Explainable AI-Driven Comparative Analysis of Machine Learning Models for Predicting HIV Viral Nonsuppression in...

Area of Science:

• Machine Learning in Public Health
• Artificial Intelligence in Medicine
• HIV/AIDS Research

Background:

• HIV viral suppression is critical for patient health and reducing transmission.
• Uganda faces significant public health challenges due to HIV/AIDS.
• Explainable AI (XAI) is underutilized, limiting transparency in HIV models.

Purpose of the Study:

• Develop and compare ML models to predict viral nonsuppression in Ugandan HIV patients on ART.
• Apply XAI techniques to identify key predictors and enhance model interpretability.
• Demonstrate population-level and individual-level insights for clinical utility.

Main Methods:

• Retrospective analysis of 1101 Ugandan patients on ART (June 2016-April 2018).
• Compared 8 ML algorithms, using XGBoost as the best-performing model.
• Employed XAI techniques (SHAP) for global and local prediction explanations.

Main Results:

• XGBoost model achieved high performance (AUC 0.80, accuracy 0.89).
• Adherence assessment was the strongest predictor of viral nonsuppression.
• Age group, urban residence, and ART duration were also significant predictors.

Conclusions:

• XGBoost model effectively predicts viral nonsuppression in Ugandan HIV patients.
• XAI identified adherence as the primary risk factor, guiding interventions.
• Transparent ML models can improve clinical decision-making in resource-limited settings.