Automated Detection of Ventricular Heartbeats from Electrocardiogram (ECG) acquired during Magnetic Resonance Imaging.

Pierre Aublin1, Jacques Felblinger2, Julien Oster2
1INSERM, 2IADI, Inserm U1254


Aim. ECG signals are distorted during Magnetic Reso- nance Imaging (MRI) by the electromagnetic environment. Automated analysis of ECG is therefore highly difficult. The detection of pathological heartbeats is currently lim- ited to excluding heartbeats with outstanding RR intervals. This study aimed at implementing a machine-learning (ML) based heartbeat classifier, using hand-crafted fea- tures, for the automatic detection of ventricular heartbeats during MRI examinations.

Method. A model was trained and assessed on the MIT- BIH Arrhythmia Database, using the AAMI recommenda- tions for class labelling and division in two equal subsets DS1 and DS2. A single lead of ECG signals was analyzed. Two classes of features were extracted for each heartbeat: (i) QRS morphological features based on Hermite func- tion decomposition, (ii) temporal features based on the lo- cal RR interval time-series around the heartbeat. A sup- port vector machine was trained to detect normal (N), and ventricular ectopic beats (V') using these features on DS1. The model was then tested on an in-house database of ECG acquired inside a 1.5T MRI scanner during standard clin- ical examination.

Results The classifier achieved F1 scores of 0.99, 0.85 on N and V' classes respectively on DS2. While the classi- fier only achieved F1 scores of 0.62 and 0.15 on the ECG signals acquired in MRI.

Conclusion A heartbeat classifier was developed on the MIT-BIH arrhythmia database focusing on temporal fea- tures and morphological features focsing mainly on the QRS complex (as they are less likely to be distorted by the MRI environment). Performance on MIT-BIH was ac- ceptable although slightly lower than state-of-the-art ap- proaches, but dropped significantly on MRI data. The re- sults highlight the need for further developments by adding MRI-related artifact reduction, while also retraining the classifier on MRI acquired datasets.