Deep learning prediction of MSI/dMMR, EBV, HER2, PD-L1, and CLDN18.2 status directly from hematoxylin and eosin-stained slides in gastric and gastroesophageal junction adenocarcinoma: a systematic review.

BACKGROUND

Microsatellite instability (MSI)/mismatch-repair deficiency (dMMR), Epstein-Barr virus (EBV), human epidermal growth factor receptor 2 (HER2), programmed death-ligand 1 (PD-L1), and claudin 18 isoform 2 (CLDN18.2) are clinically relevant biomarkers in gastric and gastroesophageal junction (GEJ) adenocarcinoma, but molecular and immunohistochemical (IHC) testing is costly and not universally available. Deep learning (DL) applied to routine hematoxylin and eosin (H&E) slides may enable inexpensive pre-screening. We assessed the accuracy of DL models predicting these five biomarkers directly from H&E.

METHODS

We searched PubMed, Embase, Cochrane CENTRAL, and Web of Science Core Collection from inception through 8 August 2026 for studies applying DL to H&E whole-slide or tissue-microarray images to predict MSI/dMMR, EBV, HER2, PD-L1, or CLDN18.2 status in gastric/GEJ adenocarcinoma. Two reviewers independently selected studies and re-extracted performance estimates from full texts, tables, footnotes, and supplementary materials. Risk of bias was assessed with Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) and Prediction model Risk Of Bias ASsessment Tool (PROBAST) and certainty was assessed with Grading of Recommendations Assessment, Development and Evaluation (GRADE). Because the requirements for independent, variance-supported area under the receiver-operating-characteristic curve (AUROC) estimates or threshold-specific 2 × 2 data were not met, study-level estimates were displayed without statistical pooling.

RESULTS

Of 1,094 records identified, 701 unique records were screened and 28 studies were included: 18 addressed MSI/dMMR, 14 addressed EBV, 2 addressed HER2, 2 addressed PD-L1, and 1 addressed CLDN18.2; marker categories overlapped. External-independent AUROCs were 0.767 (95% CI, 0.726-0.830) for MSI/dMMR, 0.941 (0.920-0.970) and 0.859 (0.823-0.919) for EBV, and 0.856 (0.775-0.926) for CLDN18.2; broader held-out AUROCs were 0.760 and 0.880 (MSI/dMMR) and 0.980 (EBV). One report each provided eligible 2 × 2 data for EBV, MSI/dMMR, and CLDN18.2; none did for HER2 or PD-L1. No pooling or final-model calibration was supported. All 28 reports (42 analytic scopes) were high risk under PROBAST; QUADAS-2 risk was high for 40 scopes and unclear for 2. Certainty was very low for MSI/dMMR, EBV, HER2, and PD-L1 and low for CLDN18.2.

CONCLUSION

Individual held-out models sometimes showed promising discrimination, particularly for EBV and MSI/dMMR, but the evidence was too sparse, dependent, and heterogeneous to define a stable summary performance or a clinically deployable triage threshold. Certainty was very low for MSI/dMMR, EBV, HER2, and PD-L1 and low for CLDN18.2; clinical utility and calibrated threshold performance remain unestablished.

SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/, identifier CRD420261449992.