Instrumentation & Measurement Magazine 24-2 - 41

belief networks (DBN). A DBN was created using stacked, restricted Boltzmann machines (RBM). An RBM is a two-layer
neural network with an input layer and a hidden layer. The approach was tested by synthetically contaminating ECG from
the MIT-BIH Atrial Fibrillation Database with a motion artifact
from the MIT-BIH Noise Stress Test Database. The MIT-BIH
Noise Stress Test Database was used to train an autoregressive
model to generate additional motion artifact to contaminate
the full 110 hours of ECG. The ECG was segmented into 129beat segments. The DBN classified nearly all the " clean " ECG
segments (no added motion artifact) as uncontaminated. This
included clean ECG segments with AF, which is crucial as
clinically relevant changes in the ECG should not be misinterpreted as contaminants. The percentage of ECG segments
classified as contaminated increased as SNR decreased, with
nearly all the ECG segments labeled as contaminated at an
SNR of -10 dB with precision and recall of 97.7% and 100%, respectively. AF detection accuracy with clean ECG was above
80% but dropped below 70% at an SNR of -20 dB. By detecting
and rejecting contaminated ECG segments, the AF detection
accuracy remained above 80%, even at an SNR of -20 dB.

Identification
The research in ECG biosignal quality analysis for the sole purpose of contaminant identification is exceedingly sparse. We
present two techniques that were developed by our research
group.
We proposed non-fiducial-based features with an ML-based
approach using random forests to identify contaminants in
ECG [23], [24]. ECG from the Normal Sinus Rhythm Database
were preprocessed using a high-pass filter to remove any existing baseline wandering and 60 Hz notch filter to remove power
line interference. To identify the contaminant type, we extracted
a set of non-fiducial-based features from 86,180 15-second ECG
segments. The features included statistical descriptions of the
raw ECG, wavelet detail and approximation coefficients, and
independent component analysis (ICA) components. The
statistical descriptions were mean, skewness, range, kurtosis, and power. The wavelet decomposition was performed
to the 7th level using the Daubechies3 mother wavelet. Since
ICA requires multiple channels, the 15-second segments were
further divided into 3-second subsegments that were aligned
and provided as an input to the FastICA algorithm to calculate four components. A total of 249 features were extracted,
which were reduced to 53 features using recursive feature elimination and cross-validated selection (RFECV) technique. The
features were fed to a 1024 random forest. Noise from the MITBIH Noise Stress Test Database was used to contaminate the
preprocessed ECG artificially. The ECG was corrupted using
baseline wandering, EMG, motion artifact, and their combinations at 0 dB, -5 dB, -10 dB, and -15 dB SNR levels. One-third of
the database was reserved for testing, while the rest was used
for training. The precision and recall were 81% and 70%, respectively, when any noise in a combination was identified.
We also used support vector machines (SVMs) to identify
contaminant types in surface EMG [12]; this SVM approach
April 2021

can be adapted to ECG. Data used in this research included
226 5-second EMG segments recorded from human subjects
during static contractions and 226 5-second simulated EMG
segments. Five types of contaminants were considered in this
research: ECG interference, motion artifact, power line interference, amplifier saturation, and additive white Gaussian
noise. At low SNR (< -10 dB), the average classification accuracy of the five contaminants was >95%. As the SNR increased,
the classification accuracy decreased with the performance
approaching random guessing for SNR = 20 dB, which was expected as the contaminants are negligible at this point and are
hard to distinguish from each other.

Quantification
Contaminants in the ECG can be quantified discretely and
qualitatively or quantified continuously and quantitatively.
A discrete scale categorizes ECG segments into a finite set of
quality bins, which may be dependent on the application. For
example, an ECG segment labeled as acceptable for heart rate
estimation, which requires the detection of the relatively large
QRS complexes in the ECG, may not be acceptable for myocardial ischemia detection, which requires the characterization
of the small changes in the ST-segment. A continuous scale
provides flexibility to appropriately map signal quality into
discrete categories based on the application. The following
two techniques demonstrate discrete and continuous scales for
the quantification of signal quality in ECG.
Clifford et al. [25] proposed 13 signal quality indices (SQIs)
to rate ECG on a scale from 0 (clean) to 4 (extreme noise): 1)
agreement between two different QRS detection algorithms;
2) skewness; 3) kurtosis; 4) power ratio of the range 5-20 Hz
to the range 5-40 Hz; 5) power ratio of the range 0-1 Hz to the
range 5-40 Hz; 6) average ratio peak to peak amplitudes of
each QRS and its respective baseline; 7) relative energy of QRS
complex; 8) standard deviation of the QRS complex; 9) the relative amplitude of the high-frequency noise; 10) signal purity;
11) sample entropy; 12) the square root of the instantaneous
power of the range 40-45 Hz after filtering using a 0.05-second
normalized Hamming window FIR filter; and 13) the periodic
component analysis (PiCA) periodicity measure of the ECG
waveform. The SQIs were fed to a support vector machine
(SVM). Testing showed that a combination of all the SQIs,
apart from the first, seventh, and eighth SQIs, provided the
optimal performance. This approach was tested on 2289 5-second ECG segments from the 2011 PhysioNet/Computing in
Cardiology (PCinC) Challenge dataset contaminated with the
different types of noise in the MIT-BIH Noise Stress Test Database and 33,979 5-second ECG segments from the MIT-BIH
arrhythmia database. The accuracy was 80.10% ± 0.54% using five-fold cross-validation on the simulated dataset (PCinC
Challenge dataset). The accuracy was 88.07% ± 0.32% on the
real dataset (Physionet MIT-BIH arrhythmia database).
We developed an SQI that rates the quality of a 30-second ECG segment on a continuous scale by estimating the
SNR [26]-[28]. SNR calculations require the power of the signal and the power of the noise. Individual PQRST waveforms

IEEE Instrumentation & Measurement Magazine 41

Instrumentation & Measurement Magazine 24-2

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