OBJECTIVE: WHO grade 4 glioma is the most common primary malignant brain tumor, with a median survival of only 14.6 months. Predicting survival outcomes remains challenging due to the tumor's heterogeneity and the influence of multiple clinical factors. Machine learning (ML) techniques have demonstrated superior predictive performance compared to traditional statistical models. Embedded feature-selection techniques such as Lasso shrinkage or Random-Forest importance scores are widely used, yet grade-4-glioma prognostic models still rely on an initial clinician-curated variable list and on ad-hoc cut-offs (e.g., "top X features" or "above certain threshold") when deciding how many ranked features to keep-choices that markedly influence model accuracy. We therefore developed a fully data-driven pipeline that begins with an unrestricted pool of clinical, functional, and biomarker variables, employs SHAP values for global importance ranking, and uses automated feature-subset optimization to identify the most optimal combination of predictors that maximizes survival-prediction performance in grade-4 glioma.
METHOD: We retrospectively analyzed clinical data from 764 patients who underwent grade 4 glioma resection at a single institution. Five ML models (XGBoost, AdaBoost, Random Forest, Decision Tree, and Neural Networks) were trained to predict survival time and classify into longer-term survival (≥ 12 months) and short-term survival (< 12 months). Feature selection was performed in two steps: (1) Shapley Additive Explanations (SHAPs) were used to identify the most important prognostic features influencing survival outcomes, and (2) feature-subset optimization was applied to determine the optimal number of top features to be included in each model. 5-fold cross-validation (CV) and holdout testing were performed to evaluate the models' performance on unseen testing data. Model evaluation was conducted using root mean square error (RMSE) for regression and area under the receiver operating characteristic curve (AUROC) for classification. Decision Curve Analysis (DCA) was performed to evaluate the clinical utility of the models.
RESULTS: Feature selection and model optimization significantly enhanced predictive accuracy across both regression and classification tasks. In regression, AdaBoost achieved the lowest RMSE of 1.69 months after feature selection, outperforming other models. In classification, XGBoost demonstrated the highest AUROC (0.85) on holdout testing, though all ensemble models (XGBoost, Random Forest, and AdaBoost) achieved comparable performance with no statistical significance. DCA revealed that XGBoost and Random Forest achieved the most net benefit of 0.24 and 0.22, respectively. Key prognostic features consistently identified included patient age, tumor location, radiation dose, extent of resection, Karnofsky Performance Score, and MGMT promoter methylation status. Biomarkers such as Ki-67, ATRX, and TP53 also emerged as important predictors of survival outcomes. The model also uncovered several cognitive and functional deficits-including preoperative and postoperative language deficits, permanent motor deficits, and perioperative seizures-that were previously underutilized in survival prediction models.
CONCLUSION: ML-based feature selection enhances survival prediction in grade 4 glioma by systematically identifying the most relevant prognostic factors while minimizing human bias. Our findings suggest that ensemble models, and particularly AdaBoost, offer robust prognostic capabilities. These insights can aid clinicians in personalized treatment planning and patient counseling.
