Artificial Neural Network–Based Method of Screening Heart
Murmurs in Children
Curt G. DeGroff, MD; Sanjay Bhatikar, PhD; Jean Hertzberg, PhD; Robin Shandas, PhD;
Lilliam Valdes-Cruz, MD; Roop L. Mahajan, PhD
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
Background—Early recognition of heart disease is an important goal in pediatrics. Efforts in developing an inexpensive
screening device that can assist in the differentiation between innocent and pathological heart murmurs have met with
limited success. Artificial neural networks (ANNs) are valuable tools used in complex pattern recognition and
classification tasks. The aim of the present study was to train an ANN to distinguish between innocent and pathological
murmurs effectively.
Methods and Results—Using an electronic stethoscope, heart sounds were recorded from 69 patients (37 pathological and
32 innocent murmurs). Sound samples were processed using digital signal analysis and fed into a custom ANN. With
optimal settings, sensitivities and specificities of 100% were obtained on the data collected with the ANN classification
system developed. For future unknowns, our results suggest the generalization would improve with better representation
of all classes in the training data.
Conclusion—We demonstrated that ANNs show significant potential in their use as an accurate diagnostic tool for the
classification of heart sound data into innocent and pathological classes. This technology offers great promise for the
development of a device for high-volume screening of children for heart disease. (Circulation. 2001;103:2711-2716.)
Key Words: heart murmurs 䡲 neural networks (computer) 䡲 child 䡲 heart defects, congenital
T
connected in sequence (Figure 1), details of which can be
found in Appendix A.
Studies of ANNs in cardiology have been mainly concerned with the evaluation of ECG signals.14 –16,18 –20 There
have been several studies on the use of ANNs on heart sounds
with limited results and applicability.7,11,12,21 The aim of the
present study was to train an ANN to distinguish between
innocent and pathological murmurs effectively.
he incidence of heart murmurs in the pediatric population
is reportedly as high as 77% to 95%. However, ⬍1% of
this population has heart disease.1 Early recognition is an
important goal,2 and equally important is avoiding misdiagnosing a pathological heart murmur in a healthy child without
heart disease.2 To acquire high-quality auscultation skills
requires the guidance of an experienced instructor using a
sizable number of patients along with frequent practice.1,2
Unfortunately, the interpretation of auscultation findings
overall remains prone to error.3– 6 Imaging technologies can
provide more direct evidence of heart disease; however, they
are generally more costly. Efforts to develop an inexpensive
screening device that can assist in the differentiation between
innocent and pathological heart murmurs have met with
limited success.7
There has been much excitement in the scientific literature
in recent years regarding artificial neural networks (ANNs)8,9
in medicine10 and, specifically, in cardiology applications.7,8,11–20 ANNs are valuable tools used in complex pattern
recognition and classification tasks. They learn complex
interactions among inputs and identify relations in input data
that may not be apparent to human analysis.14 The most
common type of ANN consists of 3 layers of processing
units: the input layer, the hidden layer, and the output layer
Methods
Study Subjects
Digital heart sound recordings were obtained from 69 pediatric
patients who were referred to our cardiology clinic for evaluation
(aged 1 week to 15 years; mean age, 2 years). Under Internal Review
Board protocol, all participants and/or parents gave written, informed
consent to participate in this study. The diagnosis was confirmed by
1 of 5 pediatric cardiologists. Echocardiographic confirmation was
performed on 100% of the patients with pathological heart murmurs
and 70% of patients with innocent heart murmurs.
Recordings
Using an electronic stethoscope (Cambridge Heart Sound Microphone; frequency range, 40 to 600 Hz) and a personal computer
(sampling at 44 kHz), heart sounds were recorded at the left
midsternal border in the supine position in a standard examination
room for all patients. Younger patients were placed in a darkened
Received December 31, 2000; revision received March 19, 2001; accepted March 21, 2001.
From the University of Colorado Health Sciences Center, the Children’s Hospital, Denver (C.G.D., R.S., L.V-C.), and the Department of Mechanical
Engineering, University of Colorado, Boulder (S.B., J.H., R.S., R.L.M.), Colo.
Correspondence to Curt G. DeGroff, MD, University of Colorado Health Science Center, the Children’s Hospital, 1056 E 19th Ave, B-100 Denver,
CO 80218. E-mail degroff.curt@tchden.org
© 2001 American Heart Association, Inc.
