A deep learning-derived risk score model from digital pathology to forecast nivolumab outcomes in gastric carcinoma.

BACKGROUND: The PD-L1 combined positive score (CPS) is a biomarker predicting responses in gastric cancer (GC) immunotherapy. OBJECTIVES: We aimed to develop a deep learning-based model to predict responses to nivolumab in GC using PD-L1 28-8 immunohistochemistry. METHODS: A cell-detection network was trained on 1927 patches from 88 whole-slide images to generate a computational positive cell ratio (cPCR). The predictive performance of cPCR was evaluated in an independent cohort of 147 patients treated with nivolumab plus chemotherapy. RESULTS: The overall objective response rate (ORR) was 49.66%. ORR was 56.12% in patients with PD-L1 CPS ≥ 5 and 36.73% in those with CPS < 5. Using cPCR, ORR was 56.03% for cPCR ≥5 and 25.81% for cPCR <5. The AUCs of CPS (0.586) and cPCR (0.601) did not differ significantly, but net reclassification analysis showed superior predictive performance for cPCR (p = 0.0269). Incorporating cPCR with tumour-stroma ratio through stepwise variable selection enabled construction of a risk-score model that improved prediction of immunotherapy benefit. CONCLUSIONS: In conclusion, we developed a cPCR-based model and risk-score system that more accurately forecasts nivolumab response in GC, offering a promising tool for optimising immunotherapy selection.