Rheumatic Heart Disease detection using Adaptive

filters and Neural Network.

Nanda Kishor Panda, Monica Subhashini M., H. Bharadwaj Milind Kejriwal
School of Electrical Engineering School of Electronics Engineering
Vellore Institute of Technology Vellore Institute of Technology
Vellore, India Vellore, India
Email: [email protected] Email: [email protected]

Abstract— An efficient and innovative method has been Envelope[9], peak location[10], are some of the most common
proposed in this paper to detect heart murmurs as a method to features considered for heart signals.
identify rheumatic fever with the use of adaptive filters,
transform techniques and Neural Network Algorithms by In this paper neural networks have been proposed to detect
considering various parameters such as number of peaks, Signal heart murmurs by identifying peaks as an early warning
to Noise Ratio (SNR) and Power Spectral Density. Under system for rheumatic fever. Among various commonly used
optimum conditions the classification returned exact outputs neural networks scaled conjugate gradient backpropagation
even when the neural network was trained under false positive algorithm has been shown to be the most adept at handling the
data thus showing its effectiveness. varying dataset required for this application and shown to have
a significant success rate in detecting false positives and false
Keywords—Hilbert Transform, Scaled Conjugate Gradient negatives.
Backpropagation, Heart murmurs, Power Spectral Density,
Adaptive Filter. II. BACKGROUND INFORMATION

I. INTRODUCTION A. Rheumatic Heart disease

Health care solutions have seen a tremendous growth in Rheumatic fever is an inflamatory disese caused in small
recent years. With improving technologies, the field of health childres in the age group 0-5 yrs. This is most predominant in
care diagnosis relies heavily on sophisticated equipment, developing nations like India, China, etc. This can cause
precise sensors and optimized processing of signals to support permanent damage to heart valves if left undetected and not
the traditional experience of medical practitioners. Yet there treated timely. The main cause of heart damage is due to
is a constant requirement to improve these factors. prolonged rheumatic fever causing inflammationation of
Auscultation has been one of the oldest and most common heart.
diagnostic tools for a variety of symptoms. B. Diagnosis
Recent trends have significantly improved this process Due to the fact that this disease occurs mostly in children
through computer aided auscultation by reducing false of developing countries early detection of heart abnormalities
negative readings for heart murmurs from 13.3% to 7.1% as are hardly accessible, resulting in premature deaths, paralysis
shown by Raymond[1]. Reducing signal noises due to and permanent damage to brain. ECG is most effective way to
environment and system is one of the most challenging detect these abnormalities but needs special instruments and
aspects of computer aided detection due to the extremely low trained technicians. This paper presents an new approach of
power of the signal requiring very high sensitivity of efficient and accurate detection of heart abnormalities using
equipment. There are various methods by which this can be phonocardiogram.
achieved for various signals. Mohammad (et. al) proposed a
method for noise reduction in coronary artery disease III. THE PHONOCARDIOGRAM
detection which broadly involves spectral analysis and The Phonocardiogram (PCG) is simply the audio signal of
segmentation to identify noisy frames and then using adaptive heart sound, which is caused due to the rhythmic contraction
filters to remove the noise[2]. Neural networks have been and expansion of heart muscles[11].
shown to improve the performance of such methods in
transformation and feature extraction with improved As shown in Fig. 1 a PCG signal can be divided into 2
algorithms as shown by Dandan (et al)[3]. Kalman filtering is major sections as S1 and S2. The S11-heart cycle is further
a widely used technique for a variety of applications. It’s a divided into 4 major components: sound due to the closure of
robust technique which is used to remove statistical noise by the mitral and tricuspid valves (S1), systolic period (S12),
conducting successive measurements over time and estimates sound corresponding to the closure of the aortic and
the error. It finds application in camera based photo pulmonary valves (S2), and the diastolic period (S21). With
plethysmography[4] which is a noninvasive technique used to correct identification of different sections of PCG heart valve
measure heart rate. abnormalities which are the major side effects of Rheumatic
fever can be acccurately determined[12]–[14].
Transformation of real time signals is an important stage
as features of the signal which are useful are not easily IV. PROPOSED TOPOLOGY
identifiable in time domain. A few common transformation A screening procedure for detection of heart abnormalities
techniques for bio-medical signals include discrete S- due to rheumatic fever using signal processing and machine
transform[5], time-frequency analysis[6], [7] and Hilbert- learning was developed. The phonocardiogram of affected
Huang Transform[8]. Upon transformation to the required people contains a lot of extra sounds along with the regular
domain, the required features of the signal need to be
extracted. Power Spectral Density (PSD), Energy
 Figure 3: Process of adaptive filtering.

