PERFORMANCE COMPARISON OF CLASSICAL ALGORITHMS AND DEEP NEURAL NETWORKS FOR TUBERCULOSIS PREDICTION

Abstract

This study compares the performance of several classical machine learning algorithms and deep neural networks for the prediction of tuberculosis in the Democratic Republic of Congo (DRC), using a sample of 1000 cases including clinical and demographic data. The sample is divided into two sets: 80% for training and 20% for testing. The algorithms evaluated include Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, Random Forest and Convolutional Neural Networks (CNN). The results show that the CNN has the best overall performance with an accuracy of 94%, an AUC of the ROC curve of 93%, an accuracy of 90%, an accuracy of 95%, a sensitivity of 88%, an F1-score of 91.3% and a Log Loss of 0.0386. The Random Forest follows closely behind with an accuracy of 92% and an AUC of 86%. The SVM and KNN models also performed strongly, but slightly less well. The Decision Tree obtained acceptable results, but inferior to the other algorithms evaluated. These results indicate that deep neural networks, and in particular the CNN, are superior for predicting tuberculosis compared with conventional machine learning algorithms. This superiority is particularly marked in terms of accuracy, sensitivity and reliability of predictions, as shown by the performance metrics obtained.