IOP PUBLISHING PHYSIOLOGICAL MEASUREMENT

Physiol. Meas. 33 (2012) 14631477 doi:10.1088/0967-3334/33/9/1463

Threshold-based system for noise detection in

multilead ECG recordings

Irena Jekova 1,3 , Vessela Krasteva 1 , Ivaylo Christov 1

and Roger Abacherli 2
1 Institute of Biophysics and Biomedical Engineering, Bulgarian Academy of Sciences,

Acad G Bonchev Str. Bl 105, 1113 Sofia, Bulgaria

2 Biomedical Research and Signal Processing, Schiller AG, Altgasse 68, CH-6341 Baar,

Switzerland

E-mail: irena@biomed.bas.bg

Received 29 February 2012, accepted for publication 26 July 2012

Published 17 August 2012
Online at stacks.iop.org/PM/33/1463

Abstract
This paper presents a system for detection of the most common noise types seen
on the electrocardiogram (ECG) in order to evaluate whether an episode from
12-lead ECG is reliable for diagnosis. It implements criteria for estimation
of the noise corruption level in specific frequency bands, aiming to identify
the main sources of ECG quality disruption, such as missing signal or limited
dynamics of the QRS components above 4 Hz; presence of high amplitude and
steep artifacts seen above 1 Hz; baseline drift estimated at frequencies below
1 Hz; powerline interference in a band 2 Hz around its central frequency;
high-frequency and electromyographic noises above 20 Hz. All noise tests
are designed to process the ECG series in the time domain, including 13
adjustable thresholds for amplitude and slope criteria which are evaluated in
adjustable time intervals, as well as number of leads. The system allows flexible
extension toward application-specific requirements for the noise levels in
acceptable quality ECGs. Training of different thresholds settings to determine
different positive noise detection rates is performed with the annotated set
of 1000 ECGs from the PhysioNet database created for the Computing in
Cardiology Challenge 2011. Two implementations are highlighted on the
receiver operating characteristic (area 0.968) to fit to different applications.
The implementation with high sensitivity (Se = 98.7%, Sp = 80.9%) appears
as a reliable alarm when there are any incidental problems with the ECG
acquisition, while the implementation with high specificity (Sp = 97.8%,
Se = 81.8%) is less susceptible to transient problems but rather validates
noisy ECGs with acceptable quality during a small portion of the recording.

3 Author to whom any correspondence should be addressed.

0967-3334/12/091463+15$33.00 2012 Institute of Physics and Engineering in Medicine Printed in the UK & the USA 1463
1464 I Jekova et al

Keywords: ECG quality metrics, noise detection, disconnected electrode,

patient motion, poor skinelectrode contact, powerline interference
(Some figures may appear in colour only in the online journal)

Introduction

Nowadays, more than 20 000 000 electrocardiogram (ECG) recordings per year are stored
in electronic format; however, about 5% are unusable due to insufficient quality (Maan et al
2011). The growth of mobile phone networks in recent years provides a basis for rapid progress
in distant ECG registration and interpretation. However, when the ECG is acquired by a
non-trained person and the conditions are not rigorously controlled, ECG quality is highly
susceptible to external noisy components and other distorting factors which may impede
the reliable manual or automated measurements, or hazard the correct diagnosis. This may
limit the application of mobile health care systems in providing access for underserved rural
populations to a reliable expert opinion available in urban hospitals. Automatic management
of a large amount of ECGs by analytical quality metrics is shown to improve the quality of
ECG annotations reducing human review and costs (Allen and Murray 1996, Vaglio et al
2010).
The potential use of portable ECG devices by persons with different amounts of training,
as well as the collection of large volumes of data for which human over-reading is unfeasible or
limited to control samples, raises the question about the need of automatic systems for signal
quality control. The participants in the Computing in Cardiology (CinC) Challenge 2011
(PhysioNet/Computing in Cardiology Challenge 2011b) have contributed to the development
of different solutions for identifying common distorting factors during ECG acquisition (e.g.
not connected electrodes, poor skinelectrode contact, external electrical interference, artifact
resulting from patient motion, etc) aiming to provide feedback about the ECG quality within
a few seconds. An accuracy range between 80% and 93% (Silva et al 2011) has been achieved
when applying some of the following ECG disturbance detection techniques.
(1) Missing lead by identification of the percentage of signal appearing as a flat or straight
line (Clifford et al 2011, Ho Chee Tat et al 2011, Liu et al 2011, Xia et al 2011), constant
voltage in 1 s (Langley et al 2011), constant section in two or more leads for at least
200 ms (Moody 2011), range (minmax) below a global disconnection threshold or signal
amplitude above low limit for at least certain percent of the total lead duration in at least
a specific number of leads (Noponen et al 2011), zero line detection for 80% of the lead
or >40% portion of samples above 2 mV (Hayn et al 2011), a too large or too small
signal (Xia et al 2011), zero sum (Chudacek et al 2011) or constant derivative of zero for
the whole lead (Johannesen 2011), low amplitude or lead saturation by assessment of the
QRS amplitude dynamics (Jekova et al 2011).
(2) Patient movements seen as baseline wander by assessment of the isoline drift (Chudacek
et al 2011), low-pass signal below 1 Hz (Jekova et al 2011, Langley et al 2011) or 3 Hz
(Starc 2011), spectral components below 0.5 Hz (Zaunseder et al 2011).
(3) Electromyogram artifacts by assessment of the root-mean-square of the high-pass filtered
signal above 20 Hz (Jekova et al 2011) and 50 Hz (Johannesen 2011), the maximal power
of 45250 Hz spectral components (Zaunseder et al 2011), the relative impact of 520 Hz
components in the power spectrum (Clifford et al 2011), the presence of strong Gaussian
noise in the power spectrum above 40 Hz (Liu et al 2011).
Threshold-based system for noise detection in multilead ECG recordings 1465

