Automated Evoked Potential Analysis Using Back Propagation Networks

A system for automated analysis of evoked potential waveforms using neural networks was developed. The system uses two separate networks. A "classification network" classifies evoked potential waveforms as absent, interpretable, or uninterpretable. A "latency measurement network" determines the latency of interpretable waveforms. Each network is a feed forward back propagation network with two hidden layers. Network performance was evaluated using single channels from 279 visual evoked potential recordings and 137 median nerve somatosensory evoked potential recordings. For each modality, data were randomly divided into two data sets, one for training and the other for testing. The system correctly classified 90% of VEPs and 93% of SEPs. For EP waveforms correctly classified as interpretable, the system determined the peak latency within +/- 3 msec for VEPs and +/- 0.5 msec for SEPs in all cases. Neural networks appear to be effective classifiers of evoked potential waveforms.