amplitude and phase compared to noise[n]. Unlike

frequency-selective filters the adaptive filter is applied to x[n]
instead of the primary input d[n]. The filter output y[n] is an
estimate of noise[n]. Subtracting this from the underlying
signal x[n] we get the error signal e[n], an estimate of
Figure 1: Sections of phonocardiogram (PCG), with components: first sound PCG[n]. Adaptive filters are preferably FIR which are stable
(S1), second sound (S2), systolic (S12), diastolic (S21) and heart cycle (S11) and have simple cost function as shown below:
periods.
J = E{e2 [n]} = (d[n] − y[n])2 → min.
systole and diastole peaks. Detection of these extra peaks can For a N order filter the above equation results in a quadratic
be considered as an early indication for people at risk. There cost-function with a global minimum. R is the autocorrelation
are many stepsThe steps involved in achieving this aresuch matrix of x[n] and P is the cross-correlation-vector between
Formatted: Indent: First line: 0.2", Keep with next
as filtering, transformation, feature extraction and x[n] and d[n].
classification performed using a variety of methods to obtain
the optimal result. The proposed process flow is as shown in J = E{e2[n]} = E{d2[n]} − 2WT P + WT RW
Fig. 2. 𝑤ℎ𝑒𝑟𝑒; 𝑊 = (𝑊0 , 𝑊1 , … , 𝑊𝑁−1 ), 𝑇: 𝑓𝑖𝑙𝑡𝑒𝑟 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡𝑠

For each iteration step the filter coefficients for the upcoming
step W(n + 1) are computed using:
W(n + 1) = W(n) + µe[n] ∗ x(n)
For guaranteed convergence for LMS-algorithm, µ taken as
Figure 2: Process flow for detection of heart abnormalities. the maximal amplitude of the reference signal (max(d[n])).

