A95-05 Deep Learning-derived Fibrotic and Inflammatory Phenotypes Predict Lung Function Across Idiopathic and Connective-tissue-related Interstitial Lung Disease

Abstract Rationale Quantitative CT (QCT) can objectively characterise the relative burden of fibrosis and inflammation in interstitial lung disease (ILD). Structural imaging phenotypes may reflect physiological impairment more directly than diagnostic labels such as idiopathic pulmonary fibrosis (IPF) or connective-tissue-disease-associated ILD (CTD-ILD). We applied a deep-learning pipeline to classify HRCTs by the relative proportions of fibrosis and inflammation and examined associations between imaging phenotype and forced vital capacity (ppFVC %) across a combined IPF and CTD-ILD cohort. Methods HRCT scans from 384 patients (350 IPF, 34 CTD-ILD) were analysed using Qureight’s automated QCT models. Fibrosis and ground glass opacities (GGOs) were quantified as percentages of total lung volume (Fibr8™ and Glass8™). Each scan was assigned to a fibrotic-predominant, inflammatory-predominant, or mixed phenotype using a ± 25 % relative difference threshold between fibrosis % and GGO %. Each HRCT was paired with the nearest ppFVC measurement; 383 unique patients comprised the evaluable cohort. Group differences in ppFVC were assessed by one-way ANOVA and multivariable linear regression including diagnostic category (CTD vs IPF). A continuous fibrosis-to-inflammation ratio was also tested for graded structure-function coupling. Results Fibrotic-predominant pattern was most frequent (78 %, n = 299), followed by inflammatory-predominant (14 %, n = 53) and mixed (8 %, n = 31). Mean ± SD ppFVC differed across phenotypes (p=0.028): fibrotic 76.9±18.8 %, inflammatory 82.1±19.3 %, and mixed 86.4±20.3 %. In multivariable regression, compared with fibrotic-predominant cases, ppFVC was +5.1 % for inflammatory-predominant (95 % CI -0.5 to + 10.7, p=0.074) and +9.4 % for mixed (95 % CI + 2.2 to + 16.5, p=0.010); CTD-ILD diagnosis was not independently associated with ppFVC (p=0.90). The continuous fibrotic-to-inflammatory ratio correlated inversely with ppFVC (r=-0.24, p0.001); each unit increase in the ratio corresponded to a 0.35 % decrease in ppFVC (β=-0.35 ± 0.07, p0.001, R²=0.056). Conclusions Across IPF and CTD-ILD, QCT-derived imaging phenotype was significantly associated with lung function, independent of diagnostic label. Increasing fibrotic predominance on HRCT corresponded to progressive physiological impairment, whereas mixed and inflammatory patterns were associated with relatively preserved ppFVC. These findings support deep learning-based imaging phenotypes as scalable, biologically meaningful biomarkers of structure-function state and suggest their potential utility for clinical-trial enrichment, enabling preferential inclusion of patients with fibrotic-predominant or inflammatory-predominant disease depending on therapeutic mechanism This abstract is funded by: None