Paper deep dive
OGA-AID: Clinician-in-the-loop AI Report Drafting Assistant for Multimodal Observational Gait Analysis in Post-Stroke Rehabilitation
Khoi T. N. Nguyen, Nghia D. Nguyen, Hui Yu Koh, Patrick W. H. Kwong, Karen Sui Geok Chua, Ananda Sidarta, Baosheng Yu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/10/2026, 3:12:31 AM
Summary
OGA-AID is a clinician-in-the-loop multi-agent AI system designed to assist in post-stroke observational gait analysis (OGA). It integrates multimodal inputsāincluding gait videos, motion-capture (MoCap) trajectories, and clinical profilesāusing three specialized agents (Recording Observer, Trajectory Analyzer, and Report Generator) to produce structured clinical reports based on the Wisconsin Gait Scale (WGS). The system improves report accuracy through human-AI collaboration, allowing clinicians to provide preliminary notes to guide the AI's assessment.
Entities (6)
Relation Signals (4)
Recording Observer ā partof ā OGA-AID
confidence 100% Ā· OGA-AID employs an agentic AI design consisting of 3 specialized agents
Trajectory Analyzer ā partof ā OGA-AID
confidence 100% Ā· OGA-AID employs an agentic AI design consisting of 3 specialized agents
Report Generator ā partof ā OGA-AID
confidence 100% Ā· OGA-AID employs an agentic AI design consisting of 3 specialized agents
OGA-AID ā utilizes ā Wisconsin Gait Scale
confidence 95% Ā· To ensure clinical interpretability, our agentic pipeline integrates WGS to guide gait analysis
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Gait analysis is essential in post-stroke rehabilitation but remains time-intensive and cognitively demanding, especially when clinicians must integrate gait videos and motion-capture data into structured reports. We present OGA-AID, a clinician-in-the-loop multi-agent large language model system for multimodal report drafting. The system coordinates 3 specialized agents to synthesize patient movement recordings, kinematic trajectories, and clinical profiles into structured assessments. Evaluated with expert physiotherapists on real patient data, OGA-AID consistently outperforms single-pass multimodal baselines with low error. In clinician-in-the-loop settings, brief expert preliminary notes further reduce error compared to reference assessments. Our findings demonstrate the feasibility of multimodal agentic systems for structured clinical gait assessment and highlight the complementary relationship between AI-assisted analysis and human clinical judgment in rehabilitation workflows.
Tags
Links
- Source: https://arxiv.org/abs/2604.05360v1
- Canonical: https://arxiv.org/abs/2604.05360v1
Trouble viewing inline? Open PDF directly ā
Full Text
44,335 characters extracted from source content.
Expand or collapse full text
OGA-AID: Clinician-in-the-loop AIReport Drafting Assistant for Multimodal O bservational Gait Analysis in Post-Stroke Rehabilitation Khoi T. N. Nguyen 1 Nghia D. Nguyen 2,4 Koh Hui Yu 1 Patrick W. H. Kwong 3 Karen Sui Geok Chua 1,5 Ananda Sidarta 1āā Baosheng Yu 1 * 1 Nanyang Technological University 2 University of Illinois Urbana-Champaign 3 The Hong Kong Polytechnic University 4 VinUni-Illinois Smart Health Center, VinUniversity 5 Institute of Rehabilitation Excellence, NHG Health Abstract Gait analysis is essential in post-stroke rehabilitation but remains time-intensive and cognitively demanding, espe- cially when clinicians must integrate gait videos and motion- capture data into structured reports. We present OGA-AID, a clinician-in-the-loop multi-agent large language model system for multimodal report drafting. The system coordi- nates 3 specialized agents to synthesize patient movement recordings, kinematic trajectories, and clinical profiles into structured assessments. Evaluated with expert physiother- apists on real patient data, OGA-AID consistently outper- forms single-pass multimodal baselines with low error. In clinician-in-the-loop settings, brief expert preliminary notes further reduce error compared to reference assessments. Our findings demonstrate the feasibility of multimodal agentic systems for structured clinical gait assessment and highlight the complementary relationship between AI-assisted analy- sis and human clinical judgment in rehabilitation workflows. 1. Introduction Stroke often causes long-term movement difficulties, and rehabilitation requires ongoing assessment to understand how a patient is recovering. In physiotherapy practice, ob- servational gait analysis (OGA), in which physiotherapists observe patientsā gait patterns from their movements, is com- monly performed to identify abnormalities and functional improvements to inform intervention planning. To support a more objective assessment beyond visual observation alone, motion capture systems (MoCap) and wearable sensors have been increasingly adopted [12,19,22] to quantify kinematic and spatiotemporal gait characteristics, such as joint angles, step symmetry, and temporal coordination. While these * Senior authors.