcontact. †Theseauthorscontributedequallytothiswork. *Correspondingauthors: ProfessorHaotianLin,M.D.,Ph.D.,Email:linht5@mail.sysu.edu.cn.ZhongshanOphthalmic Center,SunYat-senUniversity,7#JinsuiRoad,Guangzhou,China,510623.Telephone: +86-20-87330493.Fax:+86-20-87333271. AssociateProfessorWenbenChen,M.D.,Ph.D.,Email:chenwb56@mail.sysu.edu.cn.Zhongshan OphthalmicCenter,SunYat-senUniversity,7#JinsuiRoad,Guangzhou,China,510623. Telephone:+86-20-66648760.Fax:+86-20-87333271. 2 Abstract Maintainingup-to-dateclinicalpracticeguidelines(CPGs)isessentialforevidence-based medicine(EBM),yetthecurrentprocessisconstrainedbytherapidexpansionofevidenceandthe complexityofexpertconsensusformation.Toaddressthischallenge,wedevelopedGUIDE,a multi-agentcopilotsystemthatenablescomputationallyaugmentedconsensusforhuman-AI collaborativeclinicalguidelinegenerationandupdating,assigninglabor-intensivetasksto specializedAIagentswhilepreservingcriticaljudgmentandfinalarbitrationforhumanexperts. Inabenchmarkcomprising12datasetsderivedfromNICEophthalmology-specificguidelines, GUIDEshowedstrongperformanceacrossevidenceretrieval,assessment,andsynthesis, achieving88.85%literaturerecall,86.06%alignmentincertainty-of-evidenceratings,and91.67% agreementonrecommendations,onlyrequiring162.17minutesperdatasetonaverage(including 28.68%humanintervention).Performanceremainedstableinadditional20datasetsspanning cardiology,endocrinology,gastroenterology,andrespiratorymedicine,indicatingits cross-specialtygeneralizability.Wethenconductedanablationanalysistovalidatetheimportance ofhuman-AIcollaborationmechanism,asremovingexpertinterventionreducedperformancein judgment-intensivestepssuchasRisk-of-Biasassessment.Furthermore,asimulatedend-to-end evaluationusingsixpost-cutoffdatasetsdemonstratedthatGUIDEmaintainedrobustperformance, supportingitsabilitytogeneralizetogenuinelyunseenevidencewithinacontinuous guideline-updatingworkflow.Finally,prospectivereal-worldvalidationinaliveguideline updatingprocessshowedthatGUIDEsubstantiallyalleviatedexpertworkloadacrossevidence selection,extraction,andsynthesis.TheGUIDE-assistedupdatewasjudgedmoreclinically acceptablethantheoriginalguidelinerecommendationissuedadecadeearlier(meanscore,3.73 vs.1.53;agreement,66.7%vs.13.3%)andachievedoutcomescomparabletothoseofhuman-only updating(P=0.588),whilerequiringsubstantiallylesstimeandexperteffort.Together,our findingssupportGUIDEasapracticalframeworkfortimely,trustworthy,andexpert-centric clinicalguidelinegeneration,andvalidatecomputationallyaugmentedconsensusasafeasible paradigmforenhancedguidelinedevelopment. 3 Introduction Clinicalpracticeguidelines(CPGs)constitutecentralinstrumentsofevidence-basedmedicine (EBM),designedtointegratethebestavailableevidencewithclinicalexpertisetosupport informedhealthcaredecision-making[1].Bytranslatingrapidlyevolvingknowledgeinto structuredevidencebodiesandactionablepracticerecommendations,CPGsplayacriticalrolein reducingunwarrantedpracticevariation,improvingcaredeliveryquality,andenhancingthe overallefficiencyandequityofhealthcaresystems[2].Theirvalue,however,dependsontheir timeliness.Recommendationsthatarenotupdatedinlinewithemergingevidencecanweaken theirclinicalvalidityandpotentiallyleadtosuboptimalcare[3].Conventionalguideline workflowsareincreasinglyunabletokeeppacewiththeexponentialgrowthofbiomedical knowledge[4].Withtypicalupdatecyclesexceedingtwoyears[5,6],thereisaninherentdelayin translatingnewevidenceintopracticerecommendations.Thisstrainisexacerbatedbytheannual additionofoveronemillioncitationsandagrowingburdenofsecondaryresearch[7],thescale andcomplexityofwhichfarexceedcurrentcapacitiesforefficientevidencesynthesis. Yetthedifficultyofguidelinegenerationorupdatingextendsbeyondscalealone.More importantly,itisaconsensus-buildingprocessthatrequiresexpertstocollectivelyinterpret uncertainevidence,weighmethodologicallimitations,balancebenefitsandharms,andreconcile clinical,economic,andsocietalconsiderationsintoaccountableconsensus.Foreachclinical question,expertsmustdeterminestudyeligibility,interpretheterogeneousfindingsand conclusionstoestimatecertaintyofevidence,andfinallytranslateevidenceintorecommendations inbroadmedicalandsocioeconomiccontext[8].Manyofthesestepscannotbefullydelineated bydeterministicrulesbutdependoninterpretationunderuncertainty,contextualtrade-offsand deliberationamongdomainexperts.Thesetasksarenotmerelytechnicaloperations,but judgment-intensiveprocessesthatrelyontime-consumingexpertdeliberationandconsensus formationtoensurethetrustworthinessofclinicalguidelines.Consequently,continuedrelianceon traditionalupdatingworkflowsmaywidenthegapbetweentherapidgrowthofnewevidenceand theresponsivenessofexistingguidelinesystems,ultimatelylimitingtheireffectiveness[9]. Recentstudieshighlightthepotentialoflargelanguagemodels(LLMs)tostreamline literaturescreeningandevidenceextractioninmedicalguidelinedevelopment[10,11],with 4 systemslikeQuickerextendingautomationtoend-to-endrecommendationgeneration[12]. Despitethesetechnicalstrides,real-worldclinicalintegrationremainschallenging.Crucial elementssuchastransparentevidencescrutiny,theincorporationofexpertclinicaljudgment,and accountableconsensusformationunderuncertaintyarestillinadequatelyaddressed[13].As probabilisticmodels,LLMsremainunreliableintasksrequiringimplicitinferenceand quantitativereasoning.Forexample,theymayfailtodetectmethodologicalbiasindicatedby missinginformation,extractnumericaloutcomesdispersedacrossareport,orassesstheoverall credibilityofatrial[14,15].Humanexpertsarebetterequippedtoevaluatemethodological nuancesandimplementationtrade-offs,buttheircapacityisincreasinglystrainedbytherapidly expandingevidencebase.Thiscreatesanurgentneedforexpert-facingAIsystemsthatenable real-timehuman-AIcollaboration[16].Suchsystemsshouldsupportcomputationallyaugmented consensus:LLMsenhancethespeed,breadth,andconsistencyofevidenceprocessing,while expertsretainresponsibilityforresolvinguncertaintyandarbitratinghigh-stakesdecisions, consistentwiththeemergingparadigmofAI-integratedmedicine[17].Thisalsonecessitatesa well-definedhuman-AIcollaborativeframeworkthatclarifieswhattasksAIcanperform,when expertjudgmentisrequired,andhowresponsibilityshouldbeallocatedacrossbothintermediate stepsandfinalrecommendations[18]. Here,weintroduceGUIDE(GuidelineUpdateIntelligenceDecisionEngine),anLLM-based copilotframeworkforcollaborativeclinicalguidelinegeneration,builtuponahuman-AI partnershipmodelderivedfromexistingresearchandDelphi-likeexpertspanelconsultations[19]. ThedesignstrategicallyallocatesrolestooptimizethestrengthsofbothAIandhumanexperts throughouttheguidelinedevelopmentprocess.WhileAIagentsexecutehigh-volume,rule-based taskssuchasclinicalquestiondecomposition,literaturescreening,dataextraction,andevidence synthesis,humanexpertsmaintainultimateaccountability,overseeingtheprocessand contributingmedicalinsightstoensuremethodologicalrigor.Thishuman-AIcollaborationis implementedintwokeymodes:expertsinteractreactively,addressingflaggedissuesraisedbyAI reviewers,andproactively,usingtheirprofessionaljudgmenttofine-tunesystemoutputs.To evaluateGUIDE,wefirstestablishedabenchmarkcomprising12datasetsderivedfromNICE (NationalInstituteforHealthandCareExcellence)ophthalmology-specificguidelines,confirming 5 thesystem’sfoundationalabilitytoperformkeytasksinguidelinegeneration.Wethenconducted anablationstudytoquantifytheimpactofhumaninvolvement,removingreviewermechanisms andexpertparticipation,revealingthecriticalroleofhumanjudgmentinenhancingsystem performance.Toassesstheframework’sgeneralizability,weextendedtestingacrossmultiple medicalspecialtiesandclinicalquestioncomplexitieson20datasets,demonstratingitsbroad applicability.Finally,weconductedtwocriticalexperimentstoassessGUIDE’sreal-world applicability.Weperformedtemporalhold-outsimulations,applyingGUIDEtosixdatasets releasedaftertheLLM’sknowledgecutoff,totestthesystem’sabilitytointegratenewevidence intoanend-to-endguidelinedevelopmentworkflowinrealtime.Furthermore,inaprospective real-worldvalidation,weexploredhowGUIDEintegratesemergingevidenceintoanexisting guidelineonintraocularlenses(IOLs)selectionforcataractsurgerypatients.Ourfindings demonstratethattrustworthyguidelinegenerationcanbefacilitatedthroughGUIDE,aframework forcomputationallyaugmentedconsensusthatsystematicallycouplesAI-mediatedevidence processingwithexpertdeliberation,therebypreservingexpertoversight,methodologicalrigor, andclinicalaccountabilitywhilesubstantiallyreducinghumanworkload. Results DevelopmentoftheGUIDEforclinicalguidelinegeneration GUIDEisamulti-agentcopilotsystemtofacilitateclinicalguidelinegenerationbyintegrating automatedefficiencywithclinicalrigorthroughastructuredhuman-AIpartnership (SupplementaryVideo1).Thefoundationallogicofthesystemisanchoredinasetof collaborationprinciplesestablishedviaaDelphi-likeconsensusamongamultidisciplinaryexpert panel.Centeredona“human-in-the-loop”architecture,GUIDEdelegateshigh-volume,rule-based taskstoAIagentswhileensuringthatcriticalprofessionaljudgmentandfinalvalidationremain firmlyunderhumanoversight.Thiscollaborativedesignisimplementedbyexpertsviatwomodes: aproactiveinterventionapproachtoensureclinicalaccountabilityincriticaljuncturesanda condition-triggeredreviewmechanismformonitoringautomatedtasksatpredefinedcheckpoints (SupplementaryTable1).Throughouttheguidelinegenerationworkflow,thisallocationis tailoredtothedemandforprofessionalexpertiseandriskprofilesofeachspecificstep.WhileAI 6 agentsprimarilysupportrelativelystructuredandstandardizedprocessessuchassearchstrategy formulationanddataextraction,humanexpertsleadtheinterpretiveandvalue-sensitivestages: activitiesincludingPICO(Population,Intervention,Comparison,andOutcome)construction, evidencequalityassessment,andthefinalsynthesisofrecommendationsarereservedfor proactivehumaninterventiontoensureclinicalrigorandconsensus(SupplementaryTable2). Totranslatetheseprinciplesintoexecutablesystemlogic,thesystemarchitecture incorporatesspecializedagentsthatoperationalizethehuman-AIinterface.Specifically,the proactiveinterventionapproachiscoordinatedbyanassistantagentthatparsesnatural-language expertinstructionsintoexecutablecommandsforspecializedagents.Thecondition-triggered reviewmechanismisimplementedthrougharule-basedrevieweragentthatmonitorssystem outputstotriggerreactivehumaninterventionwhensuspectedanomaliesarise.Byintegrating theseagent-drivenfunctions,GUIDEestablishesadynamicdivisionoflaborthatbalances computationalpowerwiththerigorousaccountabilityrequiredforreal-worldevidencesynthesis. Theframeworkoperatesthroughaseven-stepsequentialworkflowspanningfromquestion decompositiontoclinicalrecommendationgeneration(Fig.1a,SupplementaryVideo1).The operationprocessbeginswithPICOconstruction,whereGUIDEparsestheclinicalquestionand generatesastructuredPICOdraftthatcanbedirectlyusedforevidencemapping.Giventhatthe decomposedPICOfactorsdeterminetherecommendation’sclinicalscope,expertsmustverify themandmakenecessaryamendmentsatthisstep.BasedonthefinalizedPICO,GUIDE generatesastructuredreviewprotocoldetailingthereviewtype,eligibilitycriteriaandsynthesis frameworkforexpertvalidationandadjustmentbeforeretrieval(SupplementaryVideo1). Followingprotocolconfirmation,twosearchagentsdrivenbydifferentLLMsformulatesearch strategiesandperformparallelretrievalindependently,whiletheAIreviewermonitorssignalsof potentiallyunreliableretrieval,suchasmarkeddiscrepanciesinliteraturevolumeorlimited overlapbetweentwosearchoutcomesandelicitshumanintervention.Theassistantagent facilitatesexpertsinrevisingthestrategybytranslatingnatural-languageinstructionsintosearch strings,obviatingtheneedtodirectlymodifycomplexsyntax(SupplementaryFig.1). Subsequently,twoscreeningagentsconducttitle-and-abstractandfull-texteligibilityassessments sequentially,presentingrationalesandconfidencelevelsforeachinclusionorexclusiondecision 7 (SupplementaryFig.2).Atthefull-textscreeningstage,expertsmaybebroughtinwhentheAI reviewerdetectsanydisagreementintwoparallelroundsofscreeningandjudgment.Forincluded studies,twodataextractionagentsextracttextualandnumericaldataandgenerateformats correspondingtothePICOframework(SupplementaryFig.3).Aspreviouslydescribed,any discrepancyintwoparallelextractionsidentifiedbyAIReviewerisreferredtoexpertsfor verificationagainsttheoriginaltext.Comparableoutcomesfromverifiedstudy-leveldatawere mergedintoevidencebodiesforsynthesisandcertaintyassessment,withpooledeffectsizes calculatedbytheagentwhereappropriate.Thequalityassessmentagentsthenassessthese evidencebodiesusingthewidelyadoptedGradingofRecommendationsAssessment, DevelopmentandEvaluation(GRADE)methods,andexpertsauditagents’evaluationswhere divergentgradingarises.Expertscaneitherconveytheirinsightsviatheassistantagentordirectly overrideautomatedqualityassessmentresults.Finally,theevidencesynthesisand recommendationmodule,implementedasaDelphi-likemulti-agentmodule,integratesthe processedevidenceintoEvidence-to-Recommendation(EtR)materialsanddraftrecommendation statements,whicharethensubmittedtotheexpertpanelforfinaldeliberationandconsensus decision-making(SupplementaryFig.4).Asacopilotsystemthatextendsexpertcapabilities, GUIDEsynergizesautomatedefficiencywithstringentclinicalgovernance(Fig.1b).Itsreal-time interactiveinterfaceallowsexpertstoseamlesslyintegratetheirinsightsintotheworkflow, enhancingboththeprecisionandtheclinicalaccountabilityofthefinaloutputs. GUIDE’sbenchmarkperformance ToevaluatethebaselineperformanceofGUIDE,weconductedabenchmarktestusing12 specializeddatasetsderivedfromfourophthalmologyguidelinespublishedbyNICE.NICEwas selectedasitsguidelinesdevelopmentprocessishighlytransparentandtraceable.Eachofthese 12datasetsencompassesthecompleteevidence-basedworkflow,spanningfrominitialclinical questionformulationtothedevelopmentoffinalrecommendations(SupplementaryTable3).To isolatemoduleperformanceandmitigateerrorpropagation,wetestedtheframeworkusinga modularapproach,inwhicheachstepwasindependentlyconductedandevaluatedusingoriginal NICEworkflowmaterials(SupplementaryTable4and5). Thecollaborativeworkflowforeachdatasetinvolvedaseniorspecialistwithovertenyears 8 ofclinicalexperienceintherespectivedomain.Underthisconfiguration,GUIDEperformed effectively,enhancingefficiencyacrosstheentireprocess.InthePICOconstructionphase,the agentnear-instantaneouslygeneratedstructureddraftsinameanof0.37minutesperquestion.By havingexpertverifyingitsinterpretation,thesystemachievedameansemanticsimilarityscoreof 0.83(rangingfrom0.80to0.86).Across12benchmarktasks,theretrievalmoduleidentified 25,744recordsintotal,whichwerereducedto1,369aftertitle-and-abstractscreeningand135 finalinclusionsafterfull-textscreening.Highrecallwasmaintainedthroughoutthepipeline,with meanrecallvaluesof98.81%(95%ConfidenceInterval[CI],97.42%–100.00%),93.57%(95% CI,90.07%–96.96%),and88.85%(95%CI,85.22%–92.91%)forretrieval,title-and-abstract screening,andfull-textscreening,respectively.Precisionimprovedprogressivelyfrom0.47% (95%CI,0.33%-0.64%)atretrievalto8.11%(95%CI,6.02%–10.70%)aftertitle-and-abstract screeningand84.60%(95%CI,78.28%–90.61%)afterfull-textscreening(Fig.2a).Specificity andnegativepredictivevalue(NPV)were98.56%(95%CI,98.09%–99.05%)and98.81%(95% CI,98.27%–99.32%),respectively.Inretrievalstage,theAIreviewercorrectlytriggeredhuman interventionin4of12datasetswhenretrievalcountdiscrepanciesweredetected.Atthefull-text screeningstage,discrepantscreeningdecisionswereidentifiedfor22articles(accountingfor 16.30%ofallfull-textarticlesscreened),allofwhichwerereferredtoexpertsforarbitration. Amongthem,21studieswereconfirmedtohaveinclusion/exclusiondecisionsconsistentwiththe originalguidelines.Theliteraturesearchsteprequiredanaverageof9.26minutesperdatasetin total,comprising7.95minutesofautomatedsystemprocessingand1.31minutesofhuman intervention,whereastheliteraturescreeningsteprequired76.20minutesperdatasetintotal, comprising60.40minutesofautomatedsystemprocessingand15.80minutesofhuman intervention. Thedataextractionprocessdemonstratedhighefficiencyandprecision,requiringanaverage of28.20minutesperdataset,including15.89minutesofautomatedsystemprocessingand12.31 minutesofhumanrechecking.Comparedwiththeconventionalapproach,inwhichmanual extractionofdatafromasinglearticletypicallyrequiresmorethan40minutes[20],thisworkflow substantiallyimprovedextractionefficiencywhilemaintaininghighaccuracyof98.02%. Althoughthesystemhandledthemajorityoftheworkload,22.69%oftheextractiontasks 9 requiredexpertverification,withnumericalvaluesaccountingfor88.9%oftheverifieditemsand textualdataaccountingfortheremaining11.1%.Furtheranalysisshowsthattheoverallerrorrate of1.98%wasentirelyduetoomissionsofnumericaldata.Becausetherevieweragentwas designedtoflagonlymismatchesforexpertcorrection,theseomissionsdidnottriggeralertsand thereforeremainedundetected.Subsequentassessmentofthe300profilesaveraged10.24minutes perdataset,consistingof5.14minutesofautomatedprocessingand5.10minutesofexpert intervention.Theframeworkshowedstrongagreementwiththereferencestandard,achieving 82.80%(95%CI,79.57%–86.76%)forrisk-of-biasassessment,90.16%(95%CI, 85.48%–94.33%)forindirectness,88.84%(95%CI,83.47%–94.08%)forinconsistency,92.24% (95%CI,87.90%–96.35%)forimprecision,and86.06%(95%CI,81.38%–90.96%)forfinal certaintygrading.(Fig.2b).Finally,theDelphi-likemulti-agentmoduleintegratedtheseprocessed findingsintodraftrecommendationstatementswithin37.90minutesperquestion,achievinga consistencyrateof91.67%withtherecommendationsreportedintheoriginalguidelines(Fig.2c, SupplementaryTable6).TheaveragetimeexpenditureforbothGUIDEandhumanexpertswithin eachstepisillustratedinFig.2d.Overall,GUIDEaccountedfor71.32%ofthetotaltime(115.66 minutes),whileexpertoversightandrefinementonlycomprisedtheremaining28.68%(46.51 minutes). Ablationanalysisofexpertandmulti-agentcontributionstoperformance Toquantifyperformancegainsattributabletoexpertinvolvementandthemulti-agentsystem,we conductedanablationanalysisbyexcludingrecordsfromtheAIreviewer,AIassistantandhuman expertinthebenchmarktest,retainingonlyoutputscorrespondingtoafullyautomatedworkflow. AstheAIreviewerresponsibleforidentifyingdiscrepanciesacrossparallelagentswasnot considered,werestrictedtheanalysistoasingleGPT-4.1-drivenagentateachsteptoensure workflowconsistencyandcomparability.Weobservedthattheautomatedsystemaloneperformed wellonrelativelystraightforwardtasks,achievingaPICOsemanticsimilarityof0.79andamean literatureretrievalrecallof92.58%(Fig.3a).Acrossthe12tasks,theretrievalagentidentified 29,121recordsintotal,1,645recordsaftertitle-and-abstractscreeningand155finalinclusions, reflectingrelativelybroadeligibilitycriteriaintheabsenceofexpertinput.Notably,final-stage recalldecreasedfrom88.85%(95%CI,85.22%–92.91%)inthefullconfigurationto79.80% 10 (95%CI,74.15%–86.13%)intheablationanalysis(P=0.008),whilefinal-stageprecisionfell evenmorepronouncedfrom84.60%(95%CI,78.28%–90.61%)to66.09%(95%CI, 58.92%–73.61%)(P=0.010)(Fig.3b).Specificity(96.49%(95%CI,95.36%–97.60%))andNPV (98.24%(95%CI,97.43%–98.97%))werelargelypreservedathighlevels.Thissharpdecline indicatesthatwithoutmonitoringprovidedbytheAIreviewerandhumanexpert,single automatedagentstrugglestoformulateandnavigatecomprehensiveinclusion/exclusioncriteria, leadingtoaninfluxofconfoundinginformation,especiallyirrelevantarticles,thatmayultimately dilutethequalityofevidencesynthesisandrecommendationvalue.Dataextractionaccuracy droppedfrom98.02%to85.30%(P=0.046),whiletheomissionraterosefrom1.98%to14.7% (P=0.027).Duringtheevidenceassessmentphase,consistencyinriskofbiasjudgmentsdeclined significantlyfrom82.80%(95%CI,79.57%-86.76%)to69.54%(95%CI,61.15%-77.92%)(P= 0.002),withacorrespondingdeclineinconsistencyforoverallcertainty-of-evidenceratingsfrom 86.06%(95%CI,81.38%-90.96%)to68.78%(95%CI,62.72%-74.83%)(P<0.001)(Fig.3c). Detailederroneousanalysisrevealsthatagentsfelltodetectlatentmethodologicalflowsthat requireddeductivereasoningtoidentifytheabsenceofevidence,suchaslackrandomization details.Insuchcases,theagentfrequentlydefaultedtoa“lowrisk”judgmentforallocation concealmentbasedsolelyonthepresenceoftheterm“sealed,”whereasexpertscorrectly interpretedtheabsenceofthesedetailsasindicativeofpotentialbias.Intherecommendation generationstep,theproportionofdraftedrecommendationratedas“exactlyconsistent”withthe referencestandardsalsodecreasedfrom83.33%to50.00%(Fig.3d,SupplementaryTable6). Takentogether,thesefindingsshowthatLLMsmayexecutemanystructuredcomponentsof guidelineupdatingefficiently,butthereviewer-guidedhuman-AIcollaborativemechanismis criticalformaintainingclinicallyinterpretablerecommendationgenerationinthemost judgment-intensivestages. EvaluationofGUIDEgeneralizabilityacrossclinicalspecialties WefurtherevaluatedthegeneralizabilityofGUIDEacrossclinicalspecialtiesbytestingGUIDE on20datasetsderivedfromNICEguidelinesincardiology,endocrinology,gastroenterology,and