A new Mercer sigmoid kernel for clinical data classification.

Area of Science:

• Machine Learning
• Computational Biology
• Medical Informatics

Background:

• Support Vector Machines (SVM) classification relies on Mercer kernels, like Gaussian RBF, for clinical data.
• The sigmoid kernel, though popular, is non-Mercer, posing validity challenges for clinical applications.
• Existing emulations of the sigmoid kernel have limitations.

Purpose of the Study:

• Introduce the first Mercer sigmoid kernel, ensuring trustworthiness for clinical data classification.
• Compare the Mercer sigmoid kernel with existing sigmoid kernels and Gaussian RBF kernels.
• Evaluate the performance of the Mercer sigmoid kernel on clinical and non-clinical datasets.

Main Methods:

• Developed a novel Mercer sigmoid kernel.
• Analyzed the mathematical properties and similarity to the standard sigmoid kernel.
• Implemented a normalization technique to improve sigmoid kernel accuracy.
• Performed classification tasks on three clinical datasets (melanoma detection) and non-clinical datasets.

Main Results:

• The Mercer sigmoid kernel demonstrated superior mean accuracy on clinical datasets, outperforming popular kernels in melanoma detection.
• On non-clinical datasets, the Mercer sigmoid kernel showed comparable median accuracy to the Gaussian RBF kernel.
• Both the Mercer sigmoid kernel and Gaussian RBF kernel exhibited complementary strengths in classifying specific data points.

Conclusions:

• The proposed Mercer sigmoid kernel is a valid and effective tool for clinical data classification, particularly for tasks like melanoma detection.
• The study introduces a normalization technique that enhances the performance of the standard sigmoid kernel.
• The Mercer sigmoid kernel offers a reliable alternative to non-Mercer kernels, bridging a gap in SVM applications for healthcare.