Ultra-high Frequency Deep-learning Beat Detector Delivering QRS Onsets and Offsets

Zuzana Koscova1, Radovan Smisek2, Petr Nejedly3, Josef Halamek4, Pavel Jurak1, Pavel Leinveber5, Karol Curila6, Filip Plesinger1
1Institute of Scientific Instruments of the CAS, 2Brno University of Technology, Faculty of Electrical Engineering and Communication, Department of Biomedical Engineering, 3Institute of Scientific Instruments of the Czech Academy of Science, 4Institute of Scientific Instruments, CAS, CZ, 5ICRC at St. Anne´s University Hospital, Brno, Czechia, 6Cardiocenter FNKV and 3rd Faculty of Medicine in Prague


Abstract

Background: QRS duration is a common measure linked to conduction abnormalities in heart ventricles. We hypothesized that appropriately de-signed deep-learning QRS detector could also deliver this measure. Aim: We propose a QRS detector, further able to locate QRS onset and offset. Therefore, QRS duration and QRS detection could be delivered in one inference step. Method: A 3-second window from 12 leads of UHF ECG signal (5 kHz) is standardized and processed with the UNet convolutional neural network. The output is an array of QRS probabilities, further transformed into result-ant QRS onset and offset positions using probability criterion, allowing us to determine duration and final location of QRSs. Results: The model had been trained on 2,250 ECG recordings from the FNUSA-ICRC hospital (Brno, Czechia) containing spontaneous and paced data. The model was tested on 5 different datasets: private datasets from FNUSA and FNKV hospital (Prague, Czechia) and three public datasets (Ci-pa, Strict LBBB, LUDB). In terms of QRS duration, results showed a mean absolute error of 14.14 ± 4.33 ms and mean error of -2.43 ± 6.88 ms be-tween manually annotated durations and output of the proposed model. We recieved mean F-score of 0.98 ± 0.01 for QRS detection evaluated using da-tasets with QRS annotation marks (FNUSA, FNKV). Conclusion: Our results indicate high QRS detection performance on both spontaneous and paced UHF ECG data. We also showed that QRS detection and duration could be combined in one deep learning algorithm.