Background – Microsatellite instability (MSI) occurs when the DNA mismatch repair system fails, usually leading to a high tumor mutation burden. Patients with MSI-high (MSI-H) tumors can be potentially sensitive to immunotherapy. Although MSI has been observed in many cancers, routine testing is largely limited to GI and endometrial cancers (EC) due to cost concerns. Aiming to establish a cost-effective screening tool that could ultimately be applied across cancer types, we evaluated the performance of two AI-based histogenomic models from Tempus AI to predict MSI-H status from digitized whole-slide pathology images.

Design – MSI status of 6,226 consecutive in-house NGS solid tumor tests conducted between January 2022 and August 2025 was reviewed. MSI scores above 19 and lower than 14 were classified as MSH-H and microsatellite-stable (MSS) respectively. MSI status in these cases was also confirmed by IHC. A total of 120 cases with MSI-H status were identified, including 68 lower GI, 11 upper GI, 26 gyn, 4 breast, and others. 16 MSI-H and 16 MSS cases diagnosed with EC or serous carcinoma were selected, excluding fine needle aspiration specimens and slides with low neoplastic cellularity. Two Tempus AI models (p-Endometrial and p-MSI) were evaluated for their ability to predict MSI-H status. Model performance was assessed using area under the curve (AUC), sensitivity, specificity, PPV and NPV across varying thresholds. The goal was to identify optimal thresholds to define a positive MSI-H result.

Results – An optimal balanced accuracy was achieved consistently for both models at a threshold between 0.32 to 0.35. AUCs for p-Endometrial and p-MSI model were 0.93 [95% CI: 0.84–1.00] and 0.93 [95% CI: 0.80–1.00] respectively. NPV for both models reached 0.93-0.98. These findings suggested that AI models can reliably exclude MSI-H cases, making them suitable for initial screening.

Conclusions – AI-based prediction of MSI-H status has demonstrates strong performance and potential use as a screening tool. Our current work focuses on testing MSI-H cases originating outside of the colorectal and gynecologic sites. Further assessment in larger cohorts and across additional cancer types is warranted before integrating such AI model into routine pathology workflows. Once validated, it could offer a cost-effective screening solution to evaluate patient’s MSI status across various malignancies. Combined with confirmation test, it provides a more comprehensive prediction of immunotherapy response.

VIEW THE PUBLICATION