JMIR Publications

ABSTRACT

Background:

Background:

Gait assessment is a tool for evaluating health risks in older adults but is underutilized in low-resource settings due to resource constraints. This study explores the feasibility of using a simple walking protocol, smartphone video capture, and machine learning (ML) techniques to classify age and sex—two critical health-related factors—in India and Thailand, as the first step towards creating a pipeline to assess health conditions.

Objective:

To assess whether pose parameters derived from smartphone-based gait videos can be used by ML models to classify age and sex.

Methods:

A cross-sectional study was conducted with 155 participants (59 from Thailand and 96 from India). Participants performed a simple walking protocol while being recorded with smartphones. Pose estimation was conducted using the MediaPipe algorithm to extract 109 features related to joint distances, angles, and walking speed. Machine learning models—including Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Stochastic Gradient Descent (SGD), Multi-Layer Perceptron (MLP), and an ensemble model—were trained to classify participants based on age (<65 or ≥65 years) and sex.

Results:

Pose parameters were successfully extracted from 145 of 155 video recordings (93.5%). Among 145 participants, 65% were female, and 38% were aged ≥65 years. Sex classification achieved 80–90% accuracy, with female classification sensitivity exceeding 90%, while male sensitivity ranged from 65–80%. Age classification accuracy ranged from 70–80%, with higher sensitivity for individuals aged <65 years (≥70%) than for those ≥65 years (40–60%). Performance variability was influenced by class imbalance, clothing type, and pose estimation quality. Conclusion: This study demonstrates the feasibility of using a smartphone-based video gait assessment protocol in low-resource settings for age and sex classification. The approach shows promise as a scalable, low-cost geriatric screening tool. Future research should focus on dataset expansion, ML model refinement, and exploring more clinical-related phenotypes such as frailty and fall risk.

Objective:

Objective:

To assess whether pose parameters derived from smartphone-based gait videos can be used by ML models to classify age and sex.

Methods:

Methods:

A cross-sectional study was conducted with 155 participants (59 from Thailand and 96 from India). Participants performed a simple walking protocol while being recorded with smartphones. Pose estimation was conducted using the MediaPipe algorithm to extract 109 features related to joint distances, angles, and walking speed. Machine learning models—including Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Stochastic Gradient Descent (SGD), Multi-Layer Perceptron (MLP), and an ensemble model—were trained to classify participants based on age (<65 or ≥65 years) and sex.

Results:

Results:

Pose parameters were successfully extracted from 145 of 155 video recordings (93.5%). Among 145 participants, 65% were female, and 38% were aged ≥65 years. Sex classification achieved 80–90% accuracy, with female classification sensitivity exceeding 90%, while male sensitivity ranged from 65–80%. Age classification accuracy ranged from 70–80%, with higher sensitivity for individuals aged <65 years (≥70%) than for those ≥65 years (40–60%). Performance variability was influenced by class imbalance, clothing type, and pose estimation quality.

Conclusions:

Conclusion: This study demonstrates the feasibility of using a smartphone-based video gait assessment protocol in low-resource settings for age and sex classification. The approach shows promise as a scalable, low-cost geriatric screening tool. Future research should focus on dataset expansion, ML model refinement, and exploring more clinical-related phenotypes such as frailty and fall risk.