Precision Diagnostic Algorithm for Multisubtype

Arrhythmia Classification
S R Deepa M Subramoniam Binson V A
Department of ECE Department of ECE Department of ECE
Sathyabama Institute of Science & Saintgits College of Engineering
Sathyabama Institute of Science &
Technology Kottayam, Kerala, India
Chennai, Tamilnadu, India Technology
Chennai, Tamilnadu, India [email protected]
[email protected]
[email protected]

S.Poornapushpakala S.Barani
Department of ECE Department of ECE
Sathyabama Institute of Science & Technology Sathyabama Institute of Science & Technology
Chennai, Tamilnadu, India Chennai, Tamilnadu, India

significantly impact the quality of life for individuals

Abstract - Arrhythmia, a condition often linked to lung
health, affects individuals across various age groups. With
suffering from various medical conditions [1-3]. In many
multiple distinct subtypes, categorized based on heart rate healthcare settings, traditional diagnostic procedures have
irregularities observed in the PQRST waveform, historically been carried out manually by healthcare
Arrhythmia presents a complex diagnostic challenge. professionals. While these methods are often highly
Currently, clinicians manually diagnose specific Arrhythmia reliable, they are not immune to human errors. Human
types from ECG recordings. The integration of computer errors in diagnosis can result from a variety of factors,
aided diagnosis has the potential to significantly enhance such as misinterpretation of test results, fatigue, or
diagnostic precision and subsequent treatment strategies. cognitive biases [4]. To mitigate these risks and enhance
This paper introduces a diagnostic algorithm designed to diagnostic accuracy, computer-aided diagnostic tools have
address this need. The proposed algorithm employs a
combination of wavelet feature extraction and Support
been developed by researchers and implemented in
Vector Machine (SVM) classification techniques. By clinical practice. These tools leverage the power of
extracting intricate details from ECG recordings using technology, including artificial intelligence, machine
wavelet analysis, the algorithm captures nuanced patterns learning, and data analysis, to assist healthcare providers
characteristic of Abnormal Arrhythmia (ARR), Congestive in making more accurate and efficient diagnoses [5, 6].
Heart Failure (CHF), and Normal Sinus Rhythm (NSR). This paper discusses the primary results of the computer
Subsequently, the SVM classifier deciphers these features to aided tool developed for the diagnosis of Arrhythmia.
accurately categorize ECG signals. Remarkably, the
algorithm achieves a test accuracy of 98%, demonstrating its
The conventional method of diagnosis involves analysis
effectiveness in discerning between different Arrhythmia
types. The implementation of this computer aided diagnostic
of PQRST waveform over the ECG recordings obtained
tool not only streamlines the diagnosis process but also from the subject supposed to have Arrhythmia. The
equips medical professionals with a valuable tool for methodology used in this paper involves extraction of
enhanced decision-making, leading to improved treatment signal features from the ECG recordings using wavelet
outcomes. This research highlights the potential of transform and applying SVM classifier to identify the
technology-assisted medical diagnostics, fostering more type of Arrhythmia. The algorithm was developed using
accurate and efficient patient care in the realm of the databases such as physionet, MIT-BIH (Arrhythmia,
Arrhythmia management. Normal Sinus Rhythm) and BIDMC (Congestive heart
failure). These are the standard Arrhythmia database
Keywords- Image Processing, Arrhythmia, Normal Sinus which has been used by the researchers for their study.
Rhythm, SVM Classifier, CNN Section II describes various studies done by researchers
over Arrhythmia. Section III describes the methods
I. INTRODUCTION including the selection of data from physionet database.
Section IV describes the results and discussion about the
Diagnostic tools are crucial in ensuring the effective work and section V concludes the work.
treatment of diseases. Accurate diagnosis is the
cornerstone of appropriate medical care, as it guides II. RELATED WORK
healthcare professionals in prescribing the right
treatments and interventions. A faulty diagnosis, on the Computer aided tool plays a major role in clinical
other hand, can have severe consequences, potentially diagnosis with minimum time duration and user
leading to incorrect or ineffective treatment, and this can friendliness. Researchers have developed many such
diagnostic tools that aids in diagnosis of Arrhythmia. A supports data imbalance. The craftnet increases the
few of those has been discussed below. sensitivity and results are more accurate. Ilbeigipour et.al
(2021) implemented three algorithms such as decision
Elif Izci et al. (2018) used Empherical Mode trees, random forest, and logistic regression multiclass
