Feasibility study of complete neural net based classification of signal-averaged high-resolution ECGs

Classification of signal averaged ECGs is divided into two phases: (a) QRS-onset and QRS-offset determination and (b) categorization based on three derived features: QRS duration (QRSd), root mean square of the terminal 40 ms of the QRS (RMS) and the terminal low amplitude signal of the QRS below 40 μV (LAS). Purpose of this feasibility study was the neural realization of each of these phases and the comparison of the different approaches. Both steps were realized with the neural network and the standard approach. Four combinations of the methods are possible. These were tested on 95 high-resolution signal averaged ECG recordings from 51 healthy volunteers and 44 patients with coronary artery disease. Using a neural network in the classification phase increased the sensitivity of the whole process by approximately 30% compared to the standard method without the need to visually correct the QRS-onset and-offsets. These initial results are very positive but need to be substantiated with further patient data