Truace
Subscribe

Explainable machine learning for Montreal Cognitive Assessment score estimation from gait in community-dwelling older adults: An exploratory evaluation

Background: 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. Methods: Fifty community-dwelling older adults walked overground with markerless motion capture, yielding 713 gait cycles. Gait-MoCA associations we…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1982· cite & verify
Explainable machine learning for Montreal Cognitive Assessment score estimation from gait in community-dwelling older adults: An exploratory evaluation
Growing children and awkward walking by Nunn, Thomas Percy, Sir, b. 1870. Public domain

In brief

Researchers recorded overground walking in fifty community-dwelling older adults using markerless motion capture, producing 713 gait cycles. They tested whether an explainable multimodal LSTM could estimate Montreal Cognitive Assessment scores, comparing it with simpler models and with a baseline that just returned the average training score.

The comparison matters because gait-based cognitive screening models are rarely tested against such a baseline or checked for stable explanations. Here the models gave little real advantage and their explanations shifted across validation folds, so any clinical use remains unsupported. The authors frame the work as exploratory, with gait variability as a lead for later model development.

Main points

  1. Four gait parameters were associated with MoCA after false-discovery-rate correction in unadjusted analyses, but no covariate-adjusted correlation survived correction.
  2. Stance-time and double-support-time variability had the strongest adjusted correlations, with partial r of 0.38 and 0.34.
  3. Participant-level RMSE was 3.03 for the null model, 2.83 for the best model, 2.96 for the multimodal LSTM and 2.92 for the demographics-only model.
  4. SHAP attributions concentrated at the end of the gait cycle, including 24% in the final frame, but varied between folds.

The problem

In 50 community-dwelling older adults, no gait-based or demographics-based model, including the multimodal LSTM, significantly outperformed a null model that simply predicted the training mean MoCA score.

The rundown

The study used participant-level five-fold cross-validation and compared a multimodal LSTM with summary-statistics models, reduced input sets that omitted joint-angle trajectories, and a null model returning the training-fold mean. Scaling and epoch selection were restricted to training data.

All five variability-coefficient correlations with MoCA were positive. The authors suggest variability measures capture between-cycle differences that the single-cycle models did not encode.

The authors argue that benchmarking against a mean-prediction null and measuring explanation stability can help identify which inputs merit further evaluation.

Sources

  1. Peer-reviewedClinical Biomechanics2026-10-18

The debate