Abstract
BACKGROUND: Depression among the elderly is a growing public health concern, especially in India. This study aimed to investigate the predictive validity of physiological, psychological, and functional health factors in classifying the level of depressive symptoms among the elderly using the extreme gradient boosting (XGBoost) technique. Additionally, we compared the performance of models trained on original and resampled data. METHODS: This study is entirely based on secondary data analysis of the Longitudinal Aging Study in India wave 1 data. We classified the observations into "high depressive symptom" and "low/no depressive symptom" groups based on the predictors, including physiological, psychological, and functional health factors, along with socio-demographic factors. We developed three models (Models 1, 2, and 3) trained on original, over-sampled, and under-sampled data, respectively. Model performance was evaluated using the metrics of balanced accuracy, sensitivity, specificity, and area under the receiver operating characteristics curve (AUC). RESULTS: The study included 26,065 individuals aged 60 and above. Model 3, trained on under-sampled data, demonstrated the best overall performance. It achieved a balanced accuracy of 64%, with a sensitivity of 62.8% and specificity of 65.2%. The AUC for Model 3 was 0.692. Feature importance analysis revealed that life satisfaction, instrumental activities of daily living, mobility, caste, and monthly per capita expenditure quintiles were among the most influential factors in predicting the level of depressive symptoms. CONCLUSION: The XGBoost models demonstrate promise in predicting depressive symptoms among the elderly. These findings suggest that machine learning models can be envisaged for early detection and management of depression, especially in primary care.