Machine learning model to predict the adherence of tuberculosis patients experiencing increased levels of liver enzymes in Indonesia

Abstract

Indonesia is still the second-highest tuberculosis burden country in the world. The antituberculosis adverse drug reaction and adherence may influence the success of treatment. The objective of this study is to define the model for predicting the adherence in tuberculosis patients, based on the increased level of liver enzymes. The longitudinal study using adult tuberculosis patients treated with the first line of antituberculosis was conducted prospectively. The pregnant women and patients with complications such as gout, diabetes mellitus, liver disorder and HIV were excluded. We measured the total bilirubin, aspartate aminotransferase (AST), and alanine aminotransferase (ALT) and adherence over the 2nd, 4th, and 6th months of the treatment. We used the ORANGE Data mining as the machine learning to predict the adherence. We recruited 201 patients, whereas the male participants and less than 61 years old as the dominant participants. Around 33%, 35% and 35% tuberculosis patients experienced the increase level of bilirubine, ALT and AST, respectively. There were significant differences in ALT and AST between good and poor adherence groups, especially in the female patients. The Neural Network and Random Forests were the most suitable models to predict tuberculosis patients’ adherence with good Area Under The Curve (AUC).