Case detection of tuberculosis patients using machine learning : a case study of China Friendship Hospital Naguru, Kampala

Abstract

Although early detection and treatment of Tuberculosis cases are the hallmark of successful TB control, diagnostic delay is still long and common in Uganda. Therefore, the study aimed at developing a tool that uses machine-learning techniques to detect and diagnose TB more accurately in a shorter time. Methods: This was a retrospective study that used secondary data collected between 2011 to 2018 collected from China Friendship Hospital Naguru under Infectious diseases research collaboration an (IDRC) project called MIND (Mulago Inpatient Non-invasive Diagnosis for pneumonia). The study was conducted on data from 2296 tested patients of whom 1345 had no TB and 951 had TB .The prepared data was split into training and testing data to run the models. Different measures like accuracy, recall, precision and F-measure were used to evaluate the performance of the models. Results: The WEKA tool was used to calculate the accuracy based on incorrect and correct classes produced by the confusion matrix. The results that were obtained indicate that Jrip performs well among the Rules-based detection models and J48 among the Decision tree-based detection models with accuracy of 81.4% and 80.9% from both detection models respectively. Accuracy signifies the amount of the total number of TB patient predictions that are correct. With the Jrip model, results showed 871(91.6%) true positives that is the people who actually had TB while J48 showed 828 (87.1%) true positives. Results were obtained from the ROC curve and the performance was excellent since it was 0.8481 for Jrip and 0.8393 for J48. An application prototype was developed for physicians to access and use the final machine learned diagnostic models. Conclusion/Recommendation: WEKA-based experimental Tuberculosis detection results showed promisingly high TB detection accuracies. Health informatics experts to venture into studies that are medically related to using different algorithms like J48 and Jrip in order to increase the overall performance of their healthcare delivery system.