Available via license: CC BY 3.0
Content may be subject to copyright.
Article citation info:
Bourouhou A, Jilbab A, Nacir C, Hammouch A. Classification of cardiovascular diseases using dysphonia measurement in speech.
Diagnostyka. 2021;22(1):31-37. https://doi.org/10.29354/diag/132586
31
DIAGNOSTYKA, 2021, Vol. 22, No. 1
ISSN 1641-6414
e-ISSN 2449-5220
DOI: 10.29354/diag/132586
CLASSIFICATION OF CARDIOVASCULAR DISEASES USING
DYSPHONIA MEASUREMENT IN SPEECH
Abdelhamid BOUROUHOU, Abdelilah JILBAB, Chafik NACIR, Ahmed HAMMOUCH
University Mohammed V, Ecole Normale Superieure de l'Enseignement Technique, Rabat, Morocco
e-mail: abdelhamid.bourouhou@um5s.net.ma
Abstract
Cardiovascular disease is the leading cause of death worldwide. The diagnosis is made by non-invasive
methods, but it is far from being comfortable, rapid, and accessible to everyone.
Speech analysis is an emerging non-invasive diagnostic tool, and a lot of researches have shown that it is
efficient in speech recognition and in detecting Parkinson's disease, so can it be effective for differentiating
between patients with cardiovascular disease and healthy people?
This present work answers the question posed, by collecting a database of 75 people, 35 of whom
suffering from cardiovascular diseases, and 40 are healthy. We took from each one three vocal recordings of
sustained vowels (aaaaa…, ooooo… .. and iiiiiiii… ..). By measuring dysphonia in speech, we were able to
extract 26 features, with which we will train three types of classifiers: the k-near-neighbor, the support
vectors machine classifier, and the naive Bayes classifier.
The methods were tested for accuracy and stability, and we obtained 81% accuracy as the best result
using the k-near-neighbor classifier.
Keywords: Cardiovascular disease, speech analysis, dysphonia measurement, classification methods, PCA features selection
1. INTRODUCTION
A healthy human body returns to perfect blood
circulation; the engine element behind this
circulation is the heart. The heart is a muscle that
pumps blood throughout the body.
Cardiovascular disease is abnormalities that
harm the normal functioning of the heart; they are
disorders that affect the heart and blood vessels. We
can cite, as examples, coronary heart disease,
cerebrovascular disease, peripheral arteriopathy,
rheumatic heart disease, and venous thrombosis.
The world heart federation has declared that
cardiovascular disease (CVD) is responsible for
17.5 million deaths per year worldwide. This death
rate has put CVD at the top of the world's death
causes. The world health organization (WHO)
predicts that by 2030 the number of deaths will
reach 23,6 million, and 1 of 10 people aged from 30
to 70 years old will die prematurely from
cardiovascular disease. On the other hand, 80% of
premature deaths could be avoided or delayed [1].
The following figure [2] shows a top 10 causes of
death worldwide for the year 2017, published by
the Institute for Health Metrics and Evaluation
(IHME).
The 80% that can be avoided, requires early
detection of CVD. For this reason, we need a
reliable, precise, fast, and inexpensive tool that will
make a distinction between cardiovascular patients
and healthy people.
In order to respond to this task, a lot of
researches have been carried out. Some of them
tried to develop an automatic diagnosis based on
the techniques of medical imaging,
echocardiography, or magnetic resonance imaging
(MRI). Other researches were based on the analysis
of electrocardiogram signals (ECG) [3,4], while
some others chose phono-cardiogram signals (PCG)
[5,6]. The starting information and methods are
different, but the goal remains the same – the
development of an automatic diagnostic tool. All
the previously provided methods used one or more
information from the clinical examination of the
individual concerned.
It has been suggested that the characteristics of
the voice signal are associated with a number of
different disease entities, including dyslexia,
attention deficit hyperactivity disorder, Parkinson's
disease and other neurological disorders [7, 8, 9].
Why is not it the case with CVD?
