Deep learning of H&E slides adds prognostic value beyond IASLC grading in non-mucinous lung adenocarcinoma among never-smokers

Background: Lung cancer in never smokers (LCINS) most often presents as non-mucinous adenocarcinoma. The International Association for the Study of Lung Cancer (IASLC) system is the current standard for histologic grading, but its use is limited by interobserver variability, time-intensive manual assessment, and limited scalability. We evaluated whether a deep learning model applied to routine hematoxylin and eosin (H&E) slides could predict overall survival and improve prognostic stratification beyond conventional histologic grading. Methods: We analyzed 595 stage I-III whole-slide images from the Sherlock-Lung study, each from a unique patient. For the full cohort, median follow-up was 37 months (range 1-120 months); 55 deaths occurred by 5 years and 102 deaths by 10 years from diagnosis, out of 190 events overall. Data were split into training (n=409), internal cross-validation (n=45), and held-out validation (n=141) sets. A convolutional neural network generated continuous patient-level risk scores from H&E images, which were dichotomized into high-risk (n=34) and low-risk (n=107) groups in the validation cohort using the Youden index. Prognostic discrimination for overall survival in the validation set was assessed using time-dependent AUCs and Cox models. We compared AUCs for 5- and 10-year survival probabilities estimated from IASLC grade and deep-learning risks in univariate analyses and based on the following Cox models: (1) baseline (age, sex, ancestry, tumor stage) + IASLC grade; (2) baseline + deep learning; (3) baseline + IASLC grade + deep learning. Results: Deep learning yielded higher AUCs than IASLC grade for 5 year (0.84 [0.76-0.92] vs 0.70 [0.62-0.78]; p=0.01) and 10 years overall survival (0.75 [0.61-0.90] vs 0.64 [0.48-0.79]; p=0.59). In multivariable analyses, at 5 years the AUCs were 0.79 for baseline + IASLC, 0.86 for baseline + deep learning (p<0.01 vs baseline + IASLC), and 0.87 for the full model (better than simpler models, p=0.04); at 10 years the corresponding AUCs were 0.82, 0.86, and 0.87, with no statistically significant differences overall (p=0.63), possibly reflecting fewer late events. In a Cox model for 5 years of follow-up the deep-learning low-risk group had improved overall survival versus the high-risk group (HR 0.31, 95% CI 0.13-0.76). Conclusions: Deep learning on routine H&E slides provides prognostic information independent of IASLC grade in LCINS. Integrating deep learning with grade and clinical factors improves overall survival prediction, supporting AI-augmented pathology for precision risk stratification and personalized treatment/surveillance in non-mucinous lung adenocarcinoma among never-smokers.