Abstract
Abstract Gastrointestinal cancers are leading causes of cancer-related mortality worldwide, with early detection significantly improving survival outcomes. However, current diagnostic methods are either invasive or lack sufficient sensitivity for early-stage disease. This study aimed to develop a non-invasive diagnostic model for gastric and colorectal cancers (the two most prevalent subtypes of gastrointestinal cancers) using serum glycopeptide profiling and machine learning. A case-control study was conducted with 550 participants, including 275 pathologically confirmed gastrointestinal cancer patients and 275 age- and gender-matched healthy controls. Serum levels of 28 candidate glycopeptides were quantified using a high-throughput ELISA platform. Four machine learning models (Logistic Regression, Support Vector Machine, Random Forest, and XGBoost) were constructed and validated using stratified sampling. Subgroup analyses and clinical utility evaluation were performed. Fourteen glycopeptides were significantly differentially expressed between cancer patients and controls (adjusted P < 0.05). The XGBoost model exhibited the best diagnostic performance on the independent test set, with an AUC of 0.845 (95%CI: 0.825–0.865), accuracy of 79.5%, sensitivity of 77.3%, and specificity of 81.6%. Notably, the model achieved an AUC of 0.796 for early-stage (I + II) gastrointestinal cancers. Decision curve analysis confirmed superior clinical utility compared to treat-all and treat-none strategies. A nomogram was developed for clinical translation. The serum glycopeptide-based XGBoost model provides a promising non-invasive approach for the early detection of gastrointestinal cancers, with potential to improve clinical outcomes through timely intervention.