An investigation of neural computing applied to the ambulatory monitoring of the electrocardiogram

Abstract

This thesis describes research carried out to construct an effective method of achieving data reduction of the recorded Electrocardiogram (ECG) through identification and rejection of nonnal sinus rhythm so that only abnormal heart cycles
shall be presented for recording.
Traditional ECG classification techniques are critically assessed in terms of their suitability for this project, however, it was concluded that their effectiveness is limited by their reliance on expert knowledge, and have therefore reached their theoretical maximum level of performance of between 54% to 60%. An alternative classification technique based on artificial neural networks (ANNs) was selected for further research.
It is shown that whilst techniques based on artificial neural networks are capable of performing the required pattern recognition task, results are presented to demonstrate that the sensitivity of the ANN classifier to certain wave features varies according to the method of representing the ECG data. It is concluded that to achieve maximum sensitivity in differentiating normal from abnormal ECG patterns, then hybrid data, including time and frequency domain components, must be applied to two separate ANNs, the outputs of which can be post processed to achieve improved classification
success.