Deep learning-based prognostication in idiopathic pulmonary fibrosis using chest radiographs

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Idiopathic pulmonary fibrosis (IPF) is a progressive lung disease with limited treatment options.
• Accurate prognostication is crucial for managing IPF patients and guiding treatment decisions.
• Current prognostic models often rely on clinical and physiological parameters, with limited predictive power.

Purpose of the Study:

• To develop and validate a deep learning-based prognostic model (DLPM) for IPF using chest radiographs.
• To evaluate the DLPM's performance in predicting 3-year survival.
• To compare the DLPM's prognostic accuracy against forced vital capacity (FVC) and assess its independence as a prognostic factor.

Main Methods:

• A deep learning model was trained and validated on chest radiographs from IPF patients (n=1204 training, n=117 validation).
• External validation was performed on three independent cohorts (n=152, 141, 207).
• Model performance was assessed using time-dependent AUC (TD-AUC) for 3-year survival, compared with FVC, and evaluated for independence using Cox regression. A modified GAP index (GAP-CR) was created.

Main Results:

• The DLPM demonstrated similar-to-higher performance in predicting 3-year survival compared to FVC across three external test cohorts (TD-AUCs ranging from 0.76 to 0.83 for DLPM vs. 0.68 to 0.76 for FVC).
• The DLPM was identified as an independent prognostic factor from FVC in all external cohorts (p < 0.001).
• The modified GAP-CR index showed improved prognostic performance over the original GAP index in two of three external cohorts.

Conclusions:

• A deep learning model utilizing chest radiographs can effectively predict survival in IPF patients.
• This DLPM offers prognostic performance comparable to or exceeding that of FVC and is independent of it.
• The DLPM holds potential for enhancing IPF prognostication and clinical decision-making.