Machine learning and feature selection for the analysis of Alzheimer Metabolomics Data

Metabolomics provides a highly valuable data for disease diagnosis and prediction, due to its accurate, possibly non-invasive and low cost determination of highly physiologically relevant molecular data. At the same time, early and accurate Alzheimer disease diagnosis remains a highly challenging task. Consequently, number of researchers have recently become interested in utilizing metabolomics data for the discovery of biomarker for classification of Alzheimers disease (AD). However, although many methods already exist for the determination of markers for ADs identification from high throughput data, more precise and accurate method for feature selection as well as AD classification are still needed. Various machine learning approaches have achieved successful classification of samples between cognitively healthy and AD using metabolomics data. However, they failed to achieve a good classification rates for differentiation between AD, mild cognitive disorder and cognitively healthy or normal. In this paper, we propose new machine learning approaches to select a subset of features that improve the classification rates between these three classes, thus allowing separation between different levels of cognitive disorders. Our experiment results demonstrate that the performances of several classifiers are improved when using our selected metabolic markers relative to the classification that can be obtained from the complete metabolomics dataset. The obtained results indicate that our algorithms are effective in discovering markers for ADs classification from metabolomics data.