Machine learning models for predicting metabolic syndrome to support clinical decision-making in ART-treated adults
Mpanji Siwingwa1,2, Wilbroad Mutale3,4, Sody Mweetwa Munsaka5
1School of Health Sciences, Department of Biomedical Sciences, University of Zambia, Lusaka, Zambia. mpanjisiwingwa@gmail.com.
Background:
Metabolic syndrome (MetS) is an emerging complication among people living with HIV (PLHIV) receiving long-term antiretroviral therapy (ART), particularly with protease and integrase inhibitor regimens. Early identification of high-risk individuals remains challenging, and predictive tools are limited in African settings. This study evaluated nine machine learning (ML) algorithms for predicting MetS in ART-treated adults.
Methods:
We analysed a retrospective cohort of 1,027 PLHIV without baseline MetS, followed for 144 weeks; 854 with complete data were included. Nine ML algorithms, including logistic regression, support vector machines, random forest, and XGBoost, were trained on 70% of the data using stratified repeated 10-fold cross-validation with hyperparameter tuning. Performance was assessed on a 30% test set using discrimination (AUC), calibration, and predictor importance.
Results:
Age, sex, viral load, and alcohol use were the strongest predictors of metabolic syndrome in ART-treated individuals. Discrimination was modest across models (AUC 0.49-0.64), with radial SVM and logistic regression performing best (AUC ≈ 0.64). Calibration was generally acceptable (Brier score 0.16-0.19). Sensitivity-specificity trade-offs varied: XGBoost favored sensitivity (85.2%) but had low specificity (34.4%), whereas logistic regression and random forest achieved higher specificity (~ 75%). Overall, complex models offered limited gains over logistic regression.
Conclusions:
Although internally validated ML models demonstrated acceptable calibration and modest discrimination, predictive performance remains insufficient for standalone clinical deployment and should be considered supportive rather than definitive for risk stratification in HIV care. Logistic regression and random forest provided the most consistent balance of discrimination and calibration, while complex approaches offered limited gains. Age, sex, viral load, and alcohol use emerged as key predictors. External validation and prospective evaluation are essential to establish generalisability, clinical impact, and feasibility before integration into routine practice.