Circulation is available at http://www.circulationaha.org
2711
2712
Circulation
June 5, 2001
Figure 1. An example of an ANN with an
input layer, hidden layer, and output
layer. X indicates input; w, weight; b,
bias weight; F, activation function; u,
activation; and Y, output. For further
detail, see Appendix A.
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
examination room with minimal stimuli to encourage cooperation.
No sedation was used. For each patient, 2 separate heart sound
recordings were acquired. One recording location was chosen with
the expectation that murmurs generally not heard from this location
(using standard stethoscopic methods) would be detected by the
electronic microphone used due to its sensitivity.
Two heart sound recordings for each patient of ⬇8 heart cycles in
length were recorded. Subsequently, one sound sample of 3 repre-
sentative consecutive heart cycles was chosen from these 2 recordings by one of the investigators (C.G.D.) for all 69 patients. The
investigator chose the optimal sound sample by listening to the
entirety of the 2 recordings and qualitatively determining the sound
sample (of 3 consecutive heart cycles) with the least amount of
extraneous noise, such as breathing, talking, motion artifact, and
room noises. This optimal sound sample was then processed using
digital signal analysis (Figure 2).
Figure 2. Block diagram
depicting the overall process
of our study from heart sound
recordings to the ANN’s diagnostic impression.
DeGroff et al
Artificial Neural Network to Screen Heart Murmurs
2713
TABLE 1. Structure of ANN for Frequency Ranges and
Spectral Resolutions Considered
Frequency Range,
Spectral Resolution
No. of Input
Neurons
No. of Hidden
Neurons
No. of Output
Neurons
0 –90 Hz, 3 Hz
30
10
1
0–150 Hz, 3 Hz
50
10
1
0–210 Hz, 3 Hz
70
15
1
0–300 Hz, 3 Hz
100
15
1
0–90 Hz, 1 Hz
90
15
1
0–150 Hz, 1 Hz
150
20
1
0–210 Hz, 1 Hz
210
20
1
0–255 Hz, 1 Hz
255
30
1
Signal Analysis
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
A normalized energy spectrum of the sound data was obtained by
applying a Fast Fourier Transform22 to the heart sounds. Various
spectral resolutions (1, 3, and 5 Hz) and frequency ranges (0 to 90,
0 to 150, 0 to 210, 0 to 255, and 0 to 300 Hz) were used as input into
the ANN to optimize these parameters to obtain the most favorable
results.
Artificial Neural Network
We used customized ANN software developed in our laboratory
(CUANN; see Appendix B). ANNs were trained to discriminate
between normal and pathological examples. Because of the size of
the available data, the Jack-Knifing method23 was used. The JackKnifing method is an iterative process in which one observation is
recruited each time for validation. A classifier is trained using the
remaining data and validated on the single, left-out validation point.
This ensures that the validation is unbiased, because the classifier
does not see the validation point during its training. One by one, each
available example is recruited for validation. The approach measures
the power of the classification approach rather than of one specific
classifier.
Each of the 69 data sets was recruited for validation, one at a time,
creating 69 separately trained ANNs. For each spectral resolution (1,
3, and 5 Hz) and frequency range considered (0 to 90, 0 to 150, and
0 to 210 Hz), a set of 69 neural networks with identical numbers of
input, hidden, and output neurons was created (Table 1).
Statistical Analysis
Standard equations for sensitivity and specificity were used.24
Results
Study Subjects
Pathologies included aortic and pulmonary stenosis/insufficiency, atrial and ventricular septal defects, patent ductus
arteriosus, and mitral regurgitation (Figure 3). By spectral
Doppler echocardiography, all patients with aortic stenosis
had moderate stenosis (gradient, 30 to 50 mm Hg), and all
patients with pulmonary stenosis had mild stenosis (gradient,
⬍20 mm Hg). Peripheral pulmonary stenosis murmurs were
considered pathological because of the potential necessity of
follow-up required for some of these patients. Innocent heart
murmurs consisted of Still’s murmurs and pulmonary flow
murmurs. No venous hum murmurs were included in this
study.
