Neural networks for segmentation and clustering of biomagnetical signals

Abstract

When measuring biomagnetic signals the amount of data required is very large due to modern multichannel sensor arrays. Using the example of the magnetocardiogram (MCG), the authors show how these data can be automatically segmented and clustered with the help of neural algorithms. Self-organizing maps are not suitable for this application due to the character of the measured data. The data are compressed with the help of a special neural network. A very fast learning algorithm is used in the training phase, requiring substantially less computing power than conventional methods. Combined with a hierarchical cluster algorithm, a recognition rate of 100% of extrasystoles in MCG data was achieved.<>