Classifying complex multimorbidity using latent class analysis and machine learning to generate insights into clustering of mental and cardiometabolic conditions.

Machine learning techniques earn higher accuracy and robustness in multimorbidity prediction at this moment in time. Among various forms of multimorbidity, complex multimorbidity, especially the intersection of cardiometabolic disorders and mental health conditions, poses a serious threat to the public health system and needs special priority interventions. Within the scope of this context, current study aimed to define complex multimorbidity clusters using latent class analysis (LCA), test the performance of different machine learning models for accurate classification and prediction, and identify the important features by applying three feature importance techniques. The study used an excerpt of CDC Behavioral Risk Factor Surveillance System data - BRFSS 2015. It applied LCA on 46,736 responses to identify complex multimorbidity clusters and trained six machine learning algorithms (MLR, MNB, DT, RF, XGB, and ANN) in classifying the individuals falling into a typical cluster. Performance of ML models was evaluated through AUROC, accuracy, precision, recall, and F1 score. McNemar and paired T statistics are computed to find the disagreement between the ML models to verify the suitability of model selection. RF feature importance, permutation feature importance, and SHAP values are estimated to identify risk and protective factors. Five complex multimorbidity clusters emerged from LCA, dominated by mental health conditions (30% - ~ 40%) in 1 cardiovascular cluster and 4 cardiometabolic clusters. Mental health conditions are combined with diabetes, overweight/obesity, stroke, history of heart disease, and cardiovascular risk markers. More than 60% of participants fall under complex cardiometabolic clusters who are diabetic. A greater number of overweight male/obese female with poor mental health conditions show worse CVD markers. Random Forest model outperformed other algorithms in classification task (AUROC = 0.805, 95% CI [0.800-0.809]). Mcnemar and T statistics depict significant disagreement between the results of each ML model pair (P value = 0.0000). Feature importance analyses consistently identified age, walking difficulty, socioeconomic status, general and physical health status, education, smoking habits, physical activity status and fruit/ vegetable consumption patterns as key influencing factors. Mental health plays a critical role in shaping multimorbidity clusters. AI-driven classification enables more accurate prediction of at-risk populations and can inform tailored interventions. This study can be considered as a use-case providing evidence for integrating ML into public health decision support.