Machine-Learning-Based Prediction of Clinical Outcomes in Gliomas Using Glycomic Features

Background: Gliomas are primary malignant brain tumors. Among them, grade IV astrocytomas represent the most aggressive form with limited treatment options and a poor prognosis. There is a critical need for reliable prognostic biomarkers to predict patient outcomes. Methods: In this study, we evaluated the prognostic roles of glycomics in predicting outcomes of patients with glioma. To address this need, we collected a cohort of 86 patients and conducted N-linked glycomics analysis using matrix-assisted laser desorption/ionization mass spectrometry on a tissue microarray comprising 78 gliomas and 8 controls. We performed survival analyses and machine-learning-based predictive modeling to evaluate the prognostic roles of glycomic features. Results and Conclusion: Glycomic features alone displayed strong performance in classifying mortality status, achieving a mean area under the receiver operating characteristic curve (AUROC) value of 0.880 and a mean area under the precision-recall curve (AUPRC) value of 0.925, and demonstrating consistent prognostic value in survival analyses with a mean concordance index of 0.759 across repeated cross-validation. Glycomic features also showed strong performance in predicting the incidence of seizure status, with a mean AUROC value of 0.778 and a mean AUPRC value of 0.766. Additionally, top-ranked glycans from machine-learning models can effectively stratify patients based on their survival outcomes, demonstrating their potential role as prognostic biomarkers.