Recordings
For the majority of patients, the entire recording procedure
took ⬍5 minutes (maximum, 10 minutes). All of the patient
Figure 3. Distribution of the 69 cases studied with innocent
murmurs (normal), aortic insufficiency (AI), aortic stenosis (AS),
atrial septal defects (ASD), mitral regurgitation (MR), patent ductus arteriosus (PDA), pulmonary stenosis (PS), pulmonary insufficiency (PI), peripheral pulmonary stenosis (PPS), and ventricular
septal defects (VSD).
recordings were deemed of acceptable quality by one investigator (C.G.D.).
ANN Predictions
Table 2 shows the results obtained with a spectral resolution of
3 Hz and various frequency ranges. Sensitivities and specificities
⬎90% were obtained. Figure 4A shows the predictions of the 69
ANNs (with a spectral range of 0 to 150 Hz and a spectral
resolution of 3 Hz) constituting a feature set of 50 constituent
elements. ANN prediction is plotted on the y axis versus the
corresponding validation example number on the x axis. The
data were arranged in a specific, constant order. The first 37
examples (cases 1 through 37) were pathological murmurs, and
the next 32 (cases 38 through 69) were innocent murmurs. The
horizontal dashed line represents the decision threshold. Setting
the threshold to either extremity (0 or 1) results in trivial results
so that all examples are classified as normal or pathological.
With a threshold of 0.8, a sensitivity of 100% and a specificity
of 92% were obtained. Observe the erroneous classifications
with respect to this threshold marked by bull’s-eyes. Interestingly, increasing the spectral range past 0 to 150 Hz while
keeping the spectral resolution to 3 Hz produced worsening
results (Table 2), with the ANN predictions not as tightly
distributed (Figure 4B).
To further improve the performance, the spectral resolution
was increased to 1 Hz. In the frequency range of 0 to 90 Hz,
the performance was inadequate. However, it improved as the
spectrum was expanded (Table 2). With the upper limit of the
TABLE 2. Sensitivity and Specificity of the ANN for Spectral
Resolution of 3 Hz and 1 Hz at Various Frequency Ranges
Frequency
Range
Spectral
Resolution
ANN Decision
Threshold
0 –90 Hz
3 Hz
0.75
75%
81%
0–150 Hz
3 Hz
0.8
100%
100%
0–210 Hz
3 Hz
0.7
80%
90%
0–300 Hz
3 Hz
0.75
91%
92%
0–90 Hz
1 Hz
0.8
97%
91%
0–150 Hz
1 Hz
0.8
100%
100%
0–210 Hz
1 Hz
0.7
100%
100%
0–255 Hz
1 Hz
0.8
100%
100%
Sensitivity
Specificity
2714
Circulation
June 5, 2001
a manner similar to what occurred with the spectral resolution
of 3 Hz.
Discussion
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
Figure 4. ANN classifier results. A, Classifier results with a frequency range of 0 to 150 Hz and a frequency resolution of 3 Hz
showing the predictions of the 69 ANNs. The erroneous classifications with respect to this threshold are marked by bull’s-eyes.
See text for details. B, ANN classifier results with a frequency
range of 0 to 300 Hz and a frequency resolution of 3 Hz. C,
ANN classifier results with a frequency range of 0 to 210 Hz and
a frequency resolution of 1 Hz. See text for details.
