Objective: We validated a single-camera AI pipeline to predict clinician-rated gait severity in cerebellar ataxia from standard video recordings.

Background: Gait instability is among the most disabling features of cerebellar ataxia. The Scale for the Assessment and Rating of Ataxia (SARA) is the standard measure but is limited by ordinal structure, inter-rater variability, and poor suitability for remote monitoring. Prior work has used custom pose extraction pipelines to predict SARA from gait video, but whether off-the-shelf pose estimators can match this performance with less implementation burden has not been tested.

Method: We used the University of Rochester Auto-Gait dataset: gait videos from 89 participants (65 SCA, 24 controls), labeled with SARA gait sub-scores (0-6). We applied MediaPipe Pose Landmarker Heavy to each 6-second video, detecting 33 landmarks per frame in image and metric coordinates (~180 frames, 30 FPS). We computed per-frame gait measurements (step width, trunk angle, knee/hip angles, arm swing, ankle height, speed), aggregated into statistical summaries plus lateral sway, vertical oscillation, arm swing asymmetry, and gait cycle regularity (60 features total). Spearman correlations with Benjamini-Hochberg FDR correction assessed feature-SARA associations. ML models (Random Forest, Gradient Boosting, SVR) predicted SARA via leave-one-out cross-validation. SHapley Additive exPlanations (SHAP) identified top predictive features.

Results: Of 152 videos, 148 were processed end-to-end (97.4%), with mean per-video pose detection of 97.9%. Normalized step width showed the strongest single-feature correlation with SARA gait score (rho=0.549, FDR p-value=5.3e-13). The best model (SVR, RBF kernel) achieved MAE 0.891 SARA points, R-squared 0.297, and Pearson r 0.559, comparable to prior methods. SHAP identified step width, normalized step width, knee angle, ankle height range, and gait regularity as top predictors.

Conclusion: An off-the-shelf pose estimator predicted SARA gait severity with sub-1-point accuracy from standard video, with performance comparable to prior custom pipelines but without multi-step preprocessing. Step width and gait regularity were the strongest predictors. This smartphone-only approach warrants prospective validation for clinic screening and home monitoring in cerebellar ataxia.

*Authors HK, SK, and KT contributed equally to this work.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/single-camera-ai-pose-estimation-predicts-clinician-rated-gait-severity-in-cerebellar-ataxia/