GLOBE: an explainable machine learning platform for preoperative prediction of thromboembolism and neurological deterioration in patients with glioma

BackgroundPatients with glioma are at high risk of postoperative venous thromboembolism (VTE) and postoperative neurological deterioration (PND). Conventional clinical scoring systems have limited accuracy in predicting these perioperative risks. This study aimed to develop and validate machine-learning models for individualized preoperative prediction of postoperative VTE and PND in patients with glioma.MethodsA retrospective cohort of 427 patients with glioma was included. Patients were randomly divided into training and test sets at an 8:2 ratio using stratified random sampling. Multiple machine-learning algorithms were trained and evaluated. Model performance was assessed using the area under the curve (AUC), accuracy, sensitivity, specificity, calibration curves, and decision curve analysis. An online prediction platform was developed to facilitate individualized risk assessment.ResultsAmong 427 patients, postoperative VTE and PND occurred in 34 and 35%, respectively. For VTE prediction, the final Top-10 random forest model outperformed the Caprini score alone and achieved an AUC of 0.815 (95% CI, 0.720–0.910) in the held-out test set. Performance remained strong in the clinically significant VTE sensitivity analysis (AUC, 0.923; 95% CI, 0.847–0.998). SHAP analysis indicated that older age, elevated D-dimer and fibrin degradation products (FDP), as well as lower hemoglobin levels, were associated with increased predicted VTE risk. For PND prediction, the final Top-10 logistic regression model achieved an AUC of 0.741 (95% CI, 0.627–0.854). Older age, recurrent glioma, higher Caprini score, higher neutrophil percentage, and hypertension history tended to increase predicted PND risk. Models were deployed in the GLOBE web platform (https://gliomas.shinyapps.io/GLOBE/) for real-time preoperative risk prediction.ConclusionWe developed accurate, interpretable, and clinically meaningful preoperative prediction models for postoperative VTE and PND in patients with glioma. The GLOBE online prediction system translates these models into a practical tool for individualized perioperative risk stratification.