Evaluation and Comparison of Different Machine Learning Methods to Predict Outcome of Tuberculosis Treatment Course

Abstract

Tuberculosis treatment course completion is crucial to protect patients against prolonged infectiousness, relapse, lengthened and more expensive therapy due to multidrug resistance TB. Up to 50% of all patients do not complete treatment course. To solve this problem, TB treatment with patient supervision and support as an element of the “global plan to stop TB” was considered by the World Health Organization. The plan may require a model to predict the outcome of DOTS therapy; then, this tool may be used to determine how intensive the level of providing services and supports should be. This work applied and compared machine learning techniques initially to predict the outcome of TB therapy. After feature analysis, models by six algorithms including decision tree (DT), artificial neural network (ANN), logistic regression (LR), radial basis function (RBF), Bayesian networks (BN), and support vector machine (SVM) developed and validated. Data of training (N = 4515) and testing (N = 1935) sets were applied and models evaluated by prediction accuracy, F-measure and recall. Seventeen significantly correlated features were identified ( P <= 0.004; 95% CI = 0.001 - 0.007); DT (C 4.5) was found to be the best algorithm with %74.21 prediction accuracy in comparing with ANN, BN, LR, RBF, and SVM with 62.06%, 57.88%, 57.31%, 53.74%, and 51.36% respectively. Data and distribution may create the opportunity for DT out performance. The predicted class for each TB case might be useful for improving the quality of care through making patients’ supervision and support more case—sensitive in order to enhance the quality of DOTS therapy.