respiratorymedicine,withfivedatasetsfromeachspecialty(SupplementaryTable7).Asenior domainspecialistwithovertenyearsofclinicalexperiencewasrecruitedforeachfieldtofunction 11 asacollaborativepartner.Acrossthesespecialties,GUIDEmaintainedrobustperformancedespite differencesindomain-specificterminologyandclinicalcontext.Thesystemachievedanaverage PICOsimilarityof0.85andliteratureretrievalrecallof98.88%(95%CI,97.71%–99.81%)(Fig. 4a).Thesystem’shigh-throughputefficiencyremainedconsistentduringstudyselection,where agentsprocessedatotalof38,317recordsand1,668full-textliterature.Despitetheheterogeneity ofcross-disciplinaryliterature,thecollaborativeframeworksustainedhighperformance,witha full-textscreeningrecallof88.83%(95%CI,85.45%–92.26%)andprecisionof80.11%(95%CI, 74.33%–85.90%)(Fig.4a).Thefinalscreeningyieldedaspecificityof97.24%(95%CI, 96.27%–98.14%)andanNPVof98.61%(95%CI,98.14%–99.07%).Thisstepinvolved51.79 minutesofagentprocessingandonly10.41minutesofhumanreview.Indataextractionfrom177 includedstudies,theoveralldataextractionaccuracywas81.13%,withanerrorratecausedby omissionof18.87%.Textualdataextractionreachedanaccuracyof89.91%,whereasnumerical dataextractionachievedanaccuracyof78.81%,consistentwithpreviousfindingsthatnumerical datahandingremainsachallengeforLLMs.Forevidencecertaintyassessment,thesystem achieved82.23%(95%CI,79.09%-85.55%)agreementinoverallqualitywiththereference standard,whilerequiringameanprocessingtimeof12.47minutesperarticle(8.2minutesof expertrecheck)(Fig.4b).Finally,themulti-agentmodulesynthesizedthesefindingsand,through theDelphi-likeconsensusprocess,generateddraftrecommendationstatementswithin44.0 minutesperdataset,achieving90.00%agreementinexpertevaluations(Fig.4c,Supplementary Table6).ThesefindingsindicatethatGUIDEgeneralizeswellacrossmultiplemedicalspecialties andprovidesabroadlyapplicableframeworkforguidelineupdating.Thetimeexpenditurefor bothAIagentsandhumanexpertswithineachstepisillustratedinFig.4d. End-to-endevaluationofGUIDEperformanceonpost-cutoffdatasets Toeliminatedataleakagerisksandassesstheholisticperformanceoftheframework,wecurated sixdatasetsfromNICEguidelinespublishedbetweenApril2025toMarch2026(detailsin Methods),simulatinganend-to-endprospectiveexperimentwithinacontinuousworkflow (SupplementaryTable8).Unlikemodulartestsinthebenchmark,thisend-to-endevaluation requiredGUIDEtoprocessrawoutputsfrompreviouscomponents,adesignwhereinthe frameworkmustwithstandreal-worlderrorpropagationtodemonstrateitstrueoperational 12 integrity. Theend-to-endpipelineachievedaPICOsemanticsimilarityof0.84andmaintainedstrong performancethroughliteraturescreening,withafinalscreeningrecallof87.75%(95%CI, 81.05%–94.44%),precisionof82.17%(95%CI,73.49%–91.06%),specificityof97.15%(95%CI, 95.95%–98.49%)andNPVof97.21%(95%CI,94.63%–99.35%)(Fig.5a).Whileminorshiftsin theevidencepoollimiteddirectcomparisonofintermediatemetrics,suchasevidencequality assessmentwithoriginalNICEreference,thecoreclinicalreasoningremainedconsistent. Subject-matterexpertevaluationconfirmedacomplete(100.00%)directionalagreementbetween GUIDE’soutputsandrecommendationsintheoriginalguidelines(Fig.5b,SupplementaryTable 6).Bysuccessfullyinterpretingunseenevidencetogenerateclinicallysoundrecommendations, GUIDEprovesitmovesbeyondreproducingembeddedcontent,highlightingitspotentialasa resilienttoolforthecontinuousevolutionofclinicalpractices. ProspectivecasestudyofGUIDEforcataractguidelineupdating WeconductedaprospectivevalidationstudyoftheupdateoftheNICE2017guidelineCataracts inadults:management(NG77),usingitsoriginalreviewquestion,“Whatistheoptimalstrategyto facilitatesimultaneousdistanceandnearvisionfollowingcataractsurgery?”,asacasestudy.Two seniorcataractspecialists,eachwithovertwentyyearsofclinicalexpertise,participatedinthe entireguideline-updatingprocess.Forstepsrequiringhumanjudgmentorintervention,thetwo expertsindependentlyassessedtheissueandresolveddisagreementsbyconsensus.A pre-designatedexpertthenoperatedGUIDEinterface,implementingonlytheagreedactions. IndefiningthePICOframework,the“Population”and“Comparator”criteriaremained largelyidenticaltotheoriginalversion.Notably,theAIsystemincorporatedExtendedDepthof Focus(EDOF)IOLs,acategoryofgrowingsignificanceinrecentclinicalresearch,intothe “Intervention”criteria.However,consistentwithpreviousobservationsregardingthelimitations ofautomatedsystemsincapturingsocioeconomicnuances,thesystemomittedresourceutilization andcostsfromthe“Outcomes”criteria,whichweremanuallyaddedbyexperts(Supplementary Information1). Initialsearchstrategiesgeneratedbythesystemyieldedanunmanageable9,740records, indicatinganoverlybroadsearchscope.Uponexperts’interventiontorestricttheinclusionto 13 RandomizedControlledTrials(RCTs)andsystematicreviews(followingtheoriginalguideline’s methods),thestrategieswererefinedandresultingin1,111records(1,107afterde-duplication). Followingparalleltitle-and-abstractscreening,56articlesunderwentfull-textassessment,yielding 23newinclusionspublishedaftertheoriginalguideline’scutoffdate.Furthermore,28studies identifiedfromsystematicreviewsearchesweresubjectedtoautomatedfull-textscreening,from whichfouradditionalstudieswereincluded.Theseprocessesculminatedinatotalof27eligible studiescomprisingtheupdatedbodyofevidence(SupplementaryInformation2and3).TheAI revieweridentifiedonlysix(6/56,11%)discrepanciesduringthesecondscreeningphase.These wereresolvedthroughexpertreview,whichconfirmedthatallsixshouldbeexcludedbasedon minormisalignmentswiththepredefinedPICOscope.Remarkably,filteringthe1,107retrieved recordsintothefinal27inclusionsstudiestookonly55minutes,including11minutesofhuman adjudication. Dataextractionwascomplicatedbytherelativeabundanceandnon-standardizedoutcome measures.Forexample,theoutcome“visualfunction”exhibitedsubstantialheterogeneityacross studies,encompassingmultipleassessmentinstruments,diversetestingconditions,andvarieddata formats.Theseincludedbinaryoutcomes(e.g.,spectacleindependence),continuousscalescores (e.g.,vision-relatedqualityoflifequestionnaires),continuousmeasurements(e.g.,contrast sensitivity),andperformance-basedmetrics(e.g.,reading-relatedtests).Despitesignificant variationsinnarrativephrasing,thedataextractionagentsrobustlycapturedcategoricaldata, includingstudydesign,interventiondetails,andcomplications.Conversely,whilenumericaldata wereoftenpresentedinmorestructuredtables,theAIreviewerflaggedseveralomissionsand inconsistencies.Fortheseanomalies,expertsverifiedandcorrectedthemagainsttheoriginal article.Ultimately,sevenIOLscomparisonsand21distinctoutcomecategoriesweresuccessfully extractedintostandardizeddatatableswithin38minutes,including15minutesofexpert involvement.Theagentautomaticallygroupedstudy-leveldataintocomparison-and outcome-specificevidencebodies,calculatedpooledeffectsizesusingappropriatesynthesis methods,andorganizedtheresultsintostructuredoutputsfordownstreamqualityassessment. Forevidencequalityassessment,thesystemperformancerevealeddistinctpatternsacross differentRisk-of-Biasdomains.Randomizationandblindingdomainsdemonstratedhigh 14 inter-raterconsistencybetweenthetwoassessmentagents,withonlyasingledisagreement identified(3.7%discrepancyrate).Incontrast,selectionandreportingbiaseswerepredominantly ratedas“unclear”duetoinsufficientreportingintheprimarystudies;onlytwostudiesprovided adequatedetailtowarranta“lowrisk”rating(Nodiscrepancyinthesetwodomains).Attrition biasemergedasthemostcontentiousdomain,withseveninconsistencies(25.9%)flaggedbyAI reviewerrequiringfurtheradjudication.Thesediscrepancieswereresolvedbyexpertsviewing relevantexcerptsfromsourcearticles.Followingthisrigorousreconciliationprocess,theoverall riskofbiasforallincludedstudywasappraisedasserious.Thesystemshowedmorerobust performanceapplyingGRADEcriteriaforinterventionstudies.Outof291totalqualityrating acrossevidencebodies,discrepanciesoccurredonlyintheindirectnessdomain(6cases,2.1%), and100%agreementwasobservedforimprecisionandinconsistencyowingtomoreobjective andquantitativelydefinedcriteria.Bycontrast,theseidentifieddiscrepanciesinindirectness typicallyinvolvedsophisticatedPICOinterpretationsthatdemandedcontextualunderstanding.All discrepancieswerecarefullyreviewedbytheexperts,whoprovidedthefinaladjudicatedratings (SupplementaryInformation4).Theautomatedassessmentprocesswascompletedwithin47 minutes,whileexperttookapproximately25minutestofinishjudgement. Uponcompletionoftheevidenceappraisal,avirtualDelphipanelcomprisinganAI committeechair,twoophthalmicpracticeagents,ahealthcareproviderrepresentativeagent,a commissioneragent,andavirtualpatienttogetherwithtwohumancataractspecialists,deliberated onthefindingsfromtheirrespectiveperspectivestoformulatestructuredrecommendations.From PICOformulationtorecommendationdevelopment,theentireprocesstook192minutes, including65minutesofexpertinvolvement.ByconstructingacomprehensiveEtRcontent,the virtualpanelintegratedevidencequalityandemergingtrendstobalanceclinicalefficacywithcost andsafetytrade-offs(SupplementaryInformation5).Thisevidencereviewsupportedatransition fromtheoriginal“donotoffer”statementtoamoresupportiverecommendationformultifocal andEDOFIOLs.Notably,theupdatedrecommendationadopted“consider”ratherthanmore definitivephrasing,reflectingashiftinrecommendationstrength.Thisindicatesthatwhilethese newerdesignsshowclearclinicalpromise,thefinalchoiceshouldremainhighlyindividualized anddependentonthepatient’sspecificvaluesandpreferences(SupplementaryInformation6). 15 Inparallel,thehuman-onlygroupfollowedthesameprespecifiedguideline-updating workflow,encompassingPICOrefinement,protocoldevelopment,literaturesearchingand screening,dataextractionandevidencesynthesis,evidencequalityassessment,andcompletionof theEtRframework.AsintheGUIDE-assistedgroup,theexpertsincorporatedEDOFIOLsinto theinterventionframeworkonthebasisofcontemporaryclinicalpracticeandemergingevidence (SupplementaryInformation7).Theliteraturesearchretrievedacomparable1,012records. Title-and-abstractscreeningrequiredatotalof33hours,followedby21hoursforfull-text eligibilityassessment.Oftheretrievedrecords,37articlesproceededtofull-textassessment,of which25ultimatelymettheeligibilitycriteria(SupplementaryInformation8and9).Compared withtheGUIDE-assistedgroup,fewerarticlesproceededtofull-textassessment,indicating greaterscreeningspecificityatthisstage.Dataextractionandquantitativesynthesiswasoneofthe mosttime-consumingcomponentsofthehuman-onlyworkflow,requiring76hourstoconstructa structured,long-formatdataextractiontable.ThesamesevenIOLcomparisoncategoriesand20 distinctoutcomecategorieswereextracted,andpooledeffectestimatesweresubsequently calculatedwhereappropriate.Evidencequalityassessmentrepresentedanotherlabor-intensive step,requiringanothereighthoursforindependentappraisalandconsensusresolutionof disagreements.Thedistributionofevidenceratingsandthereasonsfordowngradingwerebroadly consistentwiththoseobservedintheGUIDE-assistedgroup(SupplementaryInformation10).An eight-memberguidelinecommitteethenreviewedtheGRADEevidenceprofilesandrelevant benefits,harms,resourceuse,costsfactorstoformulatethedraftrecommendationoverthreedays (SupplementaryInformation11).Althoughphrasedslightlydifferently,theupdated recommendationsimilarlyshiftedfromastrongrecommendationagainstmultifocalIOLstoa moreconditionalrecommendationtoconsidertheiruse,consistentwiththeconclusionofthe GUIDE-assistedgroup.Overall,updatingthisreviewquestionthroughthehuman-onlyworkflow required23daysfrominitiationtocompletionwithalargelyconsistentpracticerecommendation (SupplementaryInformation12). Afterbothgroup’scompletionoftheirupdatingprocess,threerecommendations(theoriginal 2017NICEstatement,theGUIDE-assistdraftedversionandhuman-onlydevelopedversion)were presentedtoanexternalpanelof15ophthalmicexpertstoevaluateitsclinicalacceptability. 16 Expertswereblindedwithrecommendationoriginsandaskedtoratetheirlevelofagreement usingafive-pointLikertscalerangingfrom“stronglydisagree”to“stronglyagree”.Agreement wasdefinedasaratingofeither“agree”(4)or“stronglyagree”(5)(SupplementaryInformation 13). Themeanscoreswere1.53±1.06fortheoriginal2017NICEstatement,3.73±0.96forthe GUIDE-assistversion,and3.87±1.36forthehuman-onlyversion.BoththeGUIDE-assistand human-onlyversionsreceivedsignificantlyhigheragreementscoresthantheoriginal2017NICE statement(adjustedP=0.009andP=0.015),whereasnosignificantdifferencewasobserved betweentheGUIDE-assistandhuman-onlyversions(adjustedP=0.588).Agreementwas13.3%, 66.7%,and73.3%fortheoriginal2017NICE,GUIDE-assist,andhuman-onlyrecommendations, respectively,withasignificantoveralldifferenceacrossgroups(Cochran’sQ=11.23,P=0.0036), althoughpairwisecomparisonsdidnotremainsignificantaftercorrection,potentiallydueto limitedpowerfromthesmallsamplesize. Discussion Inthisstudy,wedevelopedandvalidatedGUIDE,ahuman-AIcollaborativeframeworkfor clinicalguidelinegenerationinrealisticsettings.TheGUIDEframeworksupportstheend-to-end processfromevidenceidentificationtorecommendationdraftingbyleveragingthe complementarystrengthsofLLMagentsandhumanexpertthroughastepwiseworkflow.The valueofGUIDEstemsnotonlyfromitsfunctionalautomationbutfromitsprincipledintegration intothepracticalworkflowsofguidelinesynthesis.Recognizingspecificdemandsofeachphase, atailoredapproachtohuman-AIcollaborationwasdesigned.InPICOconstruction,forinstance, thepriorityisdefiningaclinicallysharpscopeandselectingoutcomesthatbalancebenefitswith harms—aprocesswhereproactivehumanoversightisvitaltoensurerelevanceandinterpretability. Taskslikesearchformulation,screening,anddataextractionaredefinedbytheneedfortechnical precisionandhigh-volumeprocessing.Theseareidealforcondition-triggeredreview,whereLLM agentsmanagetheheavyliftingofrepetitiveoperations,whileexpertsinterveneonlywhenthe systemflagsuncertaintyorinconsistency.Finally,higher-ordersynthesisandfinaldeliberation demandholisticjudgmentsuchasintegratingrisk,cost,andpatientvalues,thatonlyexpertscan 17 provide.Thisconfigurationensuresthattheworkflowachievesmaximaloperationalefficiency withoutcompromisingtherigoroftheclinicalevidencesynthesis.Acrossbenchmarkevaluation, ablationanalysis,temporalhold-outtesting,andprospectivereal-worldvalidation,GUIDEproved tobearobustcollaborativepartner,assistingexpertsinconstructing,appraising,andsynthesizing evidencewithhighfidelityshieldingexpertsfromtheburdenofmanuallabor. PreviousstudieshaveshownthatLLMscansupportindividualcomponentsoftheprocess, suchasliteraturescreening,evidenceextraction,andqualityassessment,andmorerecentefforts suchasQuickerhavebeguntointegratethesecomponentsintoend-to-endworkflows[21-25].Yet, proceduralcontinuityalonedoesnotbyitselffulfilltherigorousdemandsofformalguideline development[26-28].Theessentialchallengeextendsbeyondtaskexecutiontothesynthesisof recommendationsthatpossessbothclinicalinterpretabilityandexpertconsensus[26].GUIDE wasdesignedtoaddressthisgapbydeeplyembeddingexpertoversightthroughouttheentire workflow.Throughareviewer-triggeredandcopilot-assistedcollaborationmechanism,experts engagebothreactivelytoresolveautomatedanomaliesandproactivelytoprovideguidanceat criticaldecisionjunctures.Thisdual-modeintegrationpreservescomputationalspeedwhile ensuringthatexpertdiscretionremainscentraltoguidelines’integrity.Thesefeaturesdistinguish GUIDEfromearliersystemsandsupportitsroleasapracticalframeworkforbringingLLM capabilitiesintoreal-worldproduction. Ourfindingsdemonstratethattheproposeddesignperformsrobustlyinmoduletesting,with itsefficacyfurthervalidatedthroughablationstudies.Inthemostlabor-intensivestagesof guidelinegenerationsuchasstudyselectionanddataextraction,thesystemachievedsignificant timesavingsbyautomatingpreliminarytasks.Furthermore,fortasksdemandingholisticjudgment andnuancedinterpretation,suchasriskofbiasassessmentandtheformulationoffinal recommendations,expertsleveragedtheirclinicalinsightstoinstructtheAIagentsorrevisetheir generations,ensuringtheoutputswerefinelycalibratedwithreal-worldconsiderations.These resultspositionGUIDEasmorethanatechnicaltoolforacceleratingguidelineproduction,instead representingaformofcomputationallyaugmentedconsensus,inwhichAI-mediatedevidence processingissystematicallycoupledwithexpertdeliberation.Byredistributinglaboracrossthe workflow,GUIDEreducestheamountofexperttimerequiredandredirectsscarceexpert 18 cognitiveresourcesawayfromlow-valuerepetitivetaskstowardcriticaljudgment,arbitration,and consensus-building.Thisreframingisparticularlyimportantinguidelinedevelopment,wherethe ultimatevalueofefficiencyliesinpreservingexpertattentionfordecisionsthatmostdirectly shapeclinicalvalidity. Importantly,wefurtherdemonstratethepracticalfeasibilityofthisparadigmthrough prospectivevalidationinreal-worldguidelineupdatingscenarios.Unlikeretrospective benchmarks,thissettingmorecloselyreflectsthecomplexity,variability,anduncertainty encounteredinrealguidelinedevelopment.Inthissetting,theindispensableroleofhumanexperts wasrepeatedlyreinforced.Beyondcorrectinganomaliesanddiscrepanciesarisingfromautomated processes,expertsremainedessentialfortasksrequiringarbitrationandcontextualjudgment, includingdefiningtheresearchscope,assessingevidencequality,interpretingoutcomes,and consideringimplementationfeasibilityinlightofbothtreatmenteffectsandbroaderpractical concerns.Suchtasksrequireclinicalexperience,normativereasoningandtheabilitytobalance competingconsiderationsthatremainbeyondthereachofcurrentLLM-basedautomation.Atthe sametime,thecopilotsystemsubstantiallyreducedtheburdenonexpertsbysupporting high-volumetaskssuchasevidenceidentification,screening,extraction,andpreliminarysynthesis [29-31].Thehuman-onlyarmfurtherhighlightedthesubstantialtimeandworkforcerequiredfor conventionalguidelineupdating.Againstthisbenchmark,theGUIDE-assistedworkflowachieved comparableevidencesynthesisandrecommendationupdatingwithmarkedlygreaterefficiency. Theoptimalmodel,therefore,isnotfullautomationbutexpert-AIcollaboration,inwhichLLMs providescalabilityandanalyticalsupport,whileexpertsensuremethodologicalrigor,clinical relevance,andcontextualappropriateness[32-36]. Severallimitationsofthisstudywarrantconsideration.First,whileourbenchmarkcovers multiplespecialties,itsscaleanddiversityremainmodest.Testingtheframeworkagainstmore complexorcontroversialclinicalquestionsisessentialtoconfirmitsgeneralizability.Second,the currentworkflowstilllacksgranularautomatedprocessinginevidencesynthesis.Specifically,it cannotyetautonomouslyextractprimarystudiesfromretrievedsystematicreviews,leavingthe finalscreeningofthesenestedreferencestohumanexperts.Third,thesystemfocusesprimarilyon clinicaloutcomesandlacksintegratedmodulesforeconomicevaluationandmodeling.Asaresult, 19 evidenceandperspectiveconcerningcost-effectivenessandresourceallocationstillrequires manualsynthesisbyexpertpanels.Furthermore,certainhigh-leveltaskssuchasrisk-of-bias assessmentandfinalrecommendationformulationremainsensitivetothereasoningnuancesof LLMs,occasionallyshowingvariabilitycomparedtohumanconsensus.Futureiterationsshould incorporatemoreexplicitdomain-specificcriteriatobridgethisgap.Additionally,thereviewer agent’sinterventionlogiccurrentlyreliesonstatic,predefinedtriggers;transitioningtoadaptive, learning-basedstrategiescouldmoreeffectivelybalanceoperationalefficiencywithquality control.Finally,althoughprospectivelyvalidated,broadermulti-centertrialsarenecessaryto evaluatethesystem’srobustnessandlong-termclinicalimpactacrossdiverseorganizational contexts. Inconclusion,thisstudydemonstratesthatwhileLLMsaloneareinsufficienttofully automateclinicalguidelinedevelopment,theirintegrationwithinastructured,human-supervised workflowcansubstantiallyenhancebothefficiencyandreliability.Byshiftingthefocusfrom automationtocollaboration,ourworkprovidesapracticalandscalableframeworkfor next-generationguidelinedevelopment.Morebroadly,thisparadigmmayrepresenta generalizableapproachforintegratingAIintootherhigh-stakes,knowledge-intensiveprocessesin medicine,wherebothcomputationalefficiencyandexpertjudgmentareindispensable. Methods Overallstudydesignanddatasets ToevaluatetheperformanceofGUIDEinclinicalguidelinegeneration,weconstructeddatasets fordifferentevaluationpurposesfrom13publiclyavailableclinicalguidelinespublishedbyNICE betweenOctober26,2017,andMarch31,2026(fordetailsofthedatasetcontents,see SupplementaryTable5).NICEguidelineswereselectedbecauseoftheirhighlytransparent documentation,whichprovidetraceablesearchstrategies,explicitinclusionandexclusioncriteria, anddetailedevidenceprofiles,witheachdatasetcorrespondingtothefullsetoforiginalmaterials forasinglereviewquestion.Wefirstestablished12datasetsderivedfromfourophthalmology guidelinescoveringdifferentsubspecialtiesforbenchmarktestingandablationstudy (SupplementaryTable3).Toassesscross-disciplinarygeneralizability,wefurtherincorporated20 20 datasetsderivedfrom6guidelinesacrossfourclinicalspecialties:cardiology,endocrinology, gastroenterology,andrespiratorymedicine(SupplementaryTable7).Toruleoutthepossibility thatthestrongperformanceofGUIDEwasdrivenbydataleakage,weconstructedsixadditional datasetsbasedonfourguidelinespublishedafterthetrainingcut-offdatesoftheLLMsusedin thisstudyandconductedend-to-endsimulatedprospectiveevaluations(SupplementaryTable8). Foreachdataset,thecorrespondingoriginalmaterials,includingPICOelements,searchstrategies, studyselectionrecords,dataextractiontables,evidencequalityassessments,andfinal recommendations,wereextractedasgroundtruthformodularevaluation(SupplementaryTable4). Inaddition,areal-worldprospectivestudywasdesignedbyidentifyingaclinicalquestionfroma 2017ophthalmologyguideline(NG77:Cataractsinadults:management)inwhichthe recommendationhadbeenremovedforfurtherconsideration,therebyenablingapractical evaluationofthesysteminarealguidelineupdatingscenario. Developmentofhuman-AIcollaborationprinciples Toestablishaframeworkforhuman-AIcollaborationinmedicalguidelinedevelopment,a multidisciplinarypanelof10expertswasconvened(fourophthalmologistswithexpertisein evidence-basedophthalmicpracticeandclinicaldecision-making,threeEBMmethodologists specializinginsystematicreviews,evidenceappraisal,andguidelinemethodology,andthree specialistswithexperienceinmedicalAI,LLMs,andclinicaldecision-supportsystems).We balancedthepanelsettingsacrossclinical,methodological,andtechnicaldomainstoensureboth scientificrigorandoperationalfeasibility.Weemployedathree-roundmodifiedDelphi-like processtodefineandrefinehuman-AIcollaborationprinciple[19].InRoundone,panelists independentlyreviewedthetraditionalmedicalguidelinedevelopmentworkflowandidentified initialareasofAIinvolvementregardingefficiencyandquality.Roundtwofocusedonrefining roleboundariesbyevaluatinganonymizedfeedbackfromthepreviousstage,withspecific emphasisonaccountability,andhumaninterventionpoints.Roundthreeinvolvedafocused reviewtoresolveremainingcontesteditemsandfinalizetheconsensus-basedsuggestions. Throughouttheprocess,responseanonymizationwasstrictlymaintainedtomitigatehierarchical biasandfacilitateexpertconvergence.Theresultingsuggestionsprovidedvaluableguidancefor thedesignofoursystem(SupplementaryTable1and2). 