Two-stage Artificial Intelligence Clinical Decision Support System for

Cardiovascular Assessment using Convolutional Neural Networks

and Decision Trees

Shahab Pasha

, Jan Lundgren

Marco Carratù

, Patrik Wreeby

and Consolatina Liguori

STC Research Centre, Mid Sweden University, Sundsvall, Sweden

Department of Industrial Engineering, University of Salerno, Fisciano, Italy

Premicare AB, Sörberge, Sweden

Keywords: Artificial Intelligence, Cardiovascular Assessment, Decision Trees, Deep Learning, Feature Selection.

Abstract: This paper describes an artificial-intelligence–assisted screening system implemented to support medical

cardiovascular examinations performed by doctors. The proposed system is a two-stage supervised classifier

comprising a convolutional neural network for heart murmur detection and a decision tree for classifying vital

signs. The classifiers are trained to prioritize higher-risk individuals for more time-efficient assessment. A

feature selection approach is applied to maximize classification accuracy by using only the most significant

vital signs correlated with heart issues. The results suggest that the trained convolutional neural network can

learn and detect heart sound anomalies from the time-domain and frequency-domain signals without using

any user-guided mathematical or statistical features. It is also concluded that the proposed two-stage approach

improves diagnostic reliability and efficiency.

1 INTRODUCTION

The healthcare sector is entering the age of global

artificial intelligence (AI) convergence (Morsy,

2018). Benefiting from mobile applications and large-

scale data, technologies such as Blue Button (website

of the Office of the National Coordinator for Health

Information Technology, n.d.) and IBM Watson

(IBM Watson, n.d.) help track treatment progress and

reduce wrong diagnoses. In healthcare, AI covers a

wide range of applications, including screening (Lin,

Chang, Lin, Tsai, & Chen, 2017), monitoring (Ivascu,

Cincar, & Negru, 2017), and diagnosis (Islam

Chowdhuryy, Sultana, Ghosh, Ahamed, &

Mahmood, 2018). Of the applied AI methods, deep

learning techniques (Gharehbaghi & Lindén, 2017;

Loh & Then, 2017) seem to be more adaptable,

accurate, and robust in a wide range of applications

and biological signals, such as lung sound

classification (Chen, Zhang, Tian, Zhang, Chen, &

Lei, 2016), cardiac auscultation (Amiriparian,

Schmitt, M., Cummins, N., Qian, K., Dong, F., &

Schuller, 2018),

phonocardiography (Thomae &

https://orcid.org/0000-0002-6805-166X

https://orcid.org/0000-0003-1819-6200

Dominik, 2016), and vital sign evaluation (Jones,

2013).

Biological signals (e.g., heart sound signals) with

cyclic characteristics often display nonstationary

behavior not only within but across cycles in cycle-

to-cycle variation. This gives the signal a high level

of complexity, making classifier development a major

challenge.

Unlike in many industrial applications, the origin

of the complexities in most biological signals has yet

to be fully understood (Gharehbaghi & Lindén,

2017).

AI-assisted methods applied to biological signals

can be categorized into four main categories: neural

network-based and deep learning classification

(Zabihi, Rad, Kiranyaz, Gabbouj, & Katsaggelos,

2016), support vector machine-based classification

(Barhatte, Ghongade, & Thakare, 2015), hidden

Markov-model–based classification (Vernekar, Nair,

Vijaysenan, & Ranjan, 2016), and clustering-based

classification (Clifford, 2016). Advanced deep

learning techniques can automatically extract salient

patterns directly from the input (i.e., training set) and

use the produced knowledge to classify unseen

samples (Shao, Wu, & Li, 2014).

This

capability

Pasha, S., Lundgren, J., Carratù, M., Wreeby, P. and Liguori, C.

Two-stage Artiﬁcial Intelligence Clinical Decision Support System for Cardiovascular Assessment using Convolutional Neural Networks and Decision Trees.