So, we came up with the idea of achieving the
same goal but using a simple information source
that does not require a clinical examination. This
source is the human speech. Successful research
into Parkinson's detection and speech recognition
has inspired us, whence the present paper.
Clinical practice uses sustained vowels to assess
the quality of the voice; it is for the speaker to
pronounce a vowel maintained as long as possible
and at a comfortable level. [10] Healthy people can
stationary produce a sustained vowel, while this is
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
32
not the case for those with vocal impairments. [11,
12, 13]
And to distinguish between healthy and sick
people using speech requires finding the difference
in spoken speech. Therefore, we must study the
speech dysfunctions of the two teams.
Fig. 1. Top 10 causes of death worldwide for
2017 from World Health Organization [2]
2. DATABASE DESCRIPTION
To conduct research on the differentiation
between healthy people and patients with
cardiovascular disease (CVD), we need a database
with these two categories of people.
The first challenge encountered was collecting
this database because, at first, it took a long time
before convincing the director of the CHU as well
as the head of the cardiology department.
Afterwards, we had to wait for the hospitalization
of cardiovascular patients, the end of their treatment
period and the arrival of a new group of patients.
Our database consists mainly of 75 people, 40
of whom are healthy and 35 suffer from
cardiovascular disease (CVD). We took 3 voice
recordings for each individual pronouncing
sustained vowels (aaaaa ..... / oooooo ...... / iiiiiii
......). The duration of each recording is 5s, and the
used device is a smartphone.
The following table summarizes the entire
contents of the database:
Table 1. Database description
Database
Total person
75
Healthy
Patient
40
35
Gender
Women
Man
48%
52%
Age average
55,5 years
Healthy
Patient
55 years
56 years
Total records
225
Sampling
frequency
44100 Hz
3. METHODS
To distinguish between the people with (CVD)
and the healthy ones using speech, we used
Dysphonia measures for each of the two categories.
Before proceeding to the measures of the
Dysphonia, we filtered and fragmented the
recordings in order to keep only the useful part.
Then, we proceeded to the Dysphonia
measurements which will be the subject of 26
characteristics extracted from speech for each
recording; these built the training base for our used
classifiers (k-NN, Naive Bayes, SVM). Validation
was ensured by the k-folds cross-validation
technique. The diagram in fig.2 presents all steps of
our approach.
3.1. Signal acquisition
In this part, we used a smartphone, equipped by
MEMS technology microphone, model (CMM-
3729 AB-38308-TR) to obtain the records. This
microphone operates in a bandwidth of [100 ....
10,000 Hz], has 65 dBA signal to noise ratio
measure at 1 kHz by 94 dBA signal, and has 0.2%
total harmonic distortion measure at 1 kHz by 94
dB signal. [14]
3.2. Segmentation & filtering
After the signal acquisition, we proceed on two
essential steps. The first one consists of signal
segmentation because our records take between 6
seconds and 10 seconds for each one, and contains
some additional useless sounds. So, we propose a
segmentation for each record to keep just the 5
seconds which represents the sustained vowel.
The second step is filtering because our records
are a speech (vocal production), and this vocal
production is made up of sounds with very specific
frequency components, often given as the
fundamental frequency (F0, corresponding to the
carrier signal), and the first formants (Fi, spikes in
the spectral amplitude due to the resonances of the
duct voice). The vowels –in our case- are very
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
33
easily characterized by their formants. The
frequencies used by human speech can therefore be
between 110 and 7 kHz (speech communications).
But we filter the signal to keep just the
bandwidth [300-3400 Hz] because our speech
records contain sustained vowels, and the chosen
bandwidth includes the fundamental frequency and
the first 3 formants, which is sufficient for us.
So, we applied a Butterworth high pass filter
with a cutoff frequency at 300Hz and seventh order
succeeded by Butterworth low pass filter with a
cutoff frequency at 3400Hz, and also with seventh
order. The Butterworth filter choice came from its
characteristics, the cutoff frequency is the same
regardless of the filter order; the response in the
bandwidth is very flat, and the attenuation slope can
be increased by increasing the filter order.
