Paper deep dive
Artificial intelligence-enabled single-lead ECG for non-invasive hyperkalemia detection: development, multicenter validation, and proof-of-concept deployment
Gongzheng Tang, Qinghao Zhao, Guangkun Nie, Yujie Xiao, Shijia Geng, Donglin Xie, Shun Huang, Deyun Zhang, Xingchen Yao, Jinwei Wang, Kangyin Chen, Luxia Zhang, Shenda Hong
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 99%
Last extracted: 3/22/2026, 5:05:55 AM
Summary
The study introduces 'Pocket-K', a single-lead AI-ECG system developed from the ECGFounder foundation model to detect hyperkalemia (serum potassium > 5.5 mmol/L). Validated across multicenter datasets (34,439 patients), the model demonstrated robust performance (AUROC 0.808â0.936) and high negative predictive value, with potential for near-real-time, non-invasive screening in handheld and wearable settings.
Entities (5)
Relation Signals (4)
Pocket-K â detects â Hyperkalemia
confidence 100% · Pocket-K, a single-lead AI-ECG system... for non-invasive hyperkalemia screening
Pocket-K â initializedfrom â ECGFounder
confidence 100% · initialized from the ECGFounder foundation model
Peking University Peopleâs Hospital â provideddatafor â Pocket-K
confidence 100% · Data from Peking University Peopleâs Hospital were divided into development and temporal validation sets
The Second Hospital of Tianjin Medical University â provideddatafor â Pocket-K
confidence 100% · data from The Second Hospital of Tianjin Medical University served as an independent external validation set
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Hyperkalemia is a life-threatening electrolyte disorder that is common in patients with chronic kidney disease and heart failure, yet frequent monitoring remains difficult outside hospital settings. We developed and validated Pocket-K, a single-lead AI-ECG system initialized from the ECGFounder foundation model for non-invasive hyperkalemia screening and handheld deployment. In this multicentre observational study using routinely collected clinical ECG and laboratory data, 34,439 patients contributed 62,290 ECG--potassium pairs. Lead I data were used to fine-tune the model. Data from Peking University People's Hospital were divided into development and temporal validation sets, and data from The Second Hospital of Tianjin Medical University served as an independent external validation set. Hyperkalemia was defined as venous serum potassium > 5.5 mmol/L. Pocket-K achieved AUROCs of 0.936 in internal testing, 0.858 in temporal validation, and 0.808 in external validation. For KDIGO-defined moderate-to-severe hyperkalemia (serum potassium >= 6.0 mmol/L), AUROCs increased to 0.940 and 0.861 in the temporal and external sets, respectively. External negative predictive value exceeded 99.3%. Model-predicted high risk below the hyperkalemia threshold was more common in patients with chronic kidney disease and heart failure. A handheld prototype enabled near-real-time inference, supporting future prospective evaluation in native handheld and wearable settings.
Tags
Links
- Source: https://arxiv.org/abs/2603.14177v2
- Canonical: https://arxiv.org/abs/2603.14177v2
Trouble viewing inline? Open PDF directly â
Full Text
52,736 characters extracted from source content.
Expand or collapse full text
Artificial intelligence-enabled single-lead ECG for non-invasive hyperkalemia detection:1 development, multicenter validation, and proof-of-concept deployment2 Gongzheng Tang 1,2,#, , Qinghao Zhao 3,# , Guangkun Nie 4,# , Yujie Xiao 1,2 , Shijia Geng 5 , Donglin3 Xie 1,2 , Shun Huang 2 , Deyun Zhang 5 , Xingchen Yao 6,7,8 , Jinwei Wang 6,7,8 , Kangyin Chen 9,* , Luxia4 Zhang 2,6,7,8,* , and Shenda Hong 1,2,10,11,*, 5 1 Institute of Medical Technology, Peking University Health Science Center, Beijing, China6 2 National Institute of Health Data Science, Peking University, Beijing, China7 3 Department of Cardiology, Peking University Peopleâs Hospital, Beijing, China8 4 School of Intelligence Science and Technology, Peking University, Beijing, China9 5 Heart Voice Medical Technology, Hefei, China10 6 Renal Division, Department of Medicine, Peking University First Hospital, Beijing, China11 7 Institute of Nephrology, Key Laboratory of Renal Disease, Ministry of Health of China, and Key12 Laboratory of Chronic Kidney Disease Prevention and Treatment (Peking University), Ministry of13 Education, Beijing, China14 8 Research Units of Diagnosis and Treatment of Immune-mediated Kidney Diseases, Chinese15 Academy of Medical Sciences, Beijing, China16 9 Tianjin Key Laboratory of Ionic-Molecular Function of Cardiovascular Disease, Department of17 Cardiology, Tianjin Institute of Cardiology, The Second Hospital of Tianjin Medical University,18 Tianjin, China19 10 State Key Laboratory of Vascular Homeostasis and Remodeling, NHC Key Laboratory of20 Cardiovascular Molecular Biology and Regulatory Peptides, Peking University, Beijing, China21 11 Institute for Artificial Intelligence, Peking University, Beijing, China22 # These authors contributed equally23 * Correspondence: hongshenda@pku.edu.cn, zhanglx@bjmu.edu.cn, chenkangyin@vip.126.com24 ABSTRACT25 Hyperkalemia is a life-threatening electrolyte disorder that is common in patients with chronic26 kidney disease and heart failure, yet frequent monitoring remains difficult outside hospital settings.27 We developed and validated Pocket-K, a single-lead AI-ECG system initialized from the ECG-28 Founder foundation model for non-invasive hyperkalemia screening and handheld deployment. In29 this multicentre observational study using routinely collected clinical ECG and laboratory data,30 34,439 patients contributed 62,290 ECGâpotassium pairs. Lead I data were used to fine-tune the31 model. Data from Peking University Peopleâs Hospital were divided into development and temporal32 validation sets, and data from The Second Hospital of Tianjin Medical University served as an33 independent external validation set. Hyperkalemia was defined as venous serum potassium>5.534 mmol/L. Pocket-K achieved AUROCs of 0.936 in internal testing, 0.858 in temporal validation,35 and 0.808 in external validation. For KDIGO-defined moderate-to-severe hyperkalemia (serum36 potassiumâ„6.0 mmol/L), AUROCs increased to 0.940 and 0.861 in the temporal and external37 sets, respectively. External negative predictive value exceeded 99.3%. Model-predicted high risk38 below the hyperkalemia threshold was more common in patients with chronic kidney disease39 and heart failure. A handheld prototype enabled near-real-time inference, supporting future40 prospective evaluation in native handheld and wearable settings.41 1 arXiv:2603.14177v2 [cs.LG] 17 Mar 2026 INTRODUCTION42 Hyperkalemia, defined as a venous serum potassium concentration above 5.5 mmol/L, is a43 common and potentially fatal electrolyte abnormality and a major risk factor for malignant ar-44 rhythmia and sudden cardiac death 1 . It is particularly common in patients with chronic kidney45 