Automatic detection of ECG cable interchange by analyzing both morphology and interlead relations
-
Upload
independent -
Category
Documents
-
view
1 -
download
0
Transcript of Automatic detection of ECG cable interchange by analyzing both morphology and interlead relations
1
Automatic Detection of ECG Cable Interchange by Analyzing Both
Morphology and Interlead Relations
Chengzong Han, PhD, Richard E. Gregg, MS, Dirk Q. Feild, MA, Saeed Babaeizadeh, PhD
Advanced Algorithm Research Center, Philips Healthcare, Andover, MA, USA
Correspondence: Chengzong Han
Institution: Philips Healthcare
Address: 3000 Minuteman Rd
City: Andover
State: Massachusetts
Country: United States
Postal Code/Zip: 01810
Telephone: 1-978-659-2277
E-mail: [email protected]
2
Abstract
Background: ECG cable interchange can generate erroneous diagnoses. For algorithms detecting
ECG cable interchange, high specificity is required to maintain a low total false positive rate because
the prevalence of interchange is low. In this study, we propose and evaluate an improved algorithm
for automatic detection and classification of ECG cable interchange.
Method: The algorithm was developed by using both ECG morphology information and redundancy
information. ECG morphology features included QRS and P-wave amplitude, frontal axis and
clockwise vector loop rotation. The redundancy features were derived based on the EASITM
lead
system transformation. The classification was implemented using linear support vector machine. The
development database came from multiple sources including both normal subjects and cardiac
patients. An independent database was used to test the algorithm performance. Common cable
interchanges were simulated by swapping either limb cables or precordial cables.
Results: For the whole validation database, the overall sensitivity and specificity for detecting
precordial cable interchange were 56.5% and 99.9%, the sensitivity and specificity for detecting limb
cable interchange (excluding left arm-left leg interchange) were 93.8% and 99.9%. Defining
precordial cable interchange or limb cable interchange as a single positive event, the total false
positive rate was 0.7%. When the algorithm was designed for higher sensitivity, the sensitivity for
detecting precordial cable interchange increased to 74.6% and the total false positive rate increased to
2.7%, while the sensitivity for detecting limb cable interchange was maintained at 93.8%. The low
total false positive rate was maintained at 0.6% for the more abnormal subset of the validation
database including only hypertrophy and infarction patients.
3
Conclusion: The proposed algorithm can detect and classify ECG cable interchanges with high
specificity and low total false positive rate, at the cost of decreased sensitivity for certain precordial
cable interchanges. The algorithm could also be configured for higher sensitivity for different
applications where a lower specificity can be tolerated.
Keywords: Electrocardiograph; lead system; cable interchange; algorithm; classification
4
Introduction
Forty million ECGs are recorded annually in the United States. Electrode placement is
considered one of the most important factors that determine ECG signal quality and misplacement can
produce incorrect ECG abnormalities and thus generate erroneous diagnostic reports (1). ECG cable
interchange is one of the most common electrode placement errors. A recent statement from the
American Heart Association recommends including cable interchange detection algorithms in ECG
devices (2). The magnitude of the cable misconnection problem has been measured between 0.4% and
4% in various clinical settings (3,4,5,6).
Different criteria and algorithms have been proposed to detect cable interchange based on ECG
analysis. Common criteria identify morphology features including P-QRS-T amplitude, area, R-wave
progression, and flat line on certain leads to detect cable interchanges (7,8,9,10). Vectorcardiogram
(VCG) analysis has also been proposed for the detection of cable interchanges (11,12). In terms of
automatic detection algorithms, most methods fall into two categories. The first category involves
morphology-based methods which extract a set of measurements from P, QRS, and T-waves and use
these measurements to derive detection criteria (13,3,14). The second category of algorithms
implements redundancy-based methods which use the redundant information contained in the eight
independent leads. Such methods use the transformation or reconstruction of ECGs from the original
lead system to an approximation and cable interchange is detected by comparing the original ECGs
with transformed or reconstructed ECGs (4,15). Batchvarov et al (1) reviewed the different types of
cable interchanges as well as criteria and algorithms for detecting cable interchanges. Most recently,
Xia et al (16) proposed a method combining morphology features from (13,3) and redundancy
features from (4) for the detection of just limb cable interchange.
