Classification of CAD severity using fusion neural net analysis of multiple exercise ECG waveform representations

A hierarchical fuzzy neural-net approach has been developed to classify averaged serial ECG waveforms gathered during exercise testing in order to determine the severity of coronary artery disease (CAD). The ST-T complex of each ECG was first transformed on a sample-by-sample basis to form multiple representations (e.g. raw amplitudes, delta-baseline, slope through J-junction) to highlight the salient features of the signal. These representations were then initially classified by individual neural-nets and the resultant CAD-group outputs used as inputs into a secondary fusion-net for final classification. Also at the fusion-net level, additional parameters (e.g., HR and phase) were added to the fusion-net´s input dimension space. Using only one lead, the processing gives nearly perfectly discrimination between angiographic normals and patients with severe 3-vessel disease. The use of FMM neural-nets is particularly relevant for this type of medical application since they allow the user to determine why and where the network decided on its results