A sparse pair-preserving centroid-based supervised learning method for high-dimensional biomedical data or images

Abstract

In various biomedical applications designed to compare two groups (e.g. patients and controls in matched case-control studies), it is often desirable to perform a dimensionality reduction in order to learn a classification rule over high-dimensional data. This paper considers a centroid-based classification method for paired data, which at the same time performs a supervised variable selection respecting the matched pairs design. We propose an algorithm for optimizing the centroid (prototype, template). A subsequent optimization of weights for the centroid ensures sparsity, robustness to outliers, and clear interpretation of the contribution of individual variables to the classification task. We apply the method to a simulated matched case-control study dataset, to a gene expression study of acute myocardial infarction, and to mouth localization in 2D facial images. The novel approach yields a comparable performance with standard classifiers and outperforms them if the data are contaminated by outliers; this robustness makes the method relevant for genomic, metabolomic or proteomic high-dimensional data (in matched case-control studies) or medical diagnostics based on images, as (excessive) noise and contamination are ubiquitous in biomedical measurements.