disease and heart failure, and its incidence increases substantially as renal function declines 2,3 .46 Previous studies have reported a U-shaped association between potassium concentration and47 mortality, with hyperkalemia linked to a higher risk of death in chronic kidney disease, heart failure,48 and acute myocardial infarction 4â6 . Timely intervention therefore depends on rapid and reliable49 recognition.50 Confirmation of hyperkalemia currently relies on venous blood sampling and laboratory mea-51 surement of potassium. Although this standard is accurate, it is invasive, intermittent, and52 dependent on access to clinical facilities, limiting frequent monitoring between routine visits for53 high-risk patients 7 . Hyperkalemia also produces characteristic electrocardiographic changes,54 including peaked T waves and QRS widening, which makes the electrocardiogram a plausible55 non-invasive screening tool 8 . However, visual interpretation alone has limited sensitivity, even56 among experienced clinicians, which reduces the reliability of the electrocardiogram for screening 9 57 and leaves hyperkalemia management an ongoing clinical challenge.58 Artificial intelligence (AI), particularly deep learning, has been increasingly applied in clinical59 research and cardiovascular medicine 10â16 . In electrocardiography, AI-based methods can detect60 subtle waveform features that are difficult to recognise by routine visual inspection and have61 shown promise in identifying cardiovascular and metabolic abnormalities 17â21 . In hyperkalemia,62 models trained on standard 12-lead ECGs have shown strong diagnostic performance and can63 capture risk signals that are subclinical or missed by conventional interpretation 22â24 . However,64 the workflow required to acquire a 12-lead ECG limits its use for repeated surveillance and65 self-screening outside the hospital, where scalable monitoring may be most valuable for high-risk66 patients. By contrast, single-lead ECG, particularly Lead I, is the dominant format used by67 handheld recorders and smartwatches, making it well suited to home monitoring and out-of-68 hospital screening. Small single-centre studies from the USA, Taiwan, and other settings have69 provided early evidence that single-lead hyperkalemia detection is feasible 25â27 . However, the70 current evidence base remains limited by modest sample size, restricted population diversity, and71 the absence of rigorous large-scale multicentre validation. Consequently, performance across72 heterogeneous populations, different health systems and acquisition platforms, and real-world73 prevalence settings remains uncertain.74 We therefore developed Pocket-K, a single-lead AI-ECG model for hyperkalemia screening,75 initialised from a pretrained ECG foundation model. We evaluated model performance in three76 settings: internal testing within the development set, temporally separated validation within the77 same health system, and independent external validation in a second health system using a78 different ECG platform. We also examined waveform-level interpretability, patient-level longitudinal79 trajectories, clinical phenotypes associated with false-positive predictions, and the technical80 feasibility of handheld deployment.81 RESULTS82 Study Overview83 We developed Pocket-K, a single-lead ECG model for non-invasive hyperkalemia screening that84 was initialized from a pretrained ECG foundation model. The overall study design is shown in85 Figure 1. The study used a multistage design. We first constructed ECGâK + pairs by linking86 2 each ECG to a venous serum potassium measurement obtained within a±1-hour window. We87 then fine-tuned the pretrained foundation model, ECGFounder, to estimate the probability of88 hyperkalemia, defined as serum K + > 5.5 mmol/L.89 We evaluated the model in 34,439 unique patients from two independent health systems. At the90 primary site, Peking University Peopleâs Hospital(PKUPH), patients were divided chronologically at91 July 2021 to examine performance over time, yielding a development set (N= 10,409;n= 26,14592 pairs) and a temporal validation set (N= 5,054;n= 10,922 pairs). We then tested the model in93 an independent external validation set from The Second Hospital of Tianjin Medical University94 (SHTMU) (N= 18,976;n= 25,223 pairs), which included different ECG management systems,95 hardware environments, and care settings.96 We also conducted a proof-of-concept evaluation to assess whether the model could support97 near-real-time, out-of-hospital screening in high-risk populations using handheld single-lead98 devices. This study design allowed us to evaluate Pocket-K both in historical clinical datasets and99 in a practical prototype deployment setting.100 ECGFounder Pretrained Foundation Model 30-sec Lead I ECG Real-time Inference Instant Risk Alert I. Development 26,145 ECG-K+ pairs Development cohort (PKUPH) Fine-tune dataset Physiologically anchored ECG-K+ pairing ± 1h Lead I ECG Venous serum K + Pocket-K Fine-tuned Model I. Validation Temporal Validation (PKUPH) Independent External Validation (SHTMU) TMUSH N = 18,976 Jan 2024-Dec 2024 PKUPH N = 5,054 Aug 2021-Nov 2024 Internal Test (PKUPH) PKUPH N = 1,041 Jan 2016-Jul 2021 from Development Set I. Proof-of-concept deployment Figure 1: Overview of Pocket-K development, validation, and proof-of-concept deployment. In the development stage, ECGâK + pairs were constructed by linking each ECG to the nearest eligible venous serum potassium measurement within a±1-hour window. These paired data were used to fine-tune ECGFounder, a pretrained ECG foundation model, to develop Pocket-K. In the validation stage, model performance was assessed in three settings: an internal test set from the PKUPH development set, a temporal validation set at PKUPH, and an independent external validation set from SHTMU. In the proof-of-concept stage, a handheld device recorded a 30-s lead I ECG, which was processed through a smartphone-based workflow for near-real-time inference and generation of a hyperkalemia risk alert. Study population and baseline characteristics101 The final study population included 34,439 unique patients and 62,290 ECGâK + pairs from two102 health systems. After quality control at PKUPH, 15,463 patients contributed 37,067 eligible pairs.103 Of these, 10,409 patients with 26,145 pairs formed the development set, and 5,054 patients with104 3 10,922 pairs formed the temporal validation set after exclusion of patients who overlapped with105 the development set. SHTMU contributed 18,976 patients with 25,223 pairs for independent106 external validation.107 Baseline clinical characteristics and technical specifications are summarised in Table 1. Across108 sets, mean age ranged from 54.08 to 58.25 years, and men accounted for 53.3%â56.0% of109 patients. Compared with the development set, the external validation set had a higher burden110 of renal dysfunction and cardiorenal comorbidity. In all sets, the