(4) Strong powerline interference in the frequency bands 4852 Hz and 5862 Hz (Jekova
et al 2011).
(5) Large and incidental ECG interruptions by identification of spikes (Hayn et al 2011), high
amplitude peaks (Chudacek et al 2011), huge impulses (Liu et al 2011) and peak artifacts
(Jekova et al 2011), steep slopes (Langley et al 2011), sudden bursts of high strength in
the amplitude (Ho Chee Tat et al 2011), frequency of the large amplitude changes (Moody
2011).
(6) Compromised quality of the QRS detection (Clifford et al 2011, Hayn et al 2011,
Johannesen 2011).
(7) ECG quality disturbances found by more complicated tests, such as ECG reconstruction
using QRS templates (Starc 2011); conversion of ECG into vectorcardiogram (VCG) with
backward ECG reconstruction and assessment of the correlation between the original and
the reconstructed ECG (Maan et al 2011); prediction of one ECG lead by the other leads
(Noponen et al2011); cross-correlation between leads and/or lead segments (Kalkstein
et al 2011, Xia et al 2011).
The classification techniques which have been applied for taking the decision about the
diagnostically acceptable ECG quality combine the measured ECG parameters in empirically
adjusted decision trees (Chudacek et al 2011, Hayn et al 2011, Jekova et al 2011, Johannesen
2011, Langley et al 2011, Liu et al 2011, Moody 2011, Noponen et al 2011, Starc 2011, Xia
et al 2011, Zaunseder et al 2011) or feed them in a more sophisticated classifiers, such as a
quasi-linear combination between Kth nearest neighbor (KNN) and an ensemble of decision
trees (Kalkstein et al 2011), combination of simple rules with support vector machine (SVM)
(Ho Chee Tat et al 2011, Kuzilek et al 2011), linear discriminant analysis (LDA), Nave Bayes,
SVM and multi-layer perceptron (MLP) artificial neural network (Clifford et al 2011).
The aim of this paper is to present a system for recognition of diagnostically useful ECG,
based on adjustable settings for detection of the common corruption factors. The focus is set on
description of specific quality metrics and their flexible extension toward application-specific
requirements for the noise levels in acceptable quality ECGs. This option is shown by training
of different thresholds settings, aiming to achieve different positive noise detection rates.
The performance of two different implementations with high sensitivity or high specificity for
noise detection is discussed.

Material and methods

ECG database
The study uses the ECG database collected for the CinC 2011 (PhysioNet/Computing in
Cardiology Challenge 2011a). It includes 10 s recordings of standard 12-lead ECGs, sampled
at 500 Hz, 5 V resolution with bandwidth 0.05100 Hz. The data have been recorded by
nurses, technicians and volunteers with varying amounts of training and comprise signals
with common problems, such as poor skinelectrode contact, not connected electrode, power
line interference (PLI), artifacts resulting from patient motion, etc. The ECG annotation is
performed manually by a group of 23 volunteer annotators2 cardiologists, 1 physician, 5
ECG analysts, 5 others with some experience in ECG reading, and 10 volunteers who had never
read ECGs previously. The quality labeling of the ECG recordings is based on the individual
assessment of 3 to 18 of the annotators working independently. Each ECG is given a score
of 0.95 (excellent), 0.85 (good), 0.75 (adequate), 0.6 (poor) or 0 (unacceptable). The average
score for the ECG recording is considered and if it is 0.7 or more, and at most one annotator
graded the ECG as being of unacceptable quality, the ECG signal is labeled as acceptable. If
1466 I Jekova et al

Noise Test Parameters

1-st
derivative
(1) FLAT TIFLAT
filter detection

4Hz HP (2) LA THRLA

filter detection

TIPA(SLOPE)
1Hz HP (3) PA THRPA(AMP)
filter detection
TIPA(AMP)
ECG leads NbLeadsPA
(I,II,II,V1-V6)

