Video-Based Gait Assessment Using Machine Learning to Classify Age and Sex in Low-Resource Settings: Cross-Sectional...

Gait assessment is an important tool for evaluating health risks in older adults but remains underused in low-resource settings. We explored the feasibility of using a low-cost, simple walking protocol with smartphone video capture to extract health-related gait signals by classifying sex and age. Sex and age are fundamental biological factors linked to most health- and aging-related outcomes. Establishing baseline classification performance provides justification for future exploration of more complex health-related conditions using this protocol.

This study aimed to assess whether pose parameters derived from smartphone-based gait videos can be used by machine learning models to classify age and sex.

A cross-sectional study was conducted with 155 participants (Thailand: n=59, 38.1%; India: n=96, 61.9%). Participants performed a simple walking protocol while being recorded using smartphones. Pose estimation was conducted using the MediaPipe algorithm to extract 109 features related to joint distances, angles, and walking speed. For feasibility assessment, we calculated the proportion of recordings for which pose estimation could be extracted. Elastic-net logistic regression and histogram-based gradient boosting classifiers were used for analysis. Model performance was evaluated using 5-fold cross-validation. Outcomes were sex (male vs female) and age group (aged<65 vs ≥65 y).

Pose parameters were successfully extracted from 145 (93.5%) of the 155 video recordings. Among the 145 participants, 94 (64.8%) were female, and 55 (37.9%) were aged 65 years or older. The 2 analytic models demonstrated comparable performance. Sex classification achieved a maximum mean area under the receiver operating characteristic curve of approximately 0.90 (SD 0.06), whereas age classification achieved a maximum mean area under the receiver operating characteristic curve of approximately 0.70 (SD 0.09). Classification performance was primarily influenced by the number of features used, clothing characteristics, and the quality of pose estimation.

This simple smartphone-based gait assessment protocol was able to extract meaningful pose parameters and classify biological features (age and sex). Further studies are warranted to evaluate its potential utility for disease screening, risk stratification, and longitudinal health monitoring.