DOI: 10.5220/0008941801990205

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 199-205

ISBN: 978-989-758-398-8; ISSN: 2184-4305

199

Figure 1: The two stage AI decision support system.

deep networks makes them the right tool with which

to extract and learn particular patterns from diverse

and relatively large training sets, such as heart sound

databases (PhysioNet, 2016). However most existing

AI-assisted heart sound analysis systems use

medically incomprehensible mathematical and

statistical features, such as wavelets (Clifford, 2016),

Stockwell transformation (Moukadem, Dieterlen, &

Brandt, 2013), Mel-frequency cepstral coefficients

(MFCCs) (Chen, 2017), and the likelihood function

(Yamashita, Himeshima, & Matsunaga, 2014).

In this research, a convolutional neural network

(CNN) is trained based on the time-domain and high-

resolution frequency-domain “normal” and

“abnormal” heart sound signals (PhysioNet, 2016).

Along with the heart sound signals, the vital signs and

patient information are analyzed by a C4.5 decision

tree as side information for a comprehensive

cardiovascular assessment. An optimization process

(Visalakshi & Radha, 2014) is applied to find the

optimal subset of vital signs yielding the highest

classification accuracy. The proposed system is

designed for making suggestions to support the

clinical assessment process. The system uses signal

processing and data-mining techniques to maximize

accuracy, using filters designed to find the most

significant frequency band for heart murmurs and to

find the optimized set of vital signs.

The remainder of this paper is organized as

follows. Section 2 briefly describes the

cardiovascular examination procedure as performed

by medical experts. Section 3 describes the proposed

AI system and the deep network structure replicating

the medical procedure described in section 2. In

section 4, the proposed two-stage system is trained

and evaluated. The paper concludes in section 5.

2 CARDIVASCULAR

EXAMINATION AND THE

HEART SOUND MODEL

During a cardiovascular examination, the doctor has

the patient assume either a lying supine or a sitting

position and starts with the general examination,

which includes measuring body temperature, blood

pressure, skin hydration, pulse, and blood oxygen

saturation. Heart sound auscultation with a

stethoscope is the next step after the general

examination. The doctor listens to the heart at the

apex, base (i.e., the part of the heart between the apex

and sternum), and in the aortic and pulmonary areas.

A normal heart sound consists of two fundamental

sounds referred to as 𝑆1 and 𝑆2. 𝑆1 occurs when the

mitral and tricuspid valves close and 𝑆2 occurs when

the aortic and pulmonic valves close. Other sounds

include the third heart sound 𝑆3, the fourth heart

sound 𝑆4, the systolic ejection click (EC), the mid-

systolic click (MC), the diastolic sound or opening

snap (OS), and heart murmurs. Heart murmurs (Vepa,

2009) made by turbulent blood flow in the heart and

blood leakage through the closed valves are a

stationary low-frequency noise that indicates

underlying heart issues. This might happen during the

delay between 𝑆1 and 𝑆2 (i.e., the systole) or

between two consecutive beats (i.e., the diastole).

The heart sound signal is a low-frequency signal

covering a range of frequencies from 20 to 500 Hz, so

a relatively low sampling frequency, i.e. 2000 Hz

according to the PhysioNet database (PhysioNet,

2016), is sufficient.

The generated heart sound traveling through the

chest and being distorted by the respiratory noise is

mathematically modeled as:

𝑥



𝑛



ℎ



𝑛 ∗𝑠



𝑛



ℎ



𝑛 ∗ 𝑙



𝑛



𝑣



𝑛



(1)

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

200

where 𝑥



𝑛



is the signal detected by the stethoscope,

𝑠



𝑛



the heart sound, 𝑙



𝑛



the bronchial/vesicular

lung sounds, and ℎ



and ℎ



the acoustic response of

the chest at the stethoscope location from the heart

and lung sounds, respectively. However, no study has

investigated the conjectured contribution of breathing

sounds to heart auscultation and so far no standard

filtering scheme has been proposed (Dalmay,

Antonini, Marquet, & Menier, 1995).

3 THE TWO-STAGE

CLASSIFICATION SYSTEM

The heart sound signal and the vital signs provide

partial indicators of the patient’s cardiovascular

health. As depicted in Figure 1, the proposed system

is a two-stage classifier that analyzes the heart sound

signals (using CNN) and the vital signs (using

decision trees) for a comprehensive initial

assessment. The scoring scheme (Table 1) labels an

individual “normal” if the heart sound signal and vital

signs are both classified as “normal.”

