A Deviation Analysis Framework for ECG Signals Using Controlled Spatial Transformation

Current automated heart monitoring tools use supervised learning methods to recognize heart disorders based on ECG signal morphology. We develop a new ECG processing algorithm that enables early prediction of disorders through a novel deviation analysis. The idea is developing a patient-specific ECG baseline and characterizing the deviation of signal morphology towards any of the abnormality classes with specific morphological features. To enable this feature, a novel controlled non-linear transformation is designed to achieve maximal symmetry in the feature space. Our results using benchmark MIT-BIH database show that the proposed method achieves a classification accuracy of 96% and can be used to trigger yellow alarms to warn patients from increased risk of upcoming heart abnormalities (5% to 10% increase with respect to normal conditions). This feature can be used in health monitoring devices to advise patients to take preventive and precaution actions before critical situations.