INNV-47. Development and validation of a machine learning model for predicting grade IV glioma recurrence

Le résumé fourni par la source

Abstract BACKGROUND Grade 4 gliomas are the most aggressive primary brain tumors, with mean survival of ~1 year and high rate of recurrence. Time to recurrence is a critical predictor of overall survival. While machine learning (ML) models have shown promise in predicting recurrence with greater accuracy than traditional statistical methods, many lack interpretability, limiting their clinical utility. METHODS This retrospective study included patients who underwent surgical resection for grade 4 glioma at Johns Hopkins between 2014 and 2019. After data preprocessing and standardization, we applied SMOTE to address class imbalance and excluded highly correlated features. Three ML models—Random Forest (RF), XGBoost, and AdaBoost—were trained to predict recurrence. Recursive feature elimination (RFE) was used for feature selection, and model performance was evaluated using 5-fold cross-validation and testing data. Shapley values were used to assess feature importance. RESULTS Among the models, Random Forest achieved the best performance with a cross-validated AUC of 0.85 and a testing AUC of 0.81. XGBoost and AdaBoost followed with slightly lower metrics. Feature importance analysis revealed that KPS score, extent of resection, radiation dose, adjuvant temozolomide, and age were among the top predictors across models. KPS score was consistently ranked as a top predictor in all three models. Interestingly, unlike any other previous ML studies, our study also revealed that having preoperative language, motor, and cognitive deficits as top predictors of recurrence. An external online platform was developed for validation and use by other institutions. CONCLUSION Our ML-based framework not only accurately predicts recurrence in grade 4 glioma but also identifies key clinical predictors, enhancing interpretability for clinical decision-making. By leveraging over 100 clinical covariates and robust feature selection techniques, this study provides a comprehensive and interpretable tool to aid in personalized risk assessment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.