Noise
(10 seconds
episode)
detection
THRBLW
1Hz LP (4) BLW TIBLW
filter detection
NbLeadsBLW

20Hz HP (5) HF THRHF(WorstQ)

filter detection THRHF(BestQ)

50Hz BP
filter
(6) PLI THRPLI(RMS)
detection THRPLI(SNR)
60Hz BP
filter

Figure 1. Flow-chart of the algorithm for noise detection based on six independent tests for the
most common noise sources in ECG. Each test generates a noise alert based on analysis of specific
parameters. At least one noise alert triggers positive noise detection. The sum of all alerts could be
used for scoring the noise corruption level between 0 (no noise) and 6 (all noise tests with positive
detection).

the average grade is less than 0.7, the ECG is labeled as unacceptable. Signals not falling into
either category are labeled as indeterminate (Silva et al 2011). The diversity in the annotators
training, as well as the used grading scale predispose to higher probability for acceptable
rhythm annotation. We used the publicly available part of the database, including a total number
of 1500 files:
Training data (set A), which contains 1000 ECG recordings with non-blinded annotations,
including 773 acceptable, 225 unacceptable and 2 indeterminate ECGs.
Test data (set B), which contains 500 ECGs with blinded annotations.

Threshold-based detection of most common noise types

The system for recognition of diagnostically useful 12-lead ECGs is based on the methods
briefly described in Jekova et al (2011) for the CinC Challenge 2011. It implements six parallel
branches which analyze ECG amplitudes and slopes in different frequency bands, aiming to
assess the corruption level of specific noises. Each noise test is performed independently and
the sum of all noise alerts could be used for scoring the noise corruption level between 0 (no
noise) and 6 (all noise tests with positive detection). As shown in the flow chart of figure 1,
13 parameters are defined to support the decision for noise presence. Different settings of the
parameters thresholds provide a flexible adaptation of the performance for each individual
Threshold-based system for noise detection in multilead ECG recordings 1467

70 10
68 Se Se
8
66
64 6

(%)
(%)

62
4 100-Sp
60

58 2
2 100-Sp
0
0
1 2 3 4 5 6 7 8 9 10 100 150 200 250 300
TI (s) (b) THRLA (uV)
(a) flat

Figure 2. Performance (Se, 100-Sp) as a function of the parameters values: (a) TIflat for FLAT
signal detection (test 1); (b) THRLA for LA detection (test 2).

noise test. Using the training data (set A), the parameters values are independently evaluated
to scan their influence on the true positive rate (Se) and the false positive rate (100-Sp), where:
Sensitivity: Se = 100 unacceptablecorrect /unacceptabletotal , (%)
Specificity: Sp = 100 acceptablecorrect /acceptabletotal , (%).
The training process is performed with nine ECG leads (I, II, III and V1V6) which
are scanned by the designed system over the full-length signal segment of interest (10 s).
Detailed description of all noise tests and the training of the relevant parameters are presented
below. The maximal sensitivity reported for each parameter is restricted by the presence of
the respective noise in the unacceptable set.
(1) Detection of flat signals (FLAT) by scanning the first derivative of the ECG lead for
zero-line segments over a moving time interval TIflat. We observe that varying the length
of TIflat from 1 to 10 s leads to gradual Se decrease by 12% (69% down to 57%) when
the false positive rate (100-Sp) keeps almost unaffected with less than 1% change (1.7%
down to 0.8%)figure 2(a).
(2) Detection of a low amplitude (LA) lead by scanning the peak-to-peak amplitude dynamics
within the QRS boundaries after 4 Hz high-pass filtering. The median amplitude range for
all QRS complexes is compared to a predefined adaptive threshold THRLA for detection
of a low amplitude ECG lead. As shown in figure 2(b), an increase of the THRLA setting
between 100 and 300 V results in a proportional increase of both Se (49.3%) and
100-Sp (17%) by about 6%.
(3) Detection of peak artifacts (PA) by scanning the output of 1 Hz high-pass filter for specific
slope and amplitude criteria over an adaptable time intervals, as follows:
Slope criterion: the first derivative of the signal is scanned to detect PA exceeding
a slope threshold of 500 V ms1, which is adjusted above the maximal slope seen
in pediatric QRS (about 400 V ms1). The detected PA is further validated by
analysis of all artifact-to-artifact intervals. Test for periodicity rejects the cases with
PA induced by a pacemaker. A second test validates only PA distanced closer than
a time interval for slope validation TIPA(SLOPE). The value of TIPA(SLOPE) is adaptable
depending on the shortest ECG segment without PA that is needed for analysis. As
shown in figure 3(a), the longer the scan interval TIPA(SLOPE) is, the higher Se for PA
detection is achieved. The maximal scan duration of 10 s provides maximal difference
between true and false PA detection rates: Se = 25%, 100-Sp = 13%.
1468 I Jekova et al

25 25
Se Se
20 20

15 15

(%)

