Feasibility of machine learning classification of depression and anxiety symptoms among college students using 3D gait and sit-to-walk biomechanics.

BACKGROUND

Depression and anxiety are common and burdensome. Objective movement assessment may complement questionnaires by providing low burden screening signals.

RESEARCH QUESTION: Can 3D gait and sit-to-walk biomechanics classify elevated depression and anxiety symptoms, and which movement yields better performance with few features?

METHODS

Thirty young adults completed the Center for Epidemiologic Studies Depression Scale (CES-D) and the Generalized Anxiety Disorder scale (GAD-7), then performed gait and sit-to-walk in a laboratory with a sixteen camera optoelectronic motion capture system. We computed 168 gait variables and 62 sit-to-walk variables spanning speed and joint angle features. Symptom labels defined a symptomatic group as CES-D≥ 16 and GAD-7 ≥ 10, and a control group as CES-D< 16 and GAD-7 < 10. Machine learning models included decision tree, k nearest neighbors, logistic regression, and multilayer perceptron. Robustness was tested with a nested participant level scheme: in the outer loop, two participants were held out at a time (one symptomatic, one control); in the inner loop, five fold cross validation tuned hyperparameters.

RESULTS

Logistic regression reached 75 % accuracy for gait, and k nearest neighbors reached 77 % for sit-to-walk. Feature selection using Gini importance improved performance, with plateaus near the top 5-10 features.

SIGNIFICANCE: Human movement, especially gait and sit-to-walk, can support objective screening of depression and anxiety with small feature sets, suggesting a feasible path toward wearable enabled monitoring. External validation and clinically actionable thresholds are needed for translation.