3.1 Deep Learning for Heart Sound

Anomaly Detection

The problem of classifying heart sounds by deep

learning is formulated as minimizing the loss

function, defined as the difference between the target

vector (i.e., ground truth labels) and predicted results

(i.e., classifier outputs) (5). Let us assume that there

are 𝑁 recordings comprising 𝑐 classes (in this

research 𝑐2, “normal” and “abnormal”) and the

recording length is represented by 𝐷 (in this research

𝐷 ranges from 4000 to 12,000, translating to 2–6

seconds at 𝑓



2 𝑘𝐻𝑧).

x



𝑥



,..,𝑥





(1)

The frequency domain signal with

𝑁



256

frequency bins calculated and averaged over 64 sample

windows with a 32-sample overlap

is represented by:

Table 1: The proposed scoring scheme.

Heart sound Vital signs Score

Normal Normal Normal

Abnormal Normal

Possible anomaly

if recurring

Normal Abnormal

Possible anomaly

if recurring

Abnormal Abnormal

Needs to be

examined by a

docto

𝑥



𝑥



𝑛



𝑒

2

𝑗

𝜋





𝑘𝑛

𝑁1

𝑛0

(2)

where 𝑘∈



1,…,𝑁







𝑥



,..,𝑥





,

(3)

It is observed that normal heart sounds have distinct

peaks (i.e., sharp edges correlated with higher

frequencies) at 𝑆1 and 𝑆2, whereas abnormal heart

sounds have weaker 𝑆1 and 𝑆2 edges distorted by

strong murmurs (Figure 2). In the PhysioNet

database, most abnormal instances are related to heart

valve defects and coronary artery disease patients

(PhysioNet, 2016). The presence of murmurs

increases the heart sound complexity (Moukadem,

Dieterlen, & Brandt, 2013). This difference leads to a

30 dB difference (Figure 2) between the “normal” and

the “ abnormal ” signal spectra. The binary

classifier is trained by the time-domain (2) and

frequency-domain (4) representations of the heart

sound signals and by the ground-truth labels y ∈

{y_Normal,y_Abnormal} provided by the applied

database (PhysioNet, 2016). For a database consisting

of N persons’ data, the matrix of the heart sound

recordings is X_hs∈R^(N×D) and the target vector

is y_hs∈Β^(N×1).

𝐗





𝑥

1,1

…𝑥

1,𝐷

⋮⋱⋮

𝑥

𝑁,1

…𝑥

𝑁,𝐷

. (4)

The predictor’s matrix 𝐗



as in (4) and the target

𝐲



are applied to train the network, which is

mathematically modeled as:

𝑚𝑖𝑛ℓ𝑌,∅



𝑋,𝑊



,…,𝑊







𝜆Ө



𝑊



,…,𝑊







𝑊









,

(5)

where ℓ



𝑌,∅



is a loss function measuring the

agreement between the true output 𝒚 and the

predicted output.

∅



𝐗



,𝐖



. The predicted label for each sample

recording 𝑥 from (1) (rows of 𝐗) is:

𝑦Ѱ𝐱

(6)

where Ѱ and 𝐖 represent the bias and weights of the

deep network, respectively. Table 2 summarizes the

applied deep network architecture of the CNN that

accepts a matrix (i.e., the transformed time-domain or

frequency-domain vector) as input. A standard sound

classification architecture is applied consisting of

three convolutional layers, each followed by batch-

normalization and max-pooling layers, followed by

two fully connected layers before the final

classification (Kim, Lee, & Nam, 2018). The applied

dataset (𝐗



) includes 3126 normal and abnormal

Two-stage Artiﬁcial Intelligence Clinical Decision Support System for Cardiovascular Assessment using Convolutional Neural Networks

and Decision Trees

201

heart sound recordings sampled at 𝑓



2000 𝐻𝑧.