(%)
10 10

5
100-Sp
5

0 0
100-Sp
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9
TIPA(SLOPE) (s) NbLeadsPA
(a) Slope criterion
35 25 25
30 Se Se
20 20
25
Se
20 15 15

(%)
(%)

(%)
15 10 10
10
5 5
5
100-Sp 100-Sp 100-Sp
0 0 0
0 10 20 30 40 50 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9

THRPA(AMP) TIPA(AMP) (s) NbLeadsPA

(ratio to median)

(b) Amplitude criterion

Figure 3. Performance (Se, 100-Sp) of PA detection (test 3) as a function of the parameters

values: (a) TIPA(SLOPE) and NbLeadsPA for the slope criterion; (b) THRPA(AMP), TIPA(AMP) and
NbLeadsPA for the amplitude criterion.

Amplitude criterion: the peak-to-peak ECG amplitudes are evaluated in non-

overlapping time intervals TIPA(AMP). PA is detected when there is at least one interval
TIPA(AMP) with peak-to-peak ECG amplitude exceeding a predefined amplitude
threshold THRPA(AMP). The amplitude threshold of the peak artifact THRPA(AMP) is
not defined in absolute units (V) but rather is proportional to the median value of the
peak-to-peak ECG amplitudes observed in all scan intervals TIPA(AMP). Figure 3(b)
illustrates that settings of lower amplitude thresholds THRPA(AMP), as well as shorter
scan intervals TIPA(AMP) result in increased PA detection rates. Maximal difference
between true and false PA detection rates is achieved at Se about 20%, 100-Sp
about 9%.
Positive PA detection is considered to influence the correct ECG analysis if PA is validated
either by slope or by amplitude criterion and if PA is seen in more than a predefined
number of leads (NbLeadsPA). Setting of bigger NbLeadsPA decreases the sensitivity for
PA detection, as illustrated in figures 3((a) and (b), right graphs), where the detection rates
of both slope and amplitude criteria fall from 20% to 25% for the decision over one lead
down to 1% for the decision validated over all nine leads.
(4) Detection of baseline wander (BLW) by scanning the output of 1 Hz low-pass filter. The
mean value of the rectified ECG signal is calculated in moving time intervals of 3 s and is
compared to a predefined amplitude threshold THRBLW. The presence of BLW is validated
if periodicity due to wide QRS complexes is not discovered and BLW is detected during
more than a predefined period (TIBLW) in a predefined number of leads (NbLeadsBLW). As
shown in figure 4, variation of the amplitude threshold THRBLW in a wide range between
250 V and 2 mV provides sensitivity between 27% down to 15%, keeping the false
positive rate between 12% and 5%. The effect of increasing the BLW detection period
TIBLW from 3 to 10 s is associated with a decrease of both true and false positive rates
by about 15% (Se range is 2710%, 100-Sp range is 152%). Validation of BLW in a
Threshold-based system for noise detection in multilead ECG recordings 1469

30 30 30
Se Se
25 25 25

20 20 20
Se

(%)

(%)
(%)

15 15 15

10 10 10
100-Sp
5 100-Sp 5 5
100-Sp
0 0 0
500 1000 1500 2000 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9
THRBLW (uV) TIBLW (s) NbLeadsBLW

Figure 4. Performance (Se, 100-Sp) of BLW detection (test 4) as a function of the parameters
values: HRBLW, TIBLW and NbLeadsBLW.

35 12

30 10
25 8
20
(%)
(%)

6
15
Se
Se 100-Sp
4
10
5
100-Sp 2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
(a) THRHF(WorstQ) (dB) (b) THRHF(BestQ) (dB)

Figure 5. Performance (Se, 100-Sp) of HF detection (test 5) as a function of the parameters values:
THRHF(WorstQ) and THRHF(BestQ).

bigger number of leads NbLeadsBLW is normally linked with a decrease of the detection
rate from 27% for detection over one lead down to 3% for BLW found in all nine leads.
(5) Detection of electromyographic and other high-frequency (HF) noises by scanning
the output of 20 Hz high-pass filter. The HF content is evaluated by the signal-to-
noise ratio (SNRHF) within moving time intervals of 3 s, according to the equation
SNRHF = max(RMSHF )/min(RMSHF ), where maximal and minimal root-mean-square
amplitudes RMSHF are extracted by analyzing all non-overlapping 200 ms intervals within
the 3 s segment, considering in rough estimation that max(RMSHF) represents the HF
signal content during QRS, while min(RMSHF) represents the HF content in zero-lines.
Out of all scanned 3 s moving intervals, the maximal and the minimal SNRHF are extracted
for assessment of the best and the worst HF signal quality, respectively. HF noise is detected
if min(SNRHF)<THRHF(WorstQ) or max(SNRHF)<THRHF(BestQ), where THRHF(WorstQ) and
THRHF(BestQ) are adjustable thresholds for setting the best and the worst signal quality.
The decision is taken for the lead with the minimal signal-to-noise ratios. As shown in
figure 5, higher settings of both thresholds THRHF(BestQ) and THRHF(WorstQ) are associated
with higher detection rates; however, settings below 0.3 dB seem as a reasonable
compromise for maximal difference between true and false positive rates.
(6) Detection of PLI by scanning the outputs of two band-pass filters BP50 (4852 Hz) and
BP60 (5862 Hz). Two measures of PLI noise levels are calculated:
RMSBP50, RMSBP60: root-mean-square noise measured outside the QRS boundaries
in the outputs of the two band-pass filters;
1470 I Jekova et al

