Composite neural network architecture for extensive long term ECG analysis
Rosaria Silipo, Carlo Marchesi · 2002
A composite ECG analyser, based on Neural Networks (NN), has been designed and carried out to recognise main cardiac diseases, like arrhythmia, ischemia, some myocardial chronic diseases. Pre-processing techniques have been introduced to enhance specific features of the different ECG abnormalities, so making more reliable further NN processing. Uncertainty management criteria gave robustness to the classifiers when dealing with new or ambiguous events. The best performances have been shown by the arrhythmia detector (error rate close to 0%; correct rejection rate of unknown patterns close to 100%). The ST-T changes detector showed a 77% sensitivity and an 85% PPA. Reliability robustness towards uncertain or missing data, wide range of cardiac abnormalities recognised by the analyser, allow the authors to consider it a step forward the commercially available systems and adequate to long term monitoring, aimed at early diagnosis, therapy assessment, post-surgical or post MI follow-up.