5
Different performance results were reported for the above-mentioned methods, depending on
the choice of database and algorithm configuration. From a clinical and practical point of view, a high
specificity algorithm for detecting cable interchange is preferred, in order to prevent generating many
false positive notifications, especially since the prevalence of cable interchange is low. To achieve this
goal, we propose a novel algorithm for automatic detection and classification of ECG cable
interchanges which uses both morphology and redundancy information. The algorithm performance
was evaluated using an independent population database for common limb cable interchanges and
precordial cable interchanges.
Methods
Database
The development database included two population sources. One source was body surface
potential maps recorded on a population including normal subjects, post-myocardial-infarction
patients, patients with a history of ventricular tachycardia but no evidence of a previous myocardial
infarction, and patients with a single-vessel coronary artery disease who underwent coronary
angioplasty (17). The second source was the Physionet PTB diagnostic ECG database (18,19) which
includes both healthy control subjects and patients with various cardiac diseases including myocardial
infarction, cardiomyopathy/heart failure, bundle branch block, dysrhythmia, myocardial hypertrophy,
valvular heart disease, myocarditis, and other miscellaneous conditions. Consecutive 10-second
snapshot ECGs were taken from each PTB recording. The validation database of 10-second records
came from the Common Standards for Electrocardiography (CSE) diagnostic database (20) which
includes healthy control subjects, ventricular hypertrophy patients, and myocardial infarction patients.
6
For both the development and validation databases, the ECGs with suspected cable interchange were
excluded. In total, there were 6272 ECGs from 1125 patients in the development database and 1166
ECGs from 1166 patients in the validation database that were considered free of any cable
interchange. To further evaluate the algorithm performance, the validation database was divided into
two subsets. Subset 1 included 527 ECGs that were confirmed by independent experts as hypertrophy
and/or infarction, while subset 2 included the remaining 639 ECGs most of which were considered
normal. The ECGs from both development database and validation database all used standard limb
electrode placement.
We investigated 9 common cable interchanges (3,4) involving 3 limb cables, right arm (RA), left
arm (LA), and left leg (LL); and 6 precordial cables, V1 to V6. For each ECG, cable interchanges
were simulated by swapping either limb cables (LA-RA and RA-LL) or precordial cables (V1-V2,
V1-V3, V2-V3, V3-V4, V4-V5, V4-V6, and V5-V6), which generated additional 9 ECGs from each
original ECG with no cable interchange. The LA-LL interchange was not considered because experts
felt that serial ECGs are required for reliable detection of the LA-LL interchange; in other words,
even an expert cannot dependably detect LA-LL interchange using only a single 10-second snapshot.
Algorithm Development
Our detection and classification algorithm uses both morphology features and redundancy
features. The Philips DXL ECG algorithm was used to automatically extract morphology features
including P-wave frontal axis, P-wave clockwise vector loop rotation direction, QRS-wave frontal
axis, QRS-wave clockwise vector loop rotation direction, R-wave amplitude and T-wave amplitude
from lead I and lead II. The P-wave features were only used when the ECG had consistent beat-to-
beat PR interval and did not show atrial fibrillation or atrial flutter. The redundancy features were
7
derived based on EASITM
transformation. In this context, redundancy means that many leads are
highly correlated. Each lead can be reconstructed from other leads with reasonable accuracy due to
this redundant information. For each ECG, after generating a separate ECG for all possible lead swaps
(including lead swaps for LA-RA, RA-LL, V1-V2, V1-V3, V2-V3, V3-V4, V4-V5, V4-V6, V5-V6,
and no lead swap), the inverse of EASITM
-lead conversion matrix (21,22) was used to transform each
swapped 12-lead ECG to the EASITM
-lead ECG and then the EASITM
-lead conversion matrix was
applied to convert back to a 12-lead ECG. The lead by lead root mean square errors (RMSE) between
swapped ECGs and the double EASITM
-transformed ECGs were calculated over the entire QRS-T
complex. The redundancy features for this study were the averaged RMSE among all 12 leads for
each lead swap.
To visualize the features (including both morphology features and redundancy features) used in
detection and classification, we used principal component analysis as the dimensionality reduction
technique to project the high-dimension feature space into a 2D feature space. The first two dominant
components with largest singular values were chosen as 2D features for visualization. After reducing
the dimensionality, we visualized the features for the development database on scatter plots. Figure
1.A shows the comparison of 2D features between LA-RA interchange and no cable interchange.