mean interval between ECG111 acquisition and potassium measurement was less than 30 min, supporting close physiological112 alignment between waveform acquisition and the laboratory reference.113 Table 1: Baseline clinical characteristics, ECGâpotassium pairing quality, and technical specifi- cations across study sets. Continuous variables are presented as mean (SD), and categorical variables as n (%). PKUPH: Peking University Peopleâs Hospital; SHTMU: The Second Hospital of Tianjin Medical University;K + : serum potassium; eGFR: estimated glomerular filtration rate; ECG: electrocardiogram. CharacteristicDevelopment Set (PKUPH) Temporal Validation Set (PKUPH) Independent External Validation Set (SHTMU) (N = 10,409; n = 26,145)(N = 5,054; n = 10,922)(N = 18,976; n = 25,223) Demographics Age, years54.08 (15.55)54.60 (14.70)58.25 (17.90) Male sex, n (%)5,642 (54.2%)2,830 (56.0%)9,994 (53.3%) Laboratory measurements Serum K + , mmol/L4.14 (0.36)4.19 (0.40)4.11 (0.42) Serum creatinine, ÎŒmol/L74.66 (95.29)78.68 (173.15)106.08 (380.26) eGFR, mL/min/1.73 m 2 95.80 (23.14)93.84 (21.91)93.92 (28.84) eGFR-based categories, n (%) â„908,757 (84.1%)3,915 (77.5%)12,523 (66.0%) 60â891,291 (12.4%)932 (18.4%)4,418 (23.3%) 30â59242 (2.3%)143 (2.8%)1,241 (6.5%) <30119 (1.1%)64 (1.3%)794 (4.2%) ECGâpotassium pairing quality ECG-to-K + interval, min28.46 (14.88)24.44 (14.41)22.39 (16.26) Comorbidities, n (%) Hypertension2,263 (21.7%)1,453 (28.7%)6,778 (35.7%) Diabetes mellitus1,041 (10.0%)690 (13.7%)3,137 (16.5%) Chronic kidney disease430 (4.1%)238 (4.7%)2,127 (11.2%) Heart failure68 (0.7%)57 (1.1%)437 (2.3%) Coronary artery disease765 (7.3%)556 (11.0%)2,463 (13.1%) Stroke444 (4.3%)262 (5.2%)1,357 (7.2%) Technical specifications ECG management systemMedExMedExNalong Dominant hardwareGE / PhilipsGE / PhilipsNalong / Nihon Kohden Sampling frequency, Hz500500500 / 1000 N denotes the number of unique patients, and n denotes the number of ECGâpotassium pairs. Technical specifications are reported at the set level rather than the patient level. 4 Pocket-K generalized across internal, temporal, and independent external114 validation sets115 Pocket-K showed stable discrimination across the validation sets. In the internal test set, the116 AUROC was 0.9364 (95% CI 0.9135â0.9575), with a sensitivity of 83.33% (95% CI 78.24%â117 87.43%) and a specificity of 96.60% (95% CI 96.06%â97.07%) at the prespecified frozen threshold.118 In the temporal validation set, despite changes in case mix over time, the AUROC was 0.8582119 (95% CI 0.8441â0.8723), with a sensitivity of 85.05% (95% CI 81.84%â87.77%), suggesting120 robustness to temporal drift. In the independent external set, the AUROC was 0.8076 (95% CI121 0.7883â0.8258) despite concurrent differences in ECG platform and lower disease prevalence, as122 shown in Figure 2.123 Although discrimination decreased across increasingly heterogeneous evaluation settings, the124 negative predictive value remained high and exceeded 99.3% in the external set. These findings125 support the role of Pocket-K as a rule-out screening tool for identifying low-risk states rather than126 a replacement for laboratory confirmation.127 0.00.20.40.60.81.0 1 - Specificity (FPR) 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity (TPR) 95% Confidence Interval AUC = 0.928 (95% CI: 0.901-0.951) Random Guess Optimal Cutoff (Sens: 0.87, Spec: 0.96) (a) Internal Test 0.00.20.40.60.81.0 1 - Specificity (FPR) 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity (TPR) 95% Confidence Interval AUC = 0.858 (95% CI: 0.844-0.872) Random Guess Optimal Cutoff (Sens: 0.85, Spec: 0.86) (b) Temporal validation 0.00.20.40.60.81.0 1 - Specificity (FPR) 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity (TPR) 95% Confidence Interval AUC = 0.808 (95% CI: 0.788-0.826) Random Guess Optimal Cutoff (Sens: 0.82, Spec: 0.67) (c) Independent external validation Figure 2: ROC curves for model performance evaluation. (a) Internal testing ROC curve of PKUPH. (b) Temporal validation ROC curve of PKUPH. (c) Independent external ROC curve of SHTMU. Pocket-K showed stronger discrimination for moderate-to-severe hyper-128 kalemia129 Model performance was better for KDIGO-defined moderate-to-severe hyperkalemia (serum130 potassiumâ„6.0 mmol/L). In the temporal validation set, the AUROC increased to 0.9399 (95%131 CI 0.9160â0.9590), with a sensitivity of 89.80% (95% CI 78.24%â95.56%). In the independent132 external validation set, the AUROC was 0.8613 (95% CI 0.8243â0.8940), with a sensitivity of133 81.52% (95% CI 72.39%â88.13%), as shown in Figure 3.134 In the independent external set, the negative predictive value for moderate-to-severe hy-135 perkalemia was 99.91% (95% CI 99.86%â99.95%), indicating a very low probability of missed136 clinically urgent events in this screening setting. These findings suggest that Pocket-K was more137 sensitive to the more marked electrophysiological abnormalities associated with larger potassium138 elevations.139 5 0.00.20.40.60.81.0 1 - Specificity (FPR) 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity (TPR) 95% Confidence Interval AUC = 0.940 (95% CI: 0.916-0.959) Random Guess Optimal Cutoff (Sens: 0.90, Spec: 0.96) (a) Temporal External Validation 0.00.20.40.60.81.0 1 - Specificity (FPR) 0.0 0.2 0.4 0.6 0.8 1.0 Sensitivity (TPR) 95% Confidence Interval AUC = 0.861 (95% CI: 0.824-0.894) Random Guess Optimal Cutoff (Sens: 0.82, Spec: 0.79) (b) Independent External Validation Figure 3: ROC curves for detection of KDIGO-defined moderate-to-severe hyperkalemia (serum potassium â„6.0 mmol/L). (a) Temporal validation ROC curve in PKUPH. (b) Independent external validation ROC curve in SHTMU. Pocket-K captured recognizable electrophysiological features of hyper-140 kalemia141 To examine the waveform features underlying model predictions, we stratified samples into high-142 risk and low-risk groups according to the predicted probability and compared the signal-averaged143 Lead I morphology between groups. Normalized heartbeat segments were extracted from the144 raw signals after temporal alignment and standardization, and group-level mean waveforms145 with standard deviation bands were then calculated. The clearest differences were observed146 in the T-wave and QRS regions, consistent with the known electrophysiological manifestations147 of hyperkalemia (Figure 4). The low-risk group shows a more typical waveform pattern. These148 findings suggest that high-risk predictions were driven by recognizable waveform features rather149 than noise alone.150 Model-predicted high risk below the hyperkalemia threshold may serve as151 a potential marker of greater cardiorenal burden152 To better understand model behaviour among samples below the hyperkalemia threshold, we153 compared the clinical profiles of reference-negative samples stratified by model-predicted risk in154 the external validation set (Figure 5). Because all samples in this analysis had serum potassium155 values below the