21 EngineeringdesignofGUIDEframework Wedevelopedauser-friendlyframeworkforhuman-AIcollaborationinclinicalguideline generation(SupplementaryVideo1).TheframeworkcomprisesmultipleLLM-basedagents,each configuredtoperformaspecificstepintheguidelinedevelopmentworkflow,anAIreviewer (ChatGPT-4.1)auditsoutputsandidentifieswhenhumaninterventionisrequired,anAIassistant (ChatGPT-4.1)enablesexpertstointeractwiththeagentsthroughnatural-languageinstructions, andmedicalexpertswhocontributewhenexpertjudgmentisrequired.Fordifferenttask requirements,wecustomizedstandardizedprompttemplatesforeachagent,includingrole definitions,detailedtaskinstructions,andspecificationsforoutputcontentandformat.During promptdevelopment,thetemplateswereiterativelyrefinedanddebuggedusingfiveadditional datasets(derivedfromNG242:Diabeticretinopathy:managementandmonitoring,NG82: Age-relatedmaculardegeneration:diagnosisandmanagement,NG98:Hearinglossinadults: assessmentandmanagement,NG257:Fertilityproblems:assessmentandtreatmentandNG250: Pneumonia:diagnosisandmanagement)thatwerenotincludedinthesubsequentevaluations. Thesedatasetswereusedformultipleroundsofdevelopmentandsmall-scaletestingtoensurethe reliabilityoftaskexecution. Theframeworkwasdesignedtofollowthecoreworkflowofguidelinedevelopment, spanningevidenceretrieval,evidenceassessment,andfinalrecommendationformulation.The processinitiateswithPICOconstruction,whereaChatGPT-4.1-drivenagentdecomposeseach clinicalquestionintostructuredPICOcomponents.Uponexpertvalidationofthesecomponents, theagentthengeneratesadraftreviewprotocoloutliningkeymethodologicalparameters,suchas studyobjectives,reviewtype,eligibilitycriteria,andtheplannedevidencesynthesisframework (SupplementaryVideo1).Thisprotocolundergoesfurtherexpertrefinementpriortoliterature retrieval.Toaccommodatecomplexclinicalquestions,thesystem’sinteractiveinterfaceallows expertstomanuallyappendadditionalcontextualelementsorbroaderconsiderations. Fromliteratureretrievaltoevidenceassessment,weadoptadual-agentparalleldesign,in whichtwoindependentagentsusingdifferentLLMmodels(ChatGPT-4.1andDeepSeek-R1) operateconcurrently,withhumaninterventiontriggeredunderpredefinedconditions.Inthe literatureretrievalstage,twoagentsindependentlygeneratesearchstringsbasedonthereview 22 protocol.Thesequeriesareusedtosearchmajorbiomedicalliteraturedatabases(PubMedvia NCBIE-utilities,andScopusviatheElsevierAPI),forretrievingarticlemetadata,includingtitles, abstracts,authorlists,publicationdates,andjournalinformation.TheAIreviewerjudgesthe consistencybypreviewingandcomparingthesetsofarticlesfetchedbythedifferentsearch stringsgeneratedbytwoagents,andtriggersexpertinterventionwhensubstantialdiscrepancies aredetected,definedaseitherarelativedifferenceofmorethan30%inthetotalnumberof retrievedrecordsorinsufficientoverlapbetweenthetworetrievalsets(indicatedbyaJaccard similaritycoefficientoflessthan0.5).Wheneitherconditionismet,humanexpertsarealerted andcaneitherdirectlyrevisethesearchstrategyorprovidenatural-languageinstructionsfor optimization.Theseinstructionsarethentranslatedbytheassistantagentintoexecutablesearch syntax,allowingrefinementoftheretrievalstrategywithoutrequiringexpertstomanually manipulatecomplexquerylogic.Tomaximizerecall,theresultsfrombothretrievalroundsare mergedtoformthecandidatesetfordownstreamscreening.Tobalancerecallandscreening efficiency,studyselectionisconductedintwostages:title-and-abstractscreeningfollowedby full-textscreening.Inthefirststage,thetwoindependentagentsevaluatethetitleandabstractof eachretrievedrecordagainstinclusionandexclusioncriteriaderivedfromthePICOelements,and eacharticleisclassifiedaseitherIncludeorExclude.Theunionofincludedrecordsretainedby thetwoagentsisthencarriedforwardtofull-textscreening.Forfulltextacquisition,GUIDE integratesopenaccessrepositories,includingSemanticScholar,OpenAlex,andCrossreftolocate freelyavailablePDFlinks.Forcopyright-protectedfulltextsobtainedthroughauthorizedchannels, GUIDEalsosupportsmanualPDFupload.GUIDEsupportsbuilt-inopticalcharacterrecognition (OCR)toextractmachinereadabletext,eliminatingmanualtranscription.Full-texteligibility assessmentislikewiseperformedindependentlybythetwoagentsandmonitoredbytheAI Revieweragent.Whendiscordantdecisionsoccur,therevieweragenttriggersexpertintervention. Screeningresultsaredisplayedinatabularinterface,allowinghumanexpertstoinspectthe AI-generateddecisionsandrationalesforeacharticleand,whennecessary,revisethe classification.Thefinalsetofincludedstudiesconstitutesthebodyofevidence. Theextractedtextwasindependentlyanalyzedbythedual-agentsfordataextraction, includingstudydesign,populationcharacteristics,interventionandcomparatordetails,sample 23 sizeforeachcomparisongroup,outcomemeasures,andcorrespondingresultvalues.These extractedfieldsfromeacharticlewereorganizedintostructuredevidencetables.When discrepanciesaroseforthesamedataelement,theAIreviewerinitiatedexpertintervention,and theexpertsverifiedtheoriginaltextandcorrectedtheextractedinformationasneeded.Underthe GRADEframework,eachclinicaloutcomecorrespondedtoabodyofevidence,anditsquality assessmentincorporatedbotheffectestimatesandcertaintyofevidence.Toestimateeffects, study-leveldatawereautomaticallycategorizedintocomparison-andoutcome-levelevidence bodies.Clinicalexpertsevaluatedtheextracteddataforconceptualandclinicalhomogeneityto determinewhetherquantitativesynthesiswasappropriate.Wherenecessary,theyrefinedthe automatedcollectionsbyexcludingorsegregatingstudiesthatinappropriateforpooling.Foreach eligibleevidencebody,thesystemselectedanappropriatepre-specifiedmeta-analyticapproach accordingtotheoutcometype,effectmeasure,reviewdesign,andcompletenessoftheavailable numericaldata.Thedatarequiredtocalculatepooledeffectestimateswerethentransformedinto structuredtabularinputs,includingstudyidentifiers,interventionandcomparatorgroups,outcome data,andrelevantvarianceinformationwhereavailable.TheseinputswerepassedtoRscripts throughaPython-basedcomputationalutility,whichinvokedtheexecutionofthestatistical analysesviasystemcommands.Aftercomputation,thepipelineextractedthepooledeffect estimatesandassociatedstatisticaloutputsandreformattedtheresultsintostructuredevidence tablessuitableforsubsequentGRADEassessment.CertaintywasevaluatedacrossRiskofbias, Indirectness,Imprecision,andInconsistency.Riskofbiaswasassessedatthestudylevelforeach outcomeusingmethodologicalassessmenttoolsmatchedtostudydesign.Forincludedstudies,the correspondinginstrumentswereapplied,includingtheCochraneRiskofBiastool(RoBorRoB2), ROBINS-I,andtheQUADAS-2checklist[37-40].Indirectnesswasassessedaccordingtothe degreeofalignmentbetweenthestudypopulation,intervention,comparator,andoutcomesand thetargetclinicalquestiondefinedbythePICOframework.Imprecisionwasevaluatedprimarily onthebasisofthewidthofconfidenceintervalsandsamplesize.Inconsistencywasassessedat thebody-of-evidencelevelbyexaminingthevariabilityofresultsacrossstudiesreportingthe sameoutcome.Thesedomain-specificassessmentsarethensynthesizedandassignanoverall certaintyratingforeachoutcomeaccordingtoGRADEupgradinganddowngradingrules.When 24 thetwoparallelagentsproduceddiscordantcertaintyratings,theAIreviewertriggeredexpert adjudication. ThefinalevidencesynthesiswasperformedbyaDelphi-likemulti-agentmodule(allagents inthismoduleweredrivenbyChatGPT-4.1model)thatintegratedtheaccumulatedevidenceand qualityassessmentintocomprehensiveEtRmaterialsandpreliminaryrecommendationdraft.To convenethevirtualcommittee,humanexpertsdynamicallyselectedAIparticipantsfromeight predefinedspecialistroles:ClinicalPracticeExpert,ResearchMethodologyExpert,Research EthicsandQualityExpert,EvidenceSynthesisExpert,HealthEconomicsExpert,Patientand PublicRepresentative,HealthPolicyandImplementationExpert,andBiostatistician.Beyond theseestablishedroles,expertsmightintroducenovelrolesbyprovidingbriefprofessional backgroundsandroledefinition,theseinputsweresubsequentlyrefinedbytheAICommittee Chairintooptimizedsystemprompts.Oncethepanelwasconstitutedbasedonthespecific decisioncontext,theAICommitteeChairprovidedeachpanelistwiththecompleteevidencebase. EachAIpanelistthengeneratedanindependent,role-specificassessment.Finally,theAI CommitteeChairsynthesizedtheseevaluations,identifiedpointsofconvergenceanddivergence, anddeterminedwhetherthepredefinedconsensuscriteriahadbeenmet.Oncethisconsensus criterionwasmet,thedeliberationprocesswasconcluded.Ifconsensuswasnotreached,theChair generatedtargetedfeedbackforeachpanelistandinitiatedasubsequentroundofdeliberation. Throughoutthisprocess,humanexpertsobservedthediscussioninrealtime,reviewedeach panelist’sdecisionandreasoning,andcouldprovidesteeringinstructionsbetweenroundsto ensurethatthedeliberationremainedmethodologicallyrigorousandclinicallygrounded.Notably, inourretrospectiveexperimentalsettings,humanexpertsdidnotparticipateinthisfinal deliberationsteptoavoidpotentialbiasfromtheirpriorknowledgeoftheoriginalguideline recommendations. Implementationofmodulartest Toassessthesystem’sperformanceateachstageandmitigateerrorpropagation,weadopteda modulartestingapproach.Thisapproachwasappliedacrossthebenchmarktest,ablationstudy, andcross-specialtygeneralizationtest.Inallthreesettings,individualcomponentswereevaluated inisolationbyinitializingintermediateresultsfromthegold-standardguidelines,enablingadirect 25 comparisonbetweenthesystem’soutputsandthereferencematerial.Inthebenchmarkand cross-specialtygeneralizationtests,eachdatasetwaspairedwithaseniordomainspecialistwho hadovertenyearsofclinicalexperienceintherelevantfieldandservedasthehuman collaborativepartner. Specifically,clinicalquestionsderivedfromguidelineswereprovidedasinputsforPICO construction.Theexpert-verifiedPICOoutputswerecomparedagainstthereferenceusing BidirectionalEncoderRepresentationsfromTransformersScore(BERTScore)toevaluate semanticsimilarity.Forliteratureretrieval,originalPICOelementsandeligibilitycriteriaderived fromprotocolutilizedassearchinputs.Toensurearigorousevaluationofthesystem’srecall,the temporalupperboundoftheliteratureretrievalperiodwasmanuallysettothepublicationdateof thesourceguideline.Followingthesearchstrategyreportedinthesourceguideline,atwo-stage screeningprocess(title-and-abstractscreeningfollowedbyfull-textscreening)wasconducted. Screeningperformancewasevaluatedusingprecision(positivepredictivevalue[PPV]),recall (sensitivity),negativepredictivevalue(NPV),andspecificity.Thesemetricswerecalculated usingthefollowingformulas: 푃푟푒푐�푃= 푇푃 푇푃+퐹푃 ，�푒푐�(푆푒�)= 푇푃 푇푃+퐹 푒푔�푒푝푟푒푑�푐�푒�푒푃= 푇 푇+퐹 ，푆푝푒푐�푐�= 푇 푇+퐹푃 Here,TPdenotesstudiesselectedinagreementwiththereference,FPdenotesstudiesselectedbut notincludedinthereference,FNdenotesstudiesincludedinthereferencebutnotselected,and TNdenotesstudiesexcludedinagreementwiththereference.Fordataextraction,full-textPDFs oftheincludedstudiesandthecorrespondingPICOframeworkswereprovided.Theaccuracyof theextractionwasvalidatedagainstthedatapointsspecifiedintheGRADEtables.Quality assessmentwasperformedbasedontheextractedoutcomesandtheeffectsizescalculatedfor eachPICOelement.FollowingtheGRADEmethodology,riskofbias,inconsistency,indirectness, andimprecisionwereassessedindividuallytodeterminetheoverallcertaintyoftheevidence. Finally,afterconsensusreachedbyDelphi-likedeliberation,thesynthesizedrecommendations wereevaluatedintothreelevelsofconsistency(exactlyconsistent,partiallyconsistent, inconsistent)relativetotheoriginalguidelines(SupplementaryTable6). Implementationofend-to-endexperimentsusingpost-cutoffdatasets 26 BeforeapplyingGUIDEinreal-worldguidelineupdatingpractice,toruleoutthepotentialrisk thatitsstrongperformancewasattributabletodataleakagearisingfrompriorLLMsexposureto guidelinescontentduringtraining,weconstructeddatasetsfromguidelinespublishedafterthe releasedatesoftheLLMsusedinGUIDE(ChatGPT-4.1(April14,2025)andDeepSeek-R1 (January20,2025))andusedthemforanend-to-endsimulatedprospectiveevaluation.As previouslydescribed,eachdatasetwaspairedwithaseniordomainspecialistwhohadoverten yearsofclinicalexperienceinthecorrespondingfieldandservedasthehumancollaborative partner.Unlikethemodularevaluationwherestepswereinitializedwithgold-standardinputs,this simulationrequiredeachsteptoprocessthedirectgenerationsofitspredecessor.Theprocess commencedwiththeclinicalquestionsasthesoleinput,fromwhichthedeterminedPICO frameworkswerepassedtotheretrievalstage.Theretrievalliteraturepoolwasthenfedintoa two-stagescreeningprocess.Allsubsequentstages,includingdataextractionandevidencequality assessment,wereperformedbasedexclusivelyonthestudiesselectedintheprecedingsteps,and thefinalrecommendationsweregeneratedrelyingentirelyontheevidencesynthesizedwithinthis specificrun. Designandimplementationoftheprospectivecasestudy WeprospectivelyevaluatedtheGUIDEhuman–AIcollaborativeframeworkinareal-worldcase studybasedontheNICE2017guidelineCataractsinadults:management(NG77).Bytargeting thereviewquestion“Whatistheoptimalstrategytofacilitatesimultaneousdistanceandnear visionfollowingcataractsurgery?”--atopicthathadbeenofficiallyprioritizedforupdatingon theNICEwebsite.Thiscasestudyaimedtodemonstratethesystem’sabilitytointegratemodern clinicaldataintoaformalupdatingprocess.Twoseniorcataractspecialists,eachwithovertwenty yearsofclinicalexpertise,participatedintheentireguideline-updatingprocess.Forsteps requiringhumanjudgmentorintervention,thetwoexpertsindependentlyassessedtheissueand discussedanydisagreementsuntilaconsensusdecisionwasreached.Oneofthetwoexpertswas pre-designatedbeforetheexperimentasthesystemoperatorandwasresponsibleforalloperations performedwithinGUIDE,implementingonlytheactionsagreeduponbybothexperts. InparallelwiththeGUIDEhuman-AIcollaborativeteam,weconvenedanindependent human-onlyguideline-updatingcommittee.Itsevidencereviewgroupcomprisedtwoadditional 27 seniorcataractspecialists(eachwithovertwentyyearsofclinicalexperience)andonetechnical analyst.Thisgroupwasresponsibleforevidenceretrieval,qualityassessment,anddataprocessing, ultimatelyproducingthebodiesofevidenceandtheircorrespondingqualityratings.Throughout theprocess,memberswereprohibitedfromusingGUIDEsystem.Thecoredecision-making committeecomprisedonecommitteechair,fourcataractspecialists(includingtwoexperts participatedintheprecedingevidencesynthesissteps),onehealthcareproviderrepresentative,one commissioner,andonelaymember.Afterreviewingthebodiesofevidence,thecommitteeuseda modifiedthree-roundDelphiprocesstocompleteanEtRtableandformulateadraft recommendationaddressingthesamereviewquestion. Tovalidatetheclinicalrelevanceoftherecommendationsgeneratedthroughthetwo updatingapproaches,anexternalconsensusprocesswasinitiatedwithabroadercohortof15 seniorphysicianswithover10yearsofclinicalexperienceincataractcare.Theobjectiveofthis surveywastoevaluatetheacceptabilityofthreerecommendationstatements:theoriginal2017 recommendation,therecommendationdraftedthroughtheGUIDEhuman-AIcollaborative process,andtherecommendationdevelopedthroughthehuman-onlyprocess.Theexternal reviewerswerenotinvolvedinanypartofeitherupdatingprocessandwerefullyblindedtothe sourceofthethreerecommendationstatements.Tominimizepotentialpreferentialevaluation basedonpresentationsequence,thethreestatementswerepresentedinarandomizedorderacross reviewers.Usingafive-pointLikertscalerangingfrom“stronglydisagree”(1)to“stronglyagree” (5),theexternalreviewersindependentlyassessedeachrecommendation.Theserattingwere dichotomizedasapproval(aratingof4or5)ornon-approval(aratingof1–3). Statisticalanalysis Inthebenchmarktest,ablationanalysis,generalizationtestandend-to-endexperiments,system performancewasevaluatedagainstguideline-derivedgoldstandards.Forclinicalquestionparsing, semanticsimilaritybetweengeneratedandreferencePICOelementswasassessedusing BERTScore.Retrievalstrategygenerationandliteraturescreeningwereevaluatedusingprecision (PPV),recall(sensitivity),NPV,andspecificity.Forevidenceextractionandqualityassessment, categoricaloutputswereevaluatedusingclassificationaccuracy,whereastheaccuracyof quantitativeresultswasassessedbytheirdeviationfromreferencevalues.Resultsarereportedas 28 meanswith95%CIs,calculatedusinganon-parametricbootstrapapproachwith1,000,000 resamples.Pairedt-testswereusedforapproximatelynormallydistributeddifferences,and Wilcoxonsigned-ranktestswereusedotherwise.Approvalforthreerecommendationstatements wasreportedasthenumberandproportionofapprovingreviewers.Overalldifferencesin approvalwereassessedusingCochran’sQtest.Iftheomnibustestwassignificant,pairwise comparisonswereperformedusingexactMcNemartestswithHolmadjustment.Alltestswere two-sided,withP < 0.05consideredstatisticallysignificant.Allstatisticalanalyseswere performedinR(v.4.2.3). Dataavailability Thepubliclyavailableguidelinedocumentsusedinthisstudyinclude:Diabeticretinopathy: managementandmonitoring(NICEguidelineReferencenumber:NG242),Cataractsinadults: management(NICEguidelineReferencenumber:NG77),Age-relatedmaculardegeneration (NICEguidelineReferencenumber:NG82),Glaucoma:diagnosisandmanagement(NICE guidelineReferencenumber:NG81),Chronicheartfailureinadults(NICEguidelineReference number:NG106),Adrenalinsufficiency:identificationandmanagement(NICEguideline Referencenumber:NG243),Cirrhosisinover16s:assessmentandmanagement(NICEguideline Referencenumber:NG50),Barrett’soesophagusandstage1adenocarcinoma(NICEguideline Referencenumber:NG231),Pneumonia:diagnosisandmanagement(NICEguidelineReference number:NG250),Asthma:diagnosisandmonitoringofasthmainadults,childrenandyoung people(NICEguidelineReferencenumber:NG80),Falls:assessmentandpreventioninolder peopleandpeople50andoverathigherrisk(NICEguidelineReferencenumber:NG249), Fertilityproblems:assessmentandtreatment(NICEguidelineReferencenumber:NG257), Suspectedsepsisinpeopleaged16orover:recognition,assessmentandearlymanagement(NICE guidelineReferencenumber:NG253),Hearinglossinadults:assessmentandmanagement(NICE guidelineReferencenumber:NG98).Alloftheseguidelinematerialscanbefoundin https://w.nice.org.uk/guidance. Codeavailability GUIDEisanLLM-basedcopilotsystemdesignedtosupportexpert-facingclinicalpractice guidelinegenerationandupdating.WearenotmakingthesourcecodeofGUIDEpublicly 29 availableatthisstageowingtothesafetyimplicationsofunmonitoreduseofsuchasystemin medicalguidelineworkflows.Inappropriatedeploymentoutsideexpert-supervisedsettingscould leadtomisuseinevidenceappraisal,recommendationformulation,orclinicaldecision-support contextsbeyondthevalidatedscopeofthisstudy.Intheinterestofresponsibleinnovation,wewill continuetoworkwithclinicalexperts,guidelinemethodologists,institutionalpartners,and relevantstakeholderstofurthervalidateGUIDEandexploresafeonwarduses.Tofacilitatepeer review,wehaveprovidedthesourcecode,testingdataset,anddemonstrationvideoas supplementarymaterialsfortheeditorsandexpertreviewersinthesubmissionsystem.The projectoverviewanddeploymentinstructionsareavailablethroughGitHub (https://github.com/KimHe-RGB/GUIDE_INTRO).Additionaldetailsmaybemadeavailableby thecorrespondingauthoruponreasonablerequest,subjecttoinstitutionalpoliciesandapproval procedures. Acknowledgments WegratefullyacknowledgetheGuangdongBasicResearchCenterofExcellenceforMajor BlindingEyeDiseasesPreventionandTreatmentforitssupportofthisstudy.Wealsosincerely thankallparticipantsfortheirvaluablecontributions.WegratefullyacknowledgeZhenjun,Tu, YunyuLian,LuyuLiu,KechenWang,YulinZheng,andQingqingZhaofortheirvaluable feedbackonthedesignanduseoftheGUIDEcopilotsystem.Thisstudywasfundedbythe NationalNaturalScienceFoundationofChina(92368205and82441003toH.L.;82401313to R.L.),GuangdongBasicandAppliedBasicResearchFoundation(2024A1515012001toW.C.; 2025A1515011616toR.L.),ChinaPostdoctoralScienceFoundation(2023M744033toR.L.),the ChinaPostdoctoralScienceFoundationSpecialFund(R.L.),thePostdoctoralFellowshipProgram ofCPSF(GZB20230896toR.L.),GuangdongBasicResearchCenterofExcellenceforMajor BlindingEyeDiseasesPreventionandTreatment(W.C.),theScienceandTechnology DevelopmentFund,MacauSAR(FileNo.0010/2023/AKPtoH.L.),Basicscientificresearch projectsofSunYat-senUniversity(23ykcxqt002toH.L.),theScienceandTechnologyPlanning ProjectsofGuangdongProvince(2018B010109008toH.L.),HainanProvinceClinicalMedical Center(H.L.).