frequency range increased from 90 to 150 Hz, the accuracy of
the classifier system improved to 100% sensitivity and
specificity. Increasing the upper limit from 150 to 210 Hz
continued to improve the classifier system (Table 2 and
Figure 4C). However, when the upper limit was increased
still further to 255 Hz, the trend was reversed. Although
100% sensitivity and specificity were maintained, the ANN
predictions were not as tightly distributed (data not shown), in
Cardiac auscultation continues to be the healthcare professional’s primary tool for distinguishing between innocent and
pathological heart murmurs in pediatric patients. The sound
characteristics for heart lesions have been previously described.25 For medical persons to acquire high-quality auscultation skills requires the guidance of an experienced
instructor using a sizable number of patients along with
frequent practice.1,2 Unfortunately, the interpretation of auscultation findings is prone to error.3– 6 In addition, the human
ear cannot appreciate many of the subtleties of heart sounds.26
Therefore, a device that can assist in the differentiation
between innocent and pathological heart murmurs would be
quite useful. Efforts to date have met with limited success.7,11,12,21,27–32 One of these studies examined the sounds
from prosthetic heart valves.11 Several of these studies are
simply observations of the time frequency analysis of innocent and pathological heart murmurs, with no emphasis on
developing automated diagnostic algorithms.29,30 All of these
previous studies (including those using ANNs) have generally focused on determining if a heart murmur exists, with a
lack of emphasis on developing diagnostic algorithms differentiating between innocent versus pathological
murmurs.7,12,21,27,28,31,32
In this study, 100% sensitivity and specificity were obtained for the ANN distinguishing between the 69 innocent
and pathological heart murmurs presented to it using a
spectral resolution of 1 Hz and a spectrum of 0 to 210 Hz. In
general, a finer resolution and larger spectrum yielded improved results. The trend in improvement of classifier accuracy with wider energy spectrum may be explained as
follows. As the energy spectrum is expanded, the ANN
classifier has more information to work with, and its accuracy
tends to increase. However, the increased computational load
on the ANN may offset this benefit, and it is seen that at a
certain point, the trend is reversed. Also, it is common
knowledge that a pattern recognition that can handle a larger
feature set is better unless the larger feature set contains
redundant information. The trend in improved classification
performance with higher Fast Fourier Transform resolution
was confirmed by decreasing the resolution to 5 Hz (data not
shown). Accuracy comparable to 1 Hz or 3 Hz levels was
never achieved at this resolution.
Although we deployed all 69 examples for validation, only
one example served for validation each time, thus creating 69
separately trained ANNs. The question arose as to whether a
classifier system comprising a set of ANNs accruing from the
Jack-Knifing approach could be field-deployed. Under such a
scheme, the individual ANNs would vote in the classification
of a previously unseen observation. To investigate this matter
and having no additional data, the data were divided into 2
sets. (1) A set of 54 examples was earmarked as training data
for the development of a system of 54 ANNs following the
Jack-Knifing scheme, and (2) a set of 15 examples was
reserved as field-testing data exclusively to validate the 54
ANNs as multiple experts, with the decision effected by a rule
DeGroff et al
Artificial Neural Network to Screen Heart Murmurs
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
of simple majority. Using the validation set, the system of
ANNs was able to classify 7 of 9 pathological examples and
5 of 6 innocent examples correctly. Both of the misclassified
pathological examples, a case with pulmonary stenosis and a
case with an atrial septal defect, belonged to pathological
classes that were grossly under-represented in the training
data. When these misclassified pathological examples and the
misclassified normal example were reintroduced into the
training data set, a threshold could be established with respect
to which the resulting classifier system of 57 ANNs was fully
accurate when field-tested on the 12 validation examples.
These results suggest that ANN generalization would improve with better representation of all classes in the training
data, for which more data would have to be collected.
Finally, the current results include training and testing of
the classifier system with only one class of aortic stenosis
(moderate) and only one class of pulmonary stenosis (mild).
Clearly, a wide spectrum of murmurs for each type of
pathological case has not yet been presented to the network.
However, the classifier system did perform well with a wide
range of sound “signatures” found to be present in the
patients with murmurs from ventricular septal defects (Appendix C).
Overall, we propose that as additional data are collected, a
point will be reached when the important pathological types
will each be adequately represented and the generalization
capability of the ANN classifier system developed by JackKnifing will stabilize at some high accuracy. At that point,
misclassifications will be rare and field deployment will be
possible. During the stabilizing process, a particular frequency range, spectral resolution, and threshold will be
selected to optimize results. Also, after this point, whether the
classifier system will be designed so that it continues to
“learn” is beyond the scope of this discussion.
Limitations and Future Work
The process of selecting the optimal 3 consecutive heart
cycles required an intelligent selection process. In developing
a screening device, automating this selection process33 will
need to be investigated. The performance of the ANN with
the addition of noise processes (eg, from an uncooperative
patient) to the input signal will need to be investigated. All of
our 69 patients were cooperative enough for us to obtain
adequate heart sound recordings. An intelligent automated
selection algorithm will also need to determine if an adequate
sample exists at all.
