
This study explores how machine learning can support better treatment decisions for people living with HIV. Using a large clinical dataset of over 283,000 patient records from treatment facilities in eastern Nigeria, four supervised learning models were compared, Random Forest, Support Vector Machine, Naïve Bayes, and Logistic Regression, to determine which best predicts medication improvement outcomes.
Rather than focusing on accuracy alone, the study emphasises clinical reliability, reproducibility, and the importance of catching positive cases consistently. The findings show that Naïve Bayes, despite being a simpler model, offered the most balanced and trustworthy performance, making it the strongest candidate for real-world clinical decision support. The study also raises important questions about overfitting and data leakage in healthcare machine learning, and highlights why model validation matters as much as model performance.
This work contributes to the growing evidence that thoughtful model selection and rigorous evaluation are essential for translating machine learning into safe, effective HIV care tools.