Deep Learning-Based Prediction Model Using Radiography in Nontuberculous Mycobacterial Pulmonary Disease

Background: Prognostic prediction of nontuberculous mycobacteria pulmonary disease using a deep learning technique has not been attempted.

Research question: Can a deep learning (DL) model using chest radiography predict the prognosis of nontuberculous mycobacteria pulmonary disease?

Study design and methods: Patients who received a diagnosis of nontuberculous mycobacteria pulmonary disease at Seoul National University Hospital (training and validation dataset) between January 2000 and December 2015 and at Seoul Metropolitan Government-Boramae Medical Center (test dataset) between January 2006 and December 2015 were included. We trained DL models to predict the 3-, 5-, and 10-year overall mortality using baseline chest radiographs at diagnosis. We tested the predictability for the corresponding mortality using only DL-driven radiographic scores and using both radiographic scores and clinical information (age, sex, BMI, and mycobacterial species).

Results: The datasets comprised 1,638 (training and validation set) and 566 (test set) chest radiographs from 1,034 and 200 patients, respectively. The Dl-driven radiographic score provided areas under the receiver operating characteristic curve (AUC) of 0.844, 0.781, and 0.792 for 10-, 5-, and 3-year mortality, respectively. The logistic regression model using both the radiographic score and clinical information provided AUCs of 0.922, 0.942, and 0.865 for the 10-, 5, and 3-year mortality, respectively.

Interpretation: The DL model we developed could predict the mid-term to-long-term mortality of patients with nontuberculous mycobacteria pulmonary disease using a baseline radiograph at diagnosis, and the predictability increased with clinical information.