A comparative study on feature selection for a risk prediction model for colorectal cancer

Area of Science:

• Computational biology
• Biostatistics
• Machine learning in healthcare

Background:

• Risk prediction models identify individuals at high risk for diseases.
• Feature selection enhances prediction model performance and identifies key risk factors.
• Assessing feature selection stability is vital for reliable cancer risk analysis.

Purpose of the Study:

• To evaluate feature ranking algorithms for colorectal cancer risk prediction.
• To assess the stability and performance of various feature selection methods.
• To compare algorithm-identified features with expert-defined factors.

Main Methods:

• Assessed feature ranking algorithms (e.g., SVM, Logistic Regression, Random Forest) for risk prediction models.
• Evaluated algorithm robustness using scalar metrics and a novel visual approach.
• Compared selected features against expert knowledge in colorectal cancer.

Main Results:

• Support Vector Machines (SVM) with top-41 features (SVM wrapper) achieved an AUC of 0.693.
• Logistic Regression with top-40 features (Pearson) achieved an AUC of 0.689.
• Feature selection improved AUC by 3.9% (SVM) and 1.9% (Logistic Regression) over full datasets. Random Forest was most stable, Neural Network-based wrapper least stable.

Conclusions:

• Stability and model performance in feature selection must be considered together.
• Random Forest offers high stability but not always top performance.
• SVM wrapper and Pearson correlation provide a balance of moderate stability and strong predictive performance for colorectal cancer risk.