Union With Recursive Feature Elimination: A Feature Selection Framework to Improve the Classification Performance of...

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Accurate modeling of cause-specific deaths in colorectal cancer (CRC) patients is challenging, impacting treatment selection.
• Existing machine learning tools often struggle with high-dimensional and imbalanced data common in cancer genomics.

Purpose of the Study:

• To develop and evaluate a novel feature selection framework, Union with Recursive Feature Elimination (U-RFE), for identifying key predictors of CRC mortality.
• To compare the performance of various classification algorithms using features selected by U-RFE for improved death classification.

Main Methods:

• Implemented the U-RFE framework using logistic regression, support vector machines, and random forest as base estimators.
• Performed union analysis on feature subsets to create a comprehensive feature set crucial for CRC progression.
• Compared five classification algorithms (LR, SVM, RF, XGBoost, Stacking) on The Cancer Genome Atlas (TCGA) dataset.

Main Results:

• The U-RFE framework significantly improved the performance of multiple classification models.
• Stacking emerged as the top-performing model, achieving high accuracy (0.864) and F1-weighted score (0.851) with 298 selected features.
• Performance for minority death categories was notably enhanced, indicating better handling of imbalanced data.

Conclusions:

• The U-RFE approach effectively selects crucial features, enhancing the performance of machine learning models for CRC mortality prediction.
• This method is beneficial for classifying deaths using clinical and omics data, especially when dealing with feature redundancy and class imbalance.
• The findings suggest U-RFE can aid in more precise treatment strategies for colorectal cancer patients.