Figure 1.C shows the comparison of 2D features between V2-V3 interchange and no cable
interchange. For each comparison plot, the histograms of corresponding model output are shown in
Figure 1.B and Figure 1.D. The feature overlap between the V2-V3 interchange and no cable
interchange, shown in Figure.1.C, is larger than the feature overlap between LA-RA interchange and
no cable interchange, shown in Figure.1.A. This indicates that detecting V2-V3 interchange is more
challenging than detecting LA-RA interchange. A larger overlap in scatter plots was also observed
8
when comparing features for the V4-V5 or V5-V6 interchanges with no cable interchange. Some
cable interchanges are difficult to detect even for expert electrocardiographers. Figure 2 shows a 12-
lead ECG example in which V5 and V6 are swapped. The difference between V5 and V6 is so small
that neither the algorithm nor the expert could detect the cable interchange with certainty. For this
particular example, the diagnostic algorithm generated the same interpretation with and without the
V5-V6 interchange.
The automatic algorithm was implemented based on a multi-class linear support vector machine
(SVM) in the form of C-support vector classification (23). As part of the standard method, binary
SVM classifiers were designed for every combination of two output classes. We used the LIBSVM
toolbox (23) to perform the training of the SVM models. A higher weight was assigned to the
parameter C in the binary SVM classifier for the output class with no cable interchange compared to
the other output classes with cable interchanges. This unequal weighting favored specificity over
sensitivity, which means fewer false positives at the cost of potentially more false negatives. To train
the SVM models, 5-fold cross validation and grid search based on overall detection accuracy were
performed to choose the optimal parameter C. Figure 1.B and Figure 1.D show the decision boundary
(vertical line) and the histogram of SVM score generated by the SVM model for the development
database. Figure 1.B shows the output of binary classifier for LA-RA interchange and no cable
interchange. Figure 1.D shows the output of binary classifier for V2-V3 interchange and no cable
interchange. The decision boundary was designed to reduce the false positive cases through the
unequal weighting of the C parameter mentioned above. There are more false negative cases in Figure
1.D than false negative cases in Figure 1.C. In order to perform multi-class classification, the “one-
against-one” approach was used, and then the final output class was chosen by comparing outputs
9
from all binary classifiers using a voting strategy. The algorithm was configured in two ways. One
algorithm configuration aimed for low false positive rate (high specificity) and the other algorithm
configuration aimed for high sensitivity.
Statistical Analysis
Algorithm performance was evaluated in terms of sensitivity and specificity from a confusion
matrix. To calculate sensitivity and specificity for each cable interchange, that particular cable
interchange was defined as a positive event and all other cable interchanges as well as no cable
interchange were defined as a negative event thereby reducing the size of the confusion matrix to 2-
by-2. Considering that there were 9 types of individual cable interchanges, the algorithm performance
was also evaluated in terms of total false positive rate, which represents the ratio of cases that were
erroneously detected as any cable interchange within the group with no cable interchange. We
calculated the statistical significance based on the 95% confidence interval (CI) of the difference in
the performance ratios (sensitivity and specificity) as recommended by Altman et al (24).
Results
Table 1 and Table 2 summarize the performance of detecting each individual cable interchange
for two configurations of the algorithm on the whole validation database. For the low false positive
rate configuration, the total false positive rate was 0.7%. The sensitivity for interchanges of LA-RA,
RA-LL, V1-V2, V1-V3, V3-V4 and V4-V6 was much higher than the sensitivity for interchanges of
V2-V3, V4-V5, and V5-V6. For the high sensitivity algorithm configuration, the sensitivity was
significantly improved, especially for precordial cable interchanges V2-V3, V4-V5, and V5-V6.
However, the total false positive rate also increased. The change in specificity for each individual
10
cable interchange was small between the two algorithm configurations.
The confusion matrix of the high sensitivity algorithm, shown in Figure 3.A, indicates that a
large number of the false positive cases (2.7% of all no cable interchange cases) came from the
erroneous detection of precordial cable interchange. In order to reduce this portion of false positive
cases (from 2.7% to 0.6%), as shown in Figure 3.B, the algorithm was configured to tradeoff true
cases of precordial cable interchange for false positive detections, which explains the decreased
sensitivity (from 74.6% to 56.5%) in the algorithm configured for low false positive rate. There were
only a small number of limb lead interchange false positive cases (0.1% of all no cable interchange
cases) for either algorithm configurations.
