Feasibility of neural network based QRS-T cancellation schemes for P-wave detection

This work reviews the possible application of artificial neural networks (ANNs) to the problem of ventricular activity (VA) cancellation. The system proposed consists of estimating a non-linear time-varying transfer function between two ECG channels using a Time Delay Neural Network (TDNN). Three different TDNN topologies (purely feed-forward, simple recurrent, and fully recurrent) were implemented and tested using record 108 of the MIT-BIH DB. In order to evaluate these networks, two performance measures were introduced in this work, namely VA energy change and Signal to Noise Ratio (SNR) improvement. Preliminary results using VA energy change and SNR improvement indicators showed that the simple recurrent topology presents the best performance.