Automated feature learning and survival prognostication in grade 4 glioma using supervised machine learning models

Area of Science:

• Neuro-oncology
• Machine Learning in Medicine
• Biostatistics

Background:

• WHO grade 4 glioma has poor prognosis (14.6 months median survival).
• Predicting outcomes is complex due to tumor heterogeneity and clinical factors.
• Machine learning (ML) models show promise over traditional methods.

Purpose of the Study:

• Develop a data-driven ML pipeline for grade 4 glioma survival prediction.
• Utilize SHAP values and automated feature optimization.
• Identify optimal predictors to maximize model accuracy.

Main Methods:

• Retrospective analysis of 764 grade 4 glioma patients.
• Trained five ML models (XGBoost, AdaBoost, Random Forest, Decision Tree, Neural Networks).
• Employed SHAP for feature importance and subset optimization, validated with cross-validation and holdout testing.

Main Results:

• Feature selection significantly improved predictive accuracy.
• AdaBoost achieved lowest RMSE (1.69 months) in regression; XGBoost highest AUROC (0.85) in classification.
• Key predictors: age, tumor location, radiation dose, resection extent, KPS, MGMT methylation, Ki-67, ATRX, TP53, and specific functional deficits.

Conclusions:

• ML-based feature selection enhances grade 4 glioma survival prediction by reducing bias.
• Ensemble models, particularly AdaBoost, demonstrate strong prognostic capabilities.
• Findings support personalized treatment planning and patient counseling.