“3510”MedicalInnovationLeadershipProgram,FundamentalResearchFundsfor 30 theCentralUniversities,SunYat-senUniversity(20263510YXON012toW.C.). AuthorContributions H.L.andW.C.contributedtotheconceptofthestudyanddesignedtheresearch;Y.L,X.H.,J.D., Z.L.,J.C.,R.L.,C.Z.,Y.S.,W.P.,J.Q.,X.W.,D.Y.,G.L.,A.X.,andS.Y.collectedthedata;W.C., Y.L,X.H.,J.D.,Z.L.,J.C.,W.P.,andG.L.conductedthestudy;Y.L,X.H.,J.D.,andJ.Q.analyzed thedata;W.C.,Y.L,X.H.,J.D.,Z.L.,andC.Z.draftedthemanuscript;H.L.,W.C.,Y.Lin,X.H., J.D.,Z.L.,R.L.,C.Z.,Y.S.,J.Q.,andS.Y.criticallyrevisedthemanuscript;H.L.,X.H.,Z.L.,and W.P.performedthetechnicalreview;andallauthorsdiscussedtheresultsandprovidedcomments regardingthemanuscript. Competinginterests Theauthorsdeclarenoconflictsofinterest. 31 References 1.CommitteeonStandardsforDevelopingTrustworthyClinicalPracticeGuidelines,Boardon HealthCareServices,InstituteofMedicine.ClinicalPracticeGuidelinesWeCanTrust. (GrahamR,MancherM,WolmanDM,GreenfieldS,SteinbergE,eds.).NationalAcademies Press;2011:13058.doi:10.17226/13058 2.LugtenbergM,BurgersJS,WestertGP.Effectsofevidence-basedclinicalpractice guidelinesonqualityofcare:asystematicreview.QualityandSafetyinHealthCare. 2009;18(5):385-392.doi:10.1136/qshc.2008.028043 3.ShekellePG,OrtizE,RhodesS,etal.ValidityoftheAgencyforHealthcareResearchand QualityClinicalPracticeGuidelines:HowQuicklyDoGuidelinesBecomeOutdated?JAMA. 2001;286(12):1461.doi:10.1001/jama.286.12.1461 4.BastianH,GlasziouP,ChalmersI.Seventy-FiveTrialsandElevenSystematicReviewsa Day:HowWillWeEverKeepUp?PLoSMed.2010;7(9):e1000326. doi:10.1371/journal.pmed.1000326 5.YaoX,SahaA,SaravananS,LowA,SussmanJ.AStudyProtocolforaComprehensive EvaluationofTwoArtificialIntelligence-BasedToolsinTitleandAbstractScreeningforthe DevelopmentofEvidence-BasedCancerGuidelines.CancerInnov.2025;4(4):e70021. Published2025Jun29.doi:10.1002/cai2.70021 6.ShekellePG.Updatingpracticeguidelines.JAMA.2014;311(20):2072-2073. doi:10.1001/jama.2014.4950 7.IoannidisJPA.TheMassProductionofRedundant,Misleading,andConflictedSystematic ReviewsandMeta-analyses.MilbankQuarterly.2016;94(3):485-514. doi:10.1111/1468-0009.12210 8.Alonso-CoelloP,SchünemannHJ,MobergJ,etal.GRADEEvidencetoDecision(EtD) frameworks:asystematicandtransparentapproachtomakingwellinformedhealthcare choices.1:Introduction.BMJ.PublishedonlineJune28,2016:i2016.doi:10.1136/bmj.i2016 9.AklEA,MeerpohlJJ,ElliottJ,etal.Livingsystematicreviews:4.Livingguideline recommendations.JournalofClinicalEpidemiology.2017;91:47-53. doi:10.1016/j.jclinepi.2017.08.009 32 10.KohaneIS.GuidelineMachines.NEJMAI.2026;3(7).doi:10.1056/AIe2600680 11.LaiH,LiuJ,BaiC,etal.Languagemodelsfordataextractionandriskofbiasassessmentin complementarymedicine.npjDigitMed.2025;8(1):74.doi:10.1038/s41746-025-01457-w 12.LiD,JiangN,HuangK,etal.Streamliningevidencebasedclinicalrecommendationswith largelanguagemodels.npjDigitMed.2025;8(1):793.doi:10.1038/s41746-025-02273-y 13.LieberumJL,ToewsM,MetzendorfMI,etal.Largelanguagemodelsforconducting systematicreviews:ontherise,butnotyetreadyforuse—ascopingreview.Journalof ClinicalEpidemiology.2025;181:111746.doi:10.1016/j.jclinepi.2025.111746 14.GuB,DesaiRJ,LinKJ,YangJ.Probabilisticmedicalpredictionsoflargelanguagemodels. npjDigitMed.2024;7(1):367.doi:10.1038/s41746-024-01366-4 15.TangL,SunZ,IdnayB,etal.Evaluatinglargelanguagemodelsonmedicalevidence summarization.npjDigitMed.2023;6(1):158.doi:10.1038/s41746-023-00896-7 16.ChristofM,ArmoundasAA.Implicationsofintegratinglargelanguagemodelsintoclinical decisionmaking.CommunMed.2025;5(1):490.doi:10.1038/s43856-025-01216-8 17.TudorCarL,WongTY,MajeedA,CollinsGS.Artificialintelligenceandthefutureof evidence-basedmedicine.Nat.Health.2026;1:268–271.doi:10.1038/s44360-025-00002-z 18.ZhengEL,JinW,HamarnehG,LeeSSJ.FromHuman-in-the-LooptoHuman-in-Power. TheAmericanJournalofBioethics.2024;24(9):84-86.doi:10.1080/15265161.2024.2377139 19.DiamondIR,GrantRC,FeldmanBM,etal.Definingconsensus:asystematicreview recommendsmethodologiccriteriaforreportingofDelphistudies.JClinEpidemiol. 2014;67(4):401-409.doi:10.1016/j.jclinepi.2013.12.002 20.Nussbaumer-StreitB,EllenM,KleringsI,etal.Resourceuseduringsystematicreview productionvarieswidely:ascopingreview.JClinEpidemiol.2021;139:287-296. doi:10.1016/j.jclinepi.2021.05.019 21.WangZ,CaoL,DanekB,JinQ,LuZ,SunJ.Acceleratingclinicalevidencesynthesiswith largelanguagemodels.NPJDigitMed.2025;8(1):509.Published2025Aug8. doi:10.1038/s41746-025-01840-7 22.CaoC,SangJ,AroraR,etal.DevelopmentofPromptTemplatesforLargeLanguage Model-DrivenScreeninginSystematicReviews.AnnInternMed.2025;178(3):389-401. 33 doi:10.7326/ANNALS-24-02189 23.HomiarA,ThomasJ,OstinelliEG,etal.Developmentandevaluationofpromptsforalarge languagemodeltoscreentitlesandabstractsinalivingsystematicreview.BMJMentHealth. 2025;28(1):e301762.Published2025Jul22.doi:10.1136/bmjment-2025-301762 24.SangheraR,ThirunavukarasuAJ,ElKhouryM,etal.High-performanceautomatedabstract screeningwithlargelanguagemodelensembles.JAmMedInformAssoc. 2025;32(5):893-904.doi:10.1093/jamia/ocaf050 25.HasanB,SaadiS,RajjoubNS,etal.Integratinglargelanguagemodelsinsystematicreviews: aframeworkandcasestudyusingROBINS-Iforriskofbiasassessment.BMJEvidBased Med.2024;29(6):394-398.Published2024Nov22.doi:10.1136/bmjebm-2023-112597 26.NationalInstituteforHealthandCareExcellence.DevelopingNICEGuidelines:The Manual.London:NationalInstituteforHealthandCareExcellence(NICE);July22,2015. 27.SchünemannHJ,WierciochW,BrozekJ,etal.GRADEEvidencetoDecision(EtD) frameworksforadoption,adaptation,anddenovodevelopmentoftrustworthy recommendations:GRADE-ADOLOPMENT.JClinEpidemiol.2017;81:101-110. doi:10.1016/j.jclinepi.2016.09.009 28.BrouwersMC,KhoME,BrowmanGP,etal.AGREEII:advancingguidelinedevelopment, reportingandevaluationinhealthcare.CMAJ.2010;182(18):E839-E842. doi:10.1503/cmaj.090449 29.MartínezGarcíaL,SanabriaAJ,GarcíaAlvarezE,etal.Thevalidityofrecommendations fromclinicalguidelines:asurvivalanalysis.CMAJ.2014;186(16):1211-1219. doi:10.1503/cmaj.140547 30.HodderRK,VogelJP,WolfendenL,TurnerT.LivingSystematicReviewsandLiving GuidelinestoMaintaintheCurrencyofPublicHealthGuidelines.AmJPublicHealth. 2024;114(1):21-26.doi:10.2105/AJPH.2023.307450 31.TendalB,VogelJP,McDonaldS,etal.Weeklyupdatesofnationallivingevidence-based guidelines:methodsfortheAustralianlivingguidelinesforcareofpeoplewithCOVID-19.J ClinEpidemiol.2021;131:11-21.doi:10.1016/j.jclinepi.2020.11.005 32.HagerP,JungmannF,HollandR,etal.Evaluationandmitigationofthelimitationsoflarge 34 languagemodelsinclinicaldecision-making.NatMed.2024;30(9):2613-2622. doi:10.1038/s41591-024-03097-1 33.WilliamsCYK,MiaoBY,KornblithAE,ButteAJ.Evaluatingtheuseoflargelanguage modelstoprovideclinicalrecommendationsintheEmergencyDepartment.NatCommun. 2024;15(1):8236.Published2024Oct8.doi:10.1038/s41467-024-52415-1 34.GohE,GalloR,HomJ,etal.LargeLanguageModelInfluenceonDiagnosticReasoning:A RandomizedClinicalTrial.JAMANetwOpen.2024;7(10):e2440969.Published2024Oct1. doi:10.1001/jamanetworkopen.2024.40969 35.GohE,GalloRJ,StrongE,etal.GPT-4assistanceforimprovementofphysician performanceonpatientcaretasks:arandomizedcontrolledtrial.NatMed. 2025;31(4):1233-1238.doi:10.1038/s41591-024-03456-y 36.EverettSS,BunningBJ,JainP,etal.FromTooltoTeammate:ARandomizedControlled TrialofClinician-AICollaborativeWorkflowsforDiagnosis.Preprint.medRxiv. 2025;2025.06.07.25329176.Published2025Jun8.doi:10.1101/2025.06.07.25329176 37.HigginsJP,AltmanDG,GøtzschePC,etal.TheCochraneCollaboration'stoolforassessing riskofbiasinrandomisedtrials.BMJ.2011;343:d5928.Published2011Oct18. doi:10.1136/bmj.d5928 38.SterneJAC,SavovićJ,PageMJ,etal.RoB2:arevisedtoolforassessingriskofbiasin randomisedtrials.BMJ.2019;366:l4898.Published2019Aug28.doi:10.1136/bmj.l4898 39.SterneJA,HernánMA,ReevesBC,etal.ROBINS-I:atoolforassessingriskofbiasin non-randomisedstudiesofinterventions.BMJ.2016;355:i4919.Published2016Oct12. doi:10.1136/bmj.i4919 40.WhitingPF,RutjesAW,WestwoodME,etal.QUADAS-2:arevisedtoolforthequality assessmentofdiagnosticaccuracystudies.AnnInternMed.2011;155(8):529-536. doi:10.7326/0003-4819-155-8-201110180-00009 35 36 Fig.1|Clinicalguidelineupdatingworkflowandhuman-GUIDEcollaborationflowchart.a, OverviewoftheGUIDEworkflowunderahuman-in-the-loopcopilotmodel.Startingfroma clinicalquestion,GUIDEsupportsPICOconstruction,literatureretrievalandscreening,evidence assembly,dataextraction,GRADE-basedqualityassessment,evidence-to-recommendation materialsynthesis,multi-agentconsensus,andgenerationofclinicalrecommendations. Throughouttheprocess,AIagentsperformstructuredevidence-synthesistasks,whereashuman expertsretainsupervisoryauthorityandintervenewhenpredefinedconditionsaremet.b,Detailed workflowshowingthetriggercriterionforexpertinvolvement.GUIDEdecomposesthereview questionintoPICOelements,coordinatesindependentagentsforsearch,screening,dataextraction andevidencequalityappraisal,andsynthesizesevidence-to-recommendationmaterialsthrougha Delphi-likemulti-agentprocess.Ateachkeystage,anAIReviewermonitorspredefinedsignalsof uncertaintyorinconsistency,suchasdiscordantretrievalresults,screeningdisagreements, inconsistentextracteddataordivergentcertaintyratings,andtriggersexpertverification, adjudicationorreassessment.Thiscondition-triggeredoversightenablesGUIDEtocombine automatedefficiencywithtransparentandaccountablehumangovernance. 37 Fig.2|Performanceofthehuman-GUIDEcollaborativeframeworkinthebenchmark evaluationusing12ophthalmologydatasets.a,RecallandprecisionofGUIDEacrossliterature retrieval,title-and-abstractscreening,andfull-textscreening.GUIDEachievedhighrecallacross allstages,whileprecisionprogressivelyincreasedacrossthethreesteps.b,Agreementbetween GUIDEandhumanreferencestandardsinevidenceassessment,includingriskofbias,indirectness, imprecision,inconsistency,andoverallcertaintygrading.Ineachboxplot,thetrianglerepresents themean,thecentrallinedenotesthemedian,theboxindicatestheinterquartilerange,andthe whiskersextendtovalueswithin1.5×theinterquartilerange.c,ConsistencybetweenGUIDE’s recommendationoutputsandoriginalrecommendationsassessedbydomainexpert.d,Time expenditureofGUIDEcopilotsystemandhumanexpertsacrosstheguidelinedevelopment workflow.GUIDEcontributedthemajorityofprocessingtime,whileexpertinvolvementwas concentratedonoversight,arbitration,andrefinement. 38 39 Fig.3|Ablationtestscomparinghuman-GUIDEcollaborationversusGUIDEalone:aThe recallrateoftheliteraturesearch.bTherecallandaccuracyratesofliteraturescreening.c–dThe consistencybetweentheGUIDEandtheNICEguidelinesfortheevidenceassessment(RoB, indirectness,imprecision,inconsistency,andtotalevidencequalityevaluations),and recommendationgenerationmodules(n=12).Theerrorbarsrepresentthe95%confidence intervals(CIs).Statisticalsignificanceisindicatedusingpairedcomparisons,withnsdenotingnot significant,**denotingP<0.01and***denotingP<0.001. 40 Fig.4|Performanceofthehuman-GUIDEcollaborativeframeworkinthegeneralizationtest using20datasetsderivedfromfourdifferentspecialties(cardiology,endocrinology, gastroenterology,andrespiratorymedicine).a,RecallandprecisionofGUIDEacrossliterature retrieval,title-and-abstractscreening,andfull-textscreening.b,ConsistencyofGUIDE-assisted evidenceassessmentacrossriskofbias,indirectness,imprecision,inconsistency,andevidence quality.c,ConsistencybetweenGUIDE’srecommendationoutputsandoriginalrecommendations assessedbydomainexpert.Ineachboxplot,thetrianglerepresentsthemean,thecentralline denotesthemedian,theboxindicatestheinterquartilerange,andthewhiskersextendtovalues within1.5×theinterquartilerange.d,Timeexpenditureacrosstheguideline-updatingworkflow, comparingGUIDEprocessingtimewithhumanexpertinvolvementforPICOconstruction, literaturesearch,literaturescreening,dataextraction,evidenceassessment,andrecommendation generation. 41 Fig.5|human-GUIDE’send-to-endperformanceinsimulatedprospectiveexperiments.a, Screeningperformanceacrosssixpost-cutoffNICEguidelinedatasets,showingconsistentlyhigh recallandprecisiondespitepotentialupstreamerrorpropagation.Ineachboxplot,thecentralline denotesthemedian,theboxindicatestheinterquartilerange,thewhiskersextendtovalueswithin 1.5×theinterquartilerange,andthetrianglerepresentsthemean.b,Assessmentofdrafted recommendationsinthefinalstep,showing100.00%directionalagreementwiththeoriginal NICErecommendations. 42 SupplementaryFig.1|Exampleofsearchstrategyrevisionthroughexpert’sinstructionsto theAIassistant.AfterGUIDEconstructsthestructuredPICOframeworkfromtheclinical question,twoindependentLLM-basedsearchagentsgenerateretrievalstrategiesinparallel.The AIReviewercomparestheresultingstrategiesandretrievaloutputstoidentifypotential unreliability,suchaslowoverlapormarkeddiscrepanciesinretrievalvolume.Whenrevisionis needed,expertsprovidenatural-languageinstructionsthroughtheassistantagent,whichtranslates theseinstructionsintoupdatedBooleansearchstrings,enablingstrategyrefinementwithoutdirect manualeditingofcomplexsyntax. 43 SupplementaryFig.2|GUIDE-assistedliteraturescreeningworkflowandrationaledisplay. Theinterfacesupportssequentialtitle-and-abstractandfull-textscreeningbyparallelscreening agents.Retrievedrecordsareindependentlyassessedforeligibility,withinclusionorexclusion labels,confidencelevels,andconciserationalesdisplayedforexpertreview.Whenagentoutputs areconsistent,thesystempresentsasummarizedscreeningjustification;whendiscrepanciesarise, detailedagent-specificreasoningisshowntofacilitatehumanadjudicationandcorrection. 44 SupplementaryFig.3|DataextractioninterfacealignedwiththePICOframework. Forstudiesincludedafterfull-textscreening,GUIDEsupportsstructuredextractionofstudy-level textualdataandoutcome-levelnumericalresults.Expertscaneditorre-extractindividualfieldsto verifydiscrepanciesagainsttheoriginalfulltextandensureconsistencywiththePICO-defined evidenceframework. 45 SupplementaryFig.4|Delphi-likemulti-agentevidencesynthesisandrecommendation drafting.ProcessedevidenceissubmittedtoaDelphi-likeconsensusmodule,inwhichhuman expertscanselectandassembleavirtualpanelwithdifferentrolespecifications.Panelmembers independentlyreviewtheevidence,providejudgments,andstatefinalrecommendationdecisions withsupportingrationales.Thecommitteeagentsynthesizestheseinputsintoan evidence-to-recommendationreportanddraftrecommendation,whichexpertscanfurtherrevise andfinalizethroughdeliberationandconsensus. 46 SupplementaryTable1.Proposedsuggestionsforhuman-AIcollaboration CategorySuggestionDescription CorePrinciplesSuggestion1Thecollaborationshouldfollowa“human-in-the-loop”model: AIexecutesrule-based,bulkprocessingandpreliminary drafting,whereashumanexpertsprovideprofessional judgementandfinaldecision-making.Thiscollaboration operateswithinadynamicframeworkthatiteratively optimizesthehuman-AIdivisionoflaborinalignmentwith thecharacteristicsandcomplexityoftasks. ProposedHuman-AI CollaborationModes Suggestion2Proactivehuman-interventionmode:Applicabletoscenarios whereAIcapabilitiesarenotyetmatureordecision-makingis criticalandrequiresnuancedjudgment. Suggestion3Condition-triggeredhumanreviewmode:Applicableto scenarioswhereAIperformanceisgenerallyrobust,but potentialanomaliesoruncertaintiesmayarise,triggering humanreviewaccordingtopredefinedcriteria. 47 SupplementaryTable2.Taskcharacteristicsacrosstheguidelinegenerationworkflowand correspondingpreferablehuman-AIcollaborationmodes StepDescriptionChallengePreferablecollaboration mode PICOConstructionFormulatingtheclinical questionintoPICO components(Population, Intervention,Comparison, Outcome). Definingafocusedscopeand selectingclinicaloutcomes thatreflectbothrisksand benefits. Proactive human-interventionmode SearchStrategyFormulationDevelopingdatabase-specific searchstringsbasedonthe PICOframework. Mappingandexpandingcore entitiesintosearchterms, formulatingsearchstrategies inaccordancewithspecific syntaxrules. Condition-triggeredhuman reviewmode LiteratureScreeningSelectingrelevantstudies fromsearchresultsbasedon inclusion/exclusioncriteria. Processinglargevolumesof literaturefollowing inclusion/exclusioncriteria. Condition-triggeredhuman reviewmode DataExtractionCollectingkeyvariablesand statisticaldataforanalysis andqualityassessment. Handlinginconsistentdata formatsandaccurately extractingfromprimary studies. Condition-triggeredhuman reviewmode EvidenceQualityAssessmentAssessingthecertaintyof evidenceandriskofbias usingtheGRADEcriteria. Detectingpotentialbiasin boththepresenceandabsence ofinformation,makinga holisticjudgmenttailoredto theuniqueconditionsof studies. Proactive human-interventionmode Evidence-to-Recommendation (EtR)Synthesis Synthesizingclinical evidenceintothe Evidence-to-Decisionreport. Accountingforpractical factorslikecost,feasibility, andpatientvaluesalongside Proactive human-interventionmode 48 clinicaleffects. Recommendation Formulation Reconcilingexpertopinions toreachafinal,unified recommendation. Resolvingexpert disagreementstoproducea clearanddeliberateclinical statement. Proactive human-interventionmode 49 SupplementaryTable3.Overviewofthebenchmarkdataset SpecialtyNICEGuidelinePublicationDateReviewquestion Ophthalmology NG242:Diabetic retinopathy: managementand monitoring 13-Aug-24 Whatistheeffectivenessofdifferentthresholdsor criteriaforstartingtreatmentfornon-proliferative diabeticretinopathy,proliferativediabetic retinopathy,anddiabeticmacularoedema? Whatistheeffectofintensivetreatmentstorapidly lowerbloodglucoselevelsonprogressionof diabeticretinopathyanddiabeticmacularoedema? Inpeoplewhoareabouttoundergoorwhohave undergonecataractsurgery,whatisthe effectivenessoftreatments(before,duringorafter surgery)formanaging: •non-proliferativediabeticretinopathy •proliferativediabeticretinopathy •diabeticmacularoedema? NG77:Cataractsin adults:management 26-Oct-17 Whatistheoptimalstrategytoaddresspre-existing regularastigmatisminpeopleundergoingcataract surgery? Aretintedlenseseffectiveinpreventingthe progressionofage-relatedmaculardegeneration comparedwithcolourlesslensesincataract surgery? Whatistheeffectivenessofsedationasanadjunct tolocalanaesthesiaduringcataractsurgery? Whatistheeffectivenessofinterventionsto 50 increasepupilsizetoimprovevisualoutcomesand reducecomplicationsduringphacoemulsification cataractsurgery? Whatistheoptimaltypeandadministrationof anesthesiaforcataractsurgery? NG82:Age-related macular degeneration: diagnosisand management 23-Jan-18 Whatistheeffectivenessofpsychologicaltherapies forAMD? Whatstrategiesandtoolsareusefulfor self-monitoringforpeoplewithAMD? Whatistheeffectivenessofadjunctivetherapiesfor thetreatmentoflateAMD(wetactive)? NG81:Glaucoma: diagnosisand management 01-Nov-17 Whatistheeffectivenessandcost-effectivenessof selectivelasertrabeculoplasty(SLT)asafirstline treatmentcomparedwithintraocularpressure loweringeyedropsinocularhypertension(OHT)or chronicopen-angleglaucoma(COAG)adult patients? 