AI-Powered Mortality Prediction for HIV/AIDS Patients on ART in Nigeria

This study compares three Machine Learning (ML) algorithmsâĂŤlogistic regression, random forest, and gradient booster classificationâĂŤand three deep learning (DL) algorithmsâĂŤartificial neural network, tabular model, and long short-term memory network (LSTM)âĂŤto predict mortality rates among HIV/AIDS patients in Nigeria. The research utilized a large electronic medical records database, merging clinical and demographic data like CD4 count, viral load, age, and ART duration. Following ethical approval from the Nigerian Ministry of Health, the data was preprocessed to address a significant 1:40 class imbalance using SMOTE oversampling and standardization. Feature engineering was also performed, including the encoding of categorical variables. Key findings indicate that ’Current_Age,’ ’MaritalStatus,’ and ’Duration on ART (Days)’ significantly impacted mortality prediction, while ’Sex’ features had minimal actual influence. SHAP analysis was used to interpret feature contributions. Although the ML and ANN models were explainable, the LSTM network achieved perfect scores (accuracy, precision, recall, F1 score of 1), suggesting potential overfitting despite controls like Loss-Based Stopping and a dropout layer. The tabular embedding model, with an accuracy of 0.9695, highlighted that not all metrics are suitable for highly imbalanced scenarios. The study emphasizes the critical importance of data integrity for policy relevance and suggests future research should use different validation techniques to ensure model reliability.