Abstract Rationale The 2025 international guidelines formally integrate diagnostic confidence into interstitial lung disease (HRCT) pattern evaluation. We developed and validated a deep learning (DL) tool, Lung-Graph-Transformer, to objectively quantify this confidence and assess its prognostic significance for mortality and progressive pulmonary fibrosis (PPF). Methods In a multicenter prospective cohort of 2,801 Chinese patients (NCT04370158), we applied the Lung-Graph-Transformer to categorize diagnostic confidence as low (50%) or high (≥50%) for specific patterns: usual interstitial pneumonia (UIP, n = 1,348), nonspecific interstitial pneumonia (NSIP, n = 971), and bronchiocentric interstitial pneumonia (BIP, n = 482). The model was developed on 70% of the cohort, with interval validation (15%) and external testing (15%). Multivariable Cox regression (adjusted for center, demographics, pulmonary function, comorbidities, and antifibrotic treatment) evaluated associations between pattern confidence and outcomes. Results Among 2,801 patients (median age 63, 33.7% female), 689 deaths and 1,440 PPF events occurred over 5 years of follow-up. The DL-based confidence tool demonstrated exceptional diagnostic performance, achieving an overall accuracy of 89.2% and a Kappa coefficient of 0.81. AUCs exceeded 0.85 across all validation sets, with particularly high discriminatory power for UIP (sensitivity 92.1%, specificity 94.3%). Prognostically, high-confidence UIP (≥50%) was independently associated with significantly increased risks of all-cause mortality (adjusted HR: 1.71, 95% CI: 1.44-2.03) and PPF, defined by relative decline in FVC% (adjusted HR: 1.74, 95% CI: 1.45-2.09), DLCO% (adjusted HR: 3.20, 95% CI: 2.68-3.81), or categorical fibrosis progression (adjusted HR: 1.24, 95% CI: 1.06-1.47). Conclusion We present a novel, rigorously validated DL tool that objectively quantifies diagnostic confidence in HRCT pattern interpretation. Our findings establish this metric as a critical, independent prognostic biomarker. Integrating this tool into multidisciplinary discussion directly enables enhanced risk stratification and personalized management, aligning with and advancing the 2025 guideline recommendations. This abstract is funded by: National Key Technologies Research and Development Program Precision Medicine Research (2021YFC2500700 and 2016YFC0901101), the National Natural Science Foundation of China (82370072), and the National High Level Hos pital Clinical Research Funding (2022-NHLHCRF-LX-01-0104)
Wang et al. (Fri,) studied this question.