Probabilistic neural network array architecture for ECG classification

Using an array of three probabilistic neural networks (PNNs), we successfully identified both normal sinus rhythm (NSR) and atrial fibrillation (AF), as well as normal and PVC waveforms. Training and test waveforms were obtained from the MIT-BIH Arrhythmia Database. We applied various preprocessing techniques to reduce the dimension of the training sets. Combining independent PNNs, each classifying based on either shape or rhythm, we enhanced the integrated system performance by diminishing PNN element misclassifications. Most notably, the percentage of correctly classified PVCs from testing record 116, the worst performance based on shape, increased from 1.8% for a shape-only classification to 84.4% when adding rhythm information. Similarly, the amount of correctly classified NSR from testing record 201, the worst performance based on rhythm, rose from 18.7% to 92.0% when shape information was added.