Detection of ventricular ectopic beats using neural networks

The authors describe a system of three artificial neural networks (ANNs) trained to detect ventricular ectopic beats (VEBs) in the AHA database. Two ANNs were trained for each record, using a learning set containing five normal beats obtained from the record and seven VEBs obtained from a larger database. These two ANNs function as pattern recognizers for each record. The third ANN was trained to arbitrate disagreements between the first two ANNs. They replaced the beat classification logic of an existing VEB detector with this system of ANNs and evaluated its performance using 69 records from the AHA database. Gross VEB sensitivity improved from 94.83% to 97.39% when the ANN system replaced the original beat classification logic, and gross VEB positive predictivity improved from 92.70% to 93.58%