The length of the database recordings (PhysioNet,

2016) varies, but to maintain consistency, all the

experiments are done with training sets and test sets

of the same lengths (i.e., 2𝑠,…,6𝑠). Having 𝑥



𝑛



)

Figure 2: Power spectrum of a) normal heart sound with

distinct peaks and b) weaker abnormal heart sound.

Table 2: Summary of the CNN configuration.

Layer Type Output

Kernel

ize

Stride

1 Convolution

12888 8 1

2 BatchNorm

12888  

Max

pooling

64  4  8 4 2

4 Convolution

64  4  8 8 1

5 BatchNorm

64  4  8  

Max

pooling

32  2  8 4 2

7 Convolution

32  2  8 8 1

8 BatchNorm

32  2  8  

Max

pooling

12  1  8 4 2

10 Flatten

96  

11 Softmax

2  

from (1), the low-pass ( 𝑔



) and high-pass ( 𝑔



)

filtered signals are given by:

𝑥





𝑛



𝑔





𝑛



∗𝑥



𝑛



(7)

The training set 𝐀



∈ℝ



,𝑀≅0.7𝑁 consists

of 70% of the dataset. The heart sound vectors are

transformed into matrices to match the 2D input layer

of the CNN (Table 2) The initial learning rate is set to

0.01 with a maximum of 250 epochs. All convolution

layers are zero padded to preserve the input

dimension.

3.2 Decision Tree for the Vital Signs

Let



𝐗



,𝐲





denote the vital signs of 𝑁 individuals

and the ground-truth targets; 𝛉 represents the

classification parameters and ∁ the cost function. The

goal of the classification problem is to find the

classifier 𝑓𝐱



,𝛉





where 𝛉



argmin

,

∁𝛉. Thirteen vital signs ( Ζ )

(University of California Irvine Machine Learning

Repository, n.d.) associated with heart defects are

considered in this research. Medical experts believe

that these 13 vital signs and additional patient

information can help in understanding the heart defect

cause and type, if measured or provided as side

information during a cardiovascular examination.

The C4.5 decision tree (i.e., the advanced version

of the ID3 algorithm) is applied in this work to create

models including both continuous numerical data

(e.g., age and serum cholesterol) and categorical data

(e.g., sex and chest pain type). The process of

inducing the decision tree includes calculating the

information gain and the gain ratio of each feature,

which shows the significance of each feature for the

classification. Finding the right threshold for each

feature depending on the data distribution is also part

of the inducing process. The applied dataset includes

thirteen features ( Ζ ) of 270 adults (i.e., 300

augmented instances). There are





⋯





1891 possible combinations of features used

in performing the classification. The feature selection

process of this research (Table 4) concludes that using

only the six features highlighted in Table 3

outperforms all other possible combinations and also

reduces the computation cost. This feature selection

improves the classification accuracy (11) by 11

percentage points from 72% to 83% compared with

the scenario in which all thirteen features are applied.

Information gain and entropy are applied to decide

the most significant vital signs. The entropy of each

vital sign is defined as:

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Normalized Frequency ( radians/sample)

-90

-80

-70

-60

-50

-40

-30

-20

Fres =0.0019536 radians/sample

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Normalized Frequency ( radians/sample)

-110

-100

-90

-80

-70

-60

-50

-40

Fres =0.0019536 radians/sample

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

202

𝐻𝐗

𝑣𝑖

𝑝



log



𝑝



(8)

where 𝑝











is the probability of each class in 𝐗



Using 𝐻, the information gain (∆𝐻) for a vital sign

(𝑧∈Ζ) is given by:

∆𝐻

𝐻



𝐗





𝐻



𝐗



,𝑧



(9)

where 𝐻



𝐗



,𝑧



is the entropy of 𝐗



classified

according to the feature 𝑧

. Accuracy and other

classification measurements are then calculated by:

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦

𝑇



𝑇



𝑇



𝑇



𝐹



𝐹



(10)

𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦𝑇𝑃𝑅













(11)

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦𝑇𝑁𝑅













(12)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑃𝑃𝑉

𝑇



𝑇



𝐹



(13)

𝐹𝑠𝑐𝑜𝑟𝑒2 





(14)

Table 3: The applied vital signs and patient data.

