Cardiovascular disease risk prediction among people living with HIV in Uganda: A comparison of traditional statistical and machine learning models

Résumé fourni par la source

Objective: To develop and compare the performance of traditional statistical models and machine learning (ML) algorithms for cardiovascular diseases (CVDs) risk prediction among people living with HIV (PLHIV) in Uganda, and identify key predictors using interpretable AI tools. Methods: Data from 1,000 adult PLHIV followed for 10 years at the Infectious Diseases Institute, Kampala were analyzed. CVD was defined as a composite of stroke, myocardial infarction, heart failure, and peripheral vascular disease. Models included Logistic regression, Cox proportional hazards, and ML algorithms (Decision Tree, Random Forests, Support Vector Machine, XGBoost, and Soft Voting Ensemble). Class imbalance was handled using the SMOTE, and hyperparameters were tuned through a grid search with 10-fold cross-validation. Performance was assessed using AUC, sensitivity, specificity, precision, recall and F1-Score. Feature importance was examined with SHAP and permutation analyses. Results: Of the participants, 61.7% were female, median age was 55 years, median ART duration 231.7 months, and median CD4 count was 504 cell/µl. Most had a viral load of <400copies/Ml (91.3%), self-employed (58.3%) and married/cohabiting (50.8%). During follow-up 31.8% developed CVDs, more commonly among those ≥55years (34.7% vs 28.3%, P=0.03), with hypertension (36.3% vs 28.6, P=0.01) or diabetes (43.2% vs 30.4%, P=0.06). No differences seen by gender, employment or substance use. The median ART duration before CVDs onset was 232.3 months. Traditional models showed moderate predictive ability (AUC < 0.65), Soft Voting Ensemble achieved best performance (AUC: 0.812) and feature importance highlighted, CD4 count, hypertension, diabetes, alcohol usage, age, marital and employment status as key predictors. Conclusion: Nearly one-third of the participants developed CVDs. ML ensemble models substantially outperformed Traditional models for CVD risk prediction, while interpretable ML identified clinically relevant predictors. These findings support integration ML tools into HIV care to strengthen CVD risk stratification and prevention.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.