Fig. 2. Diagram of the used method
3.3. Dysphonia measurement
In this step, we try to extract 26 features from
each record. These features built the training
datasets and were the keys to classify entities into
healthy and sick with (CVD).
The 26 features are linear and non-linear
parameters; we had frequency parameters,
amplitude parameters, harmonic parameters, and
some specific parameters dedicated to vocal
disorder analysis.
We used the “Praat” software to extract all
parameters, and the following table sums up all
extracted features:
Table 2. Features extracted from voice signal
Features
Description
Pitch
1. Median (Hz)
Depends on the
number of
vibrations per
second produced
by the vocal
cords.
The main
acoustic
correlate of tone
and intonation
2. Mean (Hz)
3. Standard
deviation (Hz)
4. Maximum (Hz)
5. Minimum (Hz)
Pulses
6. Number of
pulses
7. Number of
periods
8. Mean of
periods (s:
seconds)
9. Standard
deviation of
periods (s)
Jitter
10. Absolute (s)
Fundamental
frequency
variation
measurements
11. Local (%)
12. RAP (%)
13. PPQ5 (%)
14. DDP (%)
Shimmer
15. Local (%)
Amplitude
variation
measurements
16. Local (dB)
17. APQ3 (%)
18. APQ5 (%)
19. APQ11(%)
20. DDA (%)
Harmonicity
21. Mean
autocorrelation
22. Mean noise to
harmonics ratio
23. Mean
harmonics to
noise ratio (dB)
Intensity
24. Mean (dB)
25. Maximum (dB)
26. Minimum (dB)
The table below presents the used mathematical
expressions to extract features:
Table 3 Mathematics formulas of the extracted
features
Mathematical expression
Jitter
Signal acquisition
Segmentation & filtering
Dysphonia measurement
Validation & decision
Train
SVM
classifie
r
Train
k-NN
classifie
r
Train
Naive-
Bayes
classifier
Features selection by ACP
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
34
Shimmer
Harmonicity
Where, Ti represents the length of period, N
represents the number of periods and Ai represents
the Peak-to-peak amplitude.
3.4. Features selection by PCA
Now, we have a training matrix where columns
represent 26 extracted features and each row is an
observation (a record). To get reliable results in
classification, we proceeded into a features
selection technique, firstly, to reduce the matrix
and, secondly, to increase the accuracy of
discrimination.
The used technique is the Principal Component
Analysis (PCA); it’s the mostly used statistical
technique in data analysis and data compression.
The main idea consists of a data projection from the
original space of D variables to a subset
characterized by d variables uncorrelated; this
subset contains the principal components and
conserves the information contained in the original
space.
In our case, we used an algorithm of PCA
selection to reduce features number from 26 to 23.
The choice of 23 is not random, but it is determined
by running the PCA algorithm many times
changing the number of principal components to
keep in each time. Then, we compared the
classification result for all the tries. We noted that
the PCA features selection reduces the number of
features by computing automatically new “n”
features (n can be a number from 1 to 26) from the
26 original features. In our case the best projection
was for 23 new features computed automatically
from the original 26 features because it provides the
best result.
3.5. Training & Validation of Classifiers
Finally, we had to train and validate our
classifiers. We used 3 types of classifiers that are
the k near neighbor, the support vector machine,
and the Naive Bayes. All classifiers are subjected to
the same cross-validation technique.
3.5.1. Training phase
We have 75 people from whom we took 3
records each, so we collected 225 records. Then we
extracted from each record 26 features which was
reduced to 23 by the PCA features selection
technique. Thus, we built a training dataset (225
rows = 225 observations; 23 columns = 23 features)
for our used classifiers.
3.5.2. The k-NN classifier
It is one of the most used classifiers in machine
learning; it is meant to determine the k which
represents the number of neighbors to take into
consideration in the classification. A new entity will
be classified as the same class as the majority of
neighbors.
