A support vector machine-recursive feature elimination feature selection method based on artificial contrast...

Area of Science:

• Metabolomics
• Bioinformatics
• Machine Learning

Background:

• High-dimensional metabolomics data analysis requires effective feature selection.
• Support Vector Machine-Recursive Feature Elimination (SVM-RFE) is a common technique, but susceptible to noise and non-informative variables.
• Noise can negatively impact the accuracy of Support Vector Machine (SVM) learning models.

Purpose of the Study:

• To develop an improved feature selection method for metabolomics data.
• To enhance the identification of discriminative metabolites for disease classification.
• To mitigate the impact of noise and non-informative variables in SVM-RFE.

Main Methods:

• Proposed a novel Mutual Information (MI)-SVM-RFE method.
• Utilized artificial variables and MI to filter noise and non-informative features.
• Applied the method to a serum metabolomics dataset from patients with chronic hepatitis B, cirrhosis, and hepatocellular carcinoma analyzed by liquid chromatography-mass spectrometry (LC-MS).

Main Results:

• The MI-SVM-RFE method achieved an accuracy of 74.33±2.98% in distinguishing between three liver diseases.
• This accuracy is an improvement over the original SVM-RFE method, which yielded 72.00±4.15%.
• Identified 17 out of 34 significant ion features differentiating control subjects from the three liver disease groups.

Conclusions:

• The MI-SVM-RFE method is effective in filtering noise and selecting discriminative features in metabolomics data.
• This approach enhances the accuracy of disease classification compared to standard SVM-RFE.
• The identified ion features hold potential for diagnosing liver diseases.