White box radial basis function classifiers with component selection for clinical prediction models.

Area of Science:

• Machine Learning
• Computational Statistics
• Bioinformatics

Background:

• Support Vector Machines (SVMs) with Radial Basis Function (RBF) kernels offer high performance for complex classification tasks.
• However, standard SVMs lack transparency, hindering their use in domains requiring interpretable decision support systems.
• Existing feature selection methods offer limited insight into variable impact on predictions.

Purpose of the Study:

• To propose a novel, flexible, and sparse classifier that enhances interpretability in decision support systems.
• To expand the RBF kernel into interpretable and visualizable components, including main and two-way interaction effects.
• To develop a sparse model representation using an iterative l1-regularized parametric model.

Main Methods:

• Expansion of the Radial Basis Function (RBF) kernel into interpretable and visualizable components.
• Incorporation of main and two-way interaction effects for comprehensive analysis.
• Utilizing an iterative l1-regularized parametric model for sparse representation.

Main Results:

• The proposed method demonstrated superior performance on toy problems, achieving higher Area Under the Curve (AUC) than standard RBF SVMs, particularly with irrelevant variables.
• On benchmark datasets (Pima Indian diabetes, Wisconsin Breast Cancer), the method's AUC was comparable to standard RBF SVMs and existing literature.
• The method successfully identified relevant components and visualized their effects, enabling expert interpretation.

Conclusions:

• A new method for flexible and sparse risk prediction models has been developed.
• The proposed method achieves performance comparable to standard RBF SVMs.
• The key advantage is the resulting model's interpretability for domain experts.