Brain Cancer Prediction Based on Novel Interpretable Ensemble Gene Selection Algorithm and Classifier.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Human brain tumors arise from abnormal cell growth, necessitating accurate classification for prognosis and treatment.
• Cancer microarray data presents challenges due to high dimensionality (many gene expression levels) and limited samples.
• Existing classification studies often lack biological interpretability of the genes involved.

Purpose of the Study:

• To develop a novel hybrid model for brain tumor classification.
• To integrate feature selection (mRMR) and hyperparameter optimization for gradient boosting classifiers.
• To bridge the gap between cancer classification accuracy and biological interpretation of implicated genes.

Main Methods:

• Proposed a hybrid model combining Minimum Redundancy Maximum Relevance (mRMR) feature selection with distributed hyperparameter optimization for gradient boosting.
• Evaluated the model using Naive Bayes (NB), Random Forest (RF), and Support Vector Machine (SVM) classifiers.
• Compared optimized CatBoost and XGBoost performance across multiple cross-validation folds.

Main Results:

• The optimized CatBoost classifier outperformed optimized XGBoost in AUC, sensitivity, and specificity across several cross-validation folds.
• Optimized CatBoost achieved 0.91±0.12 accuracy, significantly improving upon the unoptimized version (0.81±0.24).
• Hybrid algorithms enhanced NB, RF, and SVM accuracy, with SVM and RF reaching 0.97±0.08, surpassing NB (0.91±0.12).

Conclusions:

• The proposed hybrid model significantly improves brain tumor classification accuracy.
• Optimized gradient boosting methods, particularly CatBoost, demonstrate superior performance.
• The study successfully links classification findings to relevant biological insights from gene expression data.