16 25
14
20
12
10 15
Se

(%)
(%)

8
6 10
4
Se
100-Sp 5
2
100-Sp
0 0
0 5 10 15 20 25 0 50 100 150 200
THRPLI(SNR) (dB) THRPLI(RMS) (uV)

Figure 6. Performance (Se, 100-Sp) of PLI detection (test 6) as a function of the parameters
values: THRPLI(SNR) and THRPLI(RMS).

SNRBP50, SNRBP60: signal-to-noise ratio between the QRS amplitude and the band-
pass filters outputs.

Out of all nine leads, the maximal RMSBP50, RMSBP60 and the minimal SNRBP50, SNRBP60
are taken. PLI is detected if one of the following conditions is fulfilled: RMSBP50 >
THRPLI(RMS) or RMSBP60 > THRPLI(RMS) or SNRBP50 < THRPLI(SNR) or SNRBP60 <
THRPLI(SNR). The training performance in figure 6 indicates the different effects of the
two thresholds settings on the positive detection rateshigher rates are achieved for
higher THRPLI(SNR) and lower THRPLI(RMS). Optimal thresholds ranges are observed for
THRPLI(SNR) between 8 and 17 dB and THRPLI(RMS) between 50 and 200 V, with false
positive detection rate <2.5%.

Results

Both training and testing are performed in the environment of Matlab 7.0 (MathWorks Inc.). All
noise tests are run together and at least one alert is considered as positive noise detection. Using
the training set A and different settings for the defined 13 parameters thresholds, the receiver
operating characteristic (ROC) of the designed noise detection system is built (figure 7). The
area under the ROC curve is estimated to be 0.968. Two points over the ROC are highlighted
the first one is corresponding to the implementation with the highest sensitivity (Se = 98.7%,
Sp = 80.9%), while the second one is providing high specificity at the inflection point of
the steepest ROC slope (Sp = 97.8%, Se = 81.8%). These two implementations illustrate
the opportunity to adapt the noise detection performance in respect to specific application
demands. The concept of the implementation with high Se is to detect incidental or transient
noises in at least one lead, so that ECG recordings are validated as acceptable only if they are
noise-free in their full length. In contrast, the implementation with high Sp is adjusted to detect
permanent noises in one or more leads, so that ECG recordings are alarmed as unacceptable
only if they do not contain any noise-free episode over their full length.
The implementation with high Sp is prospectively scored on the blinded annotations of
test set B by participation in Event 1 of the CinC Challenge 2011 (PhysioNet/Computing
in Cardiology Challenge 2011b). This score, shown in the first row, last column of
table 1, is calculated as the fraction of correctly classified acceptable and unacceptable
ECG recordings. It is compared to the scores given to the other participants in the CinC
Challenge 2011, Event 1.
Threshold-based system for noise detection in multilead ECG recordings 1471

100

Implementation with high Se

True positive rate (Se) (%)

80
Implementation with high Sp
60

40

20

0
0 20 40 60 80 100
False positive rate (100-Sp) (%)

Figure 7. The ROC curve of the designed noise detection system, scanned for different settings
of the 13 parameters thresholds. The red o marks highlight the two extreme implementations
of interestthe first is corresponding to the inflection point with the highest Se (Se = 98.7%,
Sp = 80.9%), the second is corresponding to the inflection point of the steepest slope with high Sp
(Sp = 97.8%, Se = 81.8%).

Different examples of 12-lead ECGs with various degradation of the channels due to
noise contamination of four noise types are illustratedPA in figure 8, BLW in figure 9, HF in
figure 10, PLI in figure 11. The examples with relatively weak noise influence have positive
detection only by the implementation with high Se, while apparently strong noises are detected
also by the implementation with high Sp.
Figure 12 presents a comparative analysis of the implementation with high Se versus
high Sp considering the positive detection rates of the specific noise tests (FLAT, LA, PA,
BLW, HF, PLI) evaluated over all 1500 files included in the PhysioNet/CinC Challenge (2011)
database. The bars corresponding to high Sp (black) are indicative for the number of cases
where strong noise influence has been detected, while the additional cases recognized by the
implementation with high Se (shaded bars) show the proportion of incidental noises in the
database. The frequency of appearance of the different noises could be scored.

