Search

In this work, a methodology for biosignal analysis (e.g. pathology diagnosis) is discussed, which is based on dynamic relevance analysis of stochastic features extracted from different decomposition techniques of biosignal recordings. Dimension reduction is carried out by adapting in time commonly used latent variable techniques, in such a way, that the data information is maximally preserved for a given relevance function. Specifically, since the maximum variance is assumed as a measure of relevance, time– adapted supervised approaches are developed. Additionally, in the case of high dimensionality data with significant correlation among the whole set, a dimensionality reduction technique is proposed, based on time–frequency relevance maps. The proposed approaches are experimentally assessed on real-world data sets, allowing to confirm whether the proposed feature selection algorithm is adequate for classification purposes. The conjunction of these advances conforms a methodology for training pattern recognition systems, which is a fully automatized dimensionality reduction method that allows the use of functional representations. The main advantage of the proposed methodology, is that preserves the maximum information among the high dimensional input data. In this terms of classifi- cation performance, the proposed methodology is efficient and competitive, outperforming other similar methods. Doctorado