Michael Kessler

Advisors
Markus Wirth (M.Sc.),  Prof. Dr. Björn Eskofier

Duration
10 / 2021 – 03 / 2022

Abstract
Cardiovascular diseases (CVDs) are the leading cause of death worldwide, with a share of 3.9 million (48%) deaths each year in Europe. It is also estimated that it costs the European economy €210 billion every year [1].

ECG is the most widely used non-invasive method for monitoring and diagnosing CVDs. Traditionally it is carried out in a stationary setting including constant supervision to enable optimized conditions. Innovation driven improvements facilitate ambulant recording of heart signals using wearable ECGs allowing for improved diagnostic. Furthermore, this method also finds application in other domains like sports to optimize training routines. However, ECG is susceptible to different types of noises, which might distort the morphological features and the interval aspects of the ECG leading to a false conclusion and diagnosis. Taking the recording procedure out of a stationary setting increases the potential exposure to noise, especially for use cases including a high dynamic.

Noise can originate from a variety of sources affecting the recording in different ways. Typical examples are muscle activity and power line interference causing abrupt alternations of the signal. The essence of considering the resulting artifacts lies in their negative influence for diagnostic evaluability and quality [1]. In literature two distinct approaches can be applied to handle artifacts: denoising and quality assessment of the recorded ECG data. Over time research enabled to develop powerful algorithms to detect and filter a variety of noise types [2]. Nevertheless, an ambulatory setting amplifies these noise sources and artifacts leading to false results or warnings in an unsupervised evaluation of the ECG data. Consequently, it is necessary to estimate the quality of the ECG [2]. The most common way to estimate the quality of a signal is the signal-to-noise ratio. This method needs a clean signal and pure noise; hence it is not suited for real world data. There are several studies evaluating methods for estimating the signal quality using adaptive filter, averaging, kurtosis, power spectrum analysis and more [2] [3] [4]. The problem with some of these approaches is the requirement of prior knowledge about the signal, such as the position of the R-peak [3]. This can be error prone in case the signal is very noisy.

Therefore, this work focuses on a combination of quality features/signal quality indices (SQIs) [5] that do not require former knowledge about certain features of the ECG signal like the position of the R-peaks. The pipeline presented in the work from Clifford et al. [5] is used and adapted for the challenges of a dynamic/sport-based ECG. For training of the algorithm an artificial dynamic dataset will be created, based on a 2-lead ECG dataset that is recorded by the Teiimo sensor shirt during a resting phase. For creating the training dataset different noise types (i.e. baseline wander, muscle artefact, power line interference and EMG noise), which are common in a dynamic/sport-based ECG recording, will be added. The signal will then be clustered binary (good and bad signal quality) using the signal-to-noise ratio. Based on this labelled dataset, R-peak independent classification approaches using feature- and deep learning-based algorithms are compared to common quality measures. In a final step a wearable ECG based dataset recorded in a dynamic sports scenario is used to do a qualitative performance analysis of the developed approach.

References:
[1] E. Wilkins, L. Wilson, K. Wickramasinghe, P. Bhatnagar, M. Rayner und N. Towsend, „European Cardiovascular Disease Statistics 2017,“ European Heart Network, 2017.
[2] J. Moeyersons, E. Smets, J. Morales, A. Villa, W. D. Raedt, D. Testelmans, B. Buyse, C. V. Hoof, R. Willems, S. V. Huffel und C. Varona, „Artefact detection and quality assessment of ambulatory ECG signals,“ Computer Methods and Programs in Biomedicine, 2019.
[3] A. S. d. Río und L. I. Romero, „Assesment of Different Methods to Estimate Electrocardiogram Signal Quality,“ Computing in Cardiology, 2011.
[4] Z. Zhao und Y. Zhang, „SQI Quality Evaluation Mechanism of Single-Lead ECG Signal Based on Simple Heuristic Fusion and Fuzyy Comprehensive Evaluation,“ Frontiers in Physiology, 2018.
[5] G. Clifford, J. Behar, Q. Li und I. Rezek, „Signal quality indices and data fusion for determining clinical acceptability of electrocardiograms,“ Phyiological Measurement, 2012.
[6] G. Clifford, A. Francisco und M. Patrick, „ECG statistics, noise, artifacts, and missing data,“ Advanced methods and tools for ECG data analysis, 2006.