Interpretable machine learning for identifying co-occurring internet addiction and maladaptive cognitive emotion regulation in older adults: a cross-sectional study.

Background: Internet addiction and maladaptive cognitive emotion regulation may co-occur in later life, yet evidence-based approaches for identifying this co-occurrence remain limited. This study developed and interpreted a machine learning classifier to distinguish community-dwelling older adults with versus without current co-occurrence. A cross-sectional survey was conducted among 1,471 adults aged 60 years or older in China. Candidate predictors included demographic characteristics, general well-being, family functioning, and social support. Methods: After preprocessing, correlation-based collinearity screening and ten-fold cross-validated LASSO regression were used for feature selection. Five models, including logistic regression, random forest, support vector machine, gradient boosting machine, and multilayer perceptron, were developed using the training set and internally validated in a randomly held-out test set. Model performance was evaluated using discrimination, calibration, and clinical utility metrics, and SHAP was applied for model interpretation. Results: Eight predictors were retained: general well-being, support utilization, age, gender, educational attainment, marital status, objective support, and living with children. The gradient boosting machine (GBM) showed the best overall performance in the test set, with an AUROC of 0.904, accuracy of 0.816, sensitivity of 0.810, specificity of 0.818, and Brier score of 0.118. SHAP analysis identified general well-being as the dominant predictor, followed by educational attainment and objective support. Lower well-being, older age, male gender, higher educational attainment, and not being married or partnered were generally associated with a higher model-predicted probability of current co-occurrence; higher support utilization was also associated with a higher predicted probability, particularly at lower levels of general well-being. Conclusion: This interpretable classifier may facilitate the identification of co-occurring conditions among community-dwelling older adults. Given its cross-sectional design, the model distinguishes current co-occurrence status rather than predicting future onset.