ā Corresponding author: ananda.sidarta@ntu.edu.sg technologies provide detailed and precise measurements, their outputs are often time-consuming and require extensive effort to interpret, especially without biomechanics exper- tise [19]. Consequently, rather than reducing workload, such systems may increase the cognitive burden for physiother- apists who must manually translate complex multidimen- sional data into clinically meaningful reports. Recent advances in large language models (LLMs) and vision language models (VLMs) have led to the rapid adop- tion of AI assistants across a wide range of medical do- mains, including clinical documentation, medical question answering, and decision support [16,23]. These models have demonstrated strong capabilities in summarizing complex information, organizing heterogeneous inputs, and gener- ating structured textual outputs. However, the application of LLM-based systems to OGA, especially for post-stroke rehabilitation, remains largely unexplored. To the best of our knowledge, no previous study has systematically examined the role of LLM-based systems in this clinical assessment setting. This gap is particularly salient given physiother- apistsā interest in AI-assisted workflows [1] that enhance efficiency while retaining human oversight to ensure outputs can be validated before being shared with patients. Motivated by these challenges, we introduce OGA-AID, an AI report assistant for post-stroke OGA. OGA-AID in- tegrates the patientās movement recordings, MoCap joint trajectories, clinical profile, and clinicianās initial observa- tions to produce structured gait assessment draft reports articulated in clear clinical language. Crucially, OGA-AID is designed to be an assistive tool rather than a replacement for clinical judgment, in which clinicians retain full control by reviewing and finalizing the generated reports. OGA-AID employs an agentic AI design consisting of 3 specialized agents with distinct roles: a Recording Ob- server that analyzes temporal movement recordings from frontal and sagittal (side) viewpoints; a Trajectory Analyzer that processes and interprets MoCap data; and a Report arXiv:2604.05360v1 [cs.HC] 7 Apr 2026 Generator that synthesizes these insights into a formatted clinical report. The resulting AI-generated reports are deliv- ered through a user-friendly interface that allows clinicians to inspect, modify, and control the final outputs. We evaluate OGA-AID through a clinician-centered study with a partner hospital using real-world post-stroke patient data. Our evaluation covers both autonomous and clinician- in-the-loop performance with deep involvement of expert physiotherapists. In summary, our contributions are: ā¢We introduce OGA-AID, a clinician-in-the-loop AI re- port assistant for post-stroke OGA. The system integrates different input modalities, including patient movement videos, MoCap trajectories, and clinical profiles. We fur- ther propose a structured decomposition of different gait factors and assign them to specialized agents to effectively analyze post-stroke gait. ā¢ We conduct clinician-centered evaluations using real- world post-stroke patient data with diverse demographics and post-stroke durations. Results show that generated re- ports achieve clinically acceptable alignment with expert physiotherapists, paving the way for assistive agentic AI in post-stroke rehabilitation practice. ā¢Through our experiments, we show that clinician involve- ment in the OGA-AID pipeline, such as providing prelimi- nary observations, improves the quality of the generated reports. This finding highlights the importance of human- AI collaboration and supports clinician-in-the-loop designs for AI-assisted analysis workflows. 2. Background & Related Work Post-stroke OGA and the Wisconsin Gait Scale. In post- stroke rehabilitation, OGA is essential for identifying move- ment asymmetry and abnormalities such as hip circumduc- tion and knee hyperextension, which are critical when tailor- ing interventions. These observations are critical for com- prehensive assessments, treatment planning, and monitoring recovery progress [4]. To mitigate subjectivity and improve consistency, several standardized clinical gait scales have been developed. Among these, the Wisconsin Gait Scale (WGS) [18] is designed to assess hemiplegic stroke gait. The WGS comprises 14 observable items that span both the stance and swing phases, each accompanied by clear descriptive criteria to guide assessment. For each factor, observed gait deviations are translated into ratings, mostly from 1 (normal) to 3 (atypical), based on predefined severity levels. The total WGS score, which is the weighted sum of all factor ratings, ranges from 13.35 to 42, with higher scores reflecting greater gait impairments. MoCap-Assisted Gait Analysis. To capture movement traits that are difficult to observe with human eyes, MoCap is adopted for quantitative gait analysis [12,19,22]. Mo- Cap systems are commonly categorized as marker-based or markerless. Marker-based systems reconstruct 3D motion using reflective markers placed on anatomical landmarks and are widely regarded as the gold standard for biomechan- ical analysis due to their accuracy. In contrast, markerless systems improve accessibility by leveraging computer vi- sion and deep learning, eliminating the need for physical markers and reducing setup cost. However, their accuracy can be affected by occlusion, lighting conditions, or atypical movement patterns. Depending on available resources and infrastructure, clinical centers may adopt either approach for gait analysis, acknowledging the inherent trade-offs. As OGA relies on visual judgment, its measurements can vary substantially across clinicians [10]. Using MoCap, certain gait factors, such as movements of the hip, knee, ankle, and trunk, can be measured more reliably. In MoCap- assisted post-stroke gait assessment, a common practice is to compare a patientās gait with normative gait patterns from healthy individuals [21] matched by demographics such as age, height, weight, and gender. This comparison allows clin- icians to contextualize observed impairments and identify functional deviations across different body segments. LLMs and Agentic AI for Healthcare. Recent advances in LLMs and VLMs have accelerated the development of AI assistants across a wide range of healthcare applications. State-of-the-art foundation models, such as GPT [20] and Gemini [3], are trained on large-scale multimodal data in- cluding medically relevant information, enabling strong ca- pabilities in medical question answering, clinical documen- tation, and decision support. Building on these capabilities, recent studies have explored agentic paradigms that coor- dinate multiple specialized models to support healthcare workflows [9,16,23]. Within stroke rehabilitation, emerg- ing work has begun to investigate AI systems for clinical education, therapy assistance, and clinicianāpatient inter- action [11,13,17]. However, these systems are primarily designed around conversational interfaces that provide guid- ance or educational feedback rather than gait analysis. They remain largely text-driven and make limited use of the pa- tientās gait or motion data. As a result, the potential of LLM-based systems to directly support clinicians in OGA remains largely unexplored. 