Machine learning nomograms for gastric cancer: addressing data limitations in SEER-based models.

The Surveillance, Epidemiology, and End Results (SEER) database is widely used to develop machine learning prognostic models for gastric cancer, yet data limitations restrict their clinical utility. We critically appraise the recent multicenter machine learning study by Guan et al., which reported superior prognostication compared with TNM staging, and identify eight methodological concerns: chemotherapy sensitivity of only 68% with systematic under capture of one-third of treated patients; complete loss of regimen-specific information that prevents distinction between curative perioperative therapy and palliative regimens; absence of performance status data despite it being the strongest predictor of outcome; missing molecular biomarkers (HER2, MSI-H, TMB, TCGA subtype) increasingly essential for precision oncology; unknown surgical technique quality metrics; and substantial cohort heterogeneity (surgery rates 26.6% vs 59.8%). These limitations collectively prevent individual treatment selection at the bedside. We conclude that SEER-based machine learning models should remain complementary to—rather than substitutes for—guideline-based decision-making, and future progress requires integration of regimen-specific treatment data, documented performance status, comorbidity indices, and molecular biomarkers.