DrGait: Biomechanically Grounded Visual Reasoning for Interpretable Clinical Gait Analysis
Xiangyu Yin, Shiqi Wang, Abrar Alamri, Yasir Aljohani, Weichen Liu, Goeran Fiedler, Wei Gao
cs.CV · cs.AI
Quick verdict
Tool-grounded TVS plausibly curbs VLM gait hallucination; ablations support it, but modest accuracy, n=3 human eval and a 69.6/49.6% inconsistency temper it.
Summary
DrGait is a training-free agentic framework that recasts the vision-language model (VLM) in clinical gait video analysis from a direct visual reasoner into a planner. A Triage-Verification-Synthesis (TVS) loop first applies a normality gate and ranks diagnostic hypotheses against a curated clinical knowledge base; the agent then calls deterministic biomechanical tools — joint angles at gait events, postural sway, foot clearance, arm swing — over reconstructed 3D SMPL mesh trajectories, segmented 2D pose tracks, and event-centered video frames, with recursive synthesis until convergence or an iteration cap. On GAVD (general pathological gaits) and ProGait (prosthetic deviations), the authors report 92.4% binary-normality and 51.4% abnormal-subtype accuracy with Gemini-3-Flash, approaching a human expert's 56.1% overall accuracy, plus the best ProGait top-3 accuracy (32.6%), though Gemini-3.1-Pro wins ProGait top-1 (18.11%). Three-rater human evaluations prefer DrGait in 73.2% (GAVD) and 69.6% of ProGait cases per Table 4 — the text separately states 49.64% for ProGait. Ablations attribute fine-grained gains mainly to tool verification (removing tools: ProGait top-3 32.6%→13.1%). Limitations: no significance testing, only zero-shot Gemini baselines, an internal inconsistency in human-eval numbers, and modest absolute accuracy cautioning against near-term clinical use.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Decoupling VLM reasoning from geometric perception via a Triage-Verification-Synthesis loop with deterministic biomechanical tools reduces hallucination in gait video diagnosis without any training.
- DrGait improves gait classification accuracy by up to 31% over zero-shot VLM baselines and reaches 92.4% binary-normality / 51.4% abnormal-subtype accuracy on GAVD with Gemini-3-Flash, which the authors frame as approaching the human expert's 56.1% overall accuracy.
- Tool verification, rather than triage alone or recursion, is the main driver of fine-grained diagnostic performance (removing tools cuts ProGait top-3 from 32.6% to 13.1%, Table 7).
- Human evaluators prefer DrGait's tool-backed reports in 73.2% (GAVD) and 69.6% (ProGait, Table 4) of cases, and DrGait on the cheaper Flash model ($0.011/video) outperforms the more expensive direct Pro baseline ($0.012/video).
Strengths
Well-motivated decoupling principle with ablation backing
Shifting the VLM from direct visual reasoner to planner over deterministic tools targets the documented failure mode of text-aligned encoders compressing spatial coordinates (§1). Ablation Table 7 substantiates it: removing tools drops GAVD abnormal-subtype accuracy 51.4%→44.9%, ProGait top-3 32.6%→13.1%, and ProGait LLM-judge 1.78→1.30.
Genuine auditability, not decoration
Every tool call carries a why_used rationale (§3.3); triage/synthesis outputs are schema-constrained JSON with an explicit convergence predicate (Algorithm 1); Appendices B-C publish full traces, including a Myopathic case where measured hip extension (180.0°) and knee angle (171.56°) overturn an initial Parkinsonian hypothesis.
Practical cost-performance argument
Selective honesty in reporting
Weaknesses
No statistical rigor anywhere
Internal inconsistency in a headline number
Incomplete baseline set for the headline claim
Modest absolute accuracy vs clinical framing
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DrGait (arXiv 2609.28796, University of Pittsburgh) is a training-free agentic framework for clinical gait video analysis. Central claims: (1) direct VLM inference hallucinates kinematic values because text-aligned encoders compress geometry, so the VLM should plan rather than measure; (2) a Triage-Verification-Synthesis (TVS) loop with deterministic biomechanical tools improves classification 'by up to 31%' (§1), reaching 92.4% binary-normality and 51.4% abnormal-subtype accuracy on GAVD with Gemini-3-Flash — framed as approaching a human expert's 56.1% overall accuracy; (3) tool-backed reports are preferred by human evaluators in 73.2% (GAVD) and 69.6% (ProGait, Table 4) of cases; (4) a small Flash model plus tools beats a larger direct Pro baseline at comparable cost (§4.4). Evaluation covers GAVD [22] and the authors' ProGait [23].
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DrGait and how does it reduce hallucinations in gait analysis?
DrGait is a training-free framework in which a vision-language model plans a diagnosis instead of reading measurements off video. In a Triage step it proposes candidate gait hypotheses; in Verification it calls deterministic biomechanical tools (joint angles at gait events, trunk sway, foot clearance, arm swing) computed from 3D mesh and 2D pose reconstructions; in Synthesis it writes an evidence-backed report. Because the VLM never invents numbers — every measurement comes from a tool with a logged rationale — the geometric hallucinations typical of direct VLM inference are largely avoided.
How accurate is DrGait compared to doctors or plain VLMs?
On the GAVD dataset the authors report 92.4% accuracy on the binary normal-vs-abnormal decision and 51.4% on naming the specific abnormal subtype, versus 85.1%/39.3% for a direct Gemini-3.1-Pro baseline; a human expert reached 56.1% overall accuracy. On ProGait (prosthetic gait), top-3 category accuracy is about 32.6% and the top-1 spot actually goes to the direct Gemini-3.1-Pro baseline (18.11%). It clearly beats zero-shot VLM baselines on most metrics, but absolute numbers remain modest and the paper does not claim clinician-grade performance.
Is DrGait's code or data publicly available?
The reviewed text does not state a code release. The paper uses two datasets — the public GAVD dataset (IEEE Access 2025) and the authors' own ProGait dataset (CVPR 2025) — and publishes prompt templates, tool declarations, knowledge bases, and full reasoning traces in the appendices, but no repository link appears in the provided text. Reproducing the exact results also requires access to the proprietary Gemini-3/3.1 models used in the experiments.
Can DrGait be used in clinics today?
Not based on this paper. Latency is roughly 80-107 seconds per video, accuracy on specific diagnoses is around 51%, and 2D-pose measurements are explicitly uncalibrated image-plane proxies. The authors position the auditable, tool-backed reports as a step toward clinical trustworthiness, but no prospective clinical validation, error-propagation analysis, or safety evaluation is presented.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The main GAVD results table (per-class and overall accuracies) is truncated in the extracted text; the headline overall accuracy is only partially visible (ends in '.4%'), so accuracy statements rely on figures repeated in §4.2/4.4/4.5 and Table 7.
- Figures 1-7 are not analyzable from text alone, and Appendix G tool declarations are cut off mid-listing; function-level verification details are unverifiable.
- arXiv preprint (v1, 23 Sep 2026), not yet peer-reviewed.
- Minor bibliographic noise: reference [26] (MotionCLIP) is miscited to a UIST proceedings while it appeared in ACM TOG, suggesting limited proofreading.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.