A. Signal Acquisition The upper bound is defined as:

2
The heart sound data base used for training and validation has µmax =
been obtained from University of Massachusetts Medical 3N · max(x)2 .
School(UMMS). The obtained data base had two sets one The LMS-algorithm only converges to a good approximation.
from digital stethoscope and the other from a mobile app. To measure the deviation of approximation from exact
Each set comprised of normal, artifacts, extrahls, murmurs, solution a misadjustment M is introduced. It depends on the
unlabeledunlabeled heart sounds. Further a second public* average power of the reference signal and the step-size µ.
µ
database was used for testing and validating the results. The M = . N. power(x)
sounds in this dataset was recorded using a digital 2
For quick convergence, a small step size is essential, but a
stethoscope DigiScope® during clinical trials.
small step-size leads to a large convergence time. The
B. Adaptive Filtering convergence time τ can be expressed as:
The filtering of the signal is essential due to the inherent
noise in systems. A digital adaptive filter is applied to the 1
τ ≈ ∗ κ(R)
noisy signal to obtain the pure signal as shown in Fig. 4. 4α
The adaptive filter used in filtering is as shown in Fig. 3. where α is the applied step-size normalized to the maximum
Starting point of the process is the unfiltered PCG, x[n]. possible µmax. λmax is the largest, λmin the smallest eigenvalue
𝑥[𝑛] = 𝑃𝐶𝐺 + 𝑛𝑜𝑖𝑠𝑒 of R.
A reference signal which is correlated with d[n] and with
uncorrelated with x[n], is used to eliminate the interference. λmax
The major components of noise in the PCG signal are: 50 Hz κ(R) = 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑛𝑢𝑚𝑏𝑒𝑟
λmin
power supply hum, lung noises, gastrointestinal noises, noise
The proposed LMS-algorithm shows slow convergence for
due to breathing. In order to suppress the above noises a
non-smooth signals and fast convergence for uniformly
signal d[n] is gripped at the power supply and from other
distributed spectrum signals.
known noisy signals. It has the same frequency, but different
Here the filter order is set as: N = 4
The unfiltered PCG signal and filtered PCG signal is as The Conjugate Gradient algorithm which has been applied
shown in Fig: 4. can be suumed up as:
1. Initialize initial weigt vector 𝑤1 .
𝑆𝑒𝑡 𝑝1 = 𝑟1 = −𝐸 ′ (𝑤1 ), 𝑘 = 1.
2. 𝑆𝑒𝑐𝑜𝑛𝑑 𝑜𝑟𝑑𝑒𝑟 𝑖𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛 𝑐𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑖𝑜:
𝑆𝑘 = 𝐸 ′′ (𝑤𝑘 )𝑝𝑘 ,
§𝑘 = 𝑝𝑘𝑇 𝑆𝑘 .
3. 𝑆𝑡𝑒𝑝 𝑠𝑖𝑧𝑒 𝑐𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑖𝑜𝑛:
µ𝑘 = 𝑝𝑘𝑇 𝑟𝑘 ,
µ𝑘
𝛼𝑘 = .
§𝑘
4. 𝑊𝑒𝑖𝑔ℎ𝑡 𝑣𝑒𝑐𝑡𝑜𝑟 𝑢𝑝𝑑𝑎𝑡𝑖𝑜𝑛:
𝑤𝑘+1 = 𝑤𝑘 + 𝛼𝑘 . 𝑝𝑘 ,
𝑟𝑘+1 = −𝐸 ′ (𝑤𝑘+1 ).
5. 𝑖𝑓 𝑘 𝑚𝑜𝑑 𝑁 = 0 ∶ 𝑝𝑘+1 = 𝑟𝑘+1 𝑒𝑙𝑠𝑒 𝑛𝑒𝑤 𝑐𝑜𝑛𝑗𝑢𝑔𝑎𝑡𝑒
𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 𝑖𝑠 𝑐𝑟𝑒𝑎𝑡𝑒𝑑 𝑎𝑠:
𝑇
|𝑟𝑘+1 |2 − 𝑟𝑘+1 ∗ 𝑟𝑘
𝛽𝑘 = ,
µ𝑘
Figure 4: Unfiltered and filtered PCG signals. 𝑝𝑘+1 = 𝑟𝑘+1 + 𝛽𝑘 ∗ 𝑝𝑘
6. 𝐼𝑓 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 𝑜𝑓 𝑠𝑡𝑒𝑒𝑝𝑒𝑠𝑡 𝑑𝑒𝑐𝑒𝑛𝑡 𝑟𝑘 ≠ 0, 𝑠𝑒𝑡
k= 𝑘 + 1 𝑎𝑛𝑑 𝑔𝑜 𝑡𝑜 2𝑛𝑑 𝑠𝑡𝑒𝑝, 𝑒𝑙𝑠𝑒 𝑡𝑒𝑟𝑚𝑖𝑛𝑎𝑡𝑒 𝑡𝑜 𝑟𝑒𝑡𝑢𝑟𝑛
C. Feature Extraction 𝑤𝑘+1 𝑎𝑠 𝑡ℎ𝑒 required minima.
After the signal is appropriately filtered transforming it
into a suitable domain is required to process it and obtain Hence, when a heart sound with abnormally large number
suitable data from it. Considering the Shannon Energy of peaks or a relatively low SNR is given as an input to the
Envelope of the signal as the required parameter, Hilbert algorithm, the system classifies it as an abnormal sound and
Transform iswas chosen. A suitable threshold is then decided thus indicates that the concerned person has a higher chance
for the detection of peaks and consequently the location of S1 of being affected by heart abnormalities.
and S2 sounds. The number of such peaks for the entire
duration of the sample is obtained and taken as a parameter for V. RESULTS AND DISCUSSIONS
training the neural network. The SNR of the signal is also A new way of effective classification of different heart
included as a parameter. abnormalities caused due to Rheumatic fever was shown in
this paper. The raw signal acquired from different sources
Shannon energy, E=-x2log(x2)
were effectively filtered as shown in Fig:4. Further a number
Apart from the above parameters the following parameters of, time domain, frequency domain and statistical features
were extracted for classification as shown in Table:1. The were extracted from the PCG signal. One such feature of peak
peaks detected were as shown in Fig:5. detection is shown in Fig:5. The extracted features were used
to train a Neural network. The results of the training were
TABLE I. LIST OF FEATURES EXTRACTED satisfactory and is shown in Fig:6.
Sl Time Statistical Frequency
no: diomain domain domain features
Features features
1 Total power Mean Peak Frequency
2 Peak Standard Intermodulation
Amplitude Deviation Distortion
3 Zero Varianes Total harmonic
Crossing rate Distortion
4 RMS Skewness Dynamic Range

D. Classification
The extracted parameters are tabulated for the entire sound
bank and fed into an algorithm that employs convolutional
neural networks for characterization of normal heart sounds.
Scaled Conjugate Gradient Backpropagation
algorithm(SCGB) is an ideal choice of neural network for this
application as it is optimized for large data sets for
classification applications [15]. Figure 5:Peak detection of rectified PCG signal.
 TABLE II. T ABLE COMPARING VARIOUS ALGORITHMS.

Algorithm Decision trees KNN SVM SCGB

Type Complex Medium Simple Fine Coarse Cubic Weighted Linear -
Without
adaptive 80.4 76.0 84.2 90.3 90.0 77.0 92.4 93.2 93.1
filter
Accuracy
With
adaptive 83.8 83.8 85.2 94.0 94.0 77.0 95.4 96.8 98.1
filter