diagnostic threshold, the aim was to determine whether model-predicted high156 risk reflected random error or preferential identification of a subgroup with greater underlying157 disease burden. We found that chronic kidney disease and heart failure were more common in158 the model high-risk group than in the model low-risk group. The prevalence of chronic kidney159 disease increased from 1.2% in the model low-risk group to 3.2% in the model high-risk group160 (P <0.001), and the prevalence of heart failure increased from 1.3% to 4.4% (P <0.001). These161 findings suggest that model-predicted high risk below the hyperkalemia threshold was not simply162 random error. Instead, it may identify a subgroup with greater cardiorenal burden, even when the163 indexed serum potassium value does not exceed the diagnostic threshold.164 6 Figure 4: Signal-averaged waveform comparison. Average Lead I ECG waveforms are shown for patients correctly classified by the model as hyperkalemic (high risk, red line) and non-hyperkalemic (low risk, blue line). Shaded areas indicate standard deviation. Reference-negative, model low risk Reference-negative, model high risk Comorbidity prevalence among reference-negative samples by model-predicted risk Figure 5: Clinical profiles of reference-negative samples stratified by model-predicted risk. In the external validation set, all samples shown had serum potassium values below the diagnostic threshold for hyper- kalemia. They were stratified into a model low-risk group and a model high-risk group according to model output. Chronic kidney disease and heart failure were more common in the model high-risk group. The prevalence of chronic kidney disease increased from 1.2% in the model low-risk group to 3.2% in the model high-risk group (P <0.001), and the prevalence of heart failure increased from 1.3% to 4.4% (P <0.001). 7 Pocket-K tracked clinically meaningful potassium trajectories over time165 To illustrate longitudinal tracking at the patient level, we selected four representative patients from166 the validation sets with repeated ECGâpotassium measurements over time (Figure 6). These167 patients illustrate several clinically relevant longitudinal patterns, including progressive potassium168 rise, early high-risk detection with subsequent recovery, recurrent severe fluctuation, and gradual169 biochemical improvement after treatment.170 For Patient A, serum potassium remained below the diagnostic threshold for most of follow-up,171 while the model assigned persistently low risk. As potassium increased late in follow-up and172 eventually crossed the hyperkalemia threshold, the predicted risk rose sharply in parallel. This173 pattern suggests that the model was sensitive to progressive potassium elevation and may be174 useful for identifying impending deterioration before or around threshold crossing.175 For Patient B, an early episode of hyperkalemia was observed, during which the predicted risk176 rose to a high level. As serum potassium subsequently declined, model output also decreased177 and remained low during later follow-up. This pattern supports the ability of the model to detect178 an acute high-risk state and to return to a low-risk range after biochemical recovery.179 For Patient C, potassium fluctuated markedly over time, including severe elevation, partial180 correction, and subsequent recurrence. Model-predicted risk changed in parallel with these181 biochemical shifts, remaining high during severe episodes, decreasing after interim improvement,182 and rising again when potassium rebounded. This pattern suggests that the model may capture183 recurrent disease activity during longitudinal surveillance.184 For Patient D, serum potassium decreased steadily over serial measurements, and the185 predicted risk declined in parallel from an initially high level to a substantially lower range. This186 pattern suggests that model output may track treatment response or clinical recovery over time in187 patients with resolving hyperkalemia.188 Pocket-K supported near-real-time inference in a handheld workflow189 To assess potential clinical use in a handheld setting, we implemented a lightweight proof-of-190 concept version of Pocket-K in a single-lead ECG workflow. The prototype device acquired a191 30-s Lead I recording and transmitted the signal to a paired smartphone application through192 Bluetooth Low Energy, where near-real-time inference was performed using a connected inference193 pipeline (Figure 7). In representative tests, the system returned a predicted risk of 0.1% for a194 normokalemic example with a serum potassium concentration of 4.1 mmol/L and 74.5% for a195 severe hyperkalemia example with a serum potassium concentration of 6.9 mmol/L within seconds.196 These results support the technical feasibility of rapid risk estimation in a handheld workflow.197 DISCUSSION198 In this multicentre study, a single-lead AI-ECG system initialized from a pretrained founda-199 tion model showed clinically useful discrimination for hyperkalemia in internal testing, temporal200 validation, and independent external validation. Although performance declined as dataset hetero-201 geneity increased, Pocket-K maintained a high negative predictive value, supporting its potential202 use as a non-invasive rule-out screening tool for people at increased cardiorenal risk who require203 repeated potassium surveillance. We also showed the technical feasibility of near-real-time204 inference in a handheld workflow, providing an initial basis for future prospective deployment205 studies.206 Prior studies using 12-lead AI-ECG have reported strong discrimination for hyperkalemia and207 suggest that deep-learning models can detect ECG signatures that are not readily captured by208 8 Longitudinal Risk Tracking Risk ProbabilitySerum PotassiumThreshold (5.5) Patient APatient B Patient C Patient D Figure 6: Longitudinal tracking of serum potassium using single-lead AI-ECG. Representative patients showing the modelâs ability to track within-patient potassium changes over time. The red line indicates laboratory serum potassium, and the blue dashed line indicates the model-predicted risk probability. Patient A shows progressive potassium rise with late threshold crossing. Patient B shows an early hyperkalemic episode followed by sustained recovery. Patient C shows recurrent severe fluctuation with corresponding changes in model output. Patient D shows gradual biochemical improvement accompanied by declining predicted risk. Overall, these examples suggest that model output tracked clinically meaningful potassium dynamics at the individual-patient level. 