The most pressing need at present is to collect more data.,
In future efforts, it would be impractical to consider that the
classifier system could be introduced to all possible heart
murmur “signatures.” Therefore, future efforts will focus on
presenting the ANN with a wide spectrum of each of the
common pathological cases studied here, both for training
and testing purposes. However, our results here are quite
promising with the limited data set available.
If, after additional training data are presented to the ANN
classifier system, the system is unable to maintain the
accuracy that has been shown here, recordings from the
standard 5 cardiac auscultation sites will be considered as
inputs. This may be necessary for murmurs not generally
2715
heard well at the one recording location chosen (left midsternal border). However, the sensitivity of the electronic microphone used will potentially grant us the economy of one
recording location.
Finally, there will most likely be classes of patients with heart
lesions for which a screening device using these algorithms may
be expected to produce false negatives. Therefore, in patients for
whom such a screening device determines that a pathological
murmur does not exist, it would simply declare: “no abnormal
cardiac murmurs detected.” Two examples are newborns with
cardiac shunt lesions in whom murmurs typically appear only
after pulmonary pressures begin to drop and those with heart
lesions whose murmurs may appear only during examination
with postural changes.
Conclusions
We demonstrated that ANNs show significant potential in
their use as an accurate diagnostic tool for the classification of
heart sound data into innocent and pathological classes. Of
course the evaluation of a child with a heart murmur is far
more inclusive that just the auscultation of that murmur.
Therefore, such a device is not seen as replacing the need for
a clinician’s assessment in a child found to have a heart
murmur. Such a device may, however, assist a clinician in
rendering an opinion concerning a murmur. In addition, this
technology offers great promise for the development of a
device for high-volume screening of children for heart disease. Eventually, this work is expected to lead to an automatic
screening device with additional capabilities of predicting
selected heart conditions.
Appendix A
An ANN is a highly interconnected system of computational nodes
or neurons. In an ANN of the feed-forward type, the neurons are
organized into layers. Typically, there is an input layer into which the
input vector (independent variables) is fed in, an output layer that
produces the output vector (dependent variables), and one or more
layers in between. A typical ANN is shown in Figure 1. It has an
input layer, a hidden layer, and an output layer. There are connections going from any one neuron in one layer to all the neurons in the
next layer. Each connection is associated with a weight. Such an
ANN can be trained to learn input-output mapping by showing it
examples of the mapping. During the training procedure, the weights
of the ANN are adjusted such that the ANN converges to the output
of the training data. Here, an input vector is presented at the input
layer and run through the ANN to obtain the output vector. The
output of any layer is propagated as the input to the following layer.
The following equation relates the output of a neuron in the first
layer to its input from the final minus first layer. The neuronal output
is also referred to as its activation.
冉冘
Nl⫺1
(1)
Yil⫽F
冊
wijlYjl⫺1⫹bil
j⫽1
Here, Yli is the activation of the ith neuron in the lth layer; wlij is the weight
of the connection from the jth neuron in the (l⫺1)th layer to the ith neuron in
the lth layer; bli is the bias connected to the ith neuron in the lth layer; and Nl–1
is the number of neurons in the (l⫺1)th layer. F(u) is the activation function,
which may be thought of as providing a nonlinear gain for the artificial
neuron. It is typically a sigmoid function and bounds the output from any
neuron in the network as follows:
(2)
1
F⫽
共1⫹e⫺u兲
2716
Circulation
June 5, 2001
where e is the base of the natural logarithm and u is the activation of
a neuron.
To train an ANN for developing input-output mapping, data are
required that are representative of the mapping. The first step of
training is the forward pass, which consists of calculating the output
vector by running the input vector through the ANN. This is
followed by a backward pass, where the error derivatives are
calculated for each weight. The error derivatives for a weight are
summed until all the data points have been run through the network
once. This constitutes an epoch. The weights are updated after each
epoch such that the ANN error decreases. For a classification task,
the output neurons of the ANN represent individual classes where,
typically, a binary classification scheme is used.
Appendix B
For this project, we used our in-house ANN software (CUANN),34,35
which is designed for efficient development and deployment of a
feed-forward ANN. It implements a feed-forward ANN using the
back-propagation learning algorithm sketched in the preceding section.
