Autoassociator structured neural network for rhythm classification of long term electrocardiogram

The authors have designed and evaluated a new Artificial Neural Network (ANN) structure, built as an autoassociator, to classify some rhythm abnormalities of a single patient. They used a standard database (BIH-MIT) for the testing phase. The network is fed to reproduce the input pattern morphology over a part of the output layer and to give the pattern class over the remaining part. The authors have defined two uncertainty criteria to reject unknown or uncertain patterns. The recognition percentages were very good for normal vs. PVB beats (99.84% normal vs. 92.96% PVB) and a little worse for normal vs. APB´s (97.71% normal vs. 93.44% APB). The error the authors found was as small as 0.15% for PVB beats, and 0.91% for APB´s