AIDS Infection Prediction from Clinical and Demographic Data Using Ensemble and Linear Models

Accurate, data-driven screening can support timely HIV/AIDS care in settings where laboratory resources are limited. This paper presents a reproducible, data-driven pipeline for predicting AIDS infection status from tabular clinical and demographic data in resource-limited settings. The workflow includes label harmonization into a binary target, class imbalance assessment, and value-annotated exploratory data analysis. Preprocessing involves median or mode imputation, standardization, and one-hot encoding. Dimensionality reduction via PCA and t-SNE complements feature ranking using ANOVA F, chi-square, mutual information, Random Forest importance, L1-logistic coefficients, and RFECV. Multiple classifiers—linear, bagged, boosted, and ensemble—are benchmarked. The Soft Voting, ExtraTrees, and Random Forest models achieved around 82.5% accuracy and an AUC of 0.817 on a stratified hold-out set. Logistic regression attained a comparable AUC of 0.818 but with lower accuracy, illustrating calibration-accuracy trade-offs. While not a diagnostic substitute, the findings suggest that transparent, lightweight tabular models can enable early screening, triage, and prioritization for confirmatory HIV/AIDS testing in low-resource healthcare environments.