3. Methodology 3.1. Overview Figure 1 illustrates OGA-AIDās architecture. To ensure clin- ical interpretability, our agentic pipeline integrates WGS to guide gait analysis and standardize the output template. Note that, instead of WGS, other OGA instruments can also be integrated. For gait analysis, individual WGS factors are assigned to the most appropriate agent instead of jointly processing all factors at once. Similar to standard prac- tice mentioned in Section 2, OGA-AID matches patientsā gait patterns against normal individuals by retrieving from - External rotation - Circumduction - Hip extension, hiking - Stance/swing phases,... Prompt Instruction Patient Profile OGA-AID Output Inputs Recording Observer LLM Gait Recording MoCap Trajectory Initial Observation Normative Lookup Trajectory Processing Trajectory Analyzer LLM Face Removal EgoBlur Final Report Clinician - Foot contact - Step length - Stance width - Stance/swing phases,... Prompt Instruction "Based on the recording and trajectory insights, provide report following the template" Prompt Instruction Report Generator LLM Clinician Editable Report UI Figure 1. The overall architecture of OGA-AID. The system is designed to integrate multi-modal gait evidence with contextual patient information and clinician preliminary observations for post-stroke gait assessment through coordinated, agent-based analysis. a database of gait trajectories collected within the hospital. This information is processed within OGA-AID for down- stream LLM-based report generation. Importantly, unlike fully automated AI-based approaches, OGA-AID is designed to support clinician-in-the-loop re- port generation grounded in multimodal gait evidence. In addition to the 3 default inputs - gait recordings, MoCap trajectories, and patient clinical profile - clinicians may pro- vide preliminary notes based on their initial observations. These lightweight observations are incorporated as context references for the agents. After the report is generated, clin- icians review and finalize the content prior to patient com- munication. In clinical environments where maintaining professional authority remains essential, we believe this col- laborative design offers a balanced and safe approach to integrating AI into practice. 3.2. Input Details In a typical MoCap-assisted OGA, the patient performs a 10-meter walking task at a comfortable pace along a clear walkway. To capture motion trajectories, reflective markers are attached to the patient during the task. We assume that the clinical environment is equipped with a functional MoCap system. OGA-AID operates with 4 inputs as follows: Gait recording. During the walking task, patient move- ments are typically captured from multiple camera views. For OGA-AID, we utilize recordings from two views, frontal and sagittal (side), which are sufficient for visual gait assess- ment while reducing video token usage. MoCap trajectory. Markerless or marker-based trajec- tories can be obtained from the patientās gait movements using third-party software. The raw MoCap data comprise time-series 3D marker coordinates, which are internally cali- brated within the software. These trajectories are provided to OGA-AID as input for downstream analysis. Patient clinical profile. This input comprises the pa- tientās clinical characteristics, including demographic at- tributes (age, sex, height, and weight) as well as the affected hemiparetic side. Identification of the hemiparetic side is particularly critical, as it directly influences accurate gait interpretation in standard gait analysis. Clinicianās initial observation. In practice, a collabo- rative agentic system like OGA-AID rarely works indepen- dently without inputs from the clinician. Clinicians may optionally provide brief free-text notes for additional con- text, reflecting their initial impressions (e.g., āno gait aidā, āprolonged stance phaseā). In particular, some gait factors can be straightforwardly observed from a patient recording without creating a cognitive burden. These observations can serve as supplementary context for the agents and reduce the overall task complexity, enabling the agent to focus on the remaining items with greater efficiency. 3.3. Input Processing In OGA-AID, the input data is processed before being sent to the LLM for analysis, as shown in Figure 2. Face removal. To meet Institutional Review Board stan- dards, we apply face removal before sending data to the (3) Patient Clinical Profile Demography Age, gender, weight, height Hemiparesis Side Left, right (4) Initial Observations Observation 1: "No gait aid" Observation 2: "Prolonged stance phase" Observation 3: "Seems slightly hesitate" Observation 4: "Right knee is not flexible" Front ViewSagittal View (1) Gait Recordings (2) Trajectory Plots Figure 2. Processed multi-modal inputs for LLM inference in OGA-AID. (1) Anonymized gait recording that blurred the faces of people, (2) Joint trajectory plots after processing the MoCap trajectory input, (3) Patient clinical profile with demography and hemiparesis information, (4) Clinicianās initial observation. LLMs. We employ EgoBlur [15], which automatically de- tects and blurs the faces