Interpretable machine learning models for predicting depression status and remission three years later in older...

Depression significantly impacts older adults, making it valuable to use machine learning to predict their future depressive status and assess whether currently depressed older adults may experience remission in the future.

This study included a total of 5310 participants from the China Health and Retirement Longitudinal Study Wave 3 and 4. We used five machine learning algorithms to construct (a) a predictive model for the future depressive status (DS) of older adults and (b) a predictive model for whether currently depressed older adults will experience remission (DR). We also used SHapley Additive exPlanations method to interpret the model and build a simplified version.

The DS group included 5310 participants (2777 males, 52.30 %), with an average age of 66.83 (5.56). The DR group consisted of 1775 participants (729 males, 41.07 %), with an average age of 66.83 (5.41). For the DS group, eXtreme Gradient Boosting (XGBoost) achieved the highest accuracy, specificity and area under the curve (0.710, 0.832, 0.738), and better net benefit under certain threshold probability. For the DR group, XGBoost also saw the highest accuracy, sensitivity and F1 score (0.702, 0.829, 0.785), and better calibration curve. The top ten predictors were used to build simplified models and the simplified models showed slight declines compared with the original ones.

The predictive performance of simplified models was modest, indicating that the models should be applied with caution in practice.

This study developed prediction and simplified models, offering significant potential for improving health management in such populations.