Performance Comparison of Deep Learning Models in Predicting HIV Incidences in Tanzania

Abstract

Human immunodeficiency virus infection/acquired immune deficiency syndrome (HIV/AIDS) is a global pandemic that has claimed more than 40 million lives since it was discovered in the late 1970s. In sub-Saharan Africa, including Tanzania. Different measures to combat the diseases have failed to be attained, like the UNAIDS  90-90-90 target, which aimed to reduce HIV by 2020, and it was moved to 2030. The availability of proper tools to control and monitor diseases and ensure proper early intervention is very important. Prediction of disease trends using Machine Learning (ML) models can improve speed towards attaining the UNAIDS targets by providing accurate insights into the disease trends. The performance of ML models depends on many factors, including datasets that influence the generalization of models. This study aims to suggest the best deep-learning model to predict HIV incidences in Tanzania. Four deep learning models, recurrent neural network (RNN), Gated Recurrent unit (GRU), Long Short-Term Memory (LSTM), and 2D convolution layer (CONV2D), have been studied. HIV data is collected from District Health Information System 2 (DHIS2), the national Health Management Information System (HMIS). The HIV data collected is for 26 regions in Mainland Tanzania, recorded from January 2015 to October 2022. The accuracy of the models was evaluated using three metrics: Mean absolute error (MAE), Mean absolute Percentage Error (MAPE), and mean square error (MSE). The results show that Conv2D achieved the lowest average training time for short-term predictions, while RNN records the highest accuracy with the lowest MAE for all considered cases. The GRU was the fastest for the long-term predictions, and the LSTM reported the best accuracy.