Table 3 and Table 4 summarize the performance of the low false positive algorithm for validation
database subset 1 and subset 2. Subset 1 included only hypertrophy and/or infarction patients, while
most of ECGs in subset 2 were normal. For these two subsets, false positive rates by the algorithm
were comparable (0.6% vs 0.8% with the difference confidence interval (CI) of [-1.0%, 1.3%]). The
sensitivity for most precordial cable interchanges was statistically different between these subsets,
except for the interchanges of V1-V2 (80.6% vs 76.2% with difference CI of [-0.3%, 9.1%]) and V3-
V4 (86.0% vs 82.6% with difference CI of [-0.9%, 7.5%]). Sensitivity was also significantly lower for
subset 1(hypertrophy and infarction) compared to subset 2 for limb cable interchanges of LA-RA
(from 98.0% to 90.3% with difference CI of [-10.6%, -5.0%]) and RA-LL (from 95.8% to 88.2% with
difference CI of [-10.8%, -4.4%]).
Discussion
In this study, we propose a new method for detecting ECG cable interchange which uses both
11
morphology and redundancy information. In cardiology departments, an algorithm with high
specificity and low false positive rate is preferred because of the low prevalence of lead-wire
interchange (3,4,5,6). Therefore, the algorithm was configured to maintain a total false positive rate
less than 1%. Even with such configuration, the algorithm was still able to achieve good sensitivity
for certain types of cable interchange. We also showed that this algorithm could be configured with
different performance for other scenarios which may need higher sensitivity in detecting cable
interchanges. For the alternate algorithm configuration, the sensitivity was much improved especially
for those difficult cases such as cable interchanges of V2-V3, V4-V5, and V5-V6, at the cost of a
higher total false positive rate. These cable interchanges are difficult to detect because in many
patients there is little difference between the ECG waveforms recorded in those interchanged leads.
The sensitivity for precordial cable interchanges varied among different types of interchange. In
general, sensitivity of cable interchanges for V2-V3, V4-V5, and V5-V6 was lower than other
precordial cable interchanges. From an algorithm perspective, this is because there are larger feature
overlaps between cases for these types of cable interchanges and cases for no cable interchange (as
shown in the scatter plot of Figure 1). Particularly, when the algorithm was configured for low total
false positive rate, the sensitivity for cable interchanges for V4-V5 and V5-V6 was very low (less than
20%). Sometimes, the difference between V4 and V5 (or V5 and V6) is so small that both automated
algorithms and ECG experts may not be able to detect the cable interchange (an example is shown in
Figure 2). For such cases, detecting interchange may not be clinically important because the
diagnostic algorithm can usually generate the same diagnostic report due to the minor difference in
waveforms.
The sensitivity for limb cable interchanges of LA-RA and RA-LL was maintained at high level
12
for both configurations of the algorithm. The specificity for two algorithm configurations was very
close and very high (larger than 99%), which however suggests the necessity of using total false
positive rate rather than specificity to compare the different algorithm configurations. The differences
between algorithms could be lost in the round off of the very high specificity numbers. Furthermore,
even under same configuration, specificity for each individual cable interchange was also very close,
which further emphasizes the importance of investigating the contribution of the individual false
positive rates for different interchanges to the total false positive rate.
During the development of the algorithm, we trained our model using databases including
various disease types and did not favor any particular disease types. To test the algorithm, we divided
the test database into two subsets where subset 1 includes ECGs from hypertrophy and/or infarction
patients, and subset 2 includes ECGs mostly from normal subjects. While the sensitivity varies
between subset 1 and subset 2, the total false positive rate and specificity are comparable and not
statistically different. Due to the limited number of test files (which require expert review and
annotation), we could not test the performance of our algorithm under every specific disease type.
This could be addressed in the future when such large databases including a variety of abnormalities
are available.
In this study, we did not investigate any cable interchange involving the right leg cable, or
interchange between limb and precordial cables. These cable interchanges cannot be simulated by
swapping the waveforms from 12-lead ECG and need specific databases for development and
validation. There are some commonly used criteria (e.g., flat line on limb lead) that could be used to
visually detect the cable interchange between arm cable and right leg cable (8,9). Compared to other
cases, an interchange between limb and precordial cables is rare since the limb cable is usually
13
physically different than the precordial cable (e.g., a limb cable is usually longer than a precordial
cable). Future studies can be conducted when such specific databases are available.
Conclusion
The present study demonstrates that our novel algorithm can be configured to achieve high
specificity and low total false positive rate, and still detects a large number of limb cable interchanges
and precordial cable interchanges. The algorithm can also be configured for higher sensitivity. Our
results suggest that this ECG cable interchange detection algorithm could complement a 12-lead ECG
algorithm for more accurate diagnosis of cardiac diseases.