9 A. Handheld ECG acquisitionB. Smartphone-connected inference workflow 30-s single-lead (Lead I) ECG recording Pocket-K AI Model Near-real-time inference via smartphone-connected workflow C. Risk outputs Risk alert Figure 7: Proof-of-concept handheld deployment workflow of Pocket-K. (A) A handheld device paired with a smartphone application records a 30-s single-lead (Lead I) ECG. (B) The recorded signal is transmitted to the smartphone through Bluetooth and processed by a connected inference pipeline for near-real-time risk estimation. (C) Representative outputs from the prototype system show a low predicted hyperkalemia risk for a normokalemic example (serum potassium 4.1 mmol/L) and a high predicted risk for a severe hyperkalemia example (serum potassium 6.9 mmol/L). routine visual interpretation 22â24 . By contrast, evidence for single-lead detection has remained209 limited to relatively small single-centre studies, often without rigorous external validation or210 evaluation across acquisition platforms 25â27 . Against this background, our study extends the211 literature by evaluating a foundation-model-based single-lead approach in a larger multicentre212 setting, with temporal separation within one health system and external validation on a different213 ECG platform. The lower discrimination observed in external validation relative to internal testing214 is therefore not unexpected, but instead reflects a more realistic test of transportability for a215 signal-limited screening model.216 These findings also help clarify the most plausible clinical role of Pocket-K. In lower-prevalence217 external settings, positive predictive value declined whereas negative predictive value remained218 high, which is consistent with the expected behaviour of a rule-out screening tool under prevalence219 shift. This pattern suggests that Pocket-K is better suited to repeated surveillance and triage220 than to definitive diagnosis. In practice, such a tool may be most useful for identifying low-risk221 states and for prompting confirmatory laboratory testing or clinical review in patients who require222 ongoing potassium surveillance, particularly those with advanced chronic kidney disease, dialysis223 dependence, or combined cardiac and renal disease.224 The stronger discrimination observed for moderate-to-severe hyperkalemia is clinically impor-225 tant. In our study, this endpoint corresponded to serum potassiumâ„6.0 mmol/L and therefore226 spanned KDIGO-defined moderate and severe hyperkalemia. This finding is biologically plau-227 sible, because larger potassium elevations are more likely to produce overt repolarization and228 conduction abnormalities, increasing signal detectability on ECG. From a clinical perspective,229 improved performance in this range is especially relevant, as the main priority of a screening tool230 in high-risk populations is to minimise missed urgent events rather than to replace biochemical231 confirmation.232 The interpretability and phenotype analyses further support the biological plausibility of the233 model. Group-level waveform differences were concentrated in the T-wave and QRS regions,234 consistent with established electrophysiological manifestations of hyperkalemia. In addition,235 among samples below the diagnostic threshold, model-predicted high risk was more common in236 patients with chronic kidney disease and heart failure. This observation raises the possibility that237 model outputs may reflect broader cardiorenal electrophysiological stress rather than potassium238 10 concentration alone at a single indexed time point. Although this interpretation remains hypothesis-239 generating, it may help explain why some apparently discordant predictions are not clinically240 trivial.241 The longitudinal case analyses and handheld prototype provide complementary evidence for242 potential clinical translation. Serial examples suggested that model-predicted risk changed in243 parallel with within-patient potassium trajectories, supporting the possibility that Pocket-K may244 capture evolving potassium-related electrophysiology over time. The proof-of-concept handheld245 workflow further showed that near-real-time inference is technically feasible in a connected246 smartphone pathway.247 Our study also provides initial evidence of translational feasibility. A handheld single-lead ECG248 workflow supported near-real-time inference through a connected smartphone-based pathway249 and may allow immediate feedback without additional invasive burden. However, this part of250 the study was designed as a proof of concept rather than an evaluation of clinical effectiveness.251 Prospective studies using native handheld and wearable recordings are therefore needed to252 assess usability, signal-quality failure modes, calibration, and the effect of AI-guided screening on253 repeat testing, clinical decision-making, and health-care use.254 This study has several limitations. First, it used historical routinely collected clinical data255 and therefore cannot establish whether model-guided screening improves clinical outcomes or256 workflow efficiency. Second, although we included temporal and independent external validation,257 both health systems were located in China, and broader validation across populations, geographic258 settings, and device form factors is still needed. Third, the single-lead input was derived from259 clinical 12-lead ECG systems rather than trained and validated primarily on native consumer-260 device recordings. Finally, model performance in the presence of rhythm abnormalities, motion261 artefact, and low-quality recordings requires more systematic evaluation before large-scale real-262 world deployment.263 In conclusion, Pocket-K provides multicentre evidence that a foundation-model-based single-264 lead AI-ECG approach can support hyperkalemia screening across temporally separated and265 externally independent validation settings. Its high negative predictive value suggests potential266 value as a scalable rule-out tool for frequent potassium surveillance in high-risk populations.267 Prospective studies should now assess how such systems can be integrated into handheld268 and wearable care pathways to support earlier detection while preserving safety and clinical269 interpretability.270 METHODS271 Study design and data sources272 This multicenter observational study used routinely collected de-identified clinical ECG and labo-273 ratory data and adhered to the TRIPOD+AI reporting guidelines for clinical artificial intelligence 28 .274 Data were collected from two geographically independent tertiary health systems in China: Peking275 University Peopleâs Hospital (PKUPH, Beijing) and The Second Hospital of Tianjin Medical Univer-276 sity (SHTMU, Tianjin). The PKUPH set was extracted from the MedEx ECG management system277 (January 2016âNovember 2024), while the SHTMU set was retrieved from the Nalong system278 (January 2024âDecember 2024). The study was approved by the Biomedical Ethics Committee of279 Peking University (IRB00001052-23152), and the requirement for informed consent was waived280 because the analysis used de-identified routinely collected clinical data.281 11 Study population assembly and data partitioning282 As illustrated in Figure 8, we first