Using genetic algorithm trained perceptrons with adaptive structure for the detection of premature ventricular contraction

Neural networks, typically multilayer perceptrons, have been widely used for the detection of heart arrhythmia based on electrocardiogram recordings. However, most state-of-the-art neural networks have a static preset structure, which tends to be over-massive and contributes to the problem of overfitting. In this study, we use genetic algorithm to adaptively decide the optimal model structure of a multilayer perceptron. The number of hidden layers and the number of nodes in each layer are dynamically obtained together with the synaptic connecting weights. Searching and cleanup stages are used in training to decide the optimal number of neurons in each layer based on crossover and mutation operators as well as cloning. Cooperative populations are chosen during evaluation. It works layer-by-layer until no more layer is required. The algorithm is applied to the MIT-BIH arrhythmia database to distinguish premature ventricular beats from normal beats. The experiment is conducted on all available records with at least 50 premature ventricular beats: records 106, 119, 200, 201, 203, 208, 210, 213, 215, 217, 219. The training starts from 500 initial neurons. As a result of adaptive structure searching, 10 out of 11 records needs only one layer except record 106 which needs 2 layers. The number of nodes on each layer is at most 3. The average recognition rate of all records is 96.96%. Results show that top performances can be obtained by multilayer perceptrons with simpler model structure than commonly reported in literature.