SVM-RFE: selection and visualization of the most relevant features through non-linear kernels.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Healthcare

Background:

• Support Vector Machines (SVM) are powerful for high-dimensional data but originally lacked predictor importance evaluation.
• Biomedical research requires identifying key variables for robust predictive models.
• Existing variable importance methods for SVM primarily focused on linear kernels, limiting applications with non-linear relationships or survival data.

Purpose of the Study:

• To extend the Recursive Feature Elimination (RFE) algorithm for variable selection in non-linear SVM and SVM survival analysis.
• To propose and evaluate novel approaches for ranking predictor variables using these advanced SVM models.

Main Methods:

• Developed three new algorithms to rank variables based on non-linear SVM and SVM for survival analysis.
• Utilized visualization of RFE iterations to identify the most relevant predictors.
• Evaluated methods using simulation studies with time-to-event data and three real biomedical datasets, comparing against the original SVM-RFE algorithm.

Main Results:

• The proposed algorithms generally outperformed the standard RFE for non-linear kernels in simulation studies.
• RFE-pseudo-samples demonstrated superior performance in identifying relevant variables, even with correlated predictors.
• The methods allow for visualization and identification of key predictors in biomedical datasets.

Conclusions:

• The proposed approaches accurately select variables and assess predictor-outcome associations for biomedical data using SVM.
• RFE-pseudo-samples is particularly effective for variable selection and interpretation in realistic biomedical data scenarios.
• These novel methods surpass classical RFE in performance for complex biomedical data structures.