H&E–based MSI/MMR testing with artificial intelligence in colorectal cancer: A blinded evaluation across diverse clinicopathological subgroups.

Abstract

3520 Background: Mismatch repair (MMR) deficiency and microsatellite instability (MSI) testing is integral to colorectal cancer management. Despite guideline recommendations, routine testing remains limited by cost, turnaround time, and pathology resources. PANProfiler Colorectal (PPC) is an artificial intelligence (AI) system that infers MSI/MMR status directly from H&E-stained whole slide images (WSIs), offering a rapid, resource-efficient alternative. This study reports blinded validation of PPC, specifically evaluating performance across clinically relevant subgroups. Methods: A total of 2,636 WSIs from St. James’s University Hospital (Leeds, UK) were retrospectively analyzed, with the dataset representing a broad range of clinicopathological characteristics including age ( < 50 to > 75 years), gender, tumor stage, differentiation, and histological subtypes. PPC returned proficient MMR ( pMMR ), deficient MMR ( dMMR ) or indeterminate (no definitive result), with results evaluated for agreement against immunohistochemistry. Results: PPC generated definitive results for 2,109 WSIs, achieving a test replacement rate (TRR) of 80%. Overall positive percent agreement (PPA) was 97.9% and negative percent agreement (NPA) 93.8% (Table). Performance was consistent across all age groups and genders. PPA and NPA were mostly robust in all cancer stages, with PPA exceeding 99% in Stage II disease, where MMR status primarily informs adjuvant therapy decisions. In Stage III/IV disease, PPA reduced to 82.6%, which may reflect greater histological heterogeneity in advanced tumors. In mucinous and poorly differentiated tumors, PPA remained high (98.7-100%), supporting reliable detection of dMMR samples, while NPA ranged from 75.4-80%, indicating more conservative classification of pMMR tumors that would require lab tests. Conclusions: PPC demonstrated a high overall performance, providing definitive results for the majority of cases. PPA and NPA were consistent in key clinical groups, including younger patients and early stage disease. In more challenging cases like mucinous and poorly differentiated tumors, PPA remained robust despite reductions in NPA, supporting safe guidance of subsequent testing. These findings suggest PPC could streamline MSI/MMR workflows, reducing reliance on conventional testing while maintaining diagnostic safety across diverse clinicopathological profiles. Performance across diverse subgroups. Overall Age <50 Age 50-65 Age 65-75 Age >75 Male Female Stage I Stage II Stage III/IV Well Diff. Moderately Diff. Poorly Diff. Adenocarcinoma Mucinous Definitive (n) 2109 155 159 698 703 1165 944 395 1015 694 245 1655 187 1879 210 PPA (%) 97.9 100 100 92.1 100 95.5 99.4 97.8 99.5 82.6 100 96 100 99.5 98.7 NPA (%) 93.8 91.7 98 91.3 94.1 95.5 91.5 95.1 91.8 95.5 92.5 95.6 80 95.3 75.4 TRR (%) 80 74.5 76.4 81.2 82.2 81.1 78.7 74.5 81 81.8 76.1 80.8 80.3 82.4 63.6