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
Appendix C
It was important to know whether the examples of a pathological
class provided an adequate representation of that class. The evaluation was performed by analyzing the extent of dissimilarity within
the examples of the same pathology in our data using the average
correlation metric.21 In general, the average correlation between 2
classes is defined as follows:
(3)
1
mn
冘冘
m
n
i
j
ui䡠vj
兩ui兩⫻兩vj兩
where ui and vj are examples from the 2 classes being compared, and
m and n are the number of cases in each class, respectively. The
metric follows from the fact that the correlation between a pair of
vectors can be calculated as their dot product. Thus, a measure of
interclass similarity is provided by the average of the normalized dot
product of each of vector ui with each vector vj. The correlation takes
on a value between 0 and 1, with 1 representing maximum correlation, which is when all vectors are identical.
The same metric, applied to just one class, provides a measure of
intraclass similarity. In this case, the metric is evaluated as follows:
(4)
1
m2
冘冘
m
m
i
j
ui䡠uj
兩ui兩⫻兩uj兩
where ui and uj are examples from the one class being compared and
m is the number of cases in the class.
We compared the set of examples representing ventricular septal
defect murmurs with the set of examples with innocent murmurs.
The intraclass correlation metric was 78% for the defects and 77%
for the innocent murmurs. Because it was known that the 32
instances of innocent murmurs in our data set included disparate
examples, we concluded that the set of ventricular septal defect
examples was also heterogeneous.
Acknowledgments
We gratefully acknowledge The Children’s Hospital Research Institute and Micro-Electronic Devices in Cardiovascular Applications
(MEDICA). Interdisciplinary Center Denver, Colorado, for partial
financial support and for facilitating this project.
References
1. Garson A, Bricker J, McNamara D. The Science and Practice of Pediatric
Cardiology. Philadelphia, Pa: Lea & Febiger; 1990:2203–2212.
2. Pelech A. The cardiac murmur. Pediatr Clin North Am. 1998;45:107–122.
3. Rangaraj M, Murthy I, Quantitative analysis of the phonocardiogram for
detection of murmurs. J Biomed Eng. 1979;1:247–252.
4. Lok C, Morgan C, Ranganathan N. The accuracy and intraobserver
agreement in detecting the ‘gallop sounds’ by cardiac auscultation. Chest.
1998;114:1283–1289.
5. VanOort A, Le Blanc-Botden M, De-Boo T, et al. The vibratory innocent
heart murmur in school children: difference in auscultatory findings between
school medical officers and a pediatric cardiologist. Pediatr Cardiol. 1994;
15:282–287.
6. Pelech A. Evaluation of the pediatric patient with a cardiac murmur. Pediatr
Clin North Am. 1999;46:167–188.
7. Cathers I. Neural network assisted cardiac auscultation. Artif Intell Med.
1995;7:53–66.
8. Crick F. The recent excitement about neural networks. Nature. 1989;337:
129–132.
9. Hinton G. How neural networks learn from experience. Sci Am. 1992;267:
145–151.
10. Tu J. Advantages and disadvantages of using artificial neural networks versus
logistic regression for predicting medical outcomes. J Clin Epidemiol. 1996;
49:1225–1231.
11. Guo Z, Durand L, Lee H, et al. Artificial neural networks in computer assisted
classification of heart sounds in patients with porcine bioprosthetic valves.
Med Biol Eng Comp. 1994;32:311–316.
12. Coehn M, Hudson D. Comparative Approaches to Medical Reasoning. River
Edge, NJ: World Scientific; 1995:271–288.
13. Binder T, Sussner M, Moertl D, et al. Artificial neural networks: theoretical
considerations and clinical applicability. Comput Biol Med. 1999;25:
1069–1076.
14. Itchhaporia D, Snow P, Almassy R, et al. Artificial neural networks: current
status in cardiovascular medicine. J Am Coll Cardiol. 1996;28:515–521.
15. Silipo R, Gori M, Taddei A, et al. Classification of arrhythmic events in
ambulatory elcetrocardiogram, using artificial neural networks. Comput
Biomed Res. 1995;28:305–318.
16. Dassen W, Mulleneers R, DenDulk K, et al. An artificial neural network to
localize atrioventricular accessory pathways in patients suffering from WolfeParkinson-White syndrome. Pacing Clin Electrophysiol. 1990;13:1792–1796.
