Best wavelet basis design for joint compression-classification of long ECG data records

Presents a novel algorithm for joint heartbeat compression-classification of long ECG data records. The joint scheme is possible by including a data adaptive analysis stage. For each heartbeat the authors design a wavelet basis that minimizes a certain cost function. As the ECG data is processed it is built into dictionaries, one of heartbeats and one of wavelets. The compressed data includes an ordered series of wavelets. The authors define two heartbeat categories, normal and abnormal, which they would like to classify, by looking only at the wavelet dictionary. A simple neural network is then trained to classify heartbeats by looking at the ordered wavelet series. When the network was trained for a given ECG the authors obtained 98% correct decisions.