Automated Drug-Resistant TB Screening: Importance of Demographic Features and Radiological Findings in Chest X-Ray

Tuberculosis (TB) is a global disease caused by the bacillus Mycobacterium tuberculosis. In recent years, great progress has been made in care and control of drug-sensitive TB, whereas drug-resistant TB continues to be a worldwide public health problem that takes a heavy toll on both patients and the health care system. Early detection of drug resistance during a patient's first visit is very important because it enables appropriate drug treatment and thus reduces the period of infectiousness. However, discrimination between drug-resistant TB (DR-TB) and drug-sensitive TB (DS-TB) using imaging and readily available demographic data is still an open problem. In this paper, we investigate the possibility of automatic discrimination between DR-TB and DS-TB with demographic data and radiological findings from chest X-rays (CXRs) using machine learning techniques as well as the importance of such features for classifier training. We use a dataset of 1311 DR-TB cases and 1311 DS-TB cases from 10 countries, collected from the NIAID TB Portals program (https://tbportals.niaid.nih.gov). We first perform a two-step preprocessing, which consists of feature quantitation and missing data imputation. Seven demographic features and 25 radiological features are selected from the dataset. Then, we train a random forest (RF) model to evaluate the ability to differentiate between DR-TB and DS-TB. An importance index calculated from the RF model is used to analyze the feature importance with respect to the discrimination task. The importance index from the RF model shows that the top ten important factors for discriminating between DR-TB and DS-TB are: number of daily contacts, BMI, patient type, education, medium density infiltrate, medium density stabilized fibrotic nodules, low ground glass density infiltrate, pleural effusion percentage of hemithorax involved, multiple nodules, small nodules. Ten-fold cross-validation using the RF model shows that automatic discrimination between DR-TB and DS-TB achieves an average accuracy of 75% and an average AUC value of 83%, when using the top ten features. Our study suggests that automatic discrimination between DR-TB and DS-TB with demographic and radiological features is possible.