Neural networks for improved automation of ventricular activation mapping

Whether or not a neural network could discriminate between electrograms from viable tissue (local myocardial activation) and nonviable tissue (distant electrical events) was investigated as a step toward improving automation of cardiac activation mapping. Cryolesions were made in canine myocardium. Five days later, plunge needles with eight electrodes were inserted through normal tissue and the lesions, and during atrial pacing unipolar electrograms were recorded, sampling at 8000 Hz with 12-bit resolution. Data were analyzed from 19 electrodes. The type of tissue from which electrograms were recorded was correctly identified for 18 of the 19 electrodes, indicating that the network was able to identify features of the electrograms that distinguished viable from nonviable tissue sources, and that this generalized to new electrograms upon which the network had not been trained. It is concluded that neural networks may have application in automated cardiac activation mapping