Enhancing clinical insights in glioma grading using Bayesian Optimization and Explainable AI

Accurate glioma grading is paramount for guiding treatment strategies and predicting patient prognosis. This study introduces a novel and clinically relevant framework for glioma grade classification, specifically distinguishing between low-grade gliomas (LGG) and glioblastomas (GBM), by leveraging the unique strengths of Generalized Additive Models (GAMs). GAMs, chosen for their ability to model complex, non-linear relationships while maintaining inherent interpretability, were optimized using Bayesian Optimization for efficient hyperparameter tuning and augmented by Explainable Artificial Intelligence (XAI) methodologies. The 'Glioma Grading Clinical and Mutation Features' dataset from Kaggle underwent meticulous preprocessing. A feature ranking algorithm identified the most informative features, reducing noise and enhancing model accuracy. The Bayesian-optimized GAM achieved an accuracy of 0.9012 and F1-score of 0.8975 on a held-out test set, demonstrating superior or competitive performance compared to other established models, including Random Forest, LogitBoost, Support Vector Machines, and Artificial Neural Networks. Notably, among these traditional methods, the Random Forest performed strongly on the training set (accuracy 0.8899); however, GAM outperformed in the test set. To elucidate model decision-making and promote clinical translation, XAI techniques, including Permutation Feature Importance (PFI), SHapley Additive exPlanations (SHAP), and Partial Dependence Plots (PDPs), were employed. PFI identified IDH1 as the most critical predictor, while SHAP values revealed that, in addition to IDH1, features like Age, PTEN, ATRX, and CIC have considerable influence on model predictions. Furthermore, PDPs demonstrated the non-linear functional relationships of these features with the predicted outcome. These techniques provided interpretable insights into both global and local feature effects, highlighted the non-linearities in the data, and fostered trust in the model's predictions. This study demonstrates that integrating GAMs, Bayesian Optimization, and XAI techniques provides a robust, accurate, and clinically interpretable framework for glioma grade classification, showcasing their potential to enhance diagnostic accuracy, improve clinicians' understanding of glioma biology, and ultimately inform more personalized treatment strategies.