Explainable machine learning for Montreal Cognitive Assessment score estimation from gait in community-dwelling older...

Machine learning for gait-based cognitive assessment is rarely benchmarked against a mean-prediction null or checked for explanation stability. We evaluated an explainable multimodal model against this null to estimate Montreal Cognitive Assessment (MoCA) score from gait. We assessed explanation stability and the contribution of joint-angle trajectories.

Fifty community-dwelling older adults walked overground with markerless motion capture, yielding 713 gait cycles. Gait-MoCA associations were assessed with Pearson and partial correlations. Under participant-level five-fold cross-validation, a multimodal long short-term memory (LSTM) model was compared with summary-statistics models, reduced input sets omitting the trajectories, and a null model returning the training-fold mean. Scaling and epoch selection were limited to training data. Shapley additive explanations (SHAP) were computed per fold.

Four gait parameters were associated with MoCA after false-discovery-rate correction in unadjusted analyses. All five variability-coefficient correlations with MoCA were positive. Stance-time and double-support-time variability had the strongest covariate-adjusted correlations (partial r=0.38 and 0.34), although no covariate-adjusted correlation survived correction. After correction, no model tested against the null improved significantly. The null-model participant-level root mean square error was 3.03, versus 2.83 for the best model, 2.96 for the multimodal LSTM and 2.92 for the selected demographics-only model. SHAP attributions concentrated at the end of the cycle, 24% in the final frame, but varied between folds.

The variability measures capture between-cycle differences that the single-cycle models did not encode and may guide future model development. Benchmarking against a mean-prediction null and quantifying explanation stability may help identify which inputs merit further evaluation.