Discussion

In this work, we have developed a system for detection of the most common noise types
seen during ECG recording in order to evaluate whether an episode from 12-lead ECG is
reliable for diagnosis. Different branches embed adequate criteria for estimation of the noise
corruption level in specific frequency bands (figure 1), identified by visual inspection of the
ECG distortions that are commonly observed in the unacceptable category of the training set
A database. In this respect, the noise tests are defined in the following frequency bands: LA
detection uses frequencies above 4 Hz to evaluate the QRS amplitude dynamics; PA detection
is applied on a wide frequency band above 1 Hz to search for steep artifacts from different
sources; BLW influence is estimated at frequencies below 1 Hz; PLI is detected in a band
2 Hz around its central frequency; HF content is evaluated above 20 Hz.
All noise tests are designed to process the ECG series in the time domain, including
adjustable thresholds for amplitude and slope criteria which are evaluated in adjustable time
intervals, as well as number of leads. The general concept for implementing both adjustable
1472 I Jekova et al

Table 1. Performance of the presented algorithm and all methods participating in the CinC
Challenge 2011, Event 1 in terms of Se, Sp and overall accuracy (Acc) for the training set A,
and challenge score on the test set B, as reported by the authors.
Training set A Test set B
Method Se (%) Sp (%) Acc (%) Score Event 1

Method Implementation with high Sp 80.9 97.8 94.2 0.908

presented in this (Jekova et al 2011)
paper Implementation with high Se 98.7 80.9 84.9
Clifford et al SVM applied on 72 features of 6 99.0 0.926
(2011) signal quality metrics
Ho Chee Tat Mimicking SVM classifier applied 0.920
et al (2011) on 4 features from QRS detector
Hayn et al Decision tree classifier with 4 basic 93.3 0.916
(2011) signal properties and 3 additional
quality tests
Kalkstein et al Combination of KNN and ensemble 74.1 98.5 93.0 0.912
(2011) of decision trees using 76 features of
global signal characteristics
Zaunseder et al Ensemble of decision trees applied 96.6 0.904
(2011) on ordinary spectral features
Xia et al (2011) Scoring system based on spectrum 93.5 0.900#
radius of 12 12 matrix of
regularity embedded in Android
platform
Noponen et al Scoring system based on robust best 80.4 96.9 93.2 0.900
(2011) subsets linear prediction error
Moody (2011) Three simple heuristic rules based 65.3 98.8 91.3 0.896
on ECG amplitude criteria
Johannesen Decision tree classifier using 3 91.0 85.0 86.4 0.880
(2011) signal quality metrics and quality of
QRS detection embedded in
Android
Langley et al Six signal quality tests with 91.4 0.857
(2011) heuristically defined thresholds
Xia et al (2011) Multistage scoring system based on 95.1 83.2 85.9 0.850#
ECG amplitude quality, self and
cross correlation between channels,
embedded in Android platform
Kuzilek et al Multistep classifier with SVM and 99.9 0.836
(2011) decision rules with feature vector
based on covariance matrixes
Chudacek et al Simple scoring system based on 5 90.3 0.828
(2011) decision rules embedded in Android
platform
Starc (2011) Simple scoring system based on 0.818
beats quality estimation by
amplitude thresholds
the accuracy is reported for own annotations, different from the publicly available annotations of the training

set A.
# the scores reported in the paper are different from the best score in Event 1 (0.932), given to Xiaopeng Zhao
(a co-author of Xia et al (2011)), as has been published in the official web page (PhysioNet/Computing in
Cardiology Challenge 2011b).
The values are not published in the papers.
Threshold-based system for noise detection in multilead ECG recordings 1473

(a) (b)

Figure 8. 12-lead ECGs corrupted by peak artifacts. (a) PA in V4 is recognized only by the
implementation with high Se. (b) PA with strong influence in all leads is also detected by the
implementation with high Sp.

(a) (b)

Figure 9. 12-lead ECGs with baseline wander. (a) Strong BLW is seen only in V1. It is recognized
only by the implementation with high Se. (b) Strong BLW is seen in all leads, so that it fulfils even
the restricted criteria of the implementation with high Sp.

thresholds and independent noise tests with individual parameters allows flexible adaptation
of the algorithm toward the required accuracy level of quality assessment, depending on the
particular application. For example some applications might have specific pre-processing for
automatic compensation of deficiencies, e.g. filter for PLI or other high-frequency noise large
enough to mask the signal, and thus being able to recover themselves an interpretable ECG
from an unacceptable recording. In this case some of the noise tests could be disabled or
the activation thresholds could be tuned toward lower sensitivity. Other applications, such
as portable computing devices or mobile phones, might be used by untrained personnel or
1474 I Jekova et al

(a) (b)

Figure 10. 12-lead ECGs with high-frequency noise. The recordings in (a) and (b) are seemly
from the same patient. (a) Strong but transient HF noise is appearing for 4 s in the limb leads. It
is recognized only by the implementation with high Se. (b) The permanent HF noise seen in the
limb leads is detected by both implementations with high Se and high Sp.

