A deep learning algorithm for predicting disease progression in idiopathic pulmonary fibrosis

Abstract

Aim: We investigated the prognostic utility a deep learning algorithm for predicting risk of progression in patients with idiopathic pulmonary fibrosis (IPF). Progression was defined as an FVC decline of 10% at 12 months, death, or transplantation. Methods: A deep learning algorithm (DL_IPF) was trained on HRCTs from The Open-Source Imaging Consortium (OSIC) and tested in Australian IPF Registry (AIPFR). A visual-based total fibrosis score was obtained for AIPF HRCTs. SOFIA UIP probability scores were obtained using a previously reported deep learning algorithm, trained in the identification of UIP features. The prognostic utility of DL_IPF (yielding a progression probability) was evaluated against conventional measures of disease severity and SOFIA-based UIP probability scores. Results: DL_IPF analysis independently predicted mortality, controlling for visual-based total fibrosis extent (n=501, HR 1.03, p<0.0001). Progression probability scores were converted to PG_PIOPED scores using PIOPED diagnostic probability thresholds. PG_PIOPED (HR 2.74, p<0.0001) and SOFIA PIOPED scores (HR 1.35, p<0.0001) independently predicted mortality. PG_PIOPED scores predicted mortality in patients with an “indeterminate” HRCT pattern (n=82, HR 8.06, p<0.0001) and patients who underwent surgical lung biopsy (SLB) (n=82, HR 3.00, p<0.0001). An increase in PFF_PIOPED score by one category, was associated with a 3.2-fold increased likelihood of developing progressive disease (OR 3.21 p<0.0001) when controlling for total fibrosis extent. Conclusion: Deep learning may be used to identify suspected IPF patients at risk of progression at 12 months