Abstract 6354: ATTUNE: optimizing the treatment journey for lung cancer patients by predicting EGFR mutation from routinely stained histology images

Abstract Background: Identifying targetable mutations through tissue-based molecular testing is part of the standard of care management of patients with NSCLC. Nevertheless, globally up to 50% of lung cancer patients are currently not being tested. Liquid biopsy can provide valuable material for molecular diagnosis; however, low sensitivity, especially in early-disease stages, may prevent its broader applicability in the clinic. Predicting EGFR status from digitalized hematoxylin and eosin (H&E) images using machine learning (ML) could be key to addressing such unmet needs. Methods: We propose an automated deep learning algorithm that retrospectively analyzes H&E images for EGFR status in a pivotal clinical trial (ADAURA), in early stage resectable EGFRm NSCLC and in a publicly available data from TCGA; totaling more than 1, 000 patients. The proposed algorithm includes automated tumor segmentation (trained using all TCGA data) and prediction of EGFR mutation from tumor regions with weakly supervised learning (trained using 80% of ADAURA). The segmentation was validated on ADAURA and the proposed EGFR mutation prediction task was validated on held-out 20% of ADAURA and all of TCGA data. Results: We outperformed existing published solutions by a margin of 10% in terms of the AUC metric. The model achieved the AUROC of 0.81 ± 0.04 and 0.75 ± 0.01 on over 100 phase III clinical trial patients and over 500 patients in TCGA, respectively. The model can be calibrated to achieve either > 90% sensitivity or specificity, depending on the clinical use case. Conclusion: The proposed approach does not require any regions marked by pathologists to predict EGFR status and our results reinforce using H&E digital pathology slides with ML to predict molecular status, thereby optimizing the treatment journey for millions of lung cancer patients. Citation Format: Talha Qaiser, Harish RaviPrakash, José Domingo Salazar, Marco Rosati, Alma Andoni, Guillem Portella, Adriana Hepner. ATTUNE: optimizing the treatment journey for lung cancer patients by predicting EGFR mutation from routinely stained histology images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6354.