Sharpley Additive Explanations for Viral Rebound Prediction in Patients on Antiretroviral Therapy

Abstract

Although Antiretroviral Therapy (ART) is an effective treatment for Human Immunodeficiency Virus (HIV) infection, pro-viral HIV-DNA remains in target cells of the body, preventing ART from curing Acquired Immunodeficiency Syndrome (AIDS) of the patient, a disease caused by HIV. In order to prevent HIV reactivation from persistent viral reservoirs, ART is necessary. When an HIV positive patient takes ART as directed by a medical professional and does not suppress after a second viral load test but had suppressed in the previous viral load test, this is known as HIV Viral Rebound(VR). Viral rebound increases the likelihood of treatment failure, resistance to ART, vulnerability to other infections among other challenges. In this research, we developed CatBoost, XgBoost and Light GBM Machine Learning (ML) models that predicted HIV positive patients with viral rebound using healthcare facilities, laboratory data, clinical and demographic information. This paper focused on one Explainable Artificial Intelligence (XAI) technique, SHAP (SHarpley Additive exPlanations) values that is used to explain a model’s output. Interpretability and transparency in ML algorithms are improved by this technique, which also helps to better understand which features have the greatest influence on predictions and helps to comprehend the decision-making process of models. After data pre-processing, the dataset was split up in portions of 70%, 15% and 15% to train, validate and test the ML models respectively. We confirmed that Light GBM performed best on test data with f1-score mean of 0.92. We also highlighted viral rebound feature interaction for the patients and makes it possible for medical professionals to understand prediction results accurately based on SHAP, foster proper user trust, and offer guidance on how to enhance a model.