of all individuals in the video, in- cluding the stroke patient and assisting personnel. This processing step reduces the risk of identity disclosure while preserving gait-relevant visual information for analysis. Normative lookup. The patient profile is used to re- trieve corresponding normative references from an estab- lished database, assuming such reference gait data are avail- able. Matching is performed using basic demographics, i.e., age, sex, height, and weight. In our setting, we are provided with gait trajectories from 50 normative subjects in advance. Trajectory processing. The raw MoCap data consist of time-series sequences of 3D marker coordinates. The tra- jectories of the stroke patient and the matched normative subject are processed as spatiotemporal plots for the hip, knee, ankle, and torso joints. To enable consistent compar- ison across walking trials, the processed joint trajectories are time-normalized to a 0-100% gait cycle, allowing the agent to localize deviations within specific gait phases. For example, a downward deviation in the ankle trajectory dur- ing the mid-swing phase (61-80%) may suggest inadequate toe clearance. These processed trajectories are then visu- alized as spatiotemporal plots, highlighting potential gait abnormalities such as reduced joint range of motion or irreg- ular temporal patterns. In Figure 2, the red curves represent the patient trajectories, while the black curves correspond to the matched normative subject. Such visualizations pro- vide interpretable motion context that supports subsequent observational analysis and report generation. 3.4. Multi-Agent Architecture and Agent Roles In OGA, physiotherapists assess multiple gait-related factors before producing a clinical report. Using the WGS as the reference standard, we incorporate 14 assessment factors. Although the factors can be examined visually, trajectory data offer a more objective and quantitative characterization of gait kinematics. In consultation with clinical experts, we use video recordings to analyze 8 visually dominant factors, whereas the remaining 6 factors are assessed using MoCap trajectories for more precise quantitative measurement. To systematically handle diverse factors and heterogeneous in- puts, we adopt a multi-agent architecture with a factor de- composition strategy. Table 1 details the description and allocation of WGS factors to each agent. To ensure consis- tency across agents, the patientās clinical profile, including age, gender, and the hemiparetic side, is provided to all agents. Details of the agents are as follows: Recording Observer. This agent analyzes gait record- ings captured from frontal and sagittal views. To process the video data within the modelās context limits, we em- ploy an equidistant frame-sampling strategy, extracting 50 frames uniformly across the 10-meter walking sequence. Since clinical recordings may include assisting personnel (e.g., therapists), the agent is prompted to perform zero- shot subject identification using the patientās clinical profile. For example, the agent may identify the stroke patient by matching the described hemiparetic side with observable gait asymmetry. After identifying the patient, the agent performs temporal reasoning across the sampled frames to generate observations for 8 visually assessed factors, analyzing the spatial relationships between the patientās limbs and the sur- rounding environment. When MoCap trajectory input is unavailable, the system switches to a video-only pipeline, in which the Recording Observer evaluates all 14 WGS factors directly from the gait recordings. Trajectory Analyzer. This agent analyzes 3D gait kine- matics by interpreting MoCap joint trajectory plots of the hip, knee, ankle, and torso along 3 orthogonal planes. To stan- dardize the input for the LLM, the agent interprets trajectory displacement along the normalized gait cycle. The agent is prompted to perform a comparative analysis between the pa- tientās impaired-side trajectory (red curve) and the matched normative baseline (black curve). By examining differences in curve magnitude, peak locations, and slope transitions across the stance phase and swing phase, the agent identifies characteristic stroke-related motion patterns, such as reduced hip extension, altered stance duration, or pelvic retraction. These detected deviations are then translated into qualitative categorical descriptors, producing structured assessments for the remaining 6 WGS factors. Report Generator. This agent aggregates the outputs from the Recording Observer and Trajectory Analyzer and reconciles with any clinicianās initial observations to produce Table 1. Assessment factors for post-stroke gait defined in the WGS. Each factor is evaluated using an ordinal scale reflecting increasing severity. OGA-AID utilizes these factors to generate structured gait analysis reports. Recording-based Factors FactorDescription Use of hand gait aid Dependence on assistance during walking reflects impaired bal- ance, weight transfer, or lower limb control. Unaffected side step length Shortened forward step distance reflects inadequate weight transfer, weakness, or instability of the affected limb. Affected side weight shift Insufficient body-weight shift toward paretic limb suggests im- paired balance or weakness in the anti-gravity muscles of the lower limb. Stance widthMediolateral distance between feet during walking. Widened dis- tance suggests balance deficits. GuardednessHesitant gestures like stiffness, reduced arm swing, or slow move- ment due to fear of falling or difficult movement coordination. Initial swing knee flexion Reduced flexion typically leads to poor foot clearance and com- pensatory strategies. Toe clearanceIf toes adequately clear the ground during swing. Inadequate clearance increases tripping risks. Initial