The nearest neighbor from the entity to classify
depends on the distance which can be Euclidean,
Cosine, Minkowski, or other types of distance. So,
the number of neighbor k and the type of distance
must be chosen carefully because it influences the
accuracy of classification.
We applied an optimization algorithm with the
aim to select the optimum parameters to train our k-
NN classifier. And the chosen parameters were
Cosine distance and k=5 as a number of the
neighbor.
3.5.3. The SVM classifier
It is a technique intended for discrimination and
regression problems. Generally, this type of
classifier is used for two-class discrimination
problems but can be extended for multiclass
problems.
The concept consists in building a decision limit
named hyperplane separator in the features space,
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
35
maximizing the margin between samples from two
different classes. These hyperplane separators are
the support vectors, which determine the class of
the new entity.
To calculate the support vectors, the SVM
classifier algorithm uses different kernels, and each
kernel has specific proprieties. Examples of these
kernels are the Gaussian kernel, the RBF kernel,
and the Linear kernel. The choice of the kernel is
crucial for prediction accuracy, so we run an
optimization algorithm to get the most appropriate
SVM classifier proprieties, and the result was the
Gaussian kernel.
3.5.4. The Naïve Bayes classifier
The Naïve Bayes classifier is a type of classifier
probabilistic based on the theorem of Bayes; it is
simple and belongs to the linear classifier family.
The prediction accuracy result for this classifier
depends on the kernel choice, Gaussian kernel,
Triangular kernel, Box kernel, or Epanechnikov
kernel. The difference between all these kernels is
determined by the formulas used in the algorithm.
The best result using the Naïve Bayes classifier was
for the Triangular kernel.
3.5.5. Validation phase
To validate our classifiers, we proceeded into k-
folds cross-validation. This type of cross-validation
consists of a random subdivision of the database by
“k”, keep one of “k” subsets for validation, and
training the classifier by all the rest “k-1” subsets.
We repeated the operation “k” times until all the
subsets may be used one time as a validation set;
we calculated the performance score each time. The
mean of the “k” squared errors average is finally
calculated to estimate the prediction error. In our
case the k=5.
After building the cross-validation model, the
judgment consists of calculating three essential
parameters – the accuracy, the sensitivity, and the
specificity; their formulas are given below:
4. RESULTS & DISCUSSION
In this section, we will present the findings.
Fig.3 below presents two extracted features, from
people with CVD and from healthy ones plotted
together to note the difference between our two
classes of people.
In the first part of the figure, we plotted the
standard deviation of the pitch for records from the
people with CVD and for the healthy one records.
In the second part of the figure, we have the
standard deviation of the period for the two classes.
We can see that the area occupation for the
people with CVD is more important and differs
significantly from that of the healthy people. So, we
come to conclude that the distinction can be done
using speech. Table 4 describes the reached results
for a first classification try, using all 26 features
extracted from records to train classifiers.
Table 4 Classifiers result without application of
features selection method
Classifier
Accuracy
(%)
Sensitivity
(%)
Specificity
(%)
K- Near
Neighbor
78.46
80.07
76.85
Support
Vectors
Machine
70.54
69.08
72.00
Naïve Bayes
63.88
65.45
62.31
Fig. 3. Comparison between healthy people
& people with CVD
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
36
We have proceeded to a first classification try,
by using all the extracted features, which provides
us a training matrix 26*225, where 225 is the
number of observations. We have subjected this
matrix to different classifiers algorithms. The best
provided result was reached by the KNN classifier
that means it is the most adapted classifier for our
database.
Fig.4 below shows the confusion matrix of the
k-near-neighbor classifier; this confusion matrix
specifies that our KNN classifier can detect the true
positive feature by 80.7%, and the true negative one
by only 76.85%. These results reveal the relation
supposed to be holding between speech and CVD,
and also that we have to improve the classification
accuracy.
Fig. 4. Confusion matrix for the k-NN
Classifier before features selection method
To increase the classification accuracy, we have
proceeded to a PCA features selection technique.
