Real Time Electrocardiogram Annotation with a Long Short Term Memory Neural Network

Continuous monitoring of electrocardiogram from wearable devices can enable early detection of heart diseases. Ubiquitous monitoring on wearable electronics requires a novel class of algorithms that are low-power and have low-memory requirements. This work proposes a wearable compatible, and automatic solution for annotating Electrocardiogram (ECG) recordings while maintaining high accuracy of detection when users are carrying daily activities such as sitting, walking, and resting. We validate our solution with two Physionet datasets: the MITDB [1] (Boston's Beth Israel Hospital and MIT Arrhythmia Database), and the EDB [2] (European ST-T Database). In addition, we validate our method on a newly recorded dataset in collaboration with the 'Ziekenhuis Oost-Limburg' Hospital
¹
that has been collected using a prototype wearable device [3]. Our solution exploits a recurrent neural network that achieves an average F1 score of 94.8% over all three datasets. Our solution achieves better generalization performance than the gold standard method Pan Tompkins which achieves an average F1 score of 93%. In addition, our method can be extended to full ECG annotation. We used the QTDB dataset [4] and we report an accuracy of 91.6% while annotating all 5 waves (P-Q-R-S-T) of the ECG complex.