foot contact How the affected foot contacts the ground. Abnormal patterns reflect impaired ankle control. Trajectory-based Factors FactorDescription Affected side stance time Reduced stance duration indicates instability in load- ing the affected side. Affected side hip extension Hip movement during terminal stance. Limited ex- tension reduces propulsion and disrupts gait rhythm. Affected leg external rotation Excessive outward rotation of the paretic hip dur- ing swing, potentially arising from weakness or im- paired motor control. Mid-swing circumduction If present, affected limb swings outward in a semi- circular motion to compensate for reduced hip flex- ion, knee flexion, or ankle dorsiflexion. Mid-swing hip hiking Elevation of the pelvis on the affected side during swing, typically to compensate for reduced hip flex- ion, knee flexion, or ankle dorsiflexion. Terminal swing pelvic rotation Limited rotation suggests trunk stiffness or impair- ments in lower limb strength or motor control. a formatted clinical draft that follows either a factor-specific WGS scoring template or a narrative clinical summary. This design ensures the generated report remains both clinically interpretable and easily editable by clinicians. 3.5. LLM Selection and Prompt Strategy We select multimodal LLMs with strong image and video understanding capabilities, which in fact only exists a few suitable candidates. In Section 5, we study 4 models: Qwen- 3 VL (qwen3-vl-plus) [2], GPT-5 [20] including GPT-5 Mini (gpt-5-mini) and GPT-5.1 (gpt-5.1), and Gem- ini 3 Flash (gemini-3-flash-preview) [5]. We in- tentionally exclude extended-reasoning or thinking variants, as the primary objective of OGA-AID is structured clinical analysis and report generation rather than exploratory rea- soning. Instruction-following models are more suitable for producing consistent, template-aligned outputs [7] grounded in provided evidence. For consistency and fair evaluation of the LLM capability, all agents within OGA-AID employ the same base LLM. Nevertheless, future work may optimize LLM combinations or integrate newer LLMs for further improvements. Prompting is standardized using structured, role-based instructions aligned with the WGS. When both MoCap tra- jectories and gait recordings are available, the Recording Observer operates under a constrained assessment prompt that focuses exclusively on the 8 visually observable WGS factors. By restricting the prompt scope to the relevant vi- sual attributes, this configuration improves the reliability of qualitative observations. If the MoCap trajectory is unavail- able, the Recording Observer evaluates all 14 WGS factors directly from recordings using a full assessment prompt. 4. Experiment Design 4.1. Expert Participation Two experienced physiotherapists were involved in both the experimental procedure and evaluation. Physiotherapist A: Senior physiotherapist with 8 years of experience in rehabilitation in both community and pri- vate sectors, working with adults presenting neurological, orthopaedic, and musculoskeletal conditions. Physiotherapist B: Principal physiotherapist with 15 years of experience in community rehabilitation and voca- tional reintegration for people with physical disabilities. Additionally, a senior biomechanist was involved to sup- port MoCap interpretation and facilitate the provision of the reference WGS made by the physiotherapists. 4.2. Experiment Details Our experiments involved 15 post-stroke survivors, with statistics demonstrated in Figure 3. The cohort includes a balanced distribution of stroke types, with ages ranging from 30 to 80 years and varying chronic stages from 8 to 216 months post-event. To ensure statistical robustness despite the clinical constraints of recruiting post-stroke participants, each participant is asked to complete the walking task 3 times. This yielded a total ofN = 45data samples, cap- turing intra-patient variability in movement execution and environmental noise (e.g., differing positions and numbers of assisting personnel). The input acquisition follows the pro- cedure described in Section 3.2. In our setting, we employed a 16-camera Qualisys Miqus-M3 [14], an industry-leading marker-based MoCap software, to collect precise kinematic trajectories. However, alternative MoCap systems can be used depending on budget conditions. Reference WGS. By analyzing the gait recordings and Figure 3. Statistics of 15 patients, including gender, stroke type, hemiparetic side, age, and duration since the stroke event (in months). the trajectory data, participating experts work together to conduct the WGS reports. To ensure a more objective mea- surement, both physiotherapists must obtain mutual agree- ment for all 15 patients. We consider the WGS reports by the physiotherapists as a reference to assess OGA-AID. LLM Configurations and Multi-run Inferences. We set the temperature to 0.0 for all LLMs to encourage de- terministic behavior, except for GPT-5 Mini, which does not support temperature control and operates with a fixed default setting. Nevertheless, multimodal LLMs may still exhibit nondeterministic behavior due to internal stochastic processes and vision processing mechanisms. Slight varia- tions may arise from token selection dynamics or the internal handling of video or image sequences. To reduce the impact of such variability, for each ofN = 45samples, we conduct M = 3 independent runs and average the results. 4.3. Autonomous Evaluation Section 5.1 examines the performance of OGA-AID in gen- erating WGS reports without clinician-in-the-loop, enabling an independent assessment of its clinical reasoning and struc- tured report generation. We compare OGA-AID with single- pass LLM baselines, which receive all 3 input modalities (gait recordings, MoCap trajectories, and the patientās clin- ical profile) within a single prompt. Using the same in- struction prompt template as OGA-AID, the model directly outputs assessments for all 14 WGS factors and produces the report in a single inference step. We further investigate the effect of different input combinations in OGA-AID to better understand how each input modality contributes to the accuracy of the generated reports. 