51 SupplementaryTable4.ReferenceAnswersandEvaluationMetricsUsedinModuleTesting StepInputGUIDEOutput ReferenceAnswers (fromguideline) EvaluationMetrics ClinicalQuestion Parsing Naturallanguageclinical question(fromsource guidelines) StructuredPICOelements PICOdefinitionsfromthereview protocol BERTScore RetrievalStrategy Generation PICOrepresentation Searchstrategy(databases, filters,studytypes, keywords) SearchprotocolRecall StudySelectionRetrievedcandidatestudiesIncludedstudylist Includedstudiesreportedinthe guideline Recall;Precision EvidenceExtractionIncludedstudies,PICO Structuredstudy-leveldata (categorical)and quantitativeresults Evidencetableswithsummaryof effectivenessevidence Typeclassification accuracy;Numerical accuracy EvidenceQuality Assessment PICO,Synthesizedevidence table Qualityratings Qualityassessmentsreportedin evidencetablesandmethods Classification accuracy Recommendation Generation PICO,Evidenceandquality assessment Guideline-style recommendation Committeefinalrecommendation Customized evaluationrubric 52 SupplementaryTable5.DescriptionoftheDatasetComposition ClinicalQuestionDefinition DataFieldDescriptionExampleDataType ClinicalQuestionTheprimaryclinicalquestionaddressedby theguideline “Whatistheeffectivenessofdifferent thresholdsforinitiatingtreatmentindiabetic retinopathy?” String PopulationTargetpatientgroupsdefinedinthereview (PinPICO) Non-proliferativeDR;proliferativeDR;DMOList[String] InterventionInterventionsbeingevaluated(IinPICO)Earlylaser;anti-VEGF;vitrectomyList[String] ComparatorComparatorconditions(CinPICO)Deferredtreatment;standardthresholdList[String] OutcomesClinicaloutcomesofinterest(OinPICO)Visualacuity;retinalthickness;qualityoflifeList[String] LiteratureRetrievalandScreening DataFieldDescriptionExampleDataType RetrievedRecordsBibliographicrecordsidentifiedduringthe initialliteraturesearch Alistofcandidatestudiesreturnedby databasesearch List[Record] RetrievedRecord Metadata Citation-levelmetadataforeachretrieved record Title;authors;publicationyear;journal; databasesource RecordObject Screened-in Records Recordsretainedaftertitle/abstract screeningforfull-textassessment Studiesjudgedpotentiallyeligibleafterinitial screening List[Record] Full-textAssessed Records Recordsundergoingfull-texteligibility assessment CandidatestudiesreviewedinfulltextList[Record] IncludedStudiesFinalstudiesincludedinevidencesynthesisBakeretal.,2019;Elmanetal.,2015List[Record] Study-LevelEvidenceExtraction DataFieldDescriptionExampleDataType StudyIdentifierUniquestudyreferenceBaker2019String StudyDesignStudymethodologyRandomizedcontrolledtrialCategorical Population Characteristics BaselinepatientcharacteristicsDMOwithgoodvisualacuityText InterventionDescriptionofinterventionPromptintravitrealanti-VEGFText 53 Details Comparator Details DescriptionofcomparatorObservationwithdeferredtreatmentText OutcomeNameSpecificclinicalendpointLossof≥15lettersinBCVAString SampleSize (Outcome-Level) Numberofobservationscontributingtothe estimate 420eyesInteger EffectTypeStatisticalmeasureusedRiskRatioCategorical EffectSizeEstimatedeffectvalue0.92Float Confidence Interval Lowerandupperboundsofestimate[0.82,1.03]Tuple EvidenceQualityAssessment DataFieldDescriptionExampleDataType RiskofBias AssessmentTool ToolusedforqualityassessmentRoB2;ROBINS-ICategorical RiskofBiasDowngradeduetostudylimitationsDowngraded(seriouslimitations)Categorical InconsistencyVariabilityacrossstudyresultsNoseriousinconsistencyCategorical IndirectnessApplicabilityofevidencetotarget population NoseriousindirectnessCategorical ImprecisionUncertaintyineffectestimatesSeriousimprecisionCategorical Other Considerations Additionalupgrading/downgradingfactorsNoneString FinalCertainty Rating OverallinterpretationofeffectFavoursearlytreatmentCategorical Recommendation-LevelExtraction DataFieldDescriptionExampleDataType Evidence Summary NarrativesummaryoffindingsEarlylaserreducesriskofseverevisuallossText BenefitSummarySummaryofbenefitsEarlytreatmentreducesprogressionText HarmSummarySummaryofharmsorrisksLimitedorunclearharmsText 54 ClinicalJudgmentExpertinterpretationbeyonddataEarlypanretinalphotocoagulationalignswith clinicalpractice Text Resource Considerations CostorfeasibilityconsiderationsIncreasedresourceuseforearlytreatmentText Recommendation Statement FinalsummarizedrecommendationOfferPRPatdiagnosisofproliferativeDRText 55 SupplementaryTable6.Evaluationcriteriafortheconsistencyofrecommendations ConsistencyLevelDescription ExactlyConsistentThesemanticmeaningofcoreelements(e.g.,recommendationdirection, targetpopulation,andinterventions)completelymatchestheoriginal guideline,withnosubstantivedifferences. PartiallyConsistentTheoverallrecommendationdirectionisconsistent,butthereare differencesoromissionsregardingthespecificscopeofthetarget population,interventiondetails,orimplementationconditions. InconsistentTherecommendationdirectionsarecontradictory,orthesystemoutputhas nocorrespondingstatementintheoriginalguideline. Note:Theoverallagreementrateiscalculatedas:AgreementRate=(ExactlyConsistent Recommendations+PartiallyConsistentRecommendations)/TotalRecommendations. 56 SupplementaryTable7.Overviewofthegeneralizationtestdataset SpecialtyNICEGuidelinePublicationDateReviewquestion Cardiovascular NG106:Chronicheart failureinadults 3-Sep-25 Whatistheclinicalandcosteffectivenessof mineralocorticoidreceptorantagonistsinpeople withheartfailurewithpreservedejection fraction? Whatistheclinicalandcosteffectivenessofsalt and/orfluidrestrictioninpeoplewithheart failure? Whatistheclinicalandcosteffectivenessof ironsupplementationinpeoplewithchronic heartfailureandirondeficiency? Whatistheclinicalandcosteffectivenessof pharmacologicalinterventionsforheartfailure inpeoplewithheartfailurewhoalsohave chronickidneydisease? Endocrinesystem NG243:Adrenal insufficiency:identification andmanagement 28-Aug-24 Whatistheclinicalandcosteffectivenessof glucocorticoidsfortheroutinemanagementof secondaryandtertiaryadrenalinsufficiency? Whatistheclinicalandcosteffectivenessof pharmacologicaltreatmentsfortheroutine managementofprimaryadrenalinsufficiency? Inpeopleatriskofadrenalinsufficiency 57 becauseofprolongedcorticosteroiduse,whatis thebestwaytomanagecorticosteroid withdrawalwhencorticosteroidsarenolonger neededtocontroldiseaseactivity? Gastrointestinal NG50:Cirrhosisinover 16s:assessmentand management 8-Sep-23 Whatistheclinicalandcost-effectivenessof non-selectivebeta-blockersfortheprimary preventionofbleedinginpeoplewith oesophagealvaricesduetocirrhosis? Whatistheclinicalandcost-effectivenessof endoscopicbandligationfortheprimary preventionofbleedinginpeoplewith oesophagealvaricesduetocirrhosis? Whatisthemostclinicalandcost-effective prophylacticantibioticfortheprimary preventionofbacterialinfectionsinpeoplewith cirrhosisanduppergastrointestinalbleeding? Whatistheclinicalandcost-effectivenessof antibioticscomparedwithplaceboforthe primarypreventionofspontaneousbacterial peritonitis(SBP)inpeoplewithcirrhosisand ascites? Whatisthemostclinicallyandcost-effective interventionforthefirst-linetreatmentofan episodeofacutehepaticencephalopathyin peoplewithcirrhosis? Whatistheclinicalandcost-effectivenessof transjugularintrahepaticportosystemicshunt (TIPS)comparedwithlarge-volume 58 paracentesis(LVP)withalbumininthe managementofdiuretic-resistantascitesdueto cirrhosis? NG231: Barrett’s oesophagusand stage1 adenocarcinoma 8-Feb-23 ForadultswithBarrett’soesophagus,whatis theclinicalandcosteffectivenessof pharmacologicalinterventions(suchasantacids, aspirin,H2receptorantagonists,protonpump inhibitors)inreducingprogressiontodysplasia orcancer? Respiratory system NG250:Pneumonia: diagnosisandmanagement 2-Sep-25 Whatistheclinicalandcosteffectivenessof careoutsideoftheacutehospitalsetting(for exampleusingintermediatecare,hospitalat home,orvirtualwards)forpeoplewith community-acquiredpneumoniawhowould otherwisebeadmittedtohospital? Areshorterdurationsofantibiotictreatment effectiveandcost-effectiveattreatingbabies, childrenandyoungpeoplewith community-acquiredpneumoniacomparedto longertreatmentcourses? Inadultswithcommunity-acquiredpneumonia, whatistheclinicalandcosteffectivenessof corticosteroidtreatmentinadditiontoantibiotic treatment,comparedwithantibiotictreatment alone? Theclinicalandcost-effectivenessof corticosteroidtreatmentinadditiontoantibiotic treatmentcomparedwithantibiotictreatment 59 aloneinchildrenwithCAPorHAP? NG80:Asthma:diagnosis andmonitoringofasthmain adults,childrenandyoung people 29-Nov-17 Inpeoplewithasthma,whatistheclinicaland cost-effectivenessofusingmeasuresof pulmonaryfunctionassessingasthmacontrol (forexample,spirometryandpeakexpiratory flow)tomonitorasthma? Inpeoplewithasthma,whatistheclinicaland cost-effectivenessofusingindirectchallenge testswithmannitolordirectchallengetestswith histamineormethacholineformonitoring asthmacontrol? 60 SupplementaryTable8.Overviewofthesimulateend-to-endprospectivetestdataset NICEGuidelinePublicationDateReviewquestion NG249:Falls:assessment andpreventioninolder peopleandpeople50and overathigherrisk 29-Apr-25 Whatarethemosteffectivemethodsfor maximisingparticipation,adherenceand continuationoffallspreventioninterventions? Whatarethemostclinicallyandcost-effective methodsforfallspreventioninolderpeoplein residentialcaresettings? NG257:Fertility problems: assessmentandtreatment 31-Mar-26 Whatistheclinicalandcosteffectivenessoftests forendometrialreceptivity(includinggene expressionanalysisandmicrobiologicalanalysis) asatreatmentadd-onforpeopleundergoing fertilitytreatment? Whatistheclinicalandcosteffectivenessof treatingDNAfragmentation(identifiedfrom screeningejaculatedsperm)onreproductive outcomesforpeoplewithmalefactorfertility problems? NG253:Suspectedsepsis inpeopleaged16or over:recognition, assessmentandearly management 19-Nov-25 Whatisthemostclinicalandcost-effective intravenousfluidforresuscitationofpeopleaged 16oroverwithsuspectedsepsis? Inpeopleaged16oroverwithsuspectedsepsis, howsafeistheperipheraladministrationof intravenousvasopressorscomparedtocentral lineadministration? 61 SupplementaryInformation1.PICOcriteria(prospectivestudy,GUIDE-assistedgroup) GenerationbyGUIDERevisionbyhumanexpert PopulationPatientsundergoingcataractsurgeryAdults(18yearsandolder)undergoingphacoemulsification cataractsurgeryandintraocularlens(IOL)implantationinone orbotheyes InterventionsAnytypeofintraocularlensestofacilitate simultaneousdistanceandnearvision(e.g., multifocalintraocularlenses,monovision withmonofocallenses,extendeddepthof focuslenses) 1.Anytypeofnon-accommodatingintraocularlensesto facilitatesimultaneousdistanceandnearvision(e.g., multifocalintraocularlenses,monovisionwithmonofocal lenses,extendeddepthoffocuslenses) 2.Standardmonofocalintraocularlenses(providingclear distancevisiononly,requiringreadingglassesfornearvision) ComparatorsStandardmonofocalintraocularlenses (providingcleardistancevisiononly, requiringreadingglassesfornearvision) Alllistedinterventionsvs.eachother Outcomes  Visualacuityatdistanceandnear  Spectacleindependence  Contrastsensitivity  Patientsatisfaction  Qualityoflife  Incidenceofvisualdisturbances(e.g., glare,halos)  Needforadditionalcorrective procedures  Visualacuityatdistanceandnear  Spectacleindependence  Contrastsensitivity  Patientsatisfaction  Qualityoflife  Incidenceofvisualdisturbances(e.g.,glare,halos)  Needforadditionalcorrectiveprocedures  Resourceuseandcosts 62 SupplementaryInformation2.Summaryofincludedstudies(prospectivestudy, GUIDE-assistedgroup) Study&locationSamplesizeMethods Cochener(2018) France 120eyes (60patients) ProspectiverandomizedcomparativestudyofbilateralAcrySofIQPanOptixand FineVisionMicroFtrifocalIOLsversustheTECNISSymfonyEDOFIOL. Vandekerckhove (2026) Switzerland 210eyes (105patients) Single-centerprospectiverandomizeddouble-blindtrialcomparingbilateral Isopure,Eyhance,andVivitypresbyopia-correctingIOLswithmini-monovision (−0.50D). McCabe(2022) USA 435eyes (218patients) Prospectivemulticenterrandomizedparallelcontrolled,patient-and assessor-maskedtrialcomparingbilateralDFT015(Vivity)nondiffractiveEDOF IOLswithSN60WFmonofocalIOLs. Turhan(2021)/ Sevik(2022) Turkey 58eyes (29patients) ProspectiverandomizedcomparativestudyofbilateralTECNISZKB00low-add bifocalIOLsversusTECNISSymfonyEDOFIOLs,bothusingmini-monovision (−0.50D). Chang(2023) USA 272patientsProspectivemulticenterbilateralrandomized,three-way-maskedcomparativetrial ofTECNISSynergyZFR00VhybridIOLsversusTECNISZCB00monofocal IOLs. Zeilinger(2025) Austria 80eyes (40patients) Single-centerrandomizeddouble-maskedcontrolledtrialcomparingbilateral RayOneEMVadvancedmonofocalIOLswithVivitynondiffractiveEDOFIOLs usingmini-monovision. Han(2025) China 100eyes (100patients) Prospectiverandomizedcase-controlstudycomparinganasphericmonofocalIOL withamultifocalIOLinmonocularcataractsurgery. Li(2023) China 99eyes (99patients) Prospectiverandomizeddouble-maskedcontrolledtrialcomparingunilateralAT LISAtoric939M/MPtrifocalIOLswithATLISAtoric909MbifocalIOLs. Webers(2020) Netherlands 54eyes (27patients) ProspectiverandomizedcontrolledtrialcomparingbilateralTECNISSymfony EDOFIOLswithmicromonovisionversusbilateralATLISAtri839MPtrifocal IOLs. Reinhard(2021) Germany,Belgium, 211patientsProspectivemulticenterrandomizedthree-armtrialcomparingbilateralATLARA andTECNISSymfonyEDOFIOLswithCTASPHINAmonofocalIOLs. 63 Portugal,Spain, France,andItaly Hida(2024) Brazil 116eyes (58patients) ProspectiveinterventionalrandomizedparallelstudycomparingbilateralTECNIS SynergyZFR00Vimplantationwithablendedstrategy(EyhanceICB00inthe dominanteyeandSynergyinthenondominanteye). Sandoval(2020) USA 76patientsProspectiveunmaskedrandomizedstudyofbilateraltoricTECNISSymfony EDOFimplantationtargetingemmetropiainbotheyesversus−0.50D mini-monovisioninthenondominanteye. Pradeep(2026) India 70patientsProspectiverandomizedcomparativestudyofbilateralPanOptixtrifocalversus VivityEDOFIOLimplantation. Gil(2022) Spain 152patientsProspectivedouble-maskedcomparativestudyinwhichsixbilateral presbyopia-correctingIOLgroupswereblock-randomized1:1:1:1:1:1(SV25T0, ATLISA809M,ATLISAtri839MP,TECNISZKB00,TECNISZLB00,and TECNISSymfonyZXR00),withanonrandomizedbilateralTECNISZA9003 monofocalcomparatorgroup. Messias(2021) Brazil 66eyes (33patients) ProspectiverandomizedstudycomparingbilateralSN6AD1diffractivemultifocal IOLswithSN60WFasphericmonofocalIOLs. Labiris(2024) Greece 90patientsProspectiverandomizedcomparativestudyofpremiummonovision(PanOptixin thenondominanteyeandVivityinthedominanteye)versusbilateralPanOptixor bilateralVivityimplantation. Bala(2022) Australia,Canada, Spain,andUK 270patientsProspectivemulticenterrandomizedparallelcontrolled,patient-and assessor-maskedtrialcomparingbilateralDFT015(Vivity)nondiffractiveEDOF IOLswithSN60WFmonofocalIOLs. Skov(2026) Denmark 180patientsSingle-centerrandomizedcontrolledpilottrialcomparingfourbilateralstrategies: partial-range-of-focusnarrowIOLstargetingemmetropia,mini-monovision(−1.25 D),monovision(−2.50D),andpartial-range-of-focusextendedIOLs. Corbett/Black (2024) AustraliaandNew 118patientsProspectivemulticenterrandomizedsubject-andevaluator-maskedtrialcomparing bilateralTECNISPureSeeZEN00VrefractiveEDOFIOLswithEyhanceICB00 enhancedmonofocalIOLs. 64 Zealand Pantanelli(2023) USA 51patientsProspectivesingle-centerdouble-armrandomized,patient-andassessor-masked trialcomparingbilateralVivitynondiffractiveEDOFIOLswithenVistaneutral asphericmonofocalIOLstargetedforemmetropia. Gil(2020) Spain 116patientsProspectiverandomizeddouble-maskedstudycomparingsixbilateral presbyopia-correctingIOLs(SV25T0,ATLISA809M,ATLISAtri839MP, TECNISZKB00,TECNISZLB00,andTECNISSymfony). Guarro(2022) Spain 88patientsProspectiverandomizeddouble-blindstudycomparingbilateralVivity,AT LARA,andTECNISSymfonyEDOFIOLswithSN60WFmonofocalIOLs. Law(2020) UK 90patientsProspectiverandomizeddouble-maskedstudycomparingbilateralBi-Flex677MY multifocalIOLswithBi-Flex677ABmonofocalIOLs. Chang(2022) USA 299patientsProspectivemulticenterbilateralrandomized,patient-andevaluator-maskedtrial comparingTECNISSymfonyZXR00EDOFIOLswithTECNISZCB00 monofocalIOLs. Palkovits(2026) Austria 35patientsProspectiverandomizedsingle-centerstudycomparingbilateralAcunexVario AN6VEDOFIOLswithAcunexAN6monofocalIOLs. 65 SupplementaryInformation3.Excludedstudiesandreasons(prospectivestudy, GUIDE-assistedgroup) StudyReasonforexclusion SaviniG,Schiano-LomorielloD,BalducciN,BarboniP.Visual performanceofanewextendeddepth-of-focusintraocularlenscompared toadistance-dominantdiffractivemultifocalintraocularlens.Journalof RefractiveSurgery.2018;34(4):228-235. Retrospectivecomparative study. Escandón-GarcíaS,RibeiroFJ,McAlindenC,QueirósA, González-MéijomeJM.Through-focusvisionperformanceandlight disturbancesof3newintraocularlensesforpresbyopiacorrection.Journal ofOphthalmology.2018;2018:6165493. Nonrandomizedcaseseries. MencucciR,FavuzzaE,CaporossiO,SavastanoA,RizzoS.Comparative analysisofvisualoutcomes,readingskills,contrastsensitivity,andpatient satisfactionwithtwomodelsoftrifocaldiffractiveintraocularlensesand anextendedrangeofvisionintraocularlens.Graefe'sArchiveforClinical andExperimentalOphthalmology.2018;256(10):1913-1922. Nonrandomizedcaseseries. PedrottiE,CaronesF,AielloF,MastropasquaR,BruniE,BonacciE,et al.Comparativeanalysisofvisualoutcomeswith4intraocularlenses: monofocal,multifocal,andextendedrangeofvision.JournalofCataract &RefractiveSurgery.2018;44(2):156-167. Participantswerenotrandomly allocatedtotheIOLgroups.* FernandesMC,NunomuraCY,MessiasA.Visualperformanceandphotic disturbanceswithdiffractiveandnondiffractiveEDOFintraocularlenses usingmini-monovision:randomizedtrial.JournalofCataract&Refractive Surgery.2024;50(2):153–159. BothgroupsreceivedEDOF IOLs. deMedeirosAL,JonesSaraivaF,IgumaCI,KniggendorfDV,AlvesG, ChavesMAPD,etal.Comparisonofvisualoutcomesafterbilateral implantationoftwointraocularlenseswithdistinctdiffractiveoptics. ClinicalOphthalmology.2019;13:1657-1663. Nonrandomizedcaseseries. RodovL,ReitblatO,LevyA,AssiaEI,KleinmannG.VisualoutcomesRetrospectivereview. 66 andpatientsatisfactionfortrifocal,extendeddepthoffocusandmonofocal intraocularlenses.JournalofRefractiveSurgery.2019;35(7):434-440. WantenJC,BauerNJC,BerendschotTTJM,vandenBiggelaarFJHM, NuijtsRMMA.Twonovelextendeddepth-of-focusintraocularlenses targetedformini-monovision:Aprospectiverandomizedcontrolledtrial. AmericanJournalofOphthalmology.2025;276:286–296. BothgroupsreceivedEDOF IOLs.* SinghB,SharmaS,DadiaS,BhartiN,BhartiS.Comparativeevaluation ofvisualoutcomesafterbilateralimplantationofadiffractivetrifocal intraocularlensandanextendeddepthoffocusintraocularlens.Eye& ContactLens.2020;46(5):314-318. Nonrandomizedcomparative study. PaikDW,ParkJS,YangCM,LimDH,ChungTY.Comparingthevisual outcome,visualquality,andsatisfactionamongthreetypesofmulti-focal intraocularlenses.ScientificReports.2020;10:14832. Nonrandomizedcomparative study. FarvardinM,JohariM,AttarzadeA,RahatF,FarvardinR,FarvardinZ. Comparisonbetweenbilateralimplantationofatrifocalintraocularlens (AlconAcrySofIQPanOptix)andextendeddepthoffocuslens(Tecnis SymfonyZXR00lens).InternationalOphthalmology. 2021;41(2):567-573. Nonrandomizedcomparative study. HayashiK,YoshidaM,SasakiH,HirataA.Binocularvisualfunctionof myopicpseudophakicmonovision.JapaneseJournalofOphthalmology. 2018;62:357–364. Interventiondidnotmeet predefinedscope.* ChaoCC,LinHY,LeeCY,MaiELC,LianIB,ChangCK.Differencein qualityofvisionoutcomeamongextendeddepthoffocus,bifocal,and monofocalintraocularlensimplantation.Healthcare.2022;10(6):1000. Retrospectivecomparative study. ZhuM,FanW,ZhangG.Stereopsisandvisualacuity:bilateraltrifocal versusblendedextendeddepthoffocusanddiffractivebifocalintraocular lenses.FrontiersinMedicine.2022;9:1042101. Nonrandomizedcomparative study. NagyZZ,Popper-SachettiA,KissHJ.Comparisonofvisualandrefractive outcomesbetweenhydrophilicandhydrophobictrifocalintraocularlenses BothIOLsweretrifocallenses. 67 sharingthesameopticaldesign.JournalofCataract&RefractiveSurgery. 2019;45(5):553-561. DickHB,AngRE,CorbettD,HoffmannP,TetzM,VillarrubiaA, PalominoC,Castillo-GomezA,TsaiL,ThomasEK,JanakiramanP. Comparisonof3-monthvisualoutcomesofanewmultifocalintraocular lensvsatrifocalintraocularlens.JournalofCataract&Refractive Surgery.2022;48(11):1270–1276. Cataract-onlypatientsoutcome datawerenotextractable.* NagyZZ,DormanP,SzalczerS,KissH.ProspectiveComparisonof ClinicalOutcomesAfterBilateralImplantationofDiffractiveTrifocal ExtendedDepth-of-FocusandDiffractiveTrifocalHydrophobic IntraocularLenses.JournalofRefractiveSurgery.2025;41(3):e102–e113. Cataract-onlypatientsoutcome datawerenotextractable.* VasavadaV,VasavadaSA,ShastriL,NathV,VasavadaAR,SrivastavaS. Visualoutcomescomparingemmetropiavsmini-monovisionafter bilateralimplantationofanondiffractiveextendedvisionintraocularlens: randomizedtrial.JournalofCataract&RefractiveSurgery. 2024;50(8):799–804. Cataract-onlypatientsoutcome datawerenotextractable.* MartínezdeCarneros-LlorenteA,MartínezdeCarnerosA,Martínezde Carneros-LlorenteP,Jiménez-AlfaroI.Comparisonofvisualqualityand subjectiveoutcomesamong3trifocalintraocularlensesand1bifocal intraocularlens.JournalofCataract&RefractiveSurgery. 2019;45(5):587-594. AllIOLsbelongedtothe multifocalclass. AsenaBS.Visualandrefractiveoutcomes,spectacleindependence,and visualdisturbancesaftercataractorrefractivelensexchangesurgery: comparisonof2trifocalintraocularlenses.JournalofCataract& RefractiveSurgery.2019;45(11):1539-1546. Bothstudygroupsreceived trifocalIOLs. WendelsteinJA,VallottonK,ZiörjenA,MüllerM,RiazKM,Pantanelli SM,LangenbucherA,SeilerTG.ReadingSpeedandVisualAcuityin PhotopicandMesopicConditionsAfterBilateralImplantationof DiffractiveMultifocalIntraocularLenses.AmericanJournalof Bothgroupsreceiveddiffractive multifocalIOLs. 