17. Griffin M, Scollan D, Moorman J. The dynamic range of neonatal heart rate
variability. J Cardiovasc Electrophysiol. 1994;5:112–124.
18. Heden B, Ohlsson M, Rittner R, et al. Agreement between artificial neural
networks and experienced electrocardiographer on electrocardiographic
diagnosis of healed myocardial infarction. J Am Coll Cardiol. 1996;28:
1012–1016.
19. Heden B, Ohlsson M, Rittner R, et al. Acute myocardial infarction detected in the
12 lead ECG by artificial neural networks. Circulation. 1997;96:1798–1802.
20. Suzuki Y, Ono K. Personal computer system for ECG ST-segment recognition based on neural networks. Med Biol Eng Comput. 1992;30:2–8.
21. Barschdorff D, Femmer U, Trowitzsch E. Automatic phonocardiogram signal
analysis in infants based on wavelet transforms and artificial neural networks.
Comput Cardiol. 1995;7:753–756.
22. Stremler FG. Introduction to Communication Systems. Reading, Mass: Addison-Wesley; 1977.
23. Masters T. Practical Neural Networks in C⫹⫹. New York: Academic Press;
1993.
24. Rosner B. Fundamentals of Biostatistics. 5th ed. New York: Duxbury Press;
2000.
25. Wartak J. Phonocardiography: Integrated Study of Heart Sounds and
Murmurs. Hagerstown, Md: Harper & Row; 1972.
26. Obaidat M. Phonocardiogram signal analysis: techniques and performance
comparison. J Med Eng Technol. 1993;17:221–227.
27. Okuni M, Hasimoto T, Kobayashi H, et al. Trial of a new cardiac mass
screening system in school children. Jpn Circ J. 1978;42:49–52.
28. Reynolds JL. Heart disease screening of preschool children. Am J Dis Child.
1970;119:488–493.
29. White PR, Collis WB, Salmon AP. Time frequency analysis of heart
murmurs in children. Proc IEEE. 1997;3:1–4.
30. El-Asir B, Khadra L, Al-Abbasi AH. Time frequency analysis of heart
sounds. Proc IEEE. 1996;12:553–558.
31. Leung TS, White PR, Collis WB, et al. Analysing pediatric heart murmurs
with discriminant analysis. Proc IEEE. 1998;18:1628–1631.
32. Sarkady AA, Clark RR, Williams R. Computer analysis techniques for
phonocardiogram diagnosis. Comput Biomed Res. 1976;9:349–363.
33. Liang H, Lukkarinen S, Hartimo I. Heart sound segmentation algorithm
based on heart sound envelogram. Proc IEEE. 1997;24:105–108.
34. Bhatikar S, Mahajan R. Neural network based diagnosis of CVD barrel
reactor. Adv Elec Pkg. 1999;1:621–626.
35. Marwah M, Li Y, Mahajan R. Integrated neural network modeling for
electronic manufacturing. J Electronic Manufacturing. 1996;6:79–91.
Artificial Neural Network−Based Method of Screening Heart Murmurs in Children
Curt G. DeGroff, Sanjay Bhatikar, Jean Hertzberg, Robin Shandas, Lilliam Valdes-Cruz and Roop L.
Mahajan
Downloaded from http://circ.ahajournals.org/ by guest on September 29, 2016
Circulation. 2001;103:2711-2716
doi: 10.1161/01.CIR.103.22.2711
Circulation is published by the American Heart Association, 7272 Greenville Avenue, Dallas, TX 75231
Copyright © 2001 American Heart Association, Inc. All rights reserved.
Print ISSN: 0009-7322. Online ISSN: 1524-4539
The online version of this article, along with updated information and services, is located on the World
Wide Web at:
http://circ.ahajournals.org/content/103/22/2711
Permissions: Requests for permissions to reproduce figures, tables, or portions of articles originally published in
Circulation can be obtained via RightsLink, a service of the Copyright Clearance Center, not the Editorial Office. Once
the online version of the published article for which permission is being requested is located, click Request Permissions in
the middle column of the Web page under Services. Further information about this process is available in thePermissions
and Rights Question and Answer document.
Reprints: Information about reprints can be found online at:
http://www.lww.com/reprints
Subscriptions: Information about subscribing to Circulation is online at:
http://circ.ahajournals.org//subscriptions/