4.4. Clinician-in-the-loop Evaluation Section 5.2 evaluates OGA-AID in a collaborative setting where clinicians provide preliminary observations to guide the analysis. To simulate this scenario, each patient is associ- ated with 5 observations that physiotherapists can straightfor- wardly identify, including knee flexion during initial swing, circumduction, stance time, hip extension, and step length. Each run uses a different sampling of these observations. The sampled observations are categorized into short (1ā2 observations), medium (3ā4 observations), and long (5 ob- servations). This evaluation measures the performance gains enabled by the clinician-in-the-loop setting. 4.5. Metrics Lety i denote the reference WGS score andĖy i denote the predicted score for patienti. We report 4 metrics as follows. Mean Absolute Error (MAE) quantifies the average magnitude of prediction error across all samples, defined by 1 NM P N n=1 P M m=1 |Ėy n,m ā y n | . We report the MAE aver- aged over 45 data samples and 3 runs per sample. Lower MAE indicates closer agreement with the reference WGS. Maximum Absolute Error (Max AE) captures the largest deviation, defined bymax n,m |Ėy n,m ā y n | .We re- port the largest Max AE acrossN Ć M = 135runs to reflect the most extreme potential error. This metric reflects worst-case discrepancies, which are clinically important, as extreme deviations may affect treatment decisions. Mean Error (Bias) measures systematic over- or under-estimation (positive or negative value), defined by 1 NM P N n=1 P M m=1 (Ėy n,m ā y n ) . We report the final Bias averaged over the 45 data samples and 3 runs per sample. Minimal Clinically Important Difference (MCID) rep- resents the smallest change in a score that is perceived as clin- ically meaningful in the patientās status. In our case, we are interested in understanding whether the conclusions made by the agentic system can be assistive to clinician workflow, i.e, do not result in large disparity compared to physiotherapistsā evaluations. Note that in OGA-AID, clinicians preserve con- trol of the final outputs, so perfect agreement is not necessary, and agreement metrics are less informative. Therefore, we instead evaluate whether the difference reported is less than a threshold defined by the MCID. Differences exceeding this threshold are considered inappropriate, as they may require the clinician to fully reevaluate the scores, which signifi- cantly impacts the usability of the system. For WGS, the MCID is defined as 2.25 [6]. 5. Results 5.1. Autonomous Evaluation Table 2 presents a comparison between OGA-AID and single-pass LLM baselines. Across all backbones, OGA- Table 2. Performance comparison between single-pass LLMs and OGA-AID for WGS score reports forN=45samples, without clinician-in-the-loop.āindicates lower values are better. āBelow MCIDā denotes if the MAE falls under the 2.25 MCID threshold. LLMMAEā Max AEā Bias Below MCID Single pass Qwen-3 VL9.1820.009.18ā GPT-5 Mini3.157.701.61ā GPT-5.12.354.750.45ā Gemini-3 Flash2.415.580.95ā Ours Qwen-3 VL6.2811.006.28ā GPT-5 Mini3.178.003.04ā GPT-5.12.056.750.38ā Gemini-3 Flash1.944.250.23ā AID generally achieves substantially lower MAE than single- pass inference, highlighting the advantage of the structured, multi-stage generation pipeline for the OGA task. Notably, with GPT-5.1 and Gemini-3 Flash, OGA-AID attains MAE values below the 2.25 MCID threshold, indicating clinically acceptable agreement with reference scoring. We find these results encouraging, as they suggest that agentic AI systems and LLM-based frameworks, such as OGA-AID, may offer a promising direction for supporting post-stroke gait analysis. More broadly, we highlight the potential role of LLM-driven tools in rehabilitation assessment workflows, an important yet relatively underexplored domain that re- quires greater research attention. Nonetheless, the relatively significant Max AE also provides a warning about cautious use of agentic systems since they may provide misaligned predictions. Regarding systematic behavior, the predominantly posi- tive Bias values among all backbones indicate a consis- tent tendency to overestimate WGS scores, which may reflect the potential miscalibration issues of modern LLMs. From this observation, we highlight the importance of careful evaluation when deploying LLM-based systems in medical applications to mitigate systematic prediction errors. In Figure 4, we examined the factor-wise MAE between the OGA-AID reports and the reference. While Gemini-3 Flash achieves the lower total MAE, GPT-5.1 is generally more consistent with the reference WGS. Nonetheless, GPT- 5.1 can also misalign significantly compared to the reference in some factors, such as guardedness. In Table 3, we further examined the contributions of dif- ferent input modalities within OGA-AID, featuring our best LLM performers GPT-5.1 and Gemini-3 Flash. We observe that removing any input component consistently increases MAE, confirming the importance of multimodal integration for accurate WGS estimation. The complete configuration achieves the strongest performance for both GPT-5.1 (MAE 0.00.20.40.60.81.01.2 MAE Gait Aid Step Length Weight Shift Stance Width Guardedness Knee Flexion Toe Clearance Foot Contact 0.00 0.33 0.32 0.38 0.68 0.80 0.28 0.37 0.00 0.33 0.53 0.38 0.25 0.87 0.15 0.37 Recording-based 0.00.20.40.60.81.01.2 MAE Stance Time Hip Extension Ext. Rotation Circumduction Hip Hiking Pelvic Rotation 0.02 0.65 0.40 0.48 0.42 