Hybrid Statistical Modeling and Machine Learning Ensemblers for Interpretable HIV Viral Load Prediction Among Clients on Antiretroviral Therapy

Abstract Elimination of HIV transmission requires a multi pronged approach and com-mitment from all stakeholders which includes machine learning engineers. The burden of HIV is highest in sub-Saharan Africa where many countries are in the fight against the disease with a high number of recorded people living with HIV, case in point Uganda with about 1.4 million clients. With the recent advent of Artificial Intelligence, the capabilities in diagnostics have grown in leaps and bounds by leveraging the Machine Learning and Deep Learning. In this study, we present an ensemble learner model that predicts which HIV client’s next viral load using HIV viral load laboratory test data. We ably trained and validated our learners on 70% of the data, and tested them on the other 30% for a 5.7 million HIV viral load dataset of ART clients for the period 2019 to 2022 from the Central Public Health Laboratory of Uganda. Our findings showed that the Staking technique outperformed the Bagging, random forest, decision tree, XG Boost, Adaboost, KNN, linear regressors with an R2 score of 0.85, MAE of 0.42, and MSE of 0.41 improving the chances of predicting next visit viral load of Antiretroviral Therapy (ART) clients. For model interpretability, we explored the LIME (Local Interpretable Model-Agnostic Explanations) to provide an insight into the variables of importance during model training. Our study demonstrates the potential of leveraging HIV viral load test data, statistical modeling and machine learning techniques to predict Antiretroviral Therapy clients next visit viral load.