Introduction:

Implanting subcutaneous loop recorders for detecting arrhythmias in syncope or cryptogenic stroke is common. However, the accuracy of current algorithms used to detect arrhythmias is low. This leads to a large number of false classifications and requires high manual effort to correct for these errors. Recent advances in deep learning (DL) show a significant improvement in time series classification. In this work we evaluated the performance of latest DL methods to classify data from patients that were monitored with the Biotronik home monitoring system. 

Method:

535  electrocardiograms (ECG)recorded by implantable loop recorders were labeled for correct diagnosis by a senior cardiologist. The labelled data were used for training various deep learning models subcategorizable into one-vs.-all artefact classifiers and 9-class single-label-classification.

Results:

With the one-vs.-all artefact classifier a high specificity of diagnosing artefacts with 95 % F1-score could be achieved and after additional model tuning, the single-label-classifier achieved the same results. For the single-label classifier acceptable rates for atrial fibrillation and supraventricular tachycardia were achieved, both with 86 % F1-score. Additional model tuning improved the respective F1-score to almost 100 % for supraventricular tachycardia, artefacts and sinus tachycardia. For ventricular arrhythmias the method failed to identify such as their occurrence is rare and the amount of training data was scarce.

Conclusion:

Using DL models greatly improves the recall and precision of ECGs classification for artefacts, sinus tachycardias and supraventricular tachycardias and therefore can assist greatly at improving specificity of currently deployed algorithms. For atrial fibrillation and differentiation from ventricular or supraventricular extrasystolie these algorithms can still be improved. Gaining high specificity for ventricular tachycardia is difficult due to its rareness for labeling and training.