Self Organizing Maps in respiratory signals classification

Abstract

Respiratory systems can be analyzed and modeled as autonomous dynamic systems. We use a Self Organizing Map (SOM) to build a suitable respiratory dynamical model, based on the reconstruction of a complex system attractor and model it using a neural network. The SOM reconstruction approach discussed in this paper is first tested on a well-known chaotic model (Hénon fractal attractor), and evaluated using both linear (mean and std) and non-linear indices (correlation dimension and Lyapunov exponents) indices. Subsequently the proposed approach is tested on a biomedical open database (MIT-BIH Polysomnographic), using respiratory signal patterns collected during the sleep stage. Classification results were analyzed both from a qualitative and a quantitative viewpoint by comparing the resulting non-linear indices obtained through the SOM-based reconstruction model with those obtained directly from inspiratory time series (TI) data. Results are in close agreement. © Springer-Verlag Berlin Heidelberg 2007.