Explainable Machine Learning Models for Glioma Subtype Classification and Survival Prediction.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Gliomas are aggressive brain tumors with poor prognoses, necessitating early and accurate diagnosis.
• Tumor classification and survival prediction are critical for effective glioma treatment strategies.

Purpose of the Study:

• To develop and validate explainable machine learning (ML) models for classifying glioma subtypes (astrocytoma, oligodendroglioma, glioblastoma).
• To predict patient survival rates using RNA-sequencing (RNA-seq) data.
• To enhance model transparency through Shapley additive explanations (SHAP) analysis.

Main Methods:

• Analysis of publicly available RNA-seq datasets.
• Application of feature selection to identify key gene biomarkers.
• Development and comparison of various ML models for classification and survival analysis.
• Interpretation of model predictions using SHAP values.

Main Results:

• Thirteen key genes (e.g., TERT, VEGFA, MMP9) were identified as significantly associated with glioma subtypes and survival.
• Support Vector Machine (SVM) achieved a balanced accuracy of 0.816 and AUC of 0.896 for classification.
• Case-Control Cox regression (CoxCC) model demonstrated strong survival prediction with a C-index of 0.809.
• SHAP analysis provided insights into gene expression's influence on model outcomes.

Conclusions:

• The developed explainable ML models offer a robust tool for glioma diagnosis and prognosis.
• These models can assist clinicians in tailoring treatment strategies for improved patient outcomes.
• The identified gene biomarkers hold potential for further research into glioma pathogenesis.