Vital sign

Information

gain

1. Age

2. Sex*

3. Chest pain*

4. Resting blood pressure

5. Serum cholesterol

6. Fasting blood sugar

7. Resting electrocardiographic results

8. Maximum heart rate achieved

9. Exercise-induced angina *

10. Old peak

11.Slope of the peak exercise ST

segment *

12. Number of major vessels colored

by fluoroscopy*

13. Thalassemia *

0.015

0.222

0.266

0.015

0.052

0.008

0.060

0.079

0.120

0.089

0.194

0.188

0.218

Table 4: Feature selection algorithm.

Calculate ∆𝐻 (9) for all the vital signs Ζ using 𝐗



and

𝐲



Sort the features based on their ∆𝐻 (Table 3) in

descending order







For 𝑖2 to 13 (length of Ζ



:

 Classify the test set using the first 𝑖 features from

Ζ



 and calculate the classification accuracy

(𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦



𝑖



) (10)

 Stop the process if 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦



𝑖



𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑖



4 EVALUATION AND

DISCUSSION

The CNN classifier and the decision tree are

evaluated using unseen samples that were excluded

from the training set. The effect of the heart sound

recording lengths, filtering and feature selection are

investigated in the evaluation process. As the two

stages consume different types of signals they are

independently trained and evaluated.

4.1 CNN Evaluation

The trained CNN was able to correctly classify an

average of 74.19%

(10) of 660 unseen samples of

different frame lengths (Figure 3) from the database

and from recordings made using a Thinklabs One

digital stethoscope (Thinklabs, n.d.) down-sampled

from 44.1 𝑘𝐻𝑧 to 2 𝑘𝐻𝑧 . The F-score

(14) is

calculated for the time-domain (1) and frequency-

domain signals (3) in Figure 3 for different test set

frame lengths. It is important to note that the false

negatives (i.e., individuals with heart issues who are

not detected by the classifier) are the main concern of

medical experts, and the goal of the assessment

process is to minimize the false negative rate, i.e.,

𝐹𝑁𝑅1  𝑇𝑃𝑅

(11). It is shown that the minimum

required heart sound auscultation length for obtaining

reliable results is 5~6 seconds (Figure 4).

Figure 3: Heart sound signal classification evaluation by

CNN.

A high-pass filter (500–1000 Hz) and a low-pass filter

(0–500 Hz) with 80𝑑𝐵 attenuation per octave are

applied to coarsely find the more discriminative

frequency band. A more thorough investigation of the

frequency bands is conducted by using low-pass and

high-pass filters with varying pass-band frequencies

22.533.544.555.56

Frame len

(

Seconds

)

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Frequecy domain-Full band

Frequecy domain-Lower band

Frequecy domain-Upper band

Time domain-Full band

Time domain-Lower band

Time domain-Upper band

Two-stage Artiﬁcial Intelligence Clinical Decision Support System for Cardiovascular Assessment using Convolutional Neural Networks

and Decision Trees

203

ranging from a minimum of 200 Hz to a maximum of

800 Hz (Figure 5). The results suggest that using a

high-pass filter with a pass-band frequency of 200 Hz

(i.e., suppressing the 0–200 Hz band) yields the

highest classification accuracy. This could be due to

removing the low-frequency noise from the

respiratory system while passing the discriminative

components of the murmur and heartbeat signals.

Figure 4: Classification results for filtered four-second

signals.

Figure 5: Vital sign classification and feature selection

evaluation.

4.2 Decision Tree Evaluation

The induced C4.5 decision tree successfully classified

83% (feature selection enabled) and 72% (using all

features) of the unseen vital signs from the test set.