A comparison showing the accuracy of our proposed method system capable of classifying and identifying human mistakes
using adaptive filtering and normal filtering followed by made when provided with sufficient data. Thus, by
classification is shown in Table: 2. It is clear from the implementing a Scaled Conjugate Gradient Backpropagation
comparison that the proposed method is more accurate and algorithm with a training set containing false positives, the
can be trained using less quantity of dataset. effectiveness of the algorithm was tested resulting in exact
output for testing data under ideal conditions.
Test results as shown in Fig:6. indicate the effectiveness of the
algorithm to identify heart murmurs even when provided with
false positives in the training. Upon setting the ideal threshold
the algorithm performance is significantly improved to obtain This required the signals to be pre-processed to remove
a 99% success rate in testing. noise achieved through a band pass filter along with the
spectral analysis of the signal using Hilbert transform. The
VI. FUTURE SCOPE effectiveness of the proposed algorithm was shown for the
parameters considered.
The whole procedure introduced in this paper can be
effectively used in the screening for heart abnormalities in
rural areas. In most rural areas pepole lack awareness about
the different heart diseases. A smart stethoscope can be VIII. REFERENCE
integrated with a Digital Signal Processor and it than can be [1] R. L. Watrous, “Computer-aided auscultation of the
used for dynamic assesing of the heart valve abnormalities. heart: From anatomy and physiology to diagnostic
This can be used in telemedicine purposes also. decision support,” Annu. Int. Conf. IEEE Eng. Med.
Biol. - Proc., pp. 140–143, 2006.
[2] M. K. Zia, B. Griffel, V. Fridman, C. Saponieri, and
J. L. Semmlow, “Noise detection in heart sound
recordings,” Proc. Annu. Int. Conf. IEEE Eng. Med.
Biol. Soc. EMBS, pp. 5880–5883, 2011.
[3] D. Song, L. Jia, Y. Lu, and L. Tao, “Heart sounds
monitor and analysis in noisy environments,” 2012
Int. Conf. Syst. Informatics, ICSAI 2012, no. Icsai, pp.
1677–1681, 2012.
[4] F. Andreotti, A. Trumpp, H. Malberg, and S.
Zaunseder, “Improved heart rate detection for
camera-based photoplethysmography by means of
Kalman filtering,” 2015 IEEE 35th Int. Conf.
Electron. Nanotechnology, ELNANO 2015 - Conf.
Proc., pp. 428–433, 2015.
[5] R. a Brown and R. Frayne, “A fast discrete S-
transform for biomedical signal processing.,” Conf.
Proc. IEEE Eng. Med. Biol. Soc., vol. 2008, pp.
2586–2589, 2008.
[6] Z. Yue, J. Liang, and Y. Shen, “Time-frequency
analysis of heart sounds in telemedicine consulting
Figure 6: Confusion Matrix validating the accuracy of training, with target system for auscultation,” 2012 IEEE Int. Conf. Inf.
class named as normal and abnormal. Autom. ICIA 2012, no. June, pp. 652–657, 2012.
[7] A. Dliou, E. Nationale, and E. Nationale,
VII. CONCLUSION “Arrhythmia ECG Signal Analysis using Non
With increasing need for improvement in health Parametric Time-Frequency Techniques,” vol. 41,
technology, doctors are relying heavily on sophisticated no. 4, pp. 25–30, 2012.
equipment as a failsafe for all treatments. This requires a [8] N. Zhang, C. Zhang, and Y. Li, “The analysis of the
 anomaly characteristic of heart sound based on heart
sounds segmentation algorithm,” Proc. - 2012 Int.
Conf. Comput. Sci. Serv. Syst. CSSS 2012, pp. 1339–
1342, 2012.
[9] H. Ali, “Heart Sound Signal Modeling and
Segmentation based on Improved Shannon Energy
Envelogram using ...,” no. October, 2007.
[10] E. F. Gomes and E. Pereira, “Classifying heart
sounds using peak location for segmentation and
feature construction,” Aistats, no. 1, pp. 1–5, 2012.
[11] A. Castro, P. Gomes, S. S. Mattos, and M. T.
Coimbra, “Comparison between users of a new
methodology for heart sound auscultation,” Proc.
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS,
vol. 2016–Octob, no. Norte 2020, pp. 5388–5391,
2016.
[12] J. Pedrosa, A. Castro, and T. T. V. Vinhoza,
“Automatic heart sound segmentation and murmur
detection in pediatric phonocardiograms,” 2014 36th
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBC
2014, vol. 2012, pp. 2294–2297, 2014.
[13] A. Castro, S. S. Mattos, and M. T. Coimbra,
“Noninvasive Blood Pressure and the Second Heart
Sound Analysis,” pp. 5494–5497, 2014.
[14] A. Castro, T. T. V. Vinhoza, S. S. Mattos, and M. T.
Coimbra, “Heart sound segmentation of pediatric
auscultations using wavelet analysis,” Proc. Annu.
Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS, pp.
3909–3912, 2013.
[15] M. F. Møller, “A scaled conjugate gradient algorithm
for fast supervised learning,” Neural Networks, vol.
6, no. 4, pp. 525–533, 1993.