0.93 0.10 0.83 0.73 0.80 0.15 0.77 Trajectory-based GPT-5.1Gemini-3 Flash Figure 4. Per-factor MAE of WGS assessments by GPT-5.1 and Gemini-3 Flash across 45 samples. Each bar represents the MAE between predicted and reference ordinal scores for a given WGS factor. Error bars indicate standard deviation across runs. = 2.05) and Gemini-3 Flash (MAE = 1.94), whereas re- lying on Recordings (R) only without MoCap trajectories (T) leads to a clear performance decline. Notably, under single-modality settings, GPT-5.1 outperforms Gemini-3 Flash, suggesting stronger standalone video and image un- derstanding. In contrast, Gemini-3 Flash demonstrates better interpretation of multiple input modalities, resulting in better performance in the full multimodal setting. However, these results may be subject to the specific prompting strategies employed. Table 3. Performance under different input combinations for OGA- AID.Rdenotes movement recordings,Tdenotes trajectory,D denotes patient clinical profile, orange and blue colors denote nega- tive and positive Bias respectively. Inputs GPT-5.1Gemini-3 Flash MAEā Max AEā Bias MAEā Max AEā Bias R3.809.50-3.34 5.4410.75-5.32 R+D2.336.75-1.53 6.2412.75-6.15 R+T2.858.50-1.27 2.065.500.48 R+T +D 2.056.750.38 1.944.250.23 WhenRis provided withoutT, both GPT-5.1 and Gemini- 3 Flash show significant negative biases (-3.34 and -5.32). This suggests that by only looking at the visual movement, the LLMs underestimate the stroke severity. Incorporat- ing MoCap trajectories helps reduce this effect. We also emphasize the importance of the patientās clinical profileD, which enables OGA-AID to identify the hemiparetic side and provides crucial information that guides the Recording Observer in interpreting the patientās gait. 5.2. Clinician-in-the-loop Evaluation Table 4 presents scenarios where cliniciansā preliminary observation notes are provided to OGA-AID as additional context. The results indicate that OGA-AID can leverage expert insights and subsequently improve the quality of the generated reports, as reflected in the overall reduction in MAE for both GPT-5.1 and Gemini-3 Flash compared to the base cases with no initial observations. Table 4. Performance with clinician preliminary observations. āObsā indicates the length of observations supplied to OGA-AID, with ā-ā referring to the base case without any initial observation. Obs GPT-5.1Gemini-3 Flash MAEā Max AEā Bias MAEā Max AEā Bias -2.056.750.38 1.944.250.23 Short 2.076.001.01 1.903.25-0.03 Med. 1.968.000.89 1.766.000.38 Long 1.674.001.00 1.584.00-0.20 Furthermore, performance improves as the richness of the clinicianās initial observations increases. When longer notes are provided, OGA-AID achieves up to 22.7% reduction in MAE (1.58 compared to 1.94). Medium observations also yield consistent performance gains, though to a lesser extent. In some cases, however, we observe a higher Max AE, suggesting that for certain outlier samples, additional clinical context may introduce ambiguity that affects how the LLM interprets gait characteristics that differ from what the clinician expected. Overall, these findings suggest that clinician-provided contextual cues help guide the analysis process of the agen- tic pipeline. OGA-AID appears to benefit from lightweight expert input that supports interpretation and reduces uncer- tainty during report generation. This observation highlights the complementary relationship between human clinical judgment and AI-assisted analysis, where brief clinician insights can meaningfully enhance automated assessment. 6. Discussions In this section, we reflect on the key findings of our study and discuss several limitations and future directions. Clinical Implications. Although the proposed system demonstrates encouraging results, substantial opportunities remain to improve performance. Future efforts may focus on more effective prompt engineering strategies and enhanced processing methods tailored to different input modalities, par- ticularly video recordings and MoCap trajectory data. For instance, additional trajectory processing could be applied to provide richer analytical contexts for LLM inferences. Nev- ertheless, the main contribution of this work is establishing the feasibility of LLM-based, agentic AI in physiotherapy, particularly in OGA and post-stroke rehabilitation. Within this framework, we investigate how LLM-based systems op- erate in realistic clinical workflows, assessing both analytical performance and the alignment between generated reports and clinician judgment. Data Constraint. Rehabilitation research is inherently constrained by limited shareable data availability. Unlike many medical domains where large-scale textual or imaging datasets are available, collecting post-stroke motion data is time-consuming due to patientsā limited mobility and often requires strict privacy considerations. Additionally, although some gait datasets exist [8,22], they typically consist of raw kinematic recordings accompanied by minimal metadata, such as demographic or diagnostic information, and rarely include high-level textual descriptions or clinically meaning- ful gait assessments. The lack of expert-annotated narrative evaluations poses a significant challenge for the develop- ment and validation of LLM-driven gait analysis systems. To mitigate data scarcity and ensure statistical robustness, we upsampled the dataset by collecting multiple trials from the stroke participants. Throughout the evaluation of OGA-AID, we collaborated closely with clinicians to ensure reliable assessment while maintaining alignment with real clinical workflows, avoiding additional workload or disruption to existing practice. 