The F-score is calculated to evaluate the classification

accuracy, including both precision and recall. The

results are compared with those of the baseline

support vector machine (SVM) usually used as a

binary classifier. It is shown that the C4.5 classifier

outperforms the baseline SVM for all training set

sizes (Figure 6).The experimental studies also

investigated the significance of feature selection

(Visalakshi & Radha, 2014) to find the optimal set

yielding the highest classification accuracy, which

could be of considerable interest for large healthcare

project data collection. It is observed that using only

the six selected features (highlighted in ) improves the

F-score by an average of 0.2 for the C4.5 classifier.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑃𝑃𝑉













(13

)

𝐹𝑠𝑐𝑜𝑟𝑒2 





(14

)

5 CONCLUSION

This research reported an implementation of an AI

system trained for cardiovascular examination

decision support. The aim of the project was to

analyze the heart sound signals and patient

information to produce a reliable, comprehensive,

and time-efficient screening. The results suggest that

the proposed two-stage approach provides accurate

suggestions that correctly classify a maximum of 91%

of the heart sound signals and 83% of the

accompanying vital signs and information. Although

using the time-domain and frequency-domain

representations of the heart sound signals did not

improve the results compared with the mathematical

and statistical features applied in the PhysioNet

challenge, using the side information and the vital

signs within a two-stage approach increased the

screening reliability. Future work will investigate

more specialized deep network architectures for

diagnosing each heart defect type.

REFERENCES

Amiriparian, S., Schmitt, M., Cummins, N., Qian, K.,

Dong, F., & Schuller, B. (2018). Deep unsupervised

representation learning for abnormal heart sound

classification. 40th Annual International Conference of

the IEEE Engineering in Medicine and Biology Society

(EMBC). 18–21 July, Honolulu, USA.

Barhatte, A., Ghongade, R., & Thakare, A. (2015). QRS

complex detection and arrhythmia classification using

SVM. 2015 Communication, Control and Intelligent

Systems (CCIS). 7–8 Nov., Mathura, India.

Chen, Q., Zhang, W., Tian , X., Zhang, X., Chen, S., & Lei,

W. (2016). Automatic heart and lung sounds

classification using convolutional neural networks.

2016 Asia-Pacific Signal and Information Processing

200 300 400 500 600 700 800

Passband frequenc

)

0.4

0.5

0.6

0.7

0.8

0.9

Frequency domain-Lowpass

Frequency domain-Highpass

Time domain-Lowpass

Time domain-Highpass

50 100 150 200 250 300

Size of the trainin

set

0.3

0.4

0.5

0.6

0.7

0.8

0.9

C4.5- All the features

C4.5 Decision tree- Feature selection

SVM- All the features

SVM- Feature selection

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

204

Association Annual Summit and Conference. 13–16

Dec., Jeju, Korea.

Chen, T. (2017). S1 and S2 heart sound recognition using

deep neural networks. Transactions on Biomedical

Engineering, 64(2), 372–380.

Clifford, G. (2016). Classification of normal/abnormal

heart sound recordings: The PhysioNet/Computing in

Cardiology Challenge 2016. 2016 Computing in

Cardiology Conference (CinC). 11–14 Sept.,

Vancouver, BC, Canada.

Dalmay, F., Antonini, M., Marquet, P., & Menier, R.

(1995). Acoustic properties of the normal chest.

European Respiratory Journal, 8(10), 1761–1769.

Gharehbaghi, A., & Lindén, M. (2017). A deep machine

learning method for classifying cyclic time series of

biological signals using time-growing neural network.

Transactions on Neural Networks and Learning

Systems, 29(9), 4102–4115.

IBM Watson. (n.d.). Retrieved 23 September 2019 from

https://www.ibm.com/watson

Islam Chowdhuryy, M., Sultana, M., Ghosh, R., Ahamed,

J., & Mahmood, M. (2018). AI assisted portable ECG

for fast and patient specific diagnosis. International

Conference on Computer, Communication, Chemical,

Material and Electronic Engineering (IC4ME2). 8–9

Feb., Rajshahi, Bangladesh.

Ivascu, T., Cincar, K., & Negru, V. (2017). Activities of

daily living and falls recognition and classification from

the wearable sensors data. 2017 E-Health and

Bioengineering Conference (EHB). 22–24 June, Sinaia,

Romania.

Jones, B. (2013). Developing a vital sign alert system. The

American Journal of Nursing, 8(113), 36-44.

