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

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pulmonary disease research

Background:

• Prognostic prediction for nontuberculous mycobacteria pulmonary disease (NTMPD) using deep learning (DL) has not been previously explored.
• NTMPD poses a significant challenge in clinical management and requires accurate prognostic tools.

Purpose of the Study:

• To investigate the capability of a DL model to predict the prognosis of NTMPD using chest radiography.
• To assess the DL model's performance in predicting 3-, 5-, and 10-year overall mortality.

Main Methods:

• A DL model was trained on chest radiographs from NTMPD patients diagnosed between 2000-2015.
• The model predicted mortality using DL-driven radiographic scores alone and in combination with clinical data (age, sex, BMI, mycobacterial species).
• Model performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

• The DL-driven radiographic score achieved AUCs ranging from 0.781 to 0.844 for 3-, 5-, and 10-year mortality.
• Incorporating clinical information with the radiographic score improved AUCs to 0.865-0.942 for the respective mortality predictions.
• The model demonstrated robust predictive capabilities for mid-term to long-term mortality.

Conclusions:

• DL models utilizing baseline chest radiographs can effectively predict mid-term to long-term mortality in NTMPD patients.
• The addition of clinical information to the DL model significantly enhances prognostic prediction accuracy.
• This approach offers a promising tool for improving patient management and outcomes in NTMPD.