Abstract
Human Immunodeficiency Virus (HIV) remains a critical public health concern, and is a significant global health challenge, particularly in developing countries. Early HIV detection supports targeted interventions, and substantially reduce the HIV burden. In many resource-limited settings, early detection of HIV is hindered by stigma, limited access to testing, and low risk awareness. This study aims to enhance HIV screening in resource-limited settings by employing machine learning models to predict HIV risk using demographic and lifestyle variables. We analyzed data from 39,295 individuals in Shiraz, Iran, identifying key predictors, including drug injection, age, having a spouse with a history of HIV, occupation, and prison record. We trained and validated an Extreme Gradient Boosting (XGBoost) model using stratified five-fold cross-validation on the dataset. The XGBoost model achieved high accuracy (0.89; Confidence Interval (CI) 95% [0.88-0.89]), very-good discriminatory ability (Area Under the ROC Curve (AUC = 0.84 [0.83-0.84], with a fair-to-good agreement (Cohen's Kappa of 0.51 [0.51-0.52]). Moreover, the performance of the proposed method (PREDICT-HIV) was consistent across test folds. Our findings align with previous studies, emphasizing the importance of socio-demographic and behavioral factors in HIV risk prediction. The model's robustness suggests its potential for practical implementation, aiding early identification and intervention in high-risk groups. Future research should incorporate additional socioeconomic variables and validate the model in diverse populations to enhance global HIV prevention efforts. The web application, implemented using the Django framework, is freely available online for public access. PREDICT-HIV may support earlier identification and intervention in underserved populations, improving the efficiency of HIV screening programs.