Artificial Intelligence for Predicting Microsatellite Instability From Haematoxylin and Eosin-Stained Histopathology...

Microsatellite instability (MSI) and mismatch-repair (MMR) deficiency are pivotal predictive biomarkers in colorectal cancer (CRC), yet reference-standard testing is resource-intensive and unevenly available. Artificial intelligence (AI) applied to routine haematoxylin and eosin (H&E)-stained histopathology offers a scalable pre-screening alternative. This Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-guided systematic review (January 2018-December 2026) identified 82 unique studies evaluating AI models for MSI/MMR prediction from H&E images, of which 51 contributed extractable area under the receiver operating characteristic curve (AUC) data. Diagnostic accuracy was synthesised descriptively and compared by subgroup using the Mann-Whitney U test. The median AUC was 0.895 (interquartile range 0.791-0.954; range 0.649-0.990). Histopathology foundation models achieved a higher median AUC than task-specific architectures (0.910 vs 0.879; U = 383.5, p = 0.124), and externally validated models performed comparably to internally validated ones (0.895 vs 0.893; U = 224.5, p = 0.652). Performance did not differ significantly between peer-reviewed articles and preprints (0.894 vs 0.910; U = 340.0, p = 0.569). AI models predict MSI/MMR status from H&E-stained CRC histopathology with consistently high discriminative accuracy. Foundation models achieved a numerically higher median AUC than task-specific architectures, although this difference did not reach statistical significance and requires confirmation in adequately powered, head-to-head evaluations. Heterogeneity in reference standards, cohorts, and reporting currently constrains formal bivariate meta-analysis; standardised reporting and prospective external validation are required before clinical deployment.