Artificial intelligence for microsatellite instability detection in precision oncology: a systematic review from 2015 to 2025

Abstract Microsatellite instability (MSI) and mismatch repair deficiency (dMMR) are clinically actionable biomarkers that guide immunotherapy selection and prognostic stratification across multiple solid tumors. Polymerase chain reaction (PCR), immunohistochemistry (IHC), and next-generation sequencing (NGS) remain reference standards for MSI/dMMR assessment. We conducted a systematic review of artificial intelligence (AI)-based approaches for MSI/dMMR detection, searching PubMed, IEEE Xplore, Springer, and ScienceDirect for studies first publicly available between January 2015 and December 2025. Study selection followed PRISMA guidelines, and risk of bias was assessed using QUADAS-2. Of 1589 records, 111 studies were included after full-text review, hierarchically sorted by validation strategy (prioritizing independent external validation), cohort size, and performance metrics. AI-based prediction has been evaluated across histopathology whole-slide images (WSIs), radiology, omics, endoscopy, clinical variables, and multimodal combinations. Deep learning applied to H&E WSIs is the most extensively studied approach, although external validation remains inconsistent. Multimodal models are promising, but superiority over single-modality models remains unproven. Most studies predict assay-defined MSI/dMMR rather than treatment outcomes; the available evidence does not yet support routine clinical use of AI for triage or assay replacement. Prospective multicenter trials should evaluate AI-derived phenotypes as complementary or standalone biomarkers for treatment selection.