|
Analysis of ultra-high frequency ECG using deep learning
Koščová, Zuzana ; Antin, Christoph Hoog (oponent) ; Plešinger, Filip (vedoucí práce)
Ultra-high-frequency ECG (UHF-ECG) analysis provides information about electrical ventricular dyssynchrony. Additionally, real-time UHF-ECG analysis enables direct optimization of the pacing electrode during pacemaker implantation. In this master thesis, we describe ventricular conduction abnormalities, the current method for UHF-ECG analysis and most importantly, we have developed several deep learning models to find out which steps of UHF-ECG analysis can be replaced by deep learning. Data used for the development and validation of the models come from 2 private hospitals (FNUSA-ICRC hospital, Brno, Czechia, and FNKV hospital Prague, Czechia) and from 3 publicly available datasets. First, we present two deep learning methods for QRS complex detection and QRS complex duration estimation in one inference step. We received an overall F1-score of 98.84 ± 0.51 \% for the detection task and a Mean Absolute Error (MAE) of 12.25 ± 2.16 ms for the QRS duration estimation task. This method enhances UHF-ECG analysis performance and therefore could significantly reduce measurement time. Furthermore, a regression model for pacing stimuli removal based on a conditional generative adversarial network was developed. The results were evaluated based on the correlation of 15 averaged high-frequency envelopes in the QRS complex region between the model output and the target signal. The results show a higher correlation on spontaneous than on paced data and a drop in correlation with the increasing frequency band. Last, two deep learning models with convolutional neural network (CNN) were created to estimate ventricular electrical dyssynchrony (VED). Specifically, one-dimensional (1D) and 2-dimensional (2D) CNN. The MAE between our solution and annotation is 12.61 ±18.95 ms and 12.27 ±17.73 ms for 1D and 2D CNN, respectively. MAE on spontaneous data is approximately 5 ms lower than on paced data for both models, indicating the need to remove the pacing stimuli. These deep learning models yield a reduction in the pre-processing pipeline while delivering output in a single inference step. For the QRS detection and QRS duration estimation model, the performance improvement over the current solution is evident and these steps of UHF-ECG analysis could be replaced by deep learning. However, for the removal of pacing stimuli and VED parameter estimation, it is required to improve the performance.
|