Identifying digital pathology imaging based biomarkers of checkpoint inhibitor response in non-small cell lung cancer.

8530 Background: Immune checkpoint inhibitors (ICIs) targeting Programmed Death 1 (PD1) or Programmed Death Ligand 1 (PDL1) have revolutionized non-small cell lung cancer (NSCLC) treatment, yet only 30% of patients respond effectively. Although approved for patient selection, using PDL1 expression as a biomarker of treatment response is controversial in predicting clinical outcomes, highlighting the need for more robust alternatives. Methods: We explore the potential of AI-based approaches using H 26 Partial/Complete response, 78 stable/Progressive disease) and the publicly available TCGA-NSCLC cohort (N=986) were-utilized for multi-task training. The remaining 40% of the SUNY cohort (N=62; 16 Partial/Complete response, 46 stable/Progressive disease) was used for testing. Training was done in 5-fold cross-validation. The final model was chosen as an ensemble of models trained over all 5 folds. Train and test splits were generated using a stratified random sampling approach, ensuring that the proportion of responders versus non-responders remained unchanged across splits. Performance was evaluated utilizing percentage accuracy (all true positive values), precision (positive predictive value), recall (sensitivity) and F1 scores (mean value of precision and recall). Attention maps generated by the AI model were used to highlight relevant spatial regions of the TME significantly associated with ICI responses. Results: Our multi-task attention-based approach achieves an overall predictive accuracy of 79%, precision of 0.53, recall of 0.24, and an F1 score of 0.33 on the held-out test set. In contrast, pathologist derived PDL1 IHC scores ≥50% exhibit an accuracy of 47%, precision of 0.09, recall of 0.125, and an F1 score of 0.11. Conclusions: Initial findings indicate that attention-based multitask learning of NSCLC H&E- images could uncover crucial tumor intrinsic and microenvironmental features that are predictive of ICI responses in patients. Inclusion of additional clinical and molecular data for training and validation holds potential to further improve ICI response prediction accuracy using AI-derived biomarkers and classifiers.