Machine learning-based prediction of glioma grading

Area of Science:

• Neuro-oncology
• Computational Biology
• Genomics

Background:

• Gliomas are primary central nervous system tumors with significant heterogeneity.
• Accurate glioma grading is crucial for treatment planning and prognosis.
• Conventional histopathology for grading gliomas suffers from subjectivity and poor reproducibility.

Purpose of the Study:

• To develop a machine learning (ML) model for improved early glioma grading.
• To integrate clinical and molecular characteristics for enhanced diagnostic accuracy.
• To support individualized treatment strategies for glioma patients.

Main Methods:

• Utilized The Cancer Genome Atlas (TCGA) dataset for model development.
• Employed recursive feature elimination (RFE) with Random Forest (RF) and Elastic Net Regression (ENR) for feature selection.
• Applied Synthetic Minority Oversampling Technique (SMOTE) for data balancing and optimized various ML algorithms using hyper-parameter optimization (HPO).
• Constructed 34 ensemble learning models using voting and stacking algorithms.
• Externally validated all models on the Chinese Glioma Genome Atlas (CGGA) dataset.
• Conducted SHapley Additive exPlanations (SHAP) analysis to interpret model predictions.

Main Results:

• Identified 11 key grading features, including TP53 and IDH1.
• The RF model achieved an Area Under Curve (AUC) of 0.916 (TCGA) and 0.797 (CGGA).
• The Voting25 ensemble model (RF, XGBoost, KNN) achieved optimal performance with AUCs of 0.928 (TCGA) and 0.794 (CGGA).

Conclusions:

• Eleven key features identified can aid in molecular detection and personalized glioma therapy.
• Optimized ML models, particularly the RF model, show promise for future research.
• The voting ensemble method integrating RF, XGBoost, and KNN demonstrated superior accuracy and robustness.
• Successful external validation on the CGGA dataset confirms the generalizability of the developed models.