Comprehensive Summary
This study, presented by Yadav and Bhutia, examines the usage of machine learning (ML) models to identify depression risk among older adults with non-communicable diseases (NCDs). Using data from the Longitudinal Ageing Study in India (LASI) Wave 1 (2017-2018; N = 58,467), the study evaluated eight supervised machine learning models which were random forest, decision tree, logistic regression, Support Vector Machine (SVM), K-Nearest Neighbor (KNN), naive bayes, neural network, and ridge classifier, in order to predict depression among older adults. Model performance concluded that random forest outperformed all other models with an Area Under the Receiver Operating Characteristics (AUROC) of 0.996 and an accuracy of 95.6% with decision tree following with AUROC of 0.915 and an accuracy of 91.5%. The study also identified key predictors of depression across SHAP values, feature importance, and regression analysis, and these predictors included poor sleep, age, body mass index (BMI), instrumental activities of daily living (IADL) limitations, monthly per capita expenditure (MPCE) quintile, religion, smoking, education, and physical inactivity. Overall, the results demonstrate that ensemble-based ML models can effectively identify depression risk in older adults using a reduced set of highly predictive and interpretable features.
Outcomes and Implications
With depression as a major public health challenge and India accounting for a significant portion of the global mental health burden, traditional mental health interventions fail to account for the complex interactions among risk factors. With the ML models that were assessed in this study, the reduced-feature random forest model is a scalable and technology-enabled approach that can be deployed by primary care and community health workers to promote early detection. The model can also be embedded into mobile health applications or electronic health records to provide efficient screening alerts and decision support for clinicians and care coordinators. By comparing algorithmic importance metrics with traditional odds ratios, this study bridges the gap between model accuracy and interpretability, which is essential for clinical and public health implementations. Therefore, the authors emphasize the model’s high potential to be integrated into India’s public health infrastructure, particularly within the frameworks of the National Program for Health Care of the Elderly (NPHCE), the National Mental Health Program (NMHP), and the Ayushman Bharat Digital Mission (ABDM). Before full integration, future research should focus on the longitudinal prediction of depression, integration with electronic health records and real-world applications of reduced-feature screening models in communities.