Binary ECG Classification Using Explainable Boosting Machines for IoT Edge Devices

This paper presents an explainable, low-complexity binary electrocardiogram (ECG) classifier to be deployed in a resource-limited wearable edge device. The presented technique could be used as stand alone on an edge device or in a two-stage distributed edge-cloud classifier, where a preliminary two-class classification is done on the edge and a more comprehensive multi-class classification is done on the cloud. Considering the importance of interpretability in clinical decision support systems, we used an Explainable Boosting Machine (EBM) classifier for the preliminary binary classification. EBMs can be implemented using a decision tree like structure and therefore complexity is much lower than deep learning models and many traditional classifiers. To further limit complexity we used only limited ECG features in this work, which include the peak amplitudes of signal segments, time intervals, and a few signal amplitudes to represent the signal morphology. In addition to these direct features, EBM is capable of learning the interactions between features to derive intermediate feature combinations and thus improve the classification performance. We used the Physionet MIT-BIH Arrhythmia dataset for performance evaluation and the EBM classifier achieves an accuracy of 96.84%, F1 score of 91.38%, and sensitivity of 96.83%. When used in a distributed edge-cloud classifier configuration, our proposed work limits cloud transmission to only 19.37% of the total data, which in turn reduces the power consumed in the wearable edge device.

Automated summary

This paper presents an explainable and low-complexity binary ECG classifier for resource-limited wearable edge devices. The classifier, based on Explainable Boosting Machine (EBM), achieves high accuracy and interpretability by using limited ECG features and a decision tree structure.