Kim, T., Lee, J., & Nam, J. (2018). Sample-level CNN

architectures for music auto-tagging using raw

waveforms. ICASSP 2018 – 2018 IEEE International

Conference on Acoustics, Speech and Signal

Processing (ICASSP). 15–20 April, Calgary, AB,

Canada.

Lin, C., Chang, Y., Lin, C., Tsai, L., & Chen, J. (2017).

Development of an AI-based non-invasive Pulse

AudioGram monitoring device for arrhythmia

screening. 2017 IEEE Healthcare Innovations and

Point of Care Technologies (HI-POCT). 6–8 Nov.,

Bethesda, MD, USA.

Loh, B., & Then, P. (2017). Deep learning for cardiac

computer-aided diagnosis: benefits, issues & solutions.

Mhealth, 3(45), doi: 10.21037/mhealth.2017.09.01.

Morsy, A. (2018). Can AI truly transform health care? A

recent IEEE Pulse on Stage forum offers some

perspective. IEEE Pulse, 9(4), 18–20.

Moukadem, A., Dieterlen, A., & Brandt, C. (2013).

Shannon Entropy based on the S-Transform

Spectrogram applied on the classification of heart

sounds. 2013 IEEE International Conference on

Acoustics, Speech and Signal Processing. 26–31 May,

Vancouver, BC, Canada.

Official Website of the Office of the National Coordinator

for Health Information Technology (ONC). (n.d.).

Retrieved 23 September 2019 from

https://www.healthit.gov/topic/health-it-

initiatives/blue-button

PhysioNet. (2016). Retrieved 26 February 2019 from

https://physionet.org/challenge/2016/

Shao, L., Wu, D., & Li, X. (2014). Learning deep and wide:

A spectral method for learning deep networks.

IEEE

Transactions on Neural Networks and Learning

Systems, 25(12), 2303–2308.

Thinklabs. (n.d.). Retrieved 3 October 2019 from

https://www.thinklabs.com/one-digital-stethoscope

Thomae, C., & Dominik, A. (2016). Using deep gated RNN

with a convolutional front end for end-to-end

classification of heart sound. 2016 Computing in

Cardiology Conference (CinC). 11–14 Sept.,

Vancouver, BC, Canada.

University of California Irvine Machine Learning

Repository. (n.d.). (Center for Machine Learning and

Intelligent Systems) Retrieved 1 October 2019 from

http://archive.ics.uci.edu/ml/datasets/statlog+(heart)

Vepa, J. (2009). Classification of heart murmurs using

cepstral features and support vector machines. 2009

Annual International Conference of the IEEE

Engineering in Medicine and Biology Society. 3–6

Sept., Minneapolis, MN, USA.

Vernekar, S., Nair, S., Vijaysenan, D., & Ranjan, R. (2016).

A novel approach for classification of normal/abnormal

phonocardiogram recordings using temporal signal

analysis and machine learning. 2016 Computing in

Cardiology Conference (CinC). 11–14 Sept.,

Vancouver, BC, Canada.

Visalakshi, S., & Radha, V. (2014). A literature review of

feature selection techniques and applications: Review

of feature selection in data mining. 2014 IEEE

International Conference on Computational

Intelligence and Computing Research. 18–20 Dec.,

Coimbatore, India.

Yamashita, M., Himeshima, M., & Matsunaga, S. (2014).

Robust classification between normal and abnormal

lung sounds using adventitious-sound and heart-sound

models. 2014 IEEE International Conference on

Acoustics, Speech and Signal Processing (ICASSP). 4–

9 May, Florence, Italy.

Zabihi, M., Rad, A., Kiranyaz, S., Gabbouj, M., &

Katsaggelos, A. (2016). Heart sound anomaly and

quality detection using ensemble of neural networks

without segmentation. 2016 Computing in Cardiology

Conference (CinC). 11–14 Sept., Vancouver, BC,

Canada.

Two-stage Artiﬁcial Intelligence Clinical Decision Support System for Cardiovascular Assessment using Convolutional Neural Networks

and Decision Trees

205