Improving Tuberculosis Prognosis with Benchmarked Machine Learning Models

Abstract

Tuberculosis remains a considerable cause of morbidity and mortality in several poor and middle-income countries. When a patient is diagnosed with tuberculosis, healthcare providers must select most appropriate treatment tailored to patient's unique situation and expected trajectory of disease, guided by clinical competence. goal is to predict chance of dying from tuberculosis, which will help doctors figure out how disease will progress and make decisions about treatment. re were 36,228 records and 130 fields in first data collection, but many of records were missing, incomplete, or wrong. After cleaning and preparing data, a new dataset was created with 24,000 entries and 37 fields. This dataset includes 22,875 reported cured tuberculosis patients and 1 140 tuberculosis-related deaths. Two controlled experiments were designed to examine impact of data imbalance on model performance, employing (1) unbalanced and (2) balanced datasets.