Immune checkpoint inhibitors (ICIs) help only about 25–30% of patients with advanced non–small cell lung cancer (NSCLC) lacking EGFR or ALK alterations. Traditional biomarkers such as PD-L1 expression and tumor mutational burden (TMB) are imperfect, necessitating better predictors. Recent advances in artificial intelligence (AI) now allow analysis of digital pathology images to identify immune features linked to treatment response.
A multicenter study developed Deep-IO, a deep learning model trained on hematoxylin-eosin–stained slides from 614 U.S. patients and validated on 344 patients receiving ICI monotherapy. By analyzing over 295,000 image tiles, Deep-IO predicted objective response rates of 26–28% and demonstrated accuracy comparable to or better than PD-L1, TMB, or tumor-infiltrating lymphocytes (TILs). Patients with higher Deep-IO scores had significantly longer progression-free and overall survival, particularly in adenocarcinoma.
Deep-IO also complements existing biomarkers: combining it with PD-L1 improved prediction, with the highest response rates in patients scoring high on both. Interpretability analyses highlighted tumor epithelial and immune regions, showing the model captures meaningful biological features. This approach offers a promising tool to guide immunotherapy decisions in advanced NSCLC.