Human Movement & Mobility Research
Sony RX0-II

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

Clinical Biomechanics
•
2026
•
Original research
Theia3D version:
Not specified in the paper
More Articles

Study at a glance

In 50 community-dwelling older adults, a set of machine learning models using Theia3D-derived gait data to estimate Montreal Cognitive Assessment scores did not outperform a simple baseline that always predicts the group average score, after accounting for multiple comparisons, underscoring how far gait-based cognitive screening still has to go before it adds value beyond existing tools.

Summary

Disclaimer: The following summary details independent academic research. Theia3D is intended for research, sports performance and educational use only. It is not cleared as a medical device and is not intended to be used for the diagnosis, treatment, mitigation or prevention of any disease, injury or medical condition.

Editor's note: this paper does not state a specific Theia3D software version in its methods; it cites Theia3D markerless motion capture generally, referencing a separate review publication. Here, Theia3D served as the data-collection tool generating gait data for a machine learning question, rather than as a system being validated against marker-based motion capture.

Study Overview

Design

An exploratory evaluation comparing fourteen machine learning models, estimating Montreal Cognitive Assessment scores from Theia3D-derived gait data, demographics, and gait variability measures, against a simple average-prediction baseline, in community-dwelling older adults.

Participants

Fifty community-dwelling older adults (20 men, 30 women; ages 60 to 84), recruited through a local employment-support organization for older adults and tested in a single university laboratory session. Participants were required to be able to walk independently and were excluded if they self-reported a diagnosis affecting movement or cognition, or used a walking aid; cognitive status was not independently verified beyond this self-report and the MoCA administered as part of the study itself.

Data Collection

Each session lasted about 30 minutes. A researcher certified in official MoCA administration gave the assessment before the walking task. Participants then walked barefoot at a comfortable pace along a roughly 7-meter path for three round trips, captured by eight synchronized Sony RX0-II cameras processed through Theia3D at 120 Hz. Joint angles were calculated at 11 sites (trunk and bilateral shoulder, elbow, hip, knee, and ankle), and gait cycles were identified between successive heel-strikes, with 713 usable cycles retained across all participants after excluding unusually slow or incomplete cycles.

Analysis

Four families of models were compared: six conventional statistical regressions using summary gait and demographic measures; a deep learning model (long short-term memory, or LSTM) using only joint-angle trajectories; a multimodal LSTM adding gait-variability measures and demographics to the joint-angle data; and six hybrid models combining LSTM-derived features with conventional regression. All fourteen models were evaluated using five-fold cross-validation and compared, using a statistical correction for multiple comparisons, against a null model that always predicted the training-group's average MoCA score. A technique called SHAP (Shapley additive explanations) was used to examine which parts of the gait cycle most influenced the deep learning model's predictions, and whether that pattern held up consistently across different subsets of participants.

Key Findings

  • None of the fourteen primary models showed a statistically significant improvement over the simple average-prediction baseline once the researchers corrected for the number of comparisons performed, even though the best-performing model showed a nominally lower prediction error before that correction was applied.
  • A model using only age, sex, and education performed comparably to, and in one direct comparison significantly outperformed, the deep learning model using detailed joint-angle trajectories alone, suggesting that in this dataset, basic demographic information carried at least as much estimation value as the detailed movement data captured by markerless motion capture.
  • Two gait-variability measures, how consistently a person's stance time and double-support time varied from stride to stride, showed the strongest associations with MoCA score among eleven gait parameters tested, but neither relationship remained statistically significant after accounting for age, sex, and education and correcting for multiple comparisons.
  • The deep learning model's predictions were disproportionately influenced by the final portion of each gait cycle, with the single last frame alone accounting for nearly a quarter of its total attention, a pattern the authors note is more consistent with a known tendency of this type of model architecture to weight its most recently processed information most heavily than with a specific, independently confirmed gait-related signal.
  • The specific gait-cycle regions the model relied on most varied considerably depending on which subset of participants it was trained on, with only moderate agreement between different training runs, indicating the model's internal reasoning was not fully stable or reproducible across the sample.

What This Means for Movement Science Researchers and Cognitive Screening Research Teams

For researchers exploring gait-based tools for cognitive screening, this study offers a candid and methodologically useful result, even though it is not a positive validation finding: none of the gait-based machine learning models tested here outperformed a simple average-prediction comparison once proper statistical correction was applied, and a model's internal reasoning, examined here using explainability analysis, did not hold up consistently across different training subsets. The authors frame this explicitly as a call for more rigorous benchmarking practices in this growing research area, specifically checking a model against a trivial baseline and verifying that its explanations are stable, rather than reporting promising-looking accuracy figures alone. For teams working with Theia3D-derived gait data for cognitive or other screening applications, this study is a useful, honest reminder that detailed movement data does not automatically translate into predictive value, and that rigorous baseline comparison is an essential step before concluding that it does.

The Role of Theia3D in Movement Analysis

Theia3D served purely as the markerless motion capture tool generating the gait data this study's machine learning models were built on; it was not itself compared against marker-based motion capture or evaluated for its own measurement accuracy in this paper. This reflects a use case in which markerless capture's practical advantages, reduced setup time and no marker application, make large-scale gait data collection for exploratory machine learning questions like this one considerably more feasible than it would be with a marker-based system.

Study Limitations and Future Directions

The authors are candid that this study was exploratory and was not designed or powered to establish definitive predictive performance: they note their sample size would have needed to be roughly twice as large to reliably detect even a single moderate correlation after correcting for the eleven gait parameters tested, and that repeated or nested cross-validation, participant-balanced sampling, and testing on an entirely separate sample were not performed. Cognitive status beyond the MoCA itself was not independently confirmed, and participants' physical activity levels, use of medications, and other conditions that can influence gait were self-reported rather than independently verified, leaving room for factors the study could not fully account for. The authors also note that gait-event timing was not checked against a force-plate or manual reference, and that their analysis was limited to front-to-back movement in the plane of walking rather than side-to-side or rotational movement. They identify testing on larger, more diverse samples, validating explainability findings against ground-truth gait mechanics rather than model architecture alone, and applying similar baseline and stability checks to related gait-based screening research as useful next steps.

Citation: Totoki, Y., Hasegawa, T., Kaminishi, K., Yozu, A., & Ota, J. (2026). Explainable machine learning for Montreal Cognitive Assessment score estimation from gait in community-dwelling older adults: An exploratory evaluation. Clinical Biomechanics, 141, 106996. https://doi.org/10.1016/j.clinbiomech.2026.106996

Link: Read the full study

Does Gait-Based Machine Learning Add Value for Cognitive Screening?

This study of 50 community-dwelling older adults finds that machine learning models using Theia3D-derived gait data did not outperform a simple average-based comparison for estimating Montreal Cognitive Assessment scores, a candid methodological finding that highlights the importance of rigorous baseline testing in gait-based screening research. If your work involves gait-based research and you want to discuss how markerless motion capture can support rigorous study design, talk to our team about applying Theia3D to your own research.

Recent Posts