68 Ophthalmology.2026;281:516–525. SchojaiM,SchultzT,JerkeC,BöckerJ,DickHB.Visualperformance comparisonof2extendeddepth-of-focusintraocularlenses.Journalof Cataract&RefractiveSurgery.2020;46(3):388–393. Bothgroupsevaluated EDOF-basedIOLs. FerreiraTB,RibeiroFJ.Prospectivecomparisonofclinicalperformance andsubjectiveoutcomesbetweentwodiffractivetrifocalintraocularlenses inbilateralcataractsurgery.JournalofRefractiveSurgery. 2019;35(7):418-425. Bothstudygroupsreceived trifocalIOLs. RibeiroFJ,FerreiraTB.Comparisonofvisualandrefractiveoutcomesof 2trifocalintraocularlenses.JournalofCataract&RefractiveSurgery. 2020;46(5):694-699. Bothstudygroupsreceived diffractivetrifocalIOLs. XunY,WanW,JiangL,HuK.Crossedversusconventional pseudophakicmonovisionforhighmyopiceyes:aprospective, randomizedpilotstudy.BMCOphthalmology.2020;20:447. Enrolledhigh-myopiccataract patients. AngRE,PicacheGCS,RiveraMCR,LopezLRL,CruzEM.A comparativeevaluationofvisual,refractive,andpatient-reportedoutcomes ofthreeextendeddepthoffocus(EDOF)intraocularlenses.Clinical Ophthalmology.2020;14:2339-2351. Retrospectivecohortstudy. HienertJ,StjepanekK,HirnschallN,RuissM,ZwicklH,FindlO.Visual performanceoftwodiffractivetrifocalintraocularlenses:arandomized trial.JournalofRefractiveSurgery.2021;37(7):460-465. Bothstudygroupsreceived diffractivetrifocalIOLs. KılıçA,ÖzpınarA,TanrıverdiC.Comparisonofvisualoutcomesoftwo trifocalIOLs.JournalofRefractiveSurgery.2023;39(8):524-530. Bothstudygroupsreceived diffractivetrifocalIOLs. DeRojasJO,SandovalHP,PotvinR,SolomonKD.Comparisonofvisual andsubjectiveoutcomesbetweentwoEDOFintraocularlenses,with emphasisonmesopicperformance.JournalofCataract&Refractive Surgery.2026;52(6):574-581. Bothstudygroupsreceived EDOFIOLs. Note:The6articlesmarkedwithasteriskswereexcludedafterreviewbyhumanexperts. 69 SupplementaryInformation4.GRADEtables(prospectivestudy,GUIDE-assistedgroup) (1)MultifocalvsMonofocal Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(lowerfavoursmultifocal) 9(Chang2023;elMaghraby 1992;Steinert1992;Percival 1993;Rossetti1994; Haaskjold1998;Kamlesh 2001;Leyland2002;Sen 2004) RCTSeriousSeriousNotseriousSerious953R1.09(0.75to 1.58) Verylow Uncorrecteddistancevisualacuityworsethan6/6(lowerfavoursmultifocal) 9(Chang2023;elMaghraby 1992;Steinert1992;Percival 1993;Rossetti1994; Haaskjold1998;Leyland 2002;Sen2004;Jusufovic 2011) RCTSeriousSeriousNotseriousNotserious943R1.01(0.89to 1.14) Low UncorrectednearvisualacuityworsethanJ3/J4oranequivalentthreshold(lowerfavoursmultifocal) 9(Chang2023;elMaghraby 1992;Percival1993;Rossetti 1994;Haaskjold1998;Javitt 2000;Leyland2002;Jusufovic 2011;Ji2013) RCTSeriousSeriousNotseriousNotserious1,043R0.18(0.08to 0.41) Low Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 9(Chang2023;Messias2021; Law2020;Leyland2002; Nijkamp2004;Harman2008; Palmer2008;Rasp2012;Peng 2012) RCTSeriousNotseriousNotseriousNotserious1,265MD0.03(0.01to 0.04) Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 4(Chang2023;Law2020; Maxwell2017;Peng2012) RCTSeriousSeriousNotseriousSerious866MD−0.12(−0.29to 0.05) Verylow Meancorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 9(Chang2023;Law2020; Javitt2000;Leyland2002; RCTSeriousVeryseriousNotseriousSerious1,667MD−0.15(−0.34to 0.05) Verylow 70 Palmer2008;Harman2008; Rasp2012;Peng2012; Maxwell2017) Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 9(Chang2023;Messias2021; Law2020;Leyland2002; Nijkamp2004;Palmer2008; Harman2008;Peng2012; Rasp2012) RCTSeriousNotseriousNotseriousNotserious1,265MD0.02(−0.01to 0.05) Moderate Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 3(Chang2023;Maxwell 2017;Peng2012) RCTSeriousSeriousNotseriousSerious776MD−0.11(−0.27to 0.05) Verylow Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 9(Chang2023;Messias2021; Law2020;Javitt2000; Leyland2002;Harman2008; Peng2012;Rasp2012; Maxwell2017) RCTSeriousSeriousNotseriousNotserious1,559MD−0.25(−0.39to −0.12) Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Anyspectacledependence(lowerfavoursmultifocal) 15(Chang2023;Han2025; Law2020;Gil2022;Steinert 1992;Percival1993;Rossetti 1994;Haaskjold1998;Javitt 2000;Leyland2002;Harman 2008;Cillino2008;Zhao 2010;Peng2012;Maxwell 2017) RCTSeriousSeriousNotseriousNotserious1,899R0.52(0.39to 0.68) Low Contrastsensitivity(higherfavoursmultifocal) 4(Harman2008;Leyland 2002;Rossetti1994;Sen 2004) RCTSeriousNotseriousNotseriousSerious288SMD−0.19(−0.57to 0.19) Low Visualfunction(higherfavoursmultifocal) 6(Han2025;Law2020; Cillino2008;Nijkamp2004; RCTSeriousSeriousNotseriousSerious663SMD0.78(−0.09to 1.65) Verylow 71 Sen2004;Zhao2010) Patientsatisfaction(numberofsatisfiedparticipants)(higherfavoursmultifocal) 3(Chang2023;Hida2024; Law2020) RCTSeriousSeriousNotseriousSerious403R2.41(0.47to 12.46) Low Patientsatisfaction(questionnairescore)(higherfavoursmultifocal) 6(Cillino2008;Nijkamp 2004;Peng2012;Sen2004; Steinert1992;Zhao2010) RCTSeriousVeryseriousNotseriousVeryserious643SMD0.26(−0.37to 0.89) Verylow Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(lowerfavoursmultifocal) 12(Chang2023;Han2025; Hida2024;Gil2022; Percival1993;Rossetti1994; Haaskjold1998;Kamlesh 2001;Sen2004;Cillino 2008;Harman2008; Maxwell2017) RCTSeriousSeriousNotseriousSerious1,405R1.34(0.98to 1.83) Verylow Halos(lowerfavoursmultifocal) 13(Chang2023;Han2025; Hida2024;Gil2022;Law 2020;Cillino2008; Haaskjold1998;Kamlesh 2001;Maxwell2017; Percival1993;Rossetti1994; Sen2004;Zhao2010) RCTSeriousSeriousNotseriousNotserious1,614R4.06(2.19to 7.55) Low Starbursts(lowerfavoursmultifocal) 2(Chang2023;Han2025)RCTSeriousNotseriousNotseriousVery serious 362R6.40(0.00to 33991.66) Verylow Nd:YAGposteriorcapsulotomy(lowerfavoursmultifocal) 1(Law2020)RCTSeriousN/ANot serious* Very serious 180R2.55(0.68to 9.55) Verylow (2)EDOFvsMonofocal Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality 72 Correcteddistancevisualacuityworsethan6/6(lowerfavoursEDOF) 2(Black2024;Chang 2022) RCTSeriousSeriousNotseriousVeryserious415R0.80(0.00, 5055.39) Verylow Correctedintermediatevisualacuitymeetingthestudy-definedthreshold(higherfavoursEDOF) 2(McCabe2022; Pantanelli2023) RCTSeriousNotseriousNotseriousNotserious269R2.91(0.43,19.82)Moderate Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 5(Reinhard2021;Bala 2022;Corbett2024; Black2024;Pantanelli 2023) RCTSeriousNotseriousNotseriousNotserious764MD0.02(−0.01, 0.05) Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 4(Reinhard2021;Bala 2022;Corbett2024; Pantanelli2023) RCTSeriousSeriousNotseriousSerious644MD−0.15(−0.28, −0.02) Verylow Meancorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 4(Reinhard2021;Bala 2022;Corbett2024; Pantanelli2023) RCTSeriousSeriousNotseriousNotserious644MD−0.13(−0.24, −0.01) Low Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 6(McCabe2022; Reinhard2021;Bala 2022;Skov2026;Black 2024;Pantanelli2023) RCTSeriousNotseriousNotseriousNotserious893MD0.03(−0.00, 0.07) Moderate Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 5(McCabe2022; Reinhard2021;Bala 2022;Skov2026; Pantanelli2023) RCTSeriousSeriousNotseriousSerious833MD−0.12(−0.21, −0.02) Verylow Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 5(McCabe2022; Reinhard2021;Bala 2022;Skov2026; Pantanelli2023) RCTSeriousSeriousNotseriousSerious833MD−0.16(−0.24, −0.08) Verylow Uncorrecteddistancevisualacuityworsethan6/6(lowerfavoursEDOF) 2(Black2024;ChangRCTSeriousNotseriousNotseriousSerious355R1.15(0.32,4.12)Low 73 2022) UncorrectednearvisualacuityworsethanJ3/J4orequivalent(lowerfavoursEDOF) 1(Chang2022)RCTSeriousN/ANotseriousNotserious295R0.11(0.05,0.25)Moderate Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality BinocularVisualFunctionquestionnairescore(scaledirectionnotreported)(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious92MD−0.16(−0.41, 0.09) Low Contrastsensitivity(higherfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousNotserious92MD−0.01(−0.06, 0.04) Moderate Meanstereoacuity(arcsec)(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVery Serious 92MD29.0(−15.4, 73.4) VeryLow Lightdistortionindex(lowerfavoursEDOF) 1(Guarro2022)RCTSeriousN/ANotserious*Serious88MD9.33(3.81, 14.85) Low Patientsatisfaction(higherfavoursEDOF) 2(Skov2026;Black 2024) RCTSeriousSeriousNotseriousSerious152R1.17(0.09,14.57)Verylow Spectacledependence,any(lowerfavoursEDOF) 8(McCabe2022;Bala 2022;Skov2026;Black 2024;Pantanelli2023; Chang2022;Palkovits 2026;Gil2022) RCTSeriousSeriousNotseriousSerious985R0.70(0.59,0.85)Verylow Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Blurredvision(lowerfavoursEDOF) 3(McCabe2022;Pantanelli 2023;Chang2022) RCTSeriousNotseriousNotseriousVeryserious565R0.50(0.12,2.06)Verylow Doublevision(lowerfavoursEDOF) 1(McCabe2022)RCTSeriousN/ANotseriousVeryserious219R2.13(0.20, 23.17) Verylow Glare(lowerfavoursEDOF) 9(McCabe2022;Bala2022;RCTSeriousNotseriousNotseriousSerious1,133R0.93(0.73,1.19)Low 74 Skov2026;Corbett2024; Pantanelli2023;Guarro 2022;Chang2022;Palkovits 2026;Gil2022) Halos(lowerfavoursEDOF) 10(McCabe2022;Gil2022; Bala2022;Skov2026; Corbett2024;Black2024; Pantanelli2023;Guarro 2022;Chang2022;Palkovits 2026) RCTSeriousSeriousNotseriousNotserious1,192R2.08(1.16,3.72)Low Hazyvision(lowerfavoursEDOF) 2(McCabe2022;Pantanelli 2023) RCTSeriousNotseriousNot serious* Veryserious270R0.78(0.01, 83.56) Verylow Negativedysphotopsia(lowerfavoursEDOF) 1(McCabe2022)RCTSeriousN/ANotseriousVeryserious219R0.68(0.27,1.69)Verylow Starbursts(lowerfavoursEDOF) 8(McCabe2022;Bala2022; Skov2026;Corbett2024; Black2024;Pantanelli2023; Chang2022;Palkovits 2026) RCTSeriousSeriousNotseriousVeryserious1,055R0.97(0.59,1.60)Verylow (3)MonovsionvsMonofocal Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousSerious137MD0.08(0.04to0.11)Low Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousSerious137MD−0.07(−0.11to−0.02)Low Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD−0.22(−0.27to−0.16)Moderate Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Contrastsensitivity(Pelli–RobsonlogCS)(higherfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD−0.01(−0.06to0.04)Moderate 75 Patientsatisfaction(higherfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious137R1.00(0.79to1.26)Verylow Spectacledependence(any)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137R0.72(0.62to0.84)Moderate Visualfunction(Catquest-7SF)(higherfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD−0.01(−0.26to0.23)Moderate Adverseevents No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious137R0.98(0.74to1.30)Verylow Halos(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousSerious137R1.25(0.78to2.01)Low Meananiseikonia,%(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious*137MD0.66(0.11to1.22)Moderate * Starbursts(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious137R0.97(0.64to1.49)Verylow (4)MultifocalvsEDOF Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(lowerfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious118R0.90(0.28,2.89)Low Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 2(AkkayaTurhan 2021;Gil2020) RCTSeriousNotseriousNotseriousNotserious145MD0.03(-0.22,0.28)Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 2(AkkayaTurhan 2021;Gil2020) RCTSeriousNotseriousNotseriousSerious145MD0.12(-0.09,0.34)Low Meancorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 2(AkkayaTurhan 2021;Gil2020) RCTSeriousVeryseriousNotseriousVeryserious145MD-0.02(-1.20,1.15)Verylow Meanuncorrecteddistancevisualacuity(decimal)(higherfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousNotserious118MD0.02(-0.05,0.10)Moderate Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 3(Sevik2022;RCTSeriousNotseriousNotseriousNotserious116MD-0.03(-0.08,0.03)Moderate 76 Webers2020; Labiris2024) Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 3(Sevik2022; Webers2020; Labiris2024) RCTSeriousNotseriousNotseriousNotserious116MD0.03(-0.04,0.10)Moderate Meanuncorrectedintermediatevisualacuity(decimal)(higherfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousNotserious118MD-0.05(-0.11,0.02)Moderate Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 3(Sevik2022; Webers2020; Labiris2024) RCTSeriousVeryseriousNotseriousSerious116MD-0.03(-0.27,0.20)Verylow Meanuncorrectednearvisualacuity(decimal)(higherfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious118MD0.09(0.03,0.16)Low Uncorrecteddistancevisualacuityworsethan6/6(lowerfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious147R1.19(0.76,1.85)Low UncorrectednearvisualacuityworsethanJ3/J4orequivalent(lowerfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious118R0.34(0.15,0.77)Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Maximumreadingspeed,words/min(higherfavoursmultifocal) 1(Akkaya Turhan2021) RCTSeriousN/ANotseriousVeryserious29MD-5.00(-26.87,16.87)Verylow Readingacuity(logMAR)(lowerfavoursmultifocal) 1(Akkaya Turhan2021) RCTSeriousN/ANotserious*Notserious29MD0.01(-0.04,0.06)Moderate Meanstereoacuity,arcseconds(lowerfavoursmultifocal) 1(Pradeep2026)RCTSeriousN/ANotserious*Serious70MD29.14(7.27,51.01)Low Visualfunction(unspecifiedinstruments)(higherfavoursmultifocal) 3(Sevik2022; Webers2020; Labiris2024) RCTSeriousNotseriousNotseriousVeryserious116SMD0.00(-0.80,0.80)Verylow Patientsatisfaction(score,0–100scale)(higherfavoursmultifocal) 1(Sevik2022)RCTSeriousN/ANotseriousVeryserious29MD-1.71(-10.67,7.25)Verylow Patientsatisfaction(binary)(higherfavoursmultifocal) 1(Webers2020)RCTSeriousN/ANotseriousSerious27R0.99(0.80,1.23)Low 77 Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(lowerfavoursmultifocal) 2(Webers2020;Gil 2022) RCTSeriousNotseriousNotseriousSerious137R1.16(0.13,10.63)Low Halos(lowerfavoursmultifocal) 2(Webers2020;Gil 2022) RCTSeriousSeriousNotseriousVeryserious138R1.05(0.02,64.58)Verylow Overallphoticphenomena(lowerfavoursmultifocal) 1(Sevik2022)RCTSeriousN/ANotseriousVeryserious29R4.69(0.24,89.88)Verylow (5)MultifocalvsMonovision Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 2(Labiris2024； Wilkins2013) RCTSeriousNotseriousNotseriousNotserious246MD0.04(-0.15,0.23)Moderate Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 2(Labiris2024； Wilkins2013) RCTSeriousNotseriousNotseriousNotserious241MD0.06(-0.14,0.26)Moderate Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 2(Labiris2024； Wilkins2013) RCTSeriousSeriousNotseriousNotserious246MD−0.00(−0.57,0.56)Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Visualfunction(lowerfavoursmultifocal) 1(Labiris2024)RCTSeriousN/ASeriousSerious60MD−3.81(−8.64,1.02)Verylow Visualfunction–VF-14(higherfavoursmultifocal) 1(Labiris2015)RCTSeriousN/ANotseriousNotserious75MD−1.47(−5.51,2.57)Moderate Spectacledependence–any(lowerfavoursmultifocal) 2(Wilkins2013; Labiris2015) RCTSeriousNotseriousNotseriousNotserious262R0.40(0.06,2.59)Moderate Contrastsensitivity–Pelli-Robsontest(higherfavoursmultifocal) 2(Wilkins2013; Labiris2015) RCTSeriousNotseriousNotseriousNotserious262MD−0.03(−0.47,0.41)Moderate Adverseevents 78 No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(lowerfavoursmultifocal) 1(Wilkins2013)RCTSeriousN/ANotseriousNotserious187R1.41(1.14,1.73)Moderate (6)EDOFvsMonovision Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(lowerfavoursEDOF) 1(Vandekerckhove 2026) RCTSeriousN/ANotseriousSerious105R5.22(1.07,25.55)Low Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 2(Vandekerckhove 2026;Zeilinger2025) RCTSeriousNotseriousNotseriousSerious185MD0.04(−0.15,0.23)Low Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 2(Vandekerckhove 2026;Zeilinger2025) RCTSeriousNotseriousNotseriousSerious185MD−0.10(−0.33, 0.12) Low Meancorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 2(Vandekerckhove 2026;Zeilinger2025) RCTSeriousSeriousNotseriousSerious185MD−0.09(−0.95, 0.77) Verylow Meanstereoacuity(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious131MD10.39(−34.01, 54.78) Low Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 5(Vandekerckhove 2026;Zeilinger2025; Sandoval2020;Labiris 2024;Skov2026) RCTSeriousNotseriousNotseriousNotserious452MD0.03(−0.02,0.08)Moderate Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 5(Vandekerckhove 2026;Zeilinger2025; Sandoval2020;Labiris 2024;Skov2026) RCTSeriousNotseriousNotseriousNotserious452MD−0.01(−0.06, 0.04) Moderate Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 5(Vandekerckhove 2026;Zeilinger2025; Sandoval2020;Labiris RCTSeriousVeryseriousNotseriousVery serious 452MD0.01(−0.17,0.19)Verylow 79 2024;Skov2026) Uncorrecteddistancevisualacuityworsethan6/6(lowerfavoursEDOF) 1(Vandekerckhove 2026) RCTSeriousN/ANotseriousSerious105R1.70(0.92,3.11)Low UncorrectednearvisualacuityworsethanJ3/J4orequivalent(lowerfavoursEDOF) 1(Vandekerckhove 2026) RCTSeriousN/ANotseriousSerious105R0.25(0.08,0.77)Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Visualfunction(VF-14)(higherfavoursEDOF) 1(Labiris2024)RCTSeriousN/ANotseriousNotserious60MD−5.59(−10.20, −0.98) Moderate Patientsatisfaction(higherfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious131R0.97(0.76,1.25)Low Spectacledependence(any)(lowerfavoursEDOF) 2(Vandekerckhove 2026;Skov2026) RCTSeriousNotseriousNotseriousSerious312R1.07(0.40,2.87)Low Contrastsensitivity(higherfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousNotserious131MD0.001(−0.044, 0.045) Moderate Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious131R1.08(0.82,1.43)Verylow Halos(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious131R0.91(0.57,1.44)Verylow Starbursts(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious131R1.29(0.87,1.90)Low (7)TrifocalmultifocalIOLvsBifocalmultifocalIOL Visualacuity NoofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNoofparticipantsEffectsize(95%CI)Quality Meancorrecteddistancevisualacuity(lowerfavourstrifocal) 3(Li2023, Gundersen2016, Jonker2015) RCTSeriousNotseriousNotseriousNotserious149MD0.00(-0.07,0.08)Moderate 80 Meancorrectedintermediatevisualacuity(lowerfavourstrifocal) 2(Li2023, Jonker2015) RCTSeriousNotseriousNotseriousSerious127MD-0.03(-0.30, 0.24) Low Meancorrectednearvisualacuity(lowerfavourstrifocal) 2(Li2023, Jonker2015) RCTSeriousNotseriousNotseriousNotserious127MD0.01(-0.26,0.29)Moderate Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavourstrifocal) 3(Li2023, Gundersen2016, Jonker2015) RCTSeriousNotseriousNotseriousNotserious149MD-0.01(-0.10, 0.07) Moderate Meanuncorrectednearvisualacuity(lowerfavourstrifocal) 2(Li2023, Jonker2015) RCTSeriousNotseriousNotseriousSerious127MD-0.00(-0.41, 0.41) Low Visualfunction NoofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNoofparticipantsEffectsize(95%CI)Quality Spectacledependence–near(lowerfavourstrifocal) 1(Jonker2015)RCTSeriousN/ANotseriousVeryserious28R0.65(0.18,2.38)Verylow Note:The6qualityindirectnessgradingandcorrespondingoverallevidencecertaintylevel markedwithasterisksweredeterminedbyhumanexperts. 