Prototype modelling of body surface maps using tangent distance

Abstract

In recent years the tangent distance approach of Simard et al (1993) to pattern recognition tasks such as handwritten character recognition has proven successful. Simard's distance measure could be made locally invariant to any set of transformations of the input and when implemented in one-nearest-neighbour classification of handwritten digits, outperformed all other classification schemes. Hastie et al (1996) propose prototype models which generalise the concept of a centroid of a set of images in the Euclidian metric to a low-dimensional hyperplane in the tangent metric, and these prototypes can be used to reduce lookup time in classification. We propose to apply and extend the tangent distance approach to classify a set of body surface maps, which are recordings of the electrical activity of the heart, of a large number of patients with various cardiac conditions. Using a grid of p electrodes attached to the anterior chest, we calculate a number of p-dimensional observation vectors for each patient and classify input maps on the basis of overall distance of map to prototypes derived from training set maps over all included observation vectors.