A novel machine learning framework for automated detection of arrhythmias in ECG segments

Arrhythmias such as Atrial Fibrillation (A(fib)), Atrial Flutter (A(fl)), and Ventricular Fibrillation (V-fib) are early indicators of Stroke and Sudden Cardiac Death, which are significant causes of death globally. Therefore, it is vital to detect patients with these conditions early. Manual inspection of ECG signals is tedious, time-consuming, and is limited by inter-observer variabilities. Further, it is challenging to accurately differentiate several types of arrhythmias in complex non-linear ECG signals. Computer-aided Decision Support Systems (CDSS) could be valuable in such a scenario. The CDSS uses machine learning techniques to learn the subtle differences in these rhythms and can be used for fast, accurate, repeated, and objective classification of arrhythmias. A novel CDSS has been proposed for the discrimination of normal rhythm (N-sr) from A(fib), A(fl,) and V-fib using machine learning techniques. Predictive models have been developed for ECG segments of two durations: 2 s and 5 s. The number of samples from each of the four classes were balanced using synthetically generated samples with the ADASYN technique. Third-order cumulant images were determined from the ECG segments. 18 non-linear features, including entropies and other texture-based features, were extracted from the cumulant images, and significant features were selected using the t-test. The selected features were used to train several classifiers.On evaluating several different classifiers with the significant features using tenfold stratified cross-validation, the Random Forest classifier consistently performed better for both two and five second ECG duration studies. An accuracy of 98.2%, sensitivity of 98.1%, and specificity of 99.4% were obtained for the 2-s dataset. For the 5-s dataset, the accuracy, sensitivity, and specificity were 98.8%, 98.8%, and 99.6%, respectively. Due to the intermittent occurrence of arrhythmia, analysis of longer duration ECG signals will help detect the onset of critical episodes of arrhythmia more accurately. Since the proposed predictive models work effectively in detecting arrhythmia in two or five second ECG segments rather than single ECG beats, they have better clinical adaptability and can be incorporated into clinical monitoring systems.