Figure 11. 12-lead ECG with 60 Hz PLI seen in V2. The PLI with varying amplitude is detected
by the implementation with high Se; however, the implementation with high Sp with restricted
thresholds does not alarm for PLI.
Threshold-based system for noise detection in multilead ECG recordings 1475

Number of ECGs with positive detection

260 248(17%) Implementation with high Se
237(16%)
240 Implementation with high Sp

(% from total of 1500 ECGs)

220
202(13%)
195(13%)
200
180
160
140 130(9%)
120
100
80 66(4%)
60 45(3%)
40 26(2%) 27(2%) 26(2%)
17(1%)
20
0
FLAT LA PA BLW HF PLI
Noise test

Figure 12. Number of ECGs with positive detection of specific noise test for both implementations
with high Se and high Sp. The ECGs are part from all 1500 files included in the database
(PhysioNet/Computing in Cardiology Challenge 2011a) database. Depending on the signal quality
in all 12-leads, one ECG recording could provoke positive detection by more than one noise test.

patients who could hardly provide conditions for acquisition of high-quality ECGs. In these
cases, activation thresholds set toward higher sensitivity are more appropriate, so that after
noise alarm the user is able to decide whether to keep a possibly faulty ECG or to record it again
according to a recommended correction of the recording conditions. Such a feasible feedback
is provided by the presented branch system with six different noise tests which indicate for
the following problems:
one or more electrodes not attachedalarm from FLAT test;
poor skinelectrode contactalarm from LA, PA, PLI or HF test;
patient motionalarm from BLW or PA test;
ambient interferencesalarm from PLI or HF test.
The present work shows the adaptable performance of the noise detection system based
on different settings of 13 parameters thresholds. Two specific points over the ROC curve
(figure 7) are highlighted to cover the above mentioned recording conditions that would require
either high sensitivity or high specificity for noise detection. The implementation with high
sensitivity (Se = 98.7%, Sp = 80.9%) is a reliable alarm when there are any incidental
problems with the ECG acquisition, while the implementation with high specificity (Sp =
97.8%, Se = 81.8%) is less susceptible to transient problems but rather validates noisy ECGs
with acceptable quality during a small portion of the recording. The examples in figures 811
show different behavior of the noise, which activates the alarm of the two implementations.
The noise in all examples is capable to activate the implementation with high Se, no matter how
incidental it is. Considering, however, only the ECGs in (a), the portion of noise-free signal in
all leads (>3 s) is enough to validate the recording as acceptable by the implementation with
high Sp. The noise influence on the recorded 1500 ECGs in the PhysioNet/CinC Challenge
(2011) database could be roughly estimated by comparing the positive noise detection rate
by the two implementations (figure 12). The most common problem (17% of the ECGs)
seems to be not connected electrodes with FLAT alarm activated by flat signal during part
or full-length recording in 4% or 13%, respectively. The next frequently seen ECG distortion
is due to PA (16% of the ECGs), which appear incidentally in 14% and permanently in 2%.
1476 I Jekova et al

Another problem is BLW (13% of the ECGs), which disturbs partially one lead in 10% and
permanently all leads in 3%. The PLI appears in 9% of the ECGs with high amplitude seen in
2%. Less frequent are the disturbances from HF noises (4% of the ECGs) which are incidental
in 3% and permanent in 1%. Only 2% of the ECGs are recognized by LA to have insufficient
QRS amplitude dynamics in at least one lead.
Seeing the large probability that ECG recordings could be corrupted during signal
acquisition, many solutions to prevent against misinterpretation by warning the noise
corruption level have been scored in the CinC Challenge 2011 (table 1). Considering
the different proportion of acceptable (773/998) and unacceptable (225/998) records in
the training set A, we participated in the challenge with the high Sp implementation,
since it provides better total accuracy 94.2% (940/998) than the accuracy of the high Se
implementation84.9% (847/998). The testing on the blinded set B yields to a score range
for all challenge participants from 0.8 to 0.93 (Silva et al 2011) where the presented algorithm is
rated at the seventh place in Event 1 with score of 0.908. Comparing to better scored solutions,
our threshold-based system could be considered as a good tradeoff between accuracy, flexible
extension of different noise tests and computational simplicity that is important for small
devices with real-time feedback. The solution of Hayn et al (2011) seems also to be resource
efficient with estimation of basic signal quality properties and smart assessment of the second
worst channel that probably have played a role for better scoring. The top rated algorithm (Xia
et al 2011) relies on relatively heavy computations, such as frequency and time-frequency
domain analyses, self and cross-correlation, entropy analysis. The other better rated solutions
use an extended set of features and/or complicated classifiers, such as: 72 features subjected to
LDA, Nave Bayes, SVM and MLP artificial neural network (Clifford et al 2011); hybridized
rule-based engine mimicking SVM (Ho Chee Tat et al 2011); 76 features fed to quasi-linear
combination between KNN and an ensemble of decision trees (Kalkstein et al 2011). From the
description of all methods in table 1, it is obvious that different approaches for noise detection
could be adopted with comparable confidence and it is up to the developers choice to fit to
specific system resources and adaptable performance.