81 SupplementaryInformation5.Evidencetorecommendations(prospectivestudy, GUIDE-assistedgroup) Relativevalueof differentoutcomes Thecommitteeagreedthatthechoiceofintraocularlensshouldbeassessedusingoutcomes thatreflecttheoverallexperienceofvisionratherthananysinglemeasureofvisualacuity. Patient-reportedvisualfunction,qualityoflifeandsatisfactionwereconsideredparticularly importantbecausetheyintegratetheabilitytoperformeverydayactivities,dependenceon spectaclesandtheacceptabilityofunwantedvisualsymptoms.Theneedforadditional treatment,seriousocularcomplicationsandthefinancialconsequencesforpatientsand healthcareserviceswerealsoconsideredimportant. Thecommitteerecognisedthattherelativeimportanceoftheseoutcomesvariesbetween stakeholders.Patientsmayplacedifferentvaluesonspectacleindependence,night-time visualqualityandexpenditure;familiesandcaregiversmayvaluethepatient’sfunctional independence;cliniciansmayprioritisepredictableoutcomesandtheavoidanceofdifficult postoperativeadaptation;andhealthcareprovidersmustalsoconsideraffordability,equity andtheconsequencesofadditionalfollow-uportreatment. Trade-offbetween benefitsandharms Thecommitteeagreedthatnoneoftheevaluatedoptionswasconsistentlysuperioracrossall relevantoutcomes.Eachoptionrepresentsadifferentbalancebetweenrangeofvision, opticalquality,unwantedvisualphenomenaandpredictabilityofthepostoperativeresult. Monofocallensesgenerallyprovidepredictabledistancevisionandareassociatedwithfewer concernsaboutglare,halosandotheropticaldisturbances.Theirprincipaldisadvantageis thatpatientsaremorelikelytorequirespectaclesforintermediateandnearactivities. Multifocallensesgenerallyofferthegreatestpotentialfornearvisionandspectacle independence.Thesebenefitsmustbebalancedagainstagreaterlikelihoodofhalos, starburstsorotherunwantedvisualphenomenaand,insomepatients,asmallreductioninthe clarityofdistancevision.Theymayremaintroublesomeforpeoplewhofrequentlydriveat night,workinlow-contrastenvironmentsorhaveparticularlyhighexpectationsofoptical quality.EDOFlensesmayprovideabroaderandmorecontinuousrangeofvisionthan monofocallenses,particularlyforintermediateactivities,whilegenerallyretaininguseful distancevision.However,theydonotguaranteespectacleindependenceforneartasks,and 82 theavailabledirectevidencedoesnotconsistentlydemonstratethattheyproducefewervisual disturbancesthanmultifocallenses.Monovisionmayimprovefunctionalintermediateand nearvisionwithoutusingpresbyopia-correctingoptics.However,itintroducesadeliberate differencebetweenthetwoeyesandmayadverselyaffectbinocularbalance,depth perceptionordistancevisualquality.Toleranceisindividualandmaybemorepredictablein peoplewhohavepreviouslyusedandacceptedmonovision.Thecommitteeconsideredthat monovisionshouldbeviewedasarefractivestrategyrequiringcarefulpatientselectionrather thanasadirectequivalentofaparticularlensdesign. Thehead-to-headevidencedidnotidentifyaclearoverallhierarchyamongdifferent strategies.Similarly,theevidencecomparingtrifocalandbifocallensesdidnotdemonstratea sufficientlyconsistentoveralladvantagetosupportonedesignforallpatients.Thecommittee emphasisedthattheabsenceofastatisticallysignificantdifferenceshouldnotbeinterpreted asevidencethattheoptionsareequivalent,particularlywherestudiesweresmallor imprecise.Thebalanceofbenefitsandharmsisthereforestronglypreferencesensitive. Overall,thecommitteeconsideredthattheevidencesupportsindividualizedselectionand informedshareddecision-makingratherthanroutinepreferenceforasingleoption.Patients shouldunderstandthatimprovedrangeofvisionandreducedspectacledependencemaybe obtainedatthecostofopticaltrade-offs,andthatnooptioncanguaranteecompletespectacle independenceorfreedomfromunwantedvisualsymptoms. Considerationof healthbenefitsand resourceuse Thecommitteenotedthattheacquisitioncostofthelens,theamountreimbursedbyapayer andtheamountultimatelypaidbythepatientaredifferentquantities.Nevertheless,they illustratethatlensesintendedtoreducepresbyopiaandspectacledependencecommonly remainsubstantiallymoreexpensivethanstandardmonofocallenses,andthattheadditional costmayfallonthehealthcaresystemortheindividualpatientdependingonthecountry. InChina,nationalvolume-basedprocurementhassubstantiallyreducedtheacquisitionprices ofartificiallenses.TheNationalHealthcareSecurityAdministrationreportedanaverage reductionofapproximately60%inthe2023nationalprocurementround.Theprincipal economicbenefitofmultifocalandotherpresbyopia-correctinglensesisthepotential reductioninfuturespectacleuse,togetherwithconvenienceandpossiblegainsinfunctional 83 independence.Thesebenefitsaccrueovertime,whereastheadditionallensorprocedurecost isgenerallyincurredatthetimeofsurgery.Thevalueoftheadditionalexpenditureis thereforelikelytobegreaterforpatientswhoplaceahighvalueonspectacleindependence, havealongerexpecteddurationofbenefitandarewillingtoaccepttheassociatedvisual trade-offs. Publishedeconomicmodelsillustratethisuncertainty.AUnitedStatespatient-perspective modelestimatedthatbilateraltrifocalimplantationproduced0.67additionalQALYsatan incrementallifetimepatientcostofUS$2,783,whilereducinglifetimespectacleexpenditure byUS$2,593.However,theresultswerehighlysensitivetotheassumedburdenofwearing glasses,patientout-of-pocketcostsandpostoperativespectacledependence.ASouthKorean modelestimatedanadditional0.60QALYsatanincrementalpatientcostofUS$4,274,but theestimatednetbenefitdecreasedsubstantiallywhenassumptionsaboutspectacleuseor willingnesstopaywerechanged.Bothstudiesweresponsoredbythemanufacturerofthe evaluatedlensesandwerebasedonmodellingassumptions,theirconclusionsshould thereforebeinterpretedcautiously.Whereadvancedlensesarepredominantlyself-funded, routinerelianceonabilitytopaymayincreaseinequityinaccess.Procurement-relatedprice reductionsimproveaffordabilitybutdonot,bythemselves,establishthattheadditional benefitsjustifyroutineuseforallpatients. Thecommitteethereforeconsideredthatthepricepremiumshouldbeassessedagainstthe benefitsthatareimportanttotheindividualpatientandagainstlocalopportunitycosts.The presentevidencedoesnotsupportassumingthatamoreexpensivelensnecessarilyprovides proportionatelygreateroverallhealthbenefit. Qualityofevidence Thecommitteenotedthatthecertaintyoftheevidencewaspredominantlylow,with moderate-certaintyevidenceavailableforonlyselectedcomparisonsandoutcomes. Importantlimitationsincludedriskofbias,imprecision,smallstudiesandlimiteddirect comparisonsbetweennewerlensdesigns.Evidenceonpatient-reportedqualityoflife, dissatisfaction,seriouscomplicationsandtheneedforadditionaltreatmentwaslesscomplete thanevidenceonvisualacuityorspectacleuse. Theresultsmayalsobeinfluencedbydifferencesinlensgeneration,surgicaltechnique, 84 refractivetargets,patientselectionandoutcomemeasurement.Consequently,theevidenceis sufficienttodescribethemaintrade-offsbutnottoestablisharobustoverallrankingofall availableoptions. Otherconsiderations Thecommitteeconsideredthatacceptabilitywoulddependheavilyonthequalityof preoperativecounsellingandwhetherthepostoperativeresultmatchedthepatient’s expectations.Patientsshouldreceivebalancedinformationaboutthelikelyrangeofvision, possiblecontinuedneedforspectacles,unwantedvisualphenomena,uncertaintyinthe evidenceandthecompletepersonalcostofeachoption. Implementationwouldrequireappropriatepatientselection,accesstoreliablebiometry, cliniciansexperiencedwiththerelevantrefractivestrategiesandsystemsformonitoring dissatisfaction,additionalproceduresandlensexchange.Providersshouldavoidpresenting premiumlensesasinherentlysuperiorandshouldensurethatfinancialincentivesdonot replacepreference-sensitiveshareddecision-making. Furtherresearchshouldprioritiseadequatelypowereddirectcomparisonsusingstandardised patient-reportedoutcomes,longerfollow-upandconsistentreportingofvisualsymptoms, additionaltreatmentandcosts. 85 SupplementaryInformation6.Recommendations(prospectivestudy,GUIDE-assisted group) Question:Whatistheoptimalstrategytofacilitatesimultaneousdistanceandnearvisionfollowingcataractsurgery? Recommendationsin2017NICEguideline (NG77:Cataractsinadults:management) RecommendationsupdatedbyGUIDE Donotoffermultifocalintraocularlensesforpeople havingcataractsurgery. Offermonovisionforuseaftercataractsurgerytopeople whohaveeitheranisometropiaormonovision preoperativelyandwouldliketoremainwithit. Considermultifocalintraocularlensesforpeoplewhowantgreater spectacleindependencefordistance,intermediateandnearvision. Whendiscussingthisoption: ensuretheperson’socularhealthissuitableformultifocal intraocularlenses explainthepotentialbenefitsinuncorrectednearand intermediatevisioncomparedwithstandardmonofocal intraocularlenses discusstheincreasedriskofvisualdisturbances(suchas halos,glareandstarbursts) takeintoaccounttheperson’svisualneedsanddailyactivities considercostimplicationsandtheperson’swillingnesstopay 86 SupplementaryInformation7.PICOcriteria(prospectivestudy,human-onlygroup) Generationbyhumanexpert PopulationAdults(18yearsandolder)undergoingphacoemulsificationcataract surgeryandintraocularlens(IOL)implantationinoneorbotheyes InterventionsNon-accommodatingintraocularlensdesignedtosupportboth nearanddistancevision(e.g.,multifocallenses, extended-depth-of-focuslensesormonovisioncorrectionstrategy) Standardmonofocalintraocularlenses ComparatorsAlllistedinterventionscomparedwithoneanother Outcomes  Distance,intermediate,andnearvisualacuity  Patientsatisfaction  Spectacleindependence  Incidenceofvisualdisturbances  Qualityoflife  Resourceuseandcosts 87 SupplementaryInformation8.Summaryofincludedstudies(prospectivestudy,human-only group) Study&locationSamplesizeMethods Cochener(2018) France 120eyes (60patients) ComparisonofAcrySofIQPanOptixandFineVision MicroFtrifocalIOLswiththeTECNISSymfony EDOFIOL. Gil(2020) Spain 232eyes (116patients) ComparisonofAcrySofReSTORSV25T0,ATLISA 809M,TECNISZKB00andZLB00bifocalIOLsand theATLISAtri839MPtrifocalIOLwiththeTECNIS SymfonyZXR00EDOFIOL. Law(2020) UnitedKingdom 180eyes (90patients) ComparisonoftheBi-Flex677MYapodized diffractivemultifocalIOLwiththeBi-Flex677AB asphericaberration-neutralmonofocalIOL. Sandoval(2020) USA 152eyes (76patients) ComparisonofbilateralTECNISSymfonytoricEDOF IOLstargetedforbinocularemmetropiaor mini-monovision. Webers(2020) TheNetherlands 54eyes (27patients) ComparisonoftheTECNISSymfony(ZXR00)EDOF IOLusingmicromonovisionwiththeATLISAtri 839MPtrifocalIOL. Messias(2021) Brazil 66eyes (33patients) ComparisonoftheAcrySofReSTORSN6AD1 apodizeddiffractivemultifocalIOLwiththeAcrySof IQSN60WFasphericmonofocalIOL. Reinhard(2021) Germany,Belgium, Portugal,Spain,France, andItaly 422eyes (211patients) ComparisonoftheATLARA829MPandTECNIS SymfonyEDOFIOLswiththeCTASPHINA409MP monofocalIOL. Turhan(2021)/Sevik (2022) Turkey 58eyes (29patients) ComparisonoftheTECNIS+2.75D(ZKB00)low-add bifocalIOLwiththeTECNISSymfony(ZXR00) EDOFIOL,bothusingmini-monovision. 88 Bala(2022) Australia,Canada,Spain, andtheUnitedKingdom 540eyes (270patients) ComparisonoftheAcrySofIQVivity(DFT015) non-diffractiveEDOFIOLwiththeAcrySofIQ SN60WFasphericmonofocalIOL. Chang(2022) USA 597eyes (299patients) ComparisonofbilateralTECNISSymfony(ZXR00) diffractiveEDOFIOLswithTECNIS1-piece(ZCB00) asphericmonofocalIOLs,bothtargetedfor emmetropia. Guarro(2022) Spain 176eyes (88patients) ComparisonoftheAcrySofIQVivitynon-diffractive EDOF,ATLARA829MPandTECNISSymfony ZXR00diffractiveEDOFIOLswiththeAcrySofIQ SN60WFmonofocalIOL. McCabe(2022) USA 435eyes (218patients) ComparisonoftheAcrySofIQVivity(DFT015) non-diffractiveEDOFIOLwiththeAcrySofIQ SN60WFasphericmonofocalIOL. Chang(2023) USA 544eyes (272patients) ComparisonoftheTECNISSynergy(ZFR00V)hybrid multifocal-EDOFIOLwiththeTECNISZCB00 monofocalIOL. Li(2023) China 99eyes (99patients) ComparisonoftheATLISAtritoric939M/MP trifocalIOLwiththeATLISAtoric909Mbifocal IOL. Pantanelli(2023) USA 102eyes (51patients) ComparisonoftheAcrySofVivity(DFT015) non-diffractiveEDOFIOLwiththeenVista(MX60E) neutralasphericmonofocalIOL,bothtargetedfor emmetropia. Corbett/Black(2024) AustraliaandNewZealand 235eyes (118patients) ComparisonoftheTECNISPureSee(ZEN00V)fully refractiveEDOFIOLwiththeTECNISEyhance (ICB00)enhancedmonofocalIOL. Hida(2024) Brazil 116eyes (58patients) ComparisonofbilateralTECNISSynergy(ZFR00V) hybridmultifocal-EDOFIOLswithblendedTECNIS 89 Eyhance(ICB00)enhancedmonofocalandTECNIS Synergyimplantation. Labiris(2024) Greece 180eyes (90patients) ComparisonofmixedAcrySofIQPanOptixtrifocal andAcrySofIQVivitynon-diffractiveEDOF implantationwithbilateralPanOptixorbilateralVivity implantation. Zeilinger(2025) Austria 80eyes (40patients) ComparisonoftheRayOneEMV(RAO200E) advancedmonofocalIOLwiththeAcrySofIQVivity (DFT015)non-diffractiveEDOFIOL,bothusing mini-monovision. Palkovits(2026) Austria 70eyes (35patients) ComparisonoftheAcunexVarioAN6Vrefractive EDOFIOLwiththeAcunexAN6monofocalIOL, bothtargetedat-0.25D. Pradeep(2026) India 140eyes (70patients) ComparisonofbilateralAcrySofIQPanOptixtrifocal IOLswithbilateralAcrySofIQVivitynon-diffractive EDOFIOLs. Skov(2026) Denmark 360eyes (180patients) Comparisonofemmetropia,mini-monovision,and monovisionusingbilateralAcrySofACU0T0or TECNISGCB00/ZCB00monofocalIOLswith bilateralAcrySofIQVivitynon-diffractiveEDOF IOLs. Vandekerckhove(2026) Switzerland 210eyes (105patients) ComparisonofIsopureandTECNISEyhance enhancedmonofocalIOLswiththeAcrySofIQVivity EDOFIOLusingmini-monovision. 90 SupplementaryInformation9.Excludedstudiesandreasons(prospectivestudy,human-only group) StudyReasonforexclusion HayashiK,YoshidaM,SasakiH,HirataA.Binocularvisualfunction ofmyopicpseudophakicmonovision.JapaneseJournalof Ophthalmology.2018;62:357–364. Interventiondidnotmeetthe prespecifiedpseudophakicmonovision definition. SchojaiM,SchultzT,JerkeC,BöckerJ,DickHB.Visual performancecomparisonof2extendeddepth-of-focusintraocular lenses.JournalofCataract&RefractiveSurgery. 2020;46(3):388–393. BothgroupsevaluatedEDOF-based IOLs. XunY,WanW,JiangL,HuK.Crossedversusconventional pseudophakicmonovisionforhighmyopiceyes:aprospective, randomizedpilotstudy.BMCOphthalmology.2020;20:447. Includedpatientswithhigh-myopic cataract. DickHB,AngRE,CorbettD,HoffmannP,TetzM,VillarrubiaA, PalominoC,Castillo-GomezA,TsaiL,ThomasEK,JanakiramanP. Comparisonof3-monthvisualoutcomesofanewmultifocal intraocularlensvsatrifocalintraocularlens.JournalofCataract& RefractiveSurgery.2022;48(11):1270–1276. Cataract-onlyoutcomedatawerenot separatelyextractable. GilMÁ,VarónC,CardonaG,BuilJA.FarandNearContrast SensitivityandQualityofVisionwithSixPresbyopiaCorrecting IntraocularLenses.JournalofClinicalMedicine.2022;11(14):4150. Participantswerenotrandomlyallocated. FernandesMC,NunomuraCY,MessiasA.Visualperformanceand photicdisturbanceswithdiffractiveandnondiffractiveEDOF intraocularlensesusingmini-monovision:randomizedtrial.Journal ofCataract&RefractiveSurgery.2024;50(2):153–159. BothgroupsreceivedEDOFIOLs. VasavadaV,VasavadaSA,ShastriL,NathV,VasavadaAR, SrivastavaS.Visualoutcomescomparingemmetropiavs mini-monovisionafterbilateralimplantationofanondiffractive extendedvisionintraocularlens:randomizedtrial.Journalof Cataract-onlyoutcomedatawerenot separatelyextractable. 91 Cataract&RefractiveSurgery.2024;50(8):799–804. HanD,MaX,GaoJ,XuZ,LiuF.Comparisonofclinicaleffectsof asphericalmonofocalandTecnisZMB00multifocalintraocular lensesinpatientswithage-relatedcataract.JournalofthePakistan MedicalAssociation.2025;75(8):1229–1234. Interventioncouldnotbereliably classified:thetitleidentifiestheTECNIS ZMB00,whereastheMethodsreportthe multifocalIOLasmodel52501TW manufacturedbyOphtec. NagyZZ,DormanP,SzalczerS,KissH.ProspectiveComparisonof ClinicalOutcomesAfterBilateralImplantationofDiffractive TrifocalExtendedDepth-of-FocusandDiffractiveTrifocal HydrophobicIntraocularLenses.JournalofRefractiveSurgery. 2025;41(2):e102–e113. Cataract-onlyoutcomedatawerenot separatelyextractable. WantenJC,BauerNJC,BerendschotTTJM,vandenBiggelaar FJHM,NuijtsRMMA.Twonovelextendeddepth-of-focus intraocularlensestargetedformini-monovision:Aprospective randomizedcontrolledtrial.AmericanJournalofOphthalmology. 2025;276:286–296. BothgroupsreceivedEDOFIOLs. WendelsteinJA,VallottonK,ZiörjenA,MüllerM,RiazKM, PantanelliSM,LangenbucherA,SeilerTG.ReadingSpeedand VisualAcuityinPhotopicandMesopicConditionsAfterBilateral ImplantationofDiffractiveMultifocalIntraocularLenses.American JournalofOphthalmology.2026;281:516–525. Bothgroupsreceiveddiffractive multifocalIOLs. 92 SupplementaryInformation10.GRADEtables(prospectivestudy,human-onlygroup) (1)MultifocalvsMonofocal Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(R<1favoursmultifocal) 9(elMaghraby1992; Steinert1992;Percival 1993;Rossetti1994; Haaskjold1998;Kamlesh 2001;Leyland2002;Sen 2004;Chang2023) RCTSeriousSeriousNotseriousSerious953R1.09(0.75to1.58)Verylow Uncorrecteddistancevisualacuityworsethan6/6(R<1favoursmultifocal) 9(elMaghraby1992; Steinert1992;Percival 1993;Rossetti1994; Haaskjold1998;Leyland 2002;Sen2004;Jusufovic 2011;Chang2023) RCTSeriousSeriousNotseriousNotserious943R1.01(0.89to1.14)Low UncorrectednearvisualacuityworsethanJ3/J4oranequivalentthreshold(R<1favoursmultifocal) 9(elMaghraby1992; Percival1993;Rossetti 1994;Haaskjold1998; Javitt2000;Leyland2002; Jusufovic2011;Ji2013; Chang2023) RCTSeriousSeriousNotseriousNotserious1,043R0.18(0.08to0.41)Low Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 8(Leyland2002;Nijkamp 2004;Harman2008; Palmer2008;Rasp2012; Peng2012;Law2020; Messias2021) RCTSeriousNotseriousNotseriousNotserious1,062MD0.03(0.01to0.05)Moderate Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 9(Leyland2002;Nijkamp 2004;Palmer2008; Harman2008;Peng2012; Rasp2012;Law2020; Messias2021;Chang2023) RCTSeriousNotseriousNotseriousNotserious1,265MD0.02(−0.01to 0.05) Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 4(Peng2012;Maxwell 2017;Law2020;Chang 2023) RCTSeriousSeriousNotseriousSerious866MD−0.12(−0.29to 0.05) Verylow Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 3(Peng2012;MaxwellRCTSeriousSeriousNotseriousSerious776MD−0.11(−0.27toVerylow 93 2017;Chang2023)0.05) Meancorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 9(Javitt2000;Leyland 2002;Palmer2008; Harman2008;Rasp2012; Peng2012;Maxwell2017; Law2020;Chang2023) RCTSeriousSeriousNotseriousNotserious1,667MD−0.15(−0.34to 0.05) Low Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 9(Javitt2000;Leyland 2002;Harman2008;Peng 2012;Rasp2012;Maxwell 2017;Law2020;Messias 2021;Chang2023) RCTSeriousSeriousNotseriousNotserious1,559MD−0.25(−0.39to −0.12) Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Contrastsensitivity(higherfavoursmultifocal) 4(Rossetti1994;Leyland 2002;Sen2004;Harman 2008) RCTSeriousNotseriousNotseriousSerious288SMD−0.19(−0.57to 0.19) Low Visualfunction(higherfavoursmultifocal) 5(Nijkamp2004;Sen 2004;Cillino2008;Zhao 2010;Law2020) RCTSeriousSeriousNotseriousSerious563SMD0.79(−0.36to 1.94) Verylow Spectacledependence(any)(R<1favoursmultifocal) 13(Steinert1992; Percival1993;Rossetti 1994;Haaskjold1998; Javitt2000;Leyland 2002;Harman2008; Cillino2008;Zhao2010; Peng2012;Maxwell 2017;Law2020;Chang 2023) RCTSeriousSeriousNotseriousNotserious1,647R0.54(0.40to0.74)Low Patientsatisfaction(numberofsatisfiedparticipants)(R>1favoursmultifocal) 3(Law2020;Chang 2023;Hida2024) RCTSeriousSeriousNotseriousSerious403R2.41(0.47to12.46)Verylow Patientsatisfaction(questionnairescore)(higherfavorsmultifocal) 6(Steinert1992; Nijkamp2004;Sen2004; RCTSeriousVeryseriousNotseriousVeryserious643SMD0.26(−0.37to 0.89) Verylow 94 Cillino2008;Zhao2010; Peng2012) Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(R<1favoursmultifocal) 12(Percival1993;Rossetti 1994;Haaskjold1998; Kamlesh2001;Sen2004; Cillino2008;Harman2008; Maxwell2017;Chang2023; Hida2024) RCTSeriousSeriousNotseriousSerious1,153R1.34(1.06to1.83)Verylow Halos(R<1favoursmultifocal) 11(Percival1993;Rossetti 1994;Haaskjold1998; Kamlesh2001;Sen2004; Cillino2008;Zhao2010; Maxwell2017;Law2020; Chang2023;Hida2024) RCTSeriousSeriousNotseriousNotserious1,362R3.70(1.89to7.23)Low Nd:YAGposteriorcapsulotomy(R<1favoursmultifocal) 1(Law2020)RCTSeriousN/ANotseriousVery serious 180R2.55(0.68to9.55)Verylow (2)EDOFvsMonofocal Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(R<1favoursEDOF) 2(Chang2022; Black2024) RCTSeriousSeriousNotseriousVery serious 415R0.80(0.21,3.08)Verylow Uncorrecteddistancevisualacuityworsethan6/6(R<1favoursEDOF) 2(Chang2022; Black2024) RCTSeriousNotseriousNotseriousSerious355R1.15(0.94,1.40)Low UncorrectednearvisualacuityworsethanJ3/J4orequivalent(R<1favoursEDOF) 1(Chang2022)RCTSeriousN/ANotseriousNotserious295R0.11(0.05,0.25)Moderate Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 5(Reinhard2021; Bala2022; Pantanelli2023; Corbett2024;Black 2024) RCTSeriousNotseriousNotseriousNotserious764MD0.02(−0.01,0.05)Moderate Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 95 6(Reinhard2021; McCabe2022;Bala 2022;Pantanelli 2023;Black2024; Skov2026) RCTSeriousNotseriousNotseriousNotserious893MD0.03(−0.00,0.07)Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 4(Reinhard2021; Bala2022; Pantanelli2023; Corbett2024) RCTSeriousSeriousNotseriousNotserious644MD−0.15(−0.28,−0.02)Low Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 5(Reinhard2021; McCabe2022;Bala 2022;Pantanelli 2023;Skov2026) RCTSeriousSeriousNotseriousSerious833MD−0.12(−0.21,−0.02)Verylow Meancorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 4(Reinhard2021; Bala2022; Pantanelli2023; Corbett2024) RCTSeriousSeriousNotseriousSerious644MD−0.13(−0.24,−0.01)Verylow Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 5(Reinhard2021; McCabe2022;Bala 2022;Pantanelli 2023;Skov2026) RCTSeriousSeriousNotseriousSerious833MD−0.16(−0.24,−0.08)Verylow Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Contrastsensitivity(higherfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousNotserious92MD−0.01(−0.06,0.04)Moderate Visualfunctionquestionnairescore(scaledirectionnotreported)(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious92MD−0.16(−0.41,0.09)Low Spectacledependence(any)(R<1favoursEDOF) 7(McCabe2022; Bala2022;Chang 2022;Pantanelli 2023;Black2024; Skov2026; Palkovits2026) RCTSeriousSeriousNotseriousSerious833R0.72(0.60,0.88)Verylow Patientsatisfaction(R>1favoursEDOF) 2(Black2024; Skov2026) RCTSeriousSeriousNotseriousSerious152R1.17(0.80,1.73)Verylow 96 Meanstereoacuity(arcsec)(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVery Serious 92MD29.0(−15.4,73.4)VeryLow Lightdistortionindex(lowerfavoursEDOF) 1(Guarro2022)RCTSeriousN/ANotseriousSerious88MD9.33(3.81,14.85)Low Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(R<1favoursEDOF) 8(McCabe2022; Bala2022;Guarro 