Abstract 2535: Uncovering bias in AI-based digital pathology analysis

Abstract As the most transformative technology in a generation, Artificial Intelligence (AI) is a tool that brings tremendous promise and significant risks, particularly in regard to healthcare. Our project aims to address two challenges faced by researchers utilizing AI in the analysis of immunohistochemical (IHC) images: the potential for covert bias and the opaque nature of “black box” systems. We have trained an AI-driven model to detect tumor and stromal regions in non-small cell lung cancer (NSCLC) tissues stained with a variety of IHC markers, including the histone methyltransferase EZH2, the histone mark H3K27me3, the antigen presentation components B2M, HLA-DR, DQ, DP, the immune evasion molecular PD-L1 and the enzyme CBS. By measuring stain intensity in each region, we observed that EZH2 and PD-L1 staining are significantly higher in squamous cell carcinoma tumor cells than in adenocarcinoma tumor cells, and that HLA-DR, DQ, DP has an inverse relationship with EZH2. These data implicate a reason why some squamous cell carcinomas fail anti-PD1 immunotherapy. However, to ensure that the system was not creating or perpetuating existing disparities, we next endeavored to assess the algorithm’s accuracy across multiple demographic bounds. Using the Cohen’s Kappa of Inter-Rater Reliability, we were able to determine that our algorithm has near perfect agreement with trained human observers in each of the six stains tested (EZH2, H3K27me3, B2M, HLA-DR, DQ, DP, PD-L1, and CBS). This was true regardless of race (Black vs Caucasian) or geographical origin (Appalachian Kentucky vs Non-Appalachian Kentucky). However, when comparing performance by gender, one category fell below the threshold for near perfect agreement (CBS in males). Next, we sought to explore what factors might drive the AI’s decision-making process by comparing its accuracy in various histologic contexts. We analyzed the size of the nuclei, the number of nuclei per mm2, the degree of immune infiltration, the width of the stroma, and the degree of randomness in the arrangement of the tumor and stroma. While the size of the nuclei was not found to have a significant impact on accuracy, the other variables did have the potential to “confuse” the AI, which could lead to less reliable results. By tackling these challenges head-on, we hope that our findings will help set a standard for the equitable and transparent application of artificial intelligence in healthcare. Work supported by Markey STRONG Scholars Program through the American Cancer Society IRG-22-152-34 (EMS), AIML Hub pilot (EMS), T32 CA165990 (DRP), R01 CA237643 (CFB), R01 HL170193 (CFB), P30 CA177558 (Markey Shared Resources) Citation Format: Erika M. Skaggs, Daniel R. Plaugher, Sally R. Ellingson, Christine F. Brainson. Uncovering bias in AI-based digital pathology analysis [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 2535.