Conclusions

The control of the conditions for multi-channel ECG acquisition should be established as an
essential practice for a reliable recording of diagnostically useful ECGs by portable systems. As
the unacceptable noise corruption level is application-specific, the presented system embedding
independent noise tests with adjustable amplitude thresholds, time intervals and number of
leads can provide different levels for quality control. Two levels, one corresponding to high
sensitivity and another to high specificity, are highlighted on the ROC curve to fit to different
applications. The implementation with high sensitivity (Se = 98.7%, Sp = 80.9%) provides an
alarm for incidental problems during the ECG acquisition, while the implementation with high
specificity (Sp = 97.8%, Se = 81.8%) validates the ECGs as acceptable when there is a small
portion of noise-free ECG suitable for interpretation. The setting could be also dependent on
the user training level or on the ECG system complexity level for automatic compensation of
deficiencies.

References

Allen J and Murray A 1996 Assessing ECG signal quality on a coronary care unit Physiol. Meas. 17 24958
Chudacek V, Zach L, Kuzilek J, Spilka J and Lhotska L 2011 Simple scoring system for ECG quality assessment on
android platform Comput. Cardiol. 38 44951
Threshold-based system for noise detection in multilead ECG recordings 1477

Clifford G, Lopez D, Li Q and Rezek I 2011 Signal quality indices and data fusion for determining acceptability of
electrocardiograms collected in noisy ambulatory environments Comput. Cardiol. 38 2858
Hayn D, Jammerbund B and Schreier G 2011 ECG quality assessment for patient empowerment in mHealth
applications Comput. Cardiol. 38 3536
Ho Chee Tat T, Xiang C and Thiam L 2011 Physionet Challenge 2011: improving the quality of electrocardiography
data collected using real time QRS-complex and T-wave detection Comput. Cardiol. 38 4414
Jekova I, Krasteva V, Dotsinsky I, Christov I and Abacherli R 2011 Recognition of diagnostically useful ECG
recordings: alert for corrupted or interchanged leads Comput. Cardiol. 38 42932
Johannesen L 2011 Assessment of ECG quality on an android platform Comput. Cardiol. 38 4336
Kalkstein N, Kinar Y, NaOaman M, Neumark N and Akiva P 2011 Using machine learning to detect problems in
ECG data collection Comput. Cardiol. 38 43740
Kuzilek J, Huptych M, Chudacek V, Spilka J and Lhotska L 2011 Data driven approach to ECG signal quality
assessment using multistep SVM classification Comput. Cardiol. 38 4535
Langley P, Di Marco L, King S, Duncan D, Di Maria C, Duan W, Bojarnejad M, Zheng D, Allen J and Murray A 2011
An algorithm for assessment of quality of ECGs acquired via mobile telephones Comput. Cardiol. 38 2814
Liu C, Li P, Zhao L, Liu F and Wang R 2011 Real-time signal quality assessment for ECGs collected using mobile
phones Comput. Cardiol. 38 35760
Maan A, van Zwet E, Man S, Oliveira-Martens S, Schalij M and Swenne C 2011 Assessment of signal quality and
electrode placement in ECGs using a reconstruction matrix Comput. Cardiol. 38 28992
Moody B 2011 Rule-based methods for ECG quality control Comput. Cardiol. 38 3613
Noponen K, Karsikas M, Tiinanen S, Kortelainen J, Huikuri H and Seppanen T 2011 Electrocardiogram quality
classification based on robust best subsets linear prediction error Comput. Cardiol. 38 3658
PhysioNet/Computing in Cardiology Challenge 2011a http://physionet.org/physiobank/database/challenge/2011/
PhysioNet/Computing in Cardiology Challenge 2011b Top scores: www.physionet.org/challenge/2011/top-scores.shtml
Silva I, Moody G and Celi L 2011 Improving the quality of ECGs collected using mobile phones: the
PhysioNet/Computing in Cardiology Challenge 2011 Comput. Cardiol. 38 2736
Starc V 2011 Could determination of equivalent dipoles from 12 Lead ECG help in detection of misplaced electrodes
Comput. Cardiol. 38 4458
Vaglio M, Isola L, Gates G and Badilini F 2010 Use of ECG quality metrics in clinical trials Comput. Cardiol.
37 5058
Xia H, Garcia G, McBride J, Sullivan A, De Bock T, Bains J, Wortham D and Zhao X 2011 Computer algorithms for
evaluating the quality of ECGs in real time Comput. Cardiol. 38 36972
Zaunseder S, Huhle R and Malberg H 2011 CinC Challengeassessing the usability of ECG by ensemble decision
trees Comput. Cardiol. 38 27780