References
1. Batchvarov VN, Malik M, Camm AJ. Incorrect electrode cable connection during electrocardiographic recording. Europace. 2007 Mov; 9(11): p. 1081-1090.
2. Kligfield P, Gettes LS, Bailey JJ, Childers R, Deal BJ, Hancock EW, et al. Recommendations for the
standardization and interpretation of the electrocardiogram: part I: The electrocardiogram and its technology: a scientific statement from the American Heart Association Electrocardiography and
Arrhythmias Committee, Council on Cli. Circulation. 2007 Mar; 115(10): p. 1306-1324.
3. Hedén B, Ohlsson M, Holst H, Mjöman M, Rittner R, Pahlm O, et al. Detection of frequently
overlooked electrocardiographic lead reversals using artificial neural networks. Am J Cardiol. 1996
Sep; 78(5): p. 600-604.
4. Kors JA, G vH. Accurate automatic detection of electrode interchange in the electrocardiogram. 2001
Aug; 88(4): p. 396-399.
5. Rudiger A, Hellermann JF, Mukherjee R, Follath F, Turina J. Electrocardiographic artifacts due to electrode misplacement and their frequency in different clinical settings. Am J Emerg Med. 2007 Feb;
25(2): p. 174-178.
6. Nilsson KR, Sewell PM, Blunden-Kasdorf P, Starkey K, Grant AO, Wagner GS. The influence of
cardiograph design and automated algorithms on the incidence and detection of electrode cable reversals in an academic electrocardiogram laboratory. J Electrocardiol. 2008 Sep-Oct; 41(5): p. 382-387.
7. Surawicz B, Knilans TK. Misplacement of leads and electrocardiographic artifacts. In Surawicz B, Knilans TK. Chow's Electrocardiography in Clinical Practice. Philadelphia: WB Saunders; 2001. p.
569-582.
8. Castellanos A, Saoudi NC, Schwartz A, Sodi-Pallares D. Electrocardiographic patterns resulting from improper connections of the right leg (ground) cable. Pacing Clin Electrophysiol. 1985 May; May(3 Pt
1): p. 364-368.
9. Haisty WKJ, Pahlm O, Edenbrandt L, Newman K. Recognition of electrocardiographic electrode misplacements involving the ground (right leg) electrode. Am J Cardiol. 1993 Jun; 71(16): p. 1490-
1495.
14
10. Hoffman I. A flatline electrocardiogram in lead II is a marker for right arm/right leg electrode switch. J Electrocardiol. 2007 Jul; 40(3): p. 226-227.
11. Abdollah H, Milliken JA. Recognition of electrocardiographic left arm/left leg lead reversal. Am J Cardiol. 1997 Nov; 80(9): p. 1247-1249.
12. Ho KK, Ho SK. Use of the sinus P wave in diagnosing electrocardiographic limb lead misplacement
not involving the right leg (ground) lead. J Electrocardiol. 2001 Apr; 34(2): p. 161-171.
13. Hedén B, Ohlsson M, Edenbrandt L, Rittner R, Pahlm O, Peterson C. Artificial neural networks for
recognition of electrocardiographic lead reversal. Am J Cardiol. 1995 May; 75(14): p. 929-933.
14. Jekova I, Krasteva V, Abacherli R. Detection of electrode interchange in precordial and orthogonal
ECG leads. In Computing in Cardiology Conference (CinC); 2013; Zaragoza. p. 519-522.
15. Feild DQ, Gregg RE. Detecting lead reversals using the EASI to 12-lead electrocardiogram transform. J Electrocardiol. 2006 Oct; 39(4): p. S129.
16. Xia H, Garcia GA, Zhao X. Automatic detection of ECG electrode misplacement: a tale of two
algorithms. Physiol Meas. 2012 Sep; 33(9): p. 1549-1561.
17. Horácek BM, Warren JW, Feild DQ, Feldman CL. Statistical and deterministic approaches to designing transformations of electrocardiographic leads. J Electrocardiol. 2002; 35: p. Suppl:41-52.
18. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank,
PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic
signals. Circulation. 2000 Jun; 101(23): p. E215-E220.
19. Bousseljot R, Kreiseler D, Schnabel A. Nutzung der EKG-Signaldatenbank CARDIODAT der PTB
über das Internet. Biomed Tech. 1995; 40: p. S317-318.
20. Willems JL, Abreu-Lima C, Arnaud P, van Bemmel JH, Brohet C, Degani R, et al. The diagnostic
performance of computer programs for the interpretation of electrocardiograms. N Engl J Med. 1991
Dec; 325(25): p. 1767-1773.