Table 5 describes the results for our second try of
classification; this time we trained our classifiers by
the matrix generated by the PCA algorithm, which
contains 23 principal components instead of our 26
extracted features.
Table 5 Classifiers result after application of features
selection method “PCA”
Classifier
Accuracy
(%)
Sensitivity
(%)
Specificity
(%)
K- Near
Neighbor
81.51
82.46
80.56
Support
Vectors
Machine
75.28
78.50
72.06
Naïve Bayes
70.45
70.00
70.09
The used features selection technique has
improved all the performances of our used
classifiers, and the KNN classifier is still the best.
Fig.5 shows the confusion matrix of the k-near-
neighbor classifier after PCA technique of features
selection.
We can also plot the Receiver Operating
Characteristic of our three classifiers to compare
the results achieved. Fig. 6 below describes that
comparison.
As shown in the figure, the area under the plot
of the KNN classifier is larger than those of the
SVM and Naïve Bayes classifiers; therefore, we
conclude that the KNN classifier is better in this
case.
Fig. 5 Confusion matrix for the k-NN
Classifier after using "PCA" algorithm
Fig. 6. ROC Curve for all classifiers result
after PCA method
5. CONCLUSION
The CVD is still the cause number one of death
worldwide, but we can avoid 80% of death if we
detect earlier people with CVD, and to ensure
detection we have to make the assessment and
diagnostic precise, fast, and inexpensive.
To achieve this, we used the speech as a tool of
diagnostic, to distinguish between people with
CVD and healthy ones. So, we have collected a
database that contains multiple records from
different people who pronounce sustained vowels
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
37
/a/, /o/ and /i/. Then, we have extracted 26 voice
features from each record.
To improve the assessment of CVD, we have
reduced the training matrix which contains 26
features and 225 observations, to another matrix
23*222 by the PCA feature selection technique.
We have used 3 classifiers, the K-near-
neighbor, the support vectors machine, and the
Naive Bayes. Choosing the K- near neighbor and
cosine distance made the KNN classifier the most
successful classifier. The best classification
accuracy result was 81.50%.
REFERENCES
1. https://www.euro.who.int/en/health-
topics/noncommunicable-diseases/cardiovascular-
diseases/data-and-statistics
2. https://ourworldindata.org/what-does-the-world-die-
from
3. Rawther NN, Cheriyan J. Detection and classification
of cardiac arrhythmias based on ECG and PCG using
temporal and wavelet features. IJARCCE. 2015; 4(4).
4. Bouguila Z, Moukadem A, Dieterlen A, Ahmed
Benyahia A, Hajjam A, Talha S, Andres E.
Autonomous cardiac diagnostic based on
synchronized ECG and PCG signal. In: 7th
International Joint Conference on Biomedical
Engineering Systems and Technologies—ESEO,
Angers. 2014
5. Ghassemian H, Kenari AR. Early detection of
pediatric heart disease by automated spectral analysis
of phonocardiogram in children. J. Inf. Syst.
Telecommun. 2015; 3(2): 66–75.
6. Nabih-Ali M, El-Dahshan E-SA, Yahia AS. Heart
diseases diagnosis using intelligent algorithm based
on PCGsignal analysis. Circuits Syst. 2017;
8(7):184–190.
7. Levanon Y, Lossos-Shifrin L. Inventors; Google,
assignee. Method and system for diagnosing
pathological phenomenon using a voice signal 2008.
US patent 7,398,213 B1.
8. Bonneh YS, Levanon Y, Dean-Pardo O, Lossos L,
Adini Y. Abnormal speech spectrum and increased
pitch variability in young autistic children. Front
Hum Neurosci. 2011;4:237.
9. Uma Rani K, Holi MS. Automatic detection of
neurological disordered voices using mel cepstral
coefficients and neural networks. In: 2013 IEEE
Point-of-Care Healthcare Technologies (PHT)
2013:76-79. Bangalore, India, 2013.
10. Titze Ingo. Phonation into a straw as a voice building
exercise. Journal of Singing. 2000; 57: 27-28.
