Late Breaking Abstract - A deep learning algorithm for predicting progression in fibrosing interstitial lung disease

<bold>Introduction:</bold> Deep learning algorithms can differentiate between stable and progressive pulmonary fibrosis. SOFIA (Systematic Objective Fibrotic Imaging Analysis Algorithm) has demonstrated superior prognostic performance compared to manual scoring in idiopathic pulmonary fibrosis (IPF). <bold>Aims and objectives:</bold> We explored the utility of SOFIA for predicting progression in patients with any fibrosing interstitial lung disease (fILD). <bold>Methods:</bold> High resolution CT images (HRCTs) of fILD patients were classified by SOFIA into definite usual interstitial pneumonia (UIP), probable UIP, indeterminate, or other. A cox proportional hazards model was employed to determine whether SOFIA predicts mortality and lung function decline. <bold>Results:</bold> HRCTs from 604 patients with fILD (44% female; age=66±12y; 35% on antifibrotics) were included. SOFIA categorised IPF most often as probable UIP (59%), connective tissue disease as indeterminate (41%), and hypersensitivity pneumonitis as other (37%). SOFIA diagnostic category, probable UIP, and indeterminate UIP were predictive of mortality in fILD (Diagnosis: HR 1.45, 95%CI 1.25-1.69; Indeterminate: HR 0.14, 95% CI 0.07-0.26; Probable: HR 2.84, 95%CI 1.76-4.59, all p<0.001). When adjusted for age, sex, and FVC, indeterminate score remained predictive of lower mortality (HR 0.29, 95%CI 0.15-0.59, p<0.001). In patients with follow up lung function (N=366), indeterminate score was most predictive of slower FVC decline, though not statistically significant (HR=0.535, 95%CI 0.265-1.080, p=0.081). <bold>Conclusions:</bold> Deep learning tools can predict disease progression from a baseline CT in fibrosing ILD patients. Further work is needed to integrate this tool into clinical practice.