screened all patients with available ECG data during the study283 periods. After exclusion of haemolysed samples, venous serum potassium measurements were284 linked to ECG recordings using an ECG-anchored pairing strategy. Specifically, for each ECG,285 we searched for potassium measurements obtained within±1 h and retained the closest eligible286 measurement to form a unique ECGâpotassium pair. ECGs without any eligible potassium287 measurement within the predefined time window were excluded.288 The PKUPH population was partitioned chronologically at July 2021. Records before this289 cutoff formed the development set, and records after this cutoff formed the temporal validation set.290 To avoid information leakage, we imposed strict patient-level non-overlap: any patient appearing291 in the development period was excluded entirely from the temporal validation set, even if later292 records were available. Within the PKUPH development set, unique patients were further divided293 at the patient level into a fine-tune set, a model-selection set, and an internal test set in an 8:1:1294 ratio according to a prespecified protocol. The geographically independent SHTMU dataset was295 reserved exclusively for external validation.296 Because individual patients could contribute multiple ECGâpotassium pairs across time, the297 number of unique patients (N ) differed from the number of paired records (n).298 Definition of clinical comorbidities299 Chronic kidney disease and heart failure were defined a priori using structured diagnosis informa-300 tion recorded on or before the index ECG. In the PKUPH sets, diagnoses were obtained from301 the hospital diagnosis table (field: diagnosis name), and in the SHTMU set, from the structured302 diagnosis string field (clinic_diag_str). Because the original diagnosis entries were recorded in303 Chinese, comorbidities were identified using prespecified rule-based keyword searches applied to304 the original diagnosis strings. The English terms reported below describe the diagnostic concepts305 used for classification rather than the literal source strings.306 Chronic kidney disease was defined using diagnosis strings corresponding to chronic kidney307 disease or chronic kidney failure in the chronic setting, including chronic kidney disease, chronic308 renal insufficiency, chronic renal failure, end-stage kidney disease, end-stage renal disease,309 uraemia, and CKD-equivalent diagnostic expressions. Isolated non-specific terms such as renal310 insufficiency or renal failure were not used unless they were explicitly documented as chronic.311 Heart failure was defined using diagnosis strings corresponding to heart failure, including312 heart failure, congestive heart failure, left heart failure, right heart failure, biventricular heart failure,313 heart failure with preserved ejection fraction, and heart failure with reduced ejection fraction.314 These comorbidity variables were used for baseline set description and for the analysis of315 reference-negative samples stratified by model-predicted risk in the external validation set.316 Reference standard and single-lead signal preprocessing317 Venous serum potassium served as the reference standard. To improve physiological correspon-318 dence between ECG waveforms and laboratory measurements, each ECG was paired with the319 nearest eligible potassium result within a±1 h window after exclusion of haemolysed samples.320 This ECG-anchored pairing strategy was designed to reduce temporal mismatch between the321 model input and the biochemical reference label.322 Lead I was extracted from 12-lead ECG recordings to construct a single-lead input. All signals323 followed a unified preprocessing pipeline. We applied a 0.5â40 Hz band-pass filter to reduce324 12 Peking University People's Hospital N = 77,785 Excluded (N = 62,322) 29,522 Missing ECG records 32,542 No serum K + test within 1h 258 Poor data quality (hemolysis or low SNR) Peking University People's Hospital N = 15,463 (n = 37,067 pairs) Chronological split into fine-tune and temporal external validation datasets (July 2021) Development Cohort N = 10,409 (n = 26,145 pairs) Temporal Validation Cohort N = 5,054 (n = 10,922 pairs) The Second Hospital of Tianjin Medical University N = 49,415 Excluded (N = 30,439) 24,011 Missing ECG records 6,361 No serum K + test within 1h 67 Poor data quality (hemolysis or low SNR) The Second Hospital of Tianjin Medical University N = 18,976 (n = 25,223 pairs) Independent External Validation Dataset N = 18,976 (n = 25,223 pairs) A. PKUPH siteB. TMUSH site Figure 8: STARD flowchart for study population selection and multicenter dataset partitioning. Panel A shows set construction at PKUPH, where 77,785 patients were initially screened. A total of 62,322 patients were excluded through a multistage quality-control pipeline: 29,522 had no ECG records, 32,542 had no eligible venous serum potassium measurement within a±1-hour window for ECG-anchored pairing, and 258 were removed because of poor data quality, including confirmed laboratory haemolysis or low signal-to-noise ratio. The remaining PKUPH population (N= 15,463;n= 37,067 pairs) was partitioned chronologically at July 2021 into a development set (N= 10,409;n= 26,145 pairs) and a temporal validation set (N= 5,054;n= 10,922 pairs). Panel B shows independent external validation at SHTMU. Of 49,415 initially screened patients, 30,439 were excluded using the same criteria (missing ECGs: n= 24,011; no eligible potassium measurement within±1 hour for ECG-anchored pairing:n= 6,361; poor data quality:n= 67), resulting in a final independent external validation set of 18,976 patients with 25,223 ECGâK + pairs. In all panels,Ndenotes the number of unique patients andndenotes the number of ECGâK + paired records. 13 baseline drift, powerline interference, and myoelectric noise, segmented the waveform into non-325 overlapping 10-s clips, and resampled each clip to 500 Hz by linear interpolation to match the326 input specification of the pretrained foundation model. Each clip was then z-score normalised to327 reduce amplitude-scale differences across acquisition systems.328 Model development329 Pocket-K was developed by fine-tuning ECGFounder, a pretrained ECG foundation model. ECG-330 Founder adopts a Net1D-style architecture with stage-wise scaling inspired by RegNet and has331 been pretrained on more than ten million ECG recordings collected across multiple countries 29â31 .332 This large-scale pretraining enables the model to capture general representations of cardiac333 depolarisation and repolarisation and reduces downstream dependence on sample size and data334 homogeneity. All fine-tuning experiments were initialised from publicly available ECGFounder335 pretrained weights.336 The model was fine-tuned using binary cross-entropy loss and the Adam optimiser with an337 initial learning rate of 1Ă10 â4 . Training was performed for up to 30 epochs. Model selection338 was based on