Interpretable machine learning for depression recognition with spatiotemporal gait features among older adults: a cross-sectional study in Xiamen, China.

OBJECTIVE

Depression in older adults is a growing public health concern, yet there is still a lack of convenient and real-time methods for depressive symptoms identification. This study aims to develop a gait-based depression recognition method for Chinese community-dwelling older adults.

METHODS

Ninety-two participants aged over 60 from Xiamen, China, were recruited for a three-week cross-sectional study. Depressive symptoms were assessed using the 10-item Center for Epidemiologic Studies Depression Scale, with a score of ≥ 10 indicating depression. From each group (depression and non-depression), twenty-five individuals were randomly selected for Kinect-based gait analysis. Gait data were recorded using a Microsoft Kinect in the indoor experimental area. χ and t-tests were used for statistical comparisons. Four machine learning techniques including Logistic Regression, Support Vector Machine, Gradient Boosting Decision Tree, and Random Forest were employed to develop predictive models for depression. SHapley Additive exPlanations were used to explain the feature importance.

RESULTS

The average age was 64.1, and 72% were female, with the prevalence of depressive symptoms was 29.34%. Older adults with depressive symptoms exhibited significant gait abnormalities, including reduced body sway (P < 0.01, 95% CI (11.81, 81.79)), right-arm swing (P < 0.05, 95% CI (4.90, 85.07)), left step length (P < 0.05, 95% CI (5.43, 154.32)), right step length (P < 0.05, 95% CI (23.89, 171.36)), left step height (P < 0.001, 95% CI (100.42, 337.85)), walking speed (P < 0.001, 95% CI (245.79, 882.54)), step width (P < 0.05, 95% CI (3.46, 172.34)), and right stride length (P < 0.01, 95% CI (3.99, 25.18)). The Random Forest algorithm achieved the best performance (AUC-ROC = 0.911, Sensitivity = 0.857) in differentiating individuals with or without depressive symptoms based on discriminated spatiotemporal gait features. The five most important gait parameters in the optimal model were left step height, walking speed, right step height, body sway, and step width.

CONCLUSIONS

Spatiotemporal gait features are associated with depressive symptoms in older adults. The developed machine learning models with high predictive accuracy, suggest the potential of Kinect-based gait assessment as a real-time and cost-effective screening tool for older adults with depressive symptoms.