11. Cnockaert L, Schoentgen J, Auzou P, Ozsancak C,
Defebvre L, Grenez F. Low-frequency vocal
modulations in vowels produced by Parkinsonian
subjects, Speech Communication. 2008;50(4):288-
300. https://doi.org/10.1016/j.specom.2007.10.003
12. Little MA, McSharry PE, Hunter EJ, Spielman J,
Ramig, LO. Suitability of dysphonia measurements
for telemonitoring of Parkinson's Disease. IEEE
Transactions on Biomedical Engineering.
2009;56(4):1015-1022.
https://doi.org/10.1109/TBME.2008.2005954
13. Tsanas A, Little MA, McSharry PE, Spielman J,
Ramig LO. Novel speech signal processing
algorithms for high-accuracy classification of
Parkinson's disease. IEEE Trans Biomed Eng.
2012;59(5):1264-1271.
https://doi.org/10.1109/TBME.2012.2183367
14. https://www.cuidevices.com/product/resource/cmm-
3729ab-38308-tr.pdf
15. Bourouhou A, Jilbab A, Nacir C, Hammouch A.
Detection and localization algorithm of the S1 and S2
heart sounds. 2017 International Conference on
Electrical and Information Technologies (ICEIT),
Rabat. 2017:1-4
https://doi.org/10.1109/EITech.2017.8255217
16. Bourouhou A, Jilbab A, Nacir C, Hammouch A.
Comparison of classification methods to detect the
Parkinson disease. 2016 International Conference on
Electrical and Information Technologies (ICEIT),
Tangiers, 2016:421-424.
https://doi.org/10.1109/EITech.2016.7519634
17. Bourouhou A, Jilbab A, Nacir C, Hammouch A.
Heart Sounds classification for a medical diagnostic
assistance. International Journal of Online and
Biomedical Engineering (iJOE) 2019; 15(11): 88–
103.
Received 2020-09-27
Accepted 2021-01-19
Available online 2021-01-25
Abdelhamid BOUROUHOU
was born in Rabat, Morocco on
December 26th, 1989. Received
the Master degree in Electrical
Engineering from ENSET,
Rabat Mohammed V
University, Morocco, in 2014
he is a research student of
Sciences and Technologies of
the Engineer in ENSIAS,
Research Laboratory in Electrical Engineering LRGE,
Research Team in Computer and Telecommunication
ERIT at ENSET, Mohammed V University, Rabat,
Morocco. His interests are in sounds classification for
medical diagnostic assistance.
Abdelilah JILBAB Professor at
ENSET Rabat, Morocco; he
graduated in electronic and
industrial computer aggregation in
1995. Since 2003, he is a member
of the laboratory LRIT (Unit
associated with the CNRST, FSR,
Mohammed V University, Rabat,
Morocco). He acquired his PhD in
Computer and Telecommunication from Mohammed V-
Agdal University, Rabat, Morocco in 2009. His domains
of interest include signal processing and embedded
systems.
NACIR Chafik Teacher
Researcher in Mathematics.
Former Head of the Department of
Mathematics and Computer
Science.
Former member of the Scientific
Commission ENSET of Rabat
Morocco.
DIAGNOSTYKA, Vol. 22, No. 1 (2021)
Bourouhou A, Jilbab A, Nacir C, Hammouch A.: Classification of cardiovascular diseases using …
38
Ahmed HAMMOUCH received
the master degree and the PhD in
Automatic, Electrical, Electronic
by the Haute Alsace University
of Mulhouse (France) in 1993
and the PhD in Signal and Image
Processing by the Mohammed V
University of Rabat in 2004.
From 1993 to 2013 he was a
professor in the Mohammed V
University in Morocco. Since 2009 he manages the
Research Laboratory in Electronic Engineering. He is an
author of several papers in international journals and
conferences. His domains of interest include multimedia
data processing and telecommunications. He is with
National Center for Scientific and Technical Research in
Rabat.