AUROC in the model-selection dataset. If validation AUROC did not improve for339 10 consecutive epochs, the learning rate was reduced by a factor of 0.1. The checkpoint with340 the best validation AUROC was retained for subsequent evaluation on the internal test dataset,341 temporal validation dataset, and external validation sets.342 Performance evaluation and statistical analyses343 The primary endpoint was discrimination for hyperkalemia, defined as venous serum potassium344 greater than 5.5 mmol/L. A prespecified secondary endpoint was discrimination for moderate-345 to-severe hyperkalemia (serum potassiumâ„6.0 mmol/L). According to the KDIGO 2024 CKD346 guideline, moderate hyperkalemia is defined as 6.0â6.4 mmol/L and severe hyperkalemia as347 â„ 6.5 mmol/L 32 . Model discrimination was primarily assessed using the area under the receiver348 operating characteristic curve (AUROC).349 Because individual patients could contribute multiple ECGâpotassium pairs, statistical un-350 certainty was quantified using clustered bootstrap resampling at the patient level with 2000351 resamples. This framework was used to estimate 95% confidence intervals for AUROC and352 threshold-dependent performance metrics and to compare model performance across internal353 test, temporal validation, and external validation sets without treating repeated paired records354 from the same patient as independent observations.355 Additional analyses and explainability356 We performed additional analyses to evaluate clinically relevant model behaviour beyond binary357 discrimination. First, moderate-to-severe hyperkalemia was analysed as a high-priority subgroup358 because of its greater clinical urgency. Second, waveform-level explainability analysis was per-359 formed by stratifying samples according to model-predicted risk and comparing signal-averaged360 heartbeat morphology between high-risk and low-risk groups, with particular attention to re-361 gions spanning the T wave and QRS complex. We also compared clinical phenotypes between362 false-positive and true-negative samples in the external validation set to determine whether363 false-positive predictions were enriched for cardiorenal comorbidity.364 To assess whether model outputs tracked disease dynamics at the individual level, we exam-365 ined representative longitudinal patients with repeated ECGâpotassium pairs over time. Finally,366 14 for proof-of-concept deployment, we evaluated inference latency and qualitative risk prompting in367 a handheld single-lead workflow.368 Proof-of-concept handheld deployment369 For proof-of-concept deployment, the handheld device acquired a 30-s single-lead ECG recording.370 To maintain consistency with the temporal scale used during model development, each recording371 was divided into three consecutive 10-s clips for inference. Clip-level outputs were then aggre-372 gated to generate a measurement-level risk probability. This design preserved temporal-scale373 consistency between model development and prototype deployment.374 Declaration statements375 Data Availability376 The data that support the findings of this study are not publicly available due to restrictions377 imposed by institutional ethics committees and data governance policies of the participating378 hospitals. Access to the data may be considered upon reasonable request to the corresponding379 author, subject to approval by the relevant ethics committees.380 Code Availability381 The code used for model development and evaluation in this study is publicly available athttps:382 //github.com/Tangoz1003/Pocket-K.383 Acknowledgements384 Shenda Hong is supported by the National Natural Science Foundation of China (62102008),385 CCF-Tencent Rhino-Bird Open Research Fund (CCF-Tencent RAGR20250108), CCF-Zhipu386 Large Model Innovation Fund (CCF-Zhipu202414), PKU-OPPO Fund (BO202301, BO202503),387 Research Project of Peking University in the State Key Laboratory of Vascular Homeostasis388 and Remodeling (2025-SKLVHR-YCTS-02), and the Beijing Municipal Science and Technology389 Commission (Z251100000725008).390 Luxia Zhang is supported by the National Natural Science Foundation of China (72125009)391 and grant from the Noncommunicable Chronic Diseases-National Science and Technology Major392 Project of China (No. 2025ZD0547500).393 Kangyin Chen is supported by the National Natural Science Foundation of China (82470527)394 and the Key Science and Technology Support Project of Tianjin Science and Technology Bureau395 (24ZXGZSY00130).396 The authors thank all collaborators and participating institutions for their support and contribu-397 tions to this research.398 Author Contributions399 Gongzheng Tang, Qinghao Zhao, and Guangkun Nie contributed equally to this work. Gongzheng400 Tang contributed to methodology development, model implementation, validation, formal analysis,401 and drafting of the manuscript. Qinghao Zhao contributed to result interpretation and manuscript402 revision. Guangkun Nie contributed to data acquisition, preprocessing, and curation. Yujie Xiao,403 15 Shijia Geng, Donglin Xie, Shun Huang, Deyun Zhang, Xingchen Yao, and Jinwei Wang contributed404 to data preparation and investigation. Kangyin Chen, Luxia Zhang, and Shenda Hong conceived405 and supervised the study, provided resources, guided study design and interpretation, and revised406 the manuscript. All authors reviewed and approved the final manuscript. Kangyin Chen, Luxia407 Zhang, and Shenda Hong are corresponding authors and take responsibility for the integrity of408 the work.409 Competing Interests410 Shenda Hong is an Associate Editor of npj Digital Medicine. Shenda Hong was not involved in411 the journalâs review of, or decisions related to, this manuscript. The other authors declare no412 competing financial or non-financial interests related to this study.413 References414 1.Kovesdy, C.P. (2016). Epidemiology of hyperkalemia: an update. Kidney international415 supplements 6, 3â6.416 2. Hougen, I., Leon, S.J., Whitlock, R., Rigatto, C., Komenda, P., Bohm, C., and Tangri, N. (2021).417 Hyperkalemia and its association with mortality, cardiovascular events, hospitalizations,418 and intensive care unit admissions in a population-based retrospective cohort. Kidney419 international reports 6, 1309â1316.420 3.Mclean, A., Nath, M., and Sawhney, S. (2022). Population epidemiology of hyperkalemia:421 cardiac and kidney long-term health outcomes. American Journal of Kidney Diseases 79,422 527â538.423 4.Kovesdy, C.P., Matsushita, K., Sang, Y., Brunskill, N.J., Carrero, J.J., Chodick, G., Hasegawa,424 T., Heerspink, H.L., Hirayama, A., Landman, G.W. et al. (2018). Serum potassium and425 adverse outcomes across the range of kidney function: a ckd prognosis consortium meta-426 analysis. European heart journal 39, 1535â1542.427 5.Goyal, A., Spertus, J.A., Gosch, K., Venkitachalam, L., Jones, P.G., Van den Berghe, G., and428 Kosiborod, M. (2012). Serum potassium levels and mortality in