An augmented ECG data based classification for arrhythmia using optimal feature set

The Electrocardiogram (ECG) is a pivotal tool for diagnosing heart conditions such as arrhythmia. Prompt detection of arrhythmias through continuous ECG monitoring is crucial to prevent life-threatening incidents during hospitalization or ambulatory monitoring. Despite advances in machine learning techniques for arrhythmia diagnosis, the analysis of complex ECG datasets poses substantial challenges. To address this, we have developed an Intelligent Arrhythmia Classification System aimed at enhancing diagnostic accuracy while maintaining a lower degree of computation.

We created a pipeline for the preprocessing of ECG data and the classification of arrhythmias. We applied low-complexity methods and incorporated domain knowledge in the preprocessing steps to improve diagnostics. We have used Multilayer Perceptron (MLP) for feature learning and classification of new ECG samples. However, because of the problem of having fewer abnormal beat samples compared to normal beats, we employed the Synthetic Minority Over-sampling Technique (SMOTE) to address this class imbalance. The experimental setup consists of a three-node cluster employed with Kafka and Spark for scalable and real-time multiclass classification.

The proposed system, tested on the benchmark MIT-BIH dataset, achieved an overall accuracy of 96.4%. Specifically, for Supra-Ventricular Ectopic Beat (SVEB), the positive predictive value is 95.2%, sensitivity is 95.3%, and F1-score is 95.24%. Similarly, for the Fusion Beat (F) class, the positive predictive value is 91.6%, sensitivity 91.4%, and F1-score is 91.49%.

Our method outperforms previous approaches with lower computational complexity despite fewer data in SVEB and F classes.