Improving ECG diagnostic classification by combining multiple neural networks

Recent research has shown that combining multiple versions of unstable classifiers such as neural networks results in reduced test set error. By resampling from the original training set, modified training sets are formed and used to train separate neural network classifiers. The outputs of these classifiers are then combined by voting. Bagging is one of the more simple techniques of resampling and involves sampling with replacement from the training set and combining the network outputs with equally weighted voting. Other more sophisticated techniques adaptively resample the training data and give additional weights to cases which have previously been misclassified. The authors applied a number of these techniques to the problem of ECG diagnostic classification and sound an improvement of greater than 10% in overall classification rate was readily achievable