acute myocardial infarction.429 Jama 307, 157â164.430 6. Lin, C., Lin, C.S., Chen, S.J., Tsai, S.H., Sung, C.C., Chen, C.C., Hsu, Y.J., Hung, Y.J., and431 Lin, S.H. (2026). Ai-enabled electrocardiogram alert for potassium imbalance treatment: a432 pragmatic randomized controlled trial. Nature Communications 17, 159.433 7. Asirvatham, J.R., Moses, V., and Bjornson, L. (2013). Errors in potassium measurement: a434 laboratory perspective for the clinician. North American journal of medical sciences 5, 255.435 8. Campese, V.M., and Adenuga, G. (2016). Electrophysiological and clinical consequences of436 hyperkalemia. Kidney international supplements 6, 16â19.437 9. Lin, C., Chau, T., Lin, C.S., Shang, H.S., Fang, W.H., Lee, D.J., Lee, C.C., Tsai, S.H., Wang,438 C.H., and Lin, S.H. (2022). Point-of-care artificial intelligence-enabled ecg for dyskalemia: a439 retrospective cohort analysis for accuracy and outcome prediction. NPJ digital medicine 5, 8.440 16 10.Zhao, Q., Geng, S., Wang, B., Sun, Y., Nie, W., Bai, B., Yu, C., Zhang, F., Tang, G., Zhang, D.441 et al. (2024). Deep learning in heart sound analysis: from techniques to clinical applications.442 Health Data Science 4, 0182.443 11. Luo, Y., Liu, X.Y., Yang, K., Huang, K., Hong, M., Zhang, J., Wu, Y., and Nie, Z. (2024).444 Toward unified ai drug discovery with multimodal knowledge. Health Data Science 4, 0113.445 12.Zhang, S., Mu, W., Dong, D., Wei, J., Fang, M., Shao, L., Zhou, Y., He, B., Zhang, S., Liu,446 Z. et al. (2023). The applications of artificial intelligence in digestive system neoplasms: a447 review. Health Data Science 3, 0005.448 13. Liu, X., Gao, K., Liu, B., Pan, C., Liang, K., Yan, L., Ma, J., He, F., Zhang, S., Pan, S. et al.449 (2021). Advances in deep learning-based medical image analysis. Health Data Science450 2021, 8786793.451 14.Xiao, Y., Tang, G., Liu, W., Li, J., Nie, G., Kan, Z., Zhang, D., Zhao, Q., and Hong, S.452 (2025). Anyecg-lab: An exploration study of fine-tuning an ecg foundation model to estimate453 laboratory values from single-lead ecg signals. arXiv preprint arXiv:2510.22301.454 15. Nie, G., Zhu, J., Tang, G., Zhang, D., Geng, S., Zhao, Q., and Hong, S. (2024). A review of455 deep learning methods for photoplethysmography data. arXiv preprint arXiv:2401.12783.456 16.Shi, T., Ma, J., Yu, Z., Xu, H., Yang, R., Xiong, M., Xiao, M., Li, Y., Zhao, H., and Kong, G.457 (2025). Large language models in critical care medicine: scoping review. JMIR Medical458 Informatics 13, e76326.459 17. Poterucha, T.J., Jing, L., Ricart, R.P., Adjei-Mosi, M., Finer, J., Hartzel, D., Kelsey, C., Long, A.,460 Rocha, D., Ruhl, J.A. et al. (2025). Detecting structural heart disease from electrocardiograms461 using ai. Nature 644, 221â230.462 18.Rawlani, M., Ieki, H., Binder, C., Yuan, V., Chiu, I.M., Bhatt, A., Ebinger, J.E., Sahashi,463 Y., Ambrosy, A.P., Usuku, H. et al. (2025). Artificial intelligence prediction of age from464 echocardiography as a marker for cardiovascular disease. npj Digital Medicine 8, 688.465 19. Nie, G., Zhao, Q., Tang, G., Li, Y., and Hong, S. (2025). Artificial intelligence-derived466 photoplethysmography age as a digital biomarker for cardiovascular health. Communications467 Medicine 5, 481.468 20.Xiao, Y., Tang, G., Zhang, D., Li, J., Nie, G., Wang, H., Huang, S., Liu, T., Zhao, Q., Chen,469 K. et al. (2025). Fine-tuning an ecg foundation model to predict coronary ct angiography470 outcomes. arXiv preprint arXiv:2512.05136.471 21.Oikonomou, E.K., Vaid, A., Holste, G., Coppi, A., McNamara, R.L., Baloescu, C., Krumholz,472 H.M., Wang, Z., Apakama, D.J., Nadkarni, G.N. et al. (2025). Artificial intelligence-guided473 detection of under-recognised cardiomyopathies on point-of-care cardiac ultrasonography: a474 multicentre study. The Lancet Digital Health 7, e113âe123.475 22. Lin, C.S., Lin, C., Fang, W.H., Hsu, C.J., Chen, S.J., Huang, K.H., Lin, W.S., Tsai, C.S., Kuo,476 C.C., Chau, T. et al. (2020). A deep-learning algorithm (ecg12net) for detecting hypokalemia477 and hyperkalemia by electrocardiography: algorithm development. JMIR medical informatics478 8, e15931.479 17 23.Harmon, D.M., Heinrich, C.K., Dillon, J.J., Carter, R.E., Kashani, K.B., Attia, Z.I., Friedman,480 P.A., and Jentzer, J.C. (2024). Mortality risk stratification utilizing artificial intelligence elec-481 trocardiogram for hyperkalemia in cardiac intensive care unit patients. JACC: Advances 3,482 101169.483 24.Chen, C.C., Lin, C., Lee, D.J., Lin, C.S., Chen, S.J., Sung, C.C., Hsu, Y.J., and Lin, S.H.484 (2025). Monitoring serum potassium concentration in patients with severe hyperkalemia: the485 role of bloodless artificial intelligence-enabled electrocardiography. Clinical Kidney Journal486 18, sfaf092.487 25.Attia, Z.I., DeSimone, C.V., Dillon, J.J., Sapir, Y., Somers, V.K., Dugan, J.L., Bruce, C.J.,488 Ackerman, M.J., Asirvatham, S.J., Striemer, B.L. et al. (2016). Novel bloodless potassium489 determination using a signal-processed single-lead ecg. Journal of the American heart490 Association 5, e002746.491 26. Urtnasan, E., Lee, J.H., Moon, B., Lee, H.Y., Lee, K., and Youk, H. (2022). Noninvasive492 screening tool for hyperkalemia using a single-lead electrocardiogram and deep learning:493 Development and usability study. JMIR medical informatics 10, e34724.494 27.Chiu, I.M., Wu, P.J., Zhang, H., Hughes, J.W., Rogers, A.J., Jalilian, L., Perez, M., Lin, C.H.R.,495 Lee, C.T., Zou, J. et al. (2024). Serum potassium monitoring using ai-enabled smartwatch496 electrocardiograms. Clinical Electrophysiology 10, 2644â2654.497 28. Collins, G.S., Moons, K.G., Dhiman, P., Riley, R.D., Beam, A.L., Van Calster, B., Ghassemi,498 M., Liu, X., Reitsma, J.B., Van Smeden, M. et al. (2024). Tripod+ ai statement: updated499 guidance for reporting clinical prediction models that use regression or machine learning500 methods. bmj 385.501 29. Li, J., Aguirre, A.D., Junior, V.M., Jin, J., Liu, C., Zhong, L., Sun, C., Clifford, G., Bran-502 don Westover, M., and Hong, S. (2025). An electrocardiogram foundation model built on over503 10 million recordings. Nejm ai 2, AIoa2401033.504 30. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., and DollĂĄr, P. (2020). Designing network505 design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern506 recognition. p. 10428â10436.507 31. Hong, S., Xu, Y., Khare, A., Priambada, S., Maher, K., Aljiffry, A., Sun, J., and Tumanov, A.508 (2020). Holmes: health online model ensemble serving for deep learning models in intensive509 care units. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge510 Discovery & Data Mining. p. 1614â1624.511 32. Levin, A., Ahmed, S.B., Carrero, J.J., Foster, B., Francis, A., Hall, R.K., Herrington, W.G., Hill,512 G., Inker, L.A., Kazancıo Ì glu, R. et al. (2024). Executive summary of the kdigo 2024 clinical513 practice guideline for the evaluation and management of chronic kidney disease: known514 knowns and known unknowns. Kidney international 105, 684â701.515 18