66P Evaluation of an artificial intelligence algorithm for automated PD-L1 CPS scoring in triple-negative breast cancer across institutions, scanners, and antibody clones

In advanced triple-negative breast cancer (TNBC), assessment of the PD-L1 status is essential to identify patients that would benefit from immunotherapy. PD-L1 is routinely evaluated by combined positivity scoring (CPS) on immunohistochemistry (IHC)-stained tissue slides. CPS scoring, however, is time-consuming for pathologists and interobserver agreement rates are low. Reliable automated analysis of PD-L1 IHC scores by artificial intelligence (AI) could increase diagnostic efficiency and concordance, but is challenging due to the different histological growth patterns of TNBC and pre-malignant lesions and the variation in PD-L1 staining across laboratories with different slide processing setups. We developed and evaluated an AI software for CPS scoring on PD-L1 stained whole-slide images (WSIs) of TNBC. Performance of the AI algorithm was evaluated on n = 127 WSIs that were stained with two different PD-L1 antibody clones (22C3, SP263) and were obtained from five institutions, with four different slide scanner models. Each WSI was scored fully autonomously by the AI software as well as by three pathologists. Unweighted Cohen's kappa (κ) was used for measuring pairwise concordance of CPS scores, binarized at the clinically relevant cut-off of CPS ≥ 10. The concordance of the AI with the majority scoring category of pathologists (κ = 0.72) was higher than average pathologist interobserver concordance (κ = 0.69). In 47.8% of cases where the AI disagreed with the pathologist majority category, there was also disagreement among pathologists. In 92.9% of cases, the AI agreed with at least one of the three pathologists. The evaluation against pathologist scores on a diverse cohort of samples showed that the investigated AI system achieves human-level concordance in PD-L1 CPS scoring for TNBC. This is, to our best knowledge, the first time an AI software demonstrates non-inferiority compared to human readers in this task. With the prospect of being used as decision support tools in diagnostic routine, this study shows the potential of AI systems to accurately identify patients that would benefit from immunotherapy in TNBC and beyond.