21. Dower GE, Yakush A, Nazzal SB, Jutzy RV, Ruiz CE. Deriving the 12-lead electrocardiogram from
four (EASI) electrodes. J Electrocardiol. 1988; 21: p. S182-187.
22. Feild DQ, Feldman CL, Horácek BM. Improved EASI coefficients: their derivation, values, and performance. J Electrocardiol. 2002; 35: p. 23-33.
23. Chang CC, Lin CJ. LIBSVM : a library for support vector machines. ACM Transactions on Intelligent Systems and Technology. 2011 Apr; 2(3): p. 1--27.
24. Altman D, Machin D, Bryant T, Gardner S. Statistics with Confidence: Confidence Intervals And
Statistical Guidelines. 2nd ed. Bristol: BMJ Books; 2000.
15
Figure Legends
Figure 1. Visualization of the decision boundary for classification of lead wire misconnection in the
development database. (A) Scatter plot of 2D features after feature dimensionality reduction for LA-
RA cable interchange (black dot) and no cable interchange (gray dot). (B) Histogram of the SVM
scores for LA-RA cable interchange (black line) and no cable interchange (gray line). The x-axis of
the histogram is the SVM score which is calculated by linear combination of all features. The y-axis
of the histogram is the number of ECG cases that have same value of SVM score. The black vertical
dashed line indicates the decision boundary, the left of which is considered as LA-RA interchange and
the right of which is considered no interchange. (C) and (D) are the scatter plot and decision boundary
for V2-V3 cable interchange and no cable interchange.
Figure 2. A 12-lead ECG example with V5-V6 cable interchange. As seen, the waveforms for leads
V5 and V6 are very similar, making it extremely difficult to detect if a cable interchange has occurred.
Figure 3. Comparison of confusion matrices for algorithms configured for high sensitivity (A) and
low total false positive rate (B). All the individual types of cable interchange were summarized into
three categories: no cable interchange, precordial cable interchange and limb cable interchange. Each
column of the matrix represents the instances of the output predicted by the algorithm, while each row
represents the instances of the actual reference. The number in each matrix element is the percentage
of the number of certain algorithm-detected instances within the total number of one actual reference
instance.
16
Tables
Table 1 Classification performance for algorithm configured for high sensitivity
V1-V2 V1-V3 V2-V3 V3-V4 V4-V5 V4-V6 V5-V6 LA-RA RA-LL
Sensitivity (%) 87.4 90.9 64.9 93.2 43.6 86.3 50.3 94.5 92.3
Specificity (%) 99.9 99.9 99.9 99.9 100.0 99.9 99.9 100.0 99.9
Note: LA: left arm; LL: left leg; RA: right arm; total false positive rate is 2.7%.
Table 2 Classification performance for algorithm configured for low total false positive rate
V1-V2 V1-V3 V2-V3 V3-V4 V4-V5 V4-V6 V5-V6 LA-RA RA-LL
Sensitivity (%) 78.2 81.7 49.1 84.1 12.7 68.1 19.1 94.5 92.4
Specificity (%) 99.9 99.9 99.9 100.0 100.0 99.9 100.0 100.0 99.9
Note: LA: left arm; LL: left leg; RA: right arm; total false positive rate is 0.7%.
Table 3 Classification performance for algorithm configured for low total false positive rate using
subset 1 of hypertrophy/infarction patients
V1-V2 V1-V3 V2-V3 V3-V4 V4-V5 V4-V6 V5-V6 LA-RA RA-LL
Sensitivity (%) 80.6 76.1 35.5 86.0 16.9 77.8 25.1 90.3 88.2
Specificity (%) 99.9 100.0 100.0 100.0 100.0 99.9 100.0 100.0 99.8
Note: LA: left arm; LL: left leg; RA: right arm; total false positive rate is 0.6%.
Table 4 Classification performance for algorithm configured for low total false positive rate using
subset 2 of validation database
V1-V2 V1-V3 V2-V3 V3-V4 V4-V5 V4-V6 V5-V6 LA-RA RA-LL
Sensitivity (%) 76.2 86.2 60.3 82.6 9.2 60.1 14.2 98.0 95.8
Specificity (%) 99.9 99.9 100.0 100.0 100.0 99.9 99.9 100.0 100.0
Note: LA: left arm; LL: left leg; RA: right arm; total false positive rate is 0.8%.