Instrumentation & Measurement Magazine 24-2 - 40

These studies highlight the adverse effects of contaminants in
ECG, across several applications. Biosignal quality analysis
can improve the performance of the systems in these studies by detecting, identifying, quantifying, and removing the
contaminants.

Biosignal Quality Analysis
Biosignal quality analysis is the formal and objective analysis of the quality of a biosignal. Fig. 4 illustrates how biosignal
quality analysis can fit within an ECG system. Biosignal quality analysis can provide a qualitative assessment of the quality
of the acquired signals and provide feedback to the user when
the quality of the acquired signal deteriorates. It can also be
used to reject ECG segments with inadequate signal quality.
Biosignal quality analysis can be split into four categories
[12]: detection; identification; quantification; and mitigation.
Detection is a binary classification of signals as contaminated

or uncontaminated. Identification is the recognition of the type
or source of the contaminant. Quantification is the estimation
of the amount of contaminant in the signal. Quantification can
be performed without identifying the contaminant and can be
reduced to detection by applying a threshold to classify the
signal as contaminated or uncontaminated. Mitigation is the
reduction or removal of the contaminant from the signal. Mitigation techniques have been extensively studied for decades
[19]. Arguably, mitigation techniques should be applied after
detection, identification, and/or quantification. When applied
unnecessarily to uncontaminated ECG, mitigation techniques
can deteriorate the ECG's quality. Also, with continuous ECG
monitoring, it is often acceptable to forgo advanced mitigation
attempts and instead simply reject ECG segments that do not
have sufficient signal quality. As such, this paper will only discuss mitigation techniques briefly.
Techniques developed for detection, identification, and
quantification are broadly comprised of two main steps: feature extraction and classification/regression [20]. The feature
extraction step isolates important information in the ECG that
can aid in the detection, identification, or quantification of a
containment. The features can be fiducial-based or non-fiducial-based. The regression/classification step synthesizes the
information to provide a quantitative assessment of the quality of the ECG. This step can be heuristic-based or machine
learning (ML)-based [20]. Techniques that employ deep learning methods may integrate these two steps.

Detection

Fig. 4. ECG acquisition and analysis system with a biosignal quality analyzer
integrated.
40

The following two techniques are presented as examples of
fiducial-based features used with heuristic rules and non-fiducial-based features used with deep learning.
Orphanidou et al. [21] proposed a set of fiducial-based
features, coupled with three heuristic-based rules, to detect
contaminated ECG segments. For the first rule, the peaks of
the R-waves were identified to determine the time intervals
between the R peaks (R-R intervals) and heart rate. If the estimated heart rate was not between 40 and 180 beats per minute
(bpm), the normal range for heart rate, the ECG segment was
labeled as contaminated. For the second rule, if the R-R intervals were more than 3 seconds or the ratio of the maximum
R-R interval to the minimum R-R interval was more than 2.2,
then the ECG segment was labeled as contaminated. For the
third rule, QRS complexes were segmented, and a template of
the QRS was created by averaging the segmented complexes.
The average correlation coefficient calculated between the segmented QRS complexes, and the template was used as the final
feature. If the average correlation coefficient was below 0.66,
then the ECG segment was also labeled as contaminated. This
approach achieved a sensitivity of 94% and specificity of 97%
when tested on 500 10-second ECG segments from the Sana/
Physionet database and a proprietary database collected at the
John Radcliffe Hospital in Oxford, England.
We developed a deep learning approach to identify contaminated ECG to reduce false positives in the detection of
AF. Taji et al. [2], [22] proposed the detection of AF using deep

IEEE Instrumentation & Measurement Magazine

April 2021

Instrumentation & Measurement Magazine 24-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-2