Recursive Support Vector Machine Biomarker Selection for Alzheimer's Disease.

Area of Science:

• Biomedical data analysis
• Machine learning applications in healthcare
• Proteomics and biomarker discovery

Background:

• Early Alzheimer's disease (AD) detection is crucial but faces diagnostic challenges.
• Machine learning (ML) offers a viable solution for developing reliable diagnostic tools.
• High-throughput proteomic data analysis requires robust feature selection methods.

Purpose of the Study:

• To present a novel Support Vector Machine Leave-One-Out Recursive Feature Elimination and Cross Validation (SVM-RFE-LOO) algorithm.
• To demonstrate the algorithm's utility in the early classification and prediction of AD.
• To compare the performance of SVM-RFE-LOO against existing methods like SVM and SVM-RFE.

Main Methods:

• Analysis of serum samples from 300 participants (150 AD, 150 controls) using a multi-plex biomarker assay.
• Application of the SVM-RFE-LOO algorithm for feature selection and model development.
• Electrochemiluminescence (ECL) assay platform for biomarker quantification.

Main Results:

• The SVM-RFE-LOO algorithm reduced 21 biomarkers to 16, achieving an Area Under the Curve (AUC) of 0.980.
• Achieved high diagnostic performance with 94.0% sensitivity and 93.3% specificity.
• SVM-RFE-LOO demonstrated comparable performance to SVM and SVM-RFE but utilized fewer biomarkers.

Conclusions:

• SVM-RFE-LOO is effective for analyzing noisy high-throughput proteomic data in AD detection.
• The algorithm shows improved robustness to noise and feature recovery compared to SVM-RFE.
• This recursive feature elimination model can be generalized for biomarker discovery in other diseases.