2022;Chang2022; Pantanelli2023; Corbett2024;Skov 2026;Palkovits2026) RCTSeriousNotseriousNotseriousSerious981R1.00(0.78,1.28)Low Halos(R<1favoursEDOF) 9(McCabe2022; Bala2022;Guarro 2022;Chang2022; Pantanelli2023; Corbett2024;Black 2024;Skov2026; Palkovits2026) RCTSeriousSeriousNotseriousNotserious1,040R1.51(1.03,2.20)Low Blurredvision(R<1favoursEDOF) 3(McCabe2022; Chang2022; Pantanelli2023) RCTSeriousNotseriousNotseriousVeryserious565R0.50(0.12,2.06)Verylow Doublevision(R<1favoursEDOF) 1(McCabe2022)RCTSeriousN/ANotseriousVeryserious219R2.13(0.20,23.17)Verylow Hazyvision(R<1favoursEDOF) 2(McCabe2022; Pantanelli2023) RCTSeriousNotseriousNotseriousVeryserious270R0.78(0.38,1.60)Verylow Negativedysphotopsia(R<1favoursEDOF) 1(McCabe2022)RCTSeriousN/ANotseriousVeryserious219R0.68(0.27,1.69)Verylow Starbursts(R<1favoursEDOF) 8(McCabe2022; Bala2022;Chang 2022;Pantanelli 2023;Corbett2024; Black2024;Skov 2026;Palkovits2026) RCTSeriousSeriousNotseriousVeryserious1,055R0.97(0.59,1.60)Verylow (3)MonovsionvsMonofocal Visualacuity 97 No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousSerious137MD0.08(0.04to0.11)Low Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousSerious137MD−0.07(−0.11to−0.02)Low Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD−0.22(−0.27to−0.16)Moderate Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Contrastsensitivity(Pelli–RobsonlogCS)(higherfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD−0.01(−0.06to0.04)Moderate Visualfunction(Catquest-7SF)(higherfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD−0.01(−0.26to0.23)Moderate Spectacledependence(any)(R<1favoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137R0.72(0.62to0.84)Moderate Patientsatisfaction(R>1favoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious137R1.00(0.79to1.26)Verylow Adverseevents No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(R<1favoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious137R0.98(0.74to1.30)Verylow Halos(R<1favoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousSerious137R1.25(0.78to2.01)Low Meananiseikonia(lowerfavoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousNotserious137MD0.66(0.11to1.22)Moderate Starbursts(R<1favoursmonovision) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious137R0.97(0.64to1.49)Verylow (4)MultifocalvsEDOF Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(R<1favoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious118R0.90(0.28,2.89)Low Uncorrecteddistancevisualacuityworsethan6/6(R<1favoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious147R1.19(0.76,1.85)Low UncorrectednearvisualacuityworsethanJ3/J4orequivalent(R<1favoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious118R0.34(0.15,0.77)Low Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 2(Gil2020;Akkaya Turhan2021) RCTSeriousNotseriousNotseriousNotserious145MD0.03(-0.01,0.07)Moderate 98 Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 3(Webers2020; Sevik2022;Labiris 2024) RCTSeriousNotseriousNotseriousNotserious116MD-0.03(-0.08,0.03)Moderate Meanuncorrecteddistancevisualacuity(decimal)(higherfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousNotserious118MD0.02(-0.05,0.10)Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 2(Gil2020;Akkaya Turhan2021) RCTSeriousNotseriousNotseriousNotserious145MD0.12(0.09,0.16)Moderate Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 3(Webers2020; Sevik2022;Labiris 2024) RCTSeriousNotseriousNotseriousNotserious116MD0.03(-0.04,0.10)Moderate Meanuncorrectedintermediatevisualacuity(decimal)(higherfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousNotserious118MD-0.05(-0.11,0.02)Moderate Meancorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 2(Gil2020;Akkaya Turhan2021) RCTSeriousVeryseriousNotseriousVeryserious145MD-0.02(-0.20,0.16)Verylow Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 3(Webers2020; Sevik2022;Labiris 2024) RCTSeriousVeryseriousNotseriousSerious116MD-0.03(-0.27,0.20)Verylow Meanuncorrectednearvisualacuity(decimal)(higherfavoursmultifocal) 1(Cochener2018)RCTSeriousN/ANotseriousSerious118MD0.09(0.03,0.16)Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Visualfunction(unspecifiedinstruments)(higherfavoursmultifocal) 3(Webers2020; Sevik2022;Labiris 2024) RCTSeriousNotseriousNotseriousVeryserious116SMD0.00(-0.80,0.80)Verylow Patientsatisfaction(score,0–100scale)(higherfavoursmultifocal) 1(Sevik2022)RCTSeriousN/ANotseriousVeryserious29MD-1.71(-10.67,7.25)Verylow Patientsatisfaction(binary)(R>1favoursmultifocal) 1(Webers2020)RCTSeriousN/ANotseriousSerious27R0.99(0.80,1.23)Low Maximumreadingspeed(words/min)(higherfavoursmultifocal) 1(AkkayaTurhan 2021) RCTSeriousN/ANotseriousVeryserious29MD-5.00(-26.87,16.87)Verylow Readingacuity(logMAR)(lowerfavoursmultifocal) 1(AkkayaTurhan 2021) RCTSeriousN/ANotseriousNotserious29MD0.01(-0.04,0.06)Moderate Meanstereoacuity,arcseconds(lowerfavoursmultifocal) 1(Pradeep2026)RCTSeriousN/ANotseriousSerious70MD29.14(7.27,51.01)Low Adverseoutcomes 99 No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(R<1favoursmultifocal) 1(Webers2020)RCTSeriousNotseriousNotseriousVeryserious27R1.08(0.07–15.51)Verylow Halos(R<1favoursmultifocal) 1(Webers2020)RCTSeriousN/ANotseriousSerious27R1.48(0.89,2.46)Low Overallphoticphenomena(R<1favoursmultifocal) 1(Sevik2022)RCTSeriousN/ANotseriousVeryserious29R4.69(0.24,89.88)Verylow (5)MultifocalvsMonovision Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursmultifocal) 2(Wilkins2013; Labiris2024) RCTSeriousNotseriousNotseriousNotserious246MD0.04(0.01,0.07)Moderate Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursmultifocal) 2(Wilkins2013; Labiris2024) RCTSeriousNotseriousNotseriousNotserious241MD0.06(0.03,0.09)Moderate Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursmultifocal) 2(Wilkins2013; Labiris2024) RCTSeriousSeriousNotseriousNotserious246MD−0.00(−0.09,0.08)Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Contrastsensitivity(Pelli-Robsontest)(higherfavoursmultifocal) 2(Wilkins2013; Labiris2015) RCTSeriousNotseriousNotseriousNotserious262MD−0.03(−0.10,0.04)Moderate Visualfunction(lowerfavoursmultifocal) 1(Labiris2024)RCTSeriousN/ASeriousSerious60MD−3.81(−8.64,1.02)Verylow Visualfunction(VF-14)(higherfavoursmultifocal) 1(Labiris2015)RCTSeriousN/ANotseriousNotserious75MD−1.47(−5.51,2.57)Moderate Spectacledependence(any)(R<1favoursmultifocal) 2(Wilkins2013; Labiris2015) RCTSeriousNotseriousNotseriousNotserious262R0.40(0.30,0.53)Moderate Adverseevents No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(R<1favoursmultifocal) 1(Wilkins2013)RCTSeriousN/ANotseriousNotserious187R1.41(1.14,1.73)Moderate (6)EDOFvsMonovision Visualacuity No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Correcteddistancevisualacuityworsethan6/6(R<1favoursEDOF) 100 1(Vandekerckhove 2026) RCTSeriousN/ANotseriousSerious105R5.22(1.07,25.55)Low Uncorrecteddistancevisualacuityworsethan6/6(R<1favoursEDOF) 1(Vandekerckhove 2026) RCTSeriousN/ANotseriousSerious105R1.70(0.92,3.11)Low UncorrectednearvisualacuityworsethanJ3/J4orequivalent(R<1favoursEDOF) 1(Vandekerckhove 2026) RCTSeriousN/ANotseriousSerious105R0.25(0.08,0.77)Low Meancorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 2(Zeilinger2025; Vandekerckhove 2026) RCTSeriousNotseriousNotseriousNotserious185MD0.04(0.01,0.07)Moderate Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavoursEDOF) 5(Sandoval2020; Labiris2024; Zeilinger2025; Vandekerckhove 2026;Skov2026) RCTSeriousNotseriousNotseriousNotserious452MD0.03(−0.02,0.08)Moderate Meancorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 2(Zeilinger2025; Vandekerckhove 2026) RCTSeriousNotseriousNotseriousSerious185MD−0.10(−0.14,−0.07)Low Meanuncorrectedintermediatevisualacuity(logMAR)(lowerfavoursEDOF) 5(Sandoval2020; Labiris2024; Zeilinger2025; Vandekerckhove 2026;Skov2026) RCTSeriousNotseriousNotseriousNotserious452MD−0.01(−0.06,0.04)Moderate Meancorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 2(Zeilinger2025; Vandekerckhove 2026) RCTSeriousSeriousNotseriousSerious185MD−0.09(−0.22,0.04)Verylow Meanuncorrectednearvisualacuity(logMAR)(lowerfavoursEDOF) 5(Sandoval2020; Labiris2024; Zeilinger2025; Vandekerckhove 2026;Skov2026) RCTSeriousVeryseriousNotseriousVery serious 452MD0.01(−0.17,0.19)Verylow Meanstereoacuity(arcsec)(lowerfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious131MD10.39(−34.01, 54.78) Low Visualfunction No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality 101 Contrastsensitivity(higherfavoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousNotserious131MD0.001(−0.044, 0.045) Moderate Visualfunction(VF-14)(higherfavoursEDOF) 1(Labiris2024)RCTSeriousN/ANotseriousNotserious60MD−5.59(−10.20, −0.98) Moderate Spectacledependence(any)(R<1favoursEDOF) 2(Vandekerckhove 2026;Skov2026) RCTSeriousNotseriousNotseriousSerious312R1.07(0.92,1.24)Low Patientsatisfaction(R>1favoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVery serious 131R0.97(0.76,1.25)Verylow Adverseoutcomes No.ofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNo.ofparticipantsEffectsize(95%CI)Quality Glare(R<1favoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious131R1.08(0.82,1.43)Verylow Halos(R<1favoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousVeryserious131R0.91(0.57,1.44)Verylow Starbursts(R<1favoursEDOF) 1(Skov2026)RCTSeriousN/ANotseriousSerious131R1.29(0.87,1.90)Low (7)TrifocalmultifocalIOLvsBifocalmultifocalIOL Visualacuity NoofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNoofparticipantsEffectsize(95%CI)Quality Meancorrecteddistancevisualacuity(lowerfavourstrifocal) 3(Jonker2015, Gundersen2016, Li2023) RCTSeriousNotseriousNotseriousNotserious149MD0.00(-0.07,0.08)Moderate Meanuncorrecteddistancevisualacuity(logMAR)(lowerfavourstrifocal) 3(Jonker2015, Gundersen2016, Li2023) RCTSeriousNotseriousNotseriousNotserious149MD-0.01(-0.10,0.07)Moderate Meancorrectedintermediatevisualacuity(lowerfavourstrifocal) 2(Jonker2015,Li 2023) RCTSeriousNotseriousNotseriousNotserious127MD-0.03(-0.07,0.01)Moderate Meancorrectednearvisualacuity(lowerfavourstrifocal) 2(Jonker2015,Li 2023) RCTSeriousNotseriousNotseriousNotserious127MD0.01(-0.03,0.06)Moderate Meanuncorrectednearvisualacuity(lowerfavourstrifocal) 2(Jonker2015,Li 2023) RCTSeriousNotseriousNotseriousNotserious127MD0.00(-0.06,0.07)Moderate Visualfunction 102 NoofstudiesDesignRiskofbiasInconsistencyIndirectnessImprecisionNoofparticipantsEffectsize(95%CI)Quality Spectacledependence(near)(R<1favourstrifocal) 1(Jonker2015)RCTSeriousN/ANotseriousVeryserious28R0.65(0.18,2.38)Verylow 103 SupplementaryInformation11.Evidencetorecommendations(prospectivestudy, human-onlygroup) Relativevalueof differentoutcomes Theguidelinedevelopmentgroupconsidereddistancevisiontobeanessentialsafetyand functionaloutcome.However,thevalueofvisualcorrectingstrategiesisdeterminedby whethertheyprovidemeaningfulimprovementsinunaidedintermediateornearvision,daily visualfunctioningandspectacleindependencewithoutimposingunacceptableoptical disturbances.Patientsatisfaction,health-relatedqualityoflife,abilitytoperform occupationaltasks,andtheeffectsofopticaldisturbancesonnight-timeactivitieswere thereforeconsideredasimportantasdifferencesinvisual-acuitymeasurements.Clinical measurementsshouldbeinterpretedinrelationtotheirperceptibleeffectoneverydaylife. Trade-offbetween benefitsandharms ThecommitteeagreedthatnoIOLstrategywasconsistentlypreferableforallpatients.For distantvision,therewasevidencethatmonofocallensesprovidedgreaterdistantvisualacuity andwereassociatedwithfewerphoticphenomena,maybebecauseopticalperformanceis concentratedatasinglefocalpoint.However,patientsusuallyrequiredspectaclesfornear tasksandsometimesforintermediateactivities.Otherstrategiescouldalsoprovidegood distancevisionandspectaclesmightstillberequiredfordemandingtaskssuchasnight driving. Forintermediateandnearvision,multifocal,especiallytrifocallensesgenerallyofferedthe widestfunctionalrangeandthegreatestreductioninspectacledependence.EDOFlenses offerarangeofvisionfromdistancetointermediate,makingthemsuitableforpatientswho frequentlyworkoncomputers.However,theydidnotreliablyprovidespectacle independenceforsmallprintorprolongednearwork. Improvementsinvisualrangewereaccompaniedbypotentialreductionsinopticalquality. Designsthatgeneratedmultiplefocalrangescouldcausehalos,glare,starburstsorreduced contrastsensitivity.Theseeffectsmightbeparticularlytroublesomeforpeoplewhodrive frequentlyatnightorworkinlowillumination.Theevidencedidnotsupporttheassumption thatallEDOFlensesproducefewervisualdisturbancesthanmultifocallenses,performance variedbetweenindividuallensplatforms. Forbinocularfunction,monovisionpresentedadistincttrade-off.Greateranisometropia 104 couldimprovenearvisionbutimpairbinocularcomfort,stereopsisordistancevisualquality. Previoussuccessfulcontact-lensmonovisionmightindicatelikelytolerance,cautionwas warrantedwhenprecisedepthperceptionwasimportantforpatientsoccupationalor recreationalrequirement. Patientsatisfactiondependsonwhetherthepostoperativevisualprofilematchedindividual priorities.Somepatientsmightaccepthalosorreducedcontrastinexchangeforgreater spectacleindependenceandothersmightprioritizeopticalqualityandreliablenightvision. Thecommitteedidnotinterpretnon-significantfindingsinsmall,heterogeneousstudiesas evidenceofequivalence.Differencesinlensgeneration,refractivetargets,follow-upand outcomemeasurementpreventedaconsistenthierarchyamongstrategies.Lensselection shouldthereforebeindividualizedthroughshareddecision-making,addressingthedistances atwhichunaidedvisionismostimportant,night-drivingandoccupationalrequirements, willingnesstousespectaclesandtoleranceofunwantedvisualphenomena.Patientsshould understandthatnocurrentstrategyguaranteesspectacleindependence,optimalvisionat everydistanceorcompletefreedomfromvisualsymptoms. Considerationof healthbenefitsand resourceuse ThecommitteeagreedthattheeconomicconsequencesofIOLselectionhavechanged substantiallyasprocurementsystemsandtherangeofavailableproductshavedeveloped. Acquisitionprice,reimbursementbythehealthcaresystemandtheamountultimatelypaidby anindividualpatientaredifferentquantitiesandshouldnotbeconsideredinterchangeable. Historically,IOLswithspecialopticaldesignswereassociatedwithasubstantialprice premiumoverstandardmonofocallenses.However,thattheabsolutepricedifferencehas narrowedinmanysettingsasmoredomesticandinternationalproductshaveenteredthe marketandcentralizedprocurementhasreducedacquisitionprices.InChina,thenational volume-basedprocurementprogrammeforintraocularlensesintroducedsubstantialprice reductionsacrossthecategory.TheNationalHealthcareSecurityAdministrationreportedan averagereductionofapproximately60%followingthe2023nationalcentralized procurementroundforintraocularlensesandrelatedhigh-valuemedicalconsumables.The additionalacquisitioncostassociatedwithamultifocal,EDOForotheradvanced-function lensmaythereforebeconsiderablysmallerthanpervious. 105 Perviousstudiessuggestthatpresbyopia-correctinglensesmayprovideadditionalQALYs andreducelifetimespectacleexpenditureinselectedpatients.However,thesefindingsare sensitivetoassumptionsaboutpostoperativespectacledependence,lenspricepremiumsand willingness-to-paythresholds,andsomestudiesweremanufacturersponsored.Their applicabilitytocontemporaryChinesepracticeshouldthereforebereassessedusingcurrent localprices,reimbursementarrangementsandpatientout-of-pocketcosts. Thecommitteealsonotedthataffordabilitymayaffectequitableaccesstoadvancedoptical functions.Althoughcentralizedprocurementandmarketcompetitionmayreducethisbarrier, higherpriceshouldnotbeinterpretedasevidenceofgreaterclinicalbenefit.Localdecisions shouldinsteadconsiderwhethertheadditionalpatient-importantbenefitsjustifythe incrementalcost,withresource-useestimatesupdatedasprocurementpricesand reimbursementpolicieschange. Qualityofevidence Allincludedcomparisonswerebasedonrandomisedtrialsandriskofbiaswasconsistently judgedtobeserious.Certaintywasfrequentlyreducedfurtherbecauseofinconsistency, imprecisionandsmallsamplesizes.Evidenceformonovision,EDOFversusmonovisionand multifocalversusEDOFwasgenerallyderivedfromrelativelyfewstudies,andseveral patient-reportedoradverseoutcomeswereinformedbyasingletrial.Theselimitations reduceconfidenceinanyconclusionthatonepresbyopia-correctingstrategyissuperior overall. Otherconsiderations Lensselectionshouldalsotakeaccountofocularcomorbidity,binocularfunction,refractive predictability,previousexperiencewithmonovisionandthepatient’sabilitytoadaptto opticalphenomena.Counsellingshouldmakeclearthatnostrategyguaranteescomplete spectacleindependenceorfreedomfromvisualsymptoms.Expectedperformanceatdistance, intermediateandnearshouldbediscussedinfunctionaltermsrelevanttotheindividual patient. 106 SupplementaryInformation12.Recommendations(prospectivestudy,human-onlygroup) Question:Whatistheoptimalstrategytofacilitatesimultaneousdistanceandnearvisionfollowingcataractsurgery? Recommendationsin2017NICEguideline (NG77:Cataractsinadults:management) Recommendationsupdatedbyhumanexperts Donotoffermultifocalintraocularlensesforpeople havingcataractsurgery. Offermonovisionforuseaftercataractsurgerytopeople whohaveeitheranisometropiaormonovision preoperativelyandwouldliketoremainwithit. Discussthechoiceofintraocularlenswiththepersonbefore cataractsurgery,takingaccountoftheirprioritiesforunaided distance,intermediateandnearvision,willingnesstouse spectaclesandtoleranceofunwantedvisualphenomena Consideramultifocalintraocularlensforpeoplewhoseocular healthissuitableandwhoplaceahighvalueonunaided intermediateandnearvisionandreducedspectacle dependence Explainthatnointraocularlensstrategyguaranteesspectacle independence,optimalvisionatalldistancesorfreedomfrom visualsymptoms,andthatspectaclesmaystillberequiredfor demandingtasks 107 SupplementaryInformation13.Externalconsensussurvey(prospectivestudy) Background Thissurveyformspartofanexpertpreferenceassessmentofcandidaterecommendationsonthe useofintraocularlenses(IOLs)inpeopleundergoingcataractsurgery. Threecandidaterecommendationshavebeendevelopedfortheclinicalquestion: Whatistheoptimalstrategytofacilitatesimultaneousdistanceandnearvisionfollowing cataractsurgery? Thissurveyseekstoassesstheclinicalacceptabilityandpracticalapplicabilityofeach recommendationamongabroadergroupofophthalmicexperts. Instructionsforexpertreviewers Pleaseevaluateeachrecommendationusingthefive-pointLikertscalebelow.Yourratingshould reflecttheextenttowhichyouagreethattherecommendationisclinicallyappropriate, evidence-aligned,andacceptableforimplementationincurrentcataractsurgerypractice. ScoreResponseoption 1Stronglydisagree 2Disagree 3Neutral 4Agree 5Stronglyagree VotingQuestions(presentedinrandomizedorder) Question1:RecommendationA Pleaseindicateyourlevelofagreementwiththefollowingrecommendation: Donotoffermultifocalintraocularlensesforpeoplehavingcataractsurgery. Offermonovisionforuseaftercataractsurgerytopeoplewhohaveeitheranisometropia ormonovisionpreoperativelyandwouldliketoremainwithit. 108 Rating: ☐ 1Stronglydisagree ☐ 2Disagree ☐ 3Neutral ☐ 4Agree ☐ 5Stronglyagree Question2:RecommendationB Pleaseindicateyourlevelofagreementwiththefollowingrecommendation: Considermultifocalintraocularlensesforpeoplewhowantgreaterspectacleindependencefor distance,intermediateandnearvision.Whendiscussingthisoption: ensuretheperson’socularhealthissuitableformultifocalintraocularlenses explainthepotentialbenefitsinuncorrectednearandintermediatevisioncomparedwith standardmonofocalintraocularlenses discusstheincreasedriskofvisualdisturbances,suchashalos,glareandstarbursts takeintoaccounttheperson’svisualneedsanddailyactivities considercostimplicationsandtheperson’swillingnesstopay. Rating: ☐ 1Stronglydisagree ☐ 2Disagree ☐ 3Neutral ☐ 4Agree ☐ 5Stronglyagree Question3:RecommendationC Pleaseindicateyourlevelofagreementwiththefollowingrecommendation: 109 Discussthechoiceofintraocularlenswiththepersonbeforecataractsurgery,taking accountoftheirprioritiesforunaideddistance,intermediateandnearvision,willingness tousespectaclesandtoleranceofunwantedvisualphenomena Consideramultifocalintraocularlensforpeoplewhoseocularhealthissuitableand whoplaceahighvalueonunaidedintermediateandnearvisionandreducedspectacle dependence Explainthatnointraocularlensstrategyguaranteesspectacleindependence,optimal visionatalldistancesorfreedomfromvisualsymptoms,andthatspectaclesmaystillbe requiredfordemandingtasks Rating: ☐ 1Stronglydisagree ☐ 2Disagree ☐ 3Neutral ☐ 4Agree ☐ 5Stronglyagree ConsensusInterpretation Responseswillbeanalyzeddescriptively.Foreachrecommendation,theproportionofexperts selecting“agree”or“stronglyagree”willbecalculated.Theresultswillbeusedtoassesswhether therevisedrecommendationisalignedwithcurrentclinicalevidenceandacceptablefor implementationinreal-worldophthalmicsurgicalpractice.

Instant research discovery