An Integrated Machine Learning Framework for Multi-Disease Prediction with Feature Selection

Abstract

The coexistence of multiple chronic diseases is common among older adults and presents substantial challenges for disease prevention, management, and healthcare planning. However, most previous studies have focused on predicting individual diseases, although multiple chronic conditions commonly coexist. This study evaluates a multi-output classification model that predicts twelve chronic disease-status labels independently and examines genetic algorithm (GA)- and particle swarm optimization (PSO)-based feature selection using a shared feature subset. Data from the Korean Longitudinal Study of Aging (KLoSA) were analyzed (n=4529). Random forest, decision tree, and XGBoost models were evaluated with five repeated nested holdout splits so that feature selection did not use the outer test set. Across runs, GA and PSO reduced the 131-feature input to 49–77 features, but changes in Composite Macro-F1 were small and model-dependent. High mean label accuracy coexisted with low sensitivity for several rare disease-positive classes. SHAP analysis on held-out data was therefore interpreted descriptively and showed lower ranking stability for selected models. These results indicate a trade-off between feature reduction, predictive performance, and computational cost rather than a general performance advantage or evidence between disease-label interactions.