7. Conclusion We presented OGA-AID, a clinician-in-the-loop, multi-agent AI system for post-stroke OGA reports. By integrating gait recordings, MoCap joint trajectories, and clinical context within a structured agentic pipeline, the system generates clinically aligned draft reports while preserving clinician oversight. This work demonstrates the feasibility of agentic LLM systems for supporting clinical assessment workflows, physiotherapy, and rehabilitation practice. Acknowledgments Ananda Sidarta, Khoi Nguyen, and Koh Hui Yu are sup- ported by tripartite funding from A*STAR/NTU/NHG, within the Rehabilitation Research Institute of Singapore. Nghia D. Nguyen is supported by the Vingroup Science and Technology Scholarship Program. References [1]Reem M. Alwhaibi, Raghad B. Alshammari, Nouf I. Alru- fayyiq, Jawaher Q. Alenzi, Amirah H. Alsumayli, Btool I. Alrushud, and Tahani J. Alahmadi. Attitudes of physical ther- apists toward ai diagnostics: barriers, enablers, and clinical implications. BMC Medical Education, 25(1), 2025. 1 [2]Shuai Bai and Yuxuan Cai et al. Qwen3-vl technical report, 2025. 5 [3]Gheorghe Comanici and Eric Bieber et al. Gemini 2.5: Push- ing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities, 2025. 2 [4] Francesco Ferrarello, Valeria Anna Maria Bianchi, Marco Baccini, Gaia Rubbieri, Enrico Mossello, Maria Chiara Cav- allini, Niccol ` o Marchionni, and Mauro Di Bari. Tools for observational gait analysis in patients with stroke: A sys- tematic review. Physical Therapy, 93(12):1673ā1685, 2013. 2 [5]Google. A new era of intelligence with Gemini 3.https: / / blog . google / products - and - platforms / products/gemini/gemini-3/, 2025. 5 [6] Agnieszka Guzik, Mariusz Dru Ģ zbicki, And Ģ zelina Wolan- Nieroda, Grzegorz Przysada, and Andrzej Kwolek. The Wisconsin gait scale ā The minimal clinically important dif- ference. Gait & Posture, 68:453ā457, 2019. 6 [7]Xiaomin Li, Zhou Yu, Zhiwei Zhang, Xupeng Chen, Ziji Zhang, Yingying Zhuang, Narayanan Sadagopan, and Anurag Beniwal. When thinking fails: The pitfalls of reasoning for instruction-following in LLMs. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 5 [8]Phyllis Liang, Wai Hang Kwong, Ananda Sidarta, Choon Kong Yap, Wee Kiat Tan, Lek Syn Lim, Pui Yee Chan, Christopher Wee Keong Kuah, Seng Kwee Wee, Karen Chua, Colin Quek, and Wei Tech Ang. An asian-centric hu- man movement database capturing activities of daily living. Scientific Data, 7(290), 2020. 8 [9]Hongjun Liu, Yinghao Zhu, Yuhui Wang, Yitao Long, Zeyu Lai, Lequan Yu, and Chen Zhao. Medmmv: A controllable multimodal multi-agent framework for reliable and verifiable clinical reasoning, 2025. 2 [10] Daiki Matsuzaka, Koki Wagatsuma, Takenori Shimada, Kenta Ikushima, and Hiroyuki Fujisawa. Reliability and validity of observational gait analysis by physical therapists: Possibility of verifying accuracy and improving technology in visual measurement of joint angles. Physical Therapy Research, 28 (2):129ā136, 2025. 2 [11]Phuc Truong Loc Nguyen and Thanh Hung Do. Congait: A clinician-centered dashboard for contestable ai in parkinsonās disease care, 2025. 2 [12]Jing Wen Pan, Ananda Sidarta, Tsung-Lin Wu, Wai Hang Patrick Kwong, Poo Lee Ong, Matthew Rong Jie Tay, Min Wee Phua, Wei Binh Chong, Wei Tech Ang, and Karen Sui Geok Chua. Unraveling stroke gait deviations with move- ment analytics, more than meets the eye: a case control study. Frontiers in Neuroscience, Volume 18 - 2024, 2024. 1, 2 [13] Shiqi Qiang, Haitao Zhang, Yang Liao, Yue Zhang, Yanfen Gu, Yiyan Wang, Zehui Xu, Hui Shi, Nuo Han, and Haiping Yu. Application of large language models in stroke rehabil- itation health education: 2-phase study. Journal of Medical Internet Research, 27:e73226āe73226, 2025. 2 [14] Qualisys AB. Qualisys track manager (qtm). 5 [15]Nikhil Raina, Guruprasad Somasundaram, Kang Zheng, Sagar Miglani, Steve Saarinen, Jeff Meissner, Mark Schwesinger, Luis Pesqueira, Ishita Prasad, Edward Miller, Prince Gupta, Mingfei Yan, Richard Newcombe, Carl Ren, and Omkar M Parkhi. Egoblur: Responsible innovation in aria, 2023. 4 [16] Zhiyao Ren, Yibing Zhan, Baosheng Yu, Liang Ding, Pingbo Xu, and Dacheng Tao. Healthcare agent: eliciting the power of large language models for medical consultation. npj Artifi- cial Intelligence, 1(24), 2025. 1, 2 [17] Alexandra Retevoi, Giada Devittori, Tobias Kowatsch, and Olivier Lambercy. Building conversational agents for stroke rehabilitation: An evaluation of large language models and retrieval augmented generation. In Proceedings of the 24th ACM International Conference on Intelligent Virtual Agents, New York, NY, USA, 2024. Association for Computing Ma- chinery. 2 [18] Arthur A. Rodriquez, Pamela O. Black, Kay A. Kile, Jack Sherman, Beth Stellberg, James McCormick, Jane Roszkowski, and Eileen Swiggum. Gait training efficacy using a home-based practice model in chronic hemiplegia. Archives of Physical Medicine and Rehabilitation, 77(8):801ā 805, 1996. 2 [19]Yashoda Sharma, Lovisa Cheung, Kara K. Patterson, and Andrea Iaboni. Factors influencing the clinical adoption of quantitative gait analysis technology with a focus on clinical efficacy and clinician perspectives: A scoping review. Gait and Posture, 108:228ā242, 2024. 1, 2 [20]Aaditya Singh and Adam Fry et al. Openai gpt-5 system card, 2025. 2, 5 [21] Dmitry V. Skvortsov, Sergey N. Kaurkin, Natalya V. Grebenk- ina, and Galina E. Ivanova. Typical changes in gait biome- chanics in patients with subacute ischemic stroke. Diagnos- tics, 15(5), 2025. 2 [22] Tamaya Van Criekinge, Wim Saeys, Steven Truijen, Luc Vereeck, Lizeth H. Sloot, and Ann Hallemans. A full-body motion capture gait dataset of 138 able-bodied adults across the life span and 50 stroke survivors. Scientific Data, 10:852, 2023. 1, 2, 8 [23] Josip Vrdoljak, Zvonimir Boban, Marino Vilovi Ģ c, Marko Kumri Ģ c, and Jo Ė sko Bo Ė zi Ģ c. A review of large language models in medical education, clinical decision support, and healthcare administration. Healthcare, 13(6):603, 2025. 1, 2