Standalone prognostication of idiopathic pulmonary fibrosis using attention-based deep learning.

Abstract

The prognosis of idiopathic pulmonary fibrosis (IPF) is heterogeneous and challenging to predict. To develop and externally test ORACLE-IPF, a fully automated deep learning model predicting 5-year mortality risk from a single baseline chest computed tomography (CT) scan. ORACLE-IPF was trained and internally tested using data from 1023 patients with IPF at one tertiary center, with external testing performed in an independent cohort of 667 patients from another tertiary center. The model integrates an attention mechanism within a 3-dimensional convolutional neural network ensemble, explicitly trained to recognize fibrosis distribution, extent, and prognostic relevance. Prognostic performance for all-cause mortality and respiratory disease mortality was evaluated using time-dependent area under the receiver operating characteristic curve (AUC) and the C-index. Additional analyses included comparative performance against clinical and imaging markers and reproducibility testing. In the external test set, ORACLE-IPF demonstrated robust prognostic performance with 1- to 5-year AUCs ranging from 0.75 to 0.81 for all-cause mortality. It outperformed both deep learning-based CT segmentation (C-index: 0.73 vs 0.69, P = .003) and the Gender, Age, and Physiology (GAP) index (0.73 vs 0.69, P = .045). Combining ORACLE-IPF with CT segmentation results or GAP further increased C-indexes from 0.69 to 0.73 and from 0.69 to 0.75, respectively (both P <.001). ORACLE-IPF also predicted respiratory disease mortality well, with AUCs ranging from 0.81 to 0.87 across the 1- to 5-year horizons. The model exhibited excellent reproducibility (same-day test-retest, r = 0.93-1.00; longitudinal, r = 0.89). ORACLE-IPF reliably predicted IPF mortality, advancing conventional prognostic markers.