Deep learning identification of high-risk lung adenocarcinoma histologic subtypes: A decision-support tool for treatment planning.

Abstract

e20022 Background: Lung adenocarcinoma (LUAD) is the most common lung cancer subtype. Solid and micropapillary patterns represent high-grade histology with significantly worse outcomes: 5-year survival of approximately 60% and 40% respectively, compared to over 90% for lepidic-predominant tumors. Accurate identification is critical for adjuvant therapy decisions, as recent evidence suggests survival benefit from chemotherapy in patients with these high-risk components. However, subtyping suffers from interobserver variability (kappa 0.38-0.55), affecting treatment consistency. We developed an AI tool to identify high-risk subtypes for treatment stratification. Methods: We analyzed 143 resected LUAD whole slide images with pathologist-confirmed subtypes: acinar (n = 60, 42%), solid (n = 55, 38%), lepidic (n = 16, 11%), micropapillary (n = 10, 7%), and papillary (n = 5, 4%). High-risk was defined as solid or micropapillary predominant (n = 65, 45%) per WHO/IASLC guidelines. Tissue patches (224×224 pixels) at 20x magnification were processed using Virchow2 foundation model, selected based on superior five-class subtyping performance. ABMIL classifiers with gated attention mechanism were trained using 5-fold stratified cross-validation with class weighting and Youden index threshold optimization. Results: The model achieved AUC of 0.951±0.06 and balanced accuracy of 86.1%±4.2% for binary high-risk classification (Table 1). At the optimized operating threshold, sensitivity was 81.7%±13.7% with specificity of 90.5%±14.7%, yielding F1 score of 0.86. Positive predictive value of 90% indicates patients flagged high-risk are likely true positives, supporting chemotherapy consideration. Among AI-classified low-risk patients, 89% were confirmed true low-risk (NPV 88.6%±8.0%), identifying candidates for observation. These operating characteristics support use as a standardized second read to reduce variability in high-risk identification. Conclusions: This tool addresses a specific clinical dilemma: which resected LUAD patients warrant adjuvant therapy intensification. The 90% PPV means patients flagged high-risk can be confidently considered for chemotherapy; the 89% NPV identifies patients where observation may be appropriate - potentially sparing treatment toxicity without compromising outcomes. In tumor boards, AI-derived risk stratification provides objective data for treatment decisions, particularly in borderline cases where pathologist interpretation varies. External validation on multi-site cohorts (including biopsies) and outcome linkage (recurrence-free survival) are needed to confirm clinical utility. Binary high-risk detection performance. Metric Value AUC 0.951 ± 0.06 Balanced Accuracy 86.1% ± 4.2% Sensitivity 81.7% ± 13.7% Specificity 90.5% ± 14.7% PPV 90% NPV 88.6% ± 8.0% F1 Score 0.86