Decomposition for the development of computer aided classifiers for identification and classification of
tool for diagnosis of Arrhythmia [7]. The tool was arrythmia disease [11]. The classifications are normal
developed using the ECG features stored in MIT-BIH beats, RBBB, and atrial fibrillation. The authors
Arrhythmia database. The developed tool aided in contribution outperforms better than the previous studies.
diagnosing six types of heart beats such as normal rhythm
(N), left bundle branch block defined as LBBB, right Bae et.al (2020) used adaptive filtering technique for the
bundle branch block defined as RBBB, premature diagnosis of arrhythmia. The authors identified that by
ventricular contraction defined as PVC, paced beat and varying the sampling rate signal analysis can be done for
atrial premature beats defined as APB. The accuracy rate the diagnosis of Arrhythmia [12]. In this type of analysis
achieved with this tool is 87 percent. the R point plays a vital role.Using the equation this R
wave is separated using the median filter. If the
Yağmur and Şen (2019) developed a neural network abnormality is observed for more than one minute of time
framework for the classification of Arrhythmia [8]. Two then it is classified under Bradycardia and Tachycardia.
approaches with time series signals and times frequency Ullah et.al (2021) propsed a 1D and 2D Convolution
were made in this study to classify heartbeats into normal Neural Network algorithm for classification of Arrythmia
rhythm (N), premature ventricular contractions defined as disease [13]. The samples were taken from MIT-BIH data
PVC, right bundle branch block defined as RBBB. The base. Both the proposed algorithm outperforms better
developed algorithm was evaluated using statistical with 97.38% and 99.02% of accuracy.
measures. The evaluated results concluded that time series
approach shows better performance as compared to the Çınar and Tuncer (2021) proposed two step process for
other approach done in this study. Ji et al. (2021), classification. As a first step classification is done by
suggests that a proper classification techniques are conventional algorithms SVM and KNN [14]. This
essential to classify the various types of Arrythmia at is algorithm results in 68.75 % and 65.63 % of accuracy
one of the serious threats for humans [9]. The author only. In the second step this classified input is further
proposed stacking – DWKNN algorithm to implement given and classified using LSTM network achieveing
multifeatured classification. The sample is denoised and 96.77 5 of accuracy. In this paper an Hybrid Alexnet-
segmented initially. Then various features such as P, T, SVM is applied. The sample size is 192 images.
QRS length, PR, QT, RR Interval, R and T amplitude and
ST segments are the extracted features. The extracted The literature survey defines the various methodologies
features are analyze during DWKNN algorithm for and challenges faced by the researchers in the
further classification. The author achieves 89.42% of development of various tools for the diagnosis of
sensitivity and 99.01 % of accuracy. 94.90% and 97.07 % Arrhythmia. Based on the study made the algorithm for
of S type and V type features are predicted accurately. the proposed study is formulated. Features are extracted
from the ECG signal using wavelet transform. With the
Though ECG is a popular non – invasive technique, the extracted features classification is performed using SVM
identification of disease requires manual examination of classifier.
signals. Chances of wrong identification is high in this
case. Hence an automated tool is essential to examine the III. METHODOLOGY
signal and identify the disease. Ullah et al. (2022) has
attempted to implement a computer aided automated tool
which would analyze the signal using Convolution Neural The block diagram of the system developed classification
Network [10]. The model implemented is a dense net of Arrhythmia is shown in the following figure 1.
model for arrythmia identification. The author has given
Pre-
97720 images as training images. The datasets were processing
Signal Multiclass
collected from Massachusetts Institute of Technology- and feature SVM
Data
Beth Israel Hospital (MIT-BIH) arrhythmia and St. extraction
Petersburg Institute of Cardiological Technics (INCART) using
datasets. Four classifications were done such as
Normal(N), Ventricular Ectopic(V), Super ventricular Fig. 1. Schematic representation of the proposed system
Ectopic(S) and Fusion (F). The proposed model
outperforms better with 99.80 % of accuracy, 98.34% of A. Dataset
precision and 98.91% of F1 Score on MIT-BIH Dataset
and 99.63 % of accuracy, 98.94% of precision and The ECG signal samples from physionet data base is used
98.91% of F1 Score on INCART dataset. In addition to for the study. ECG Dataset contains Data and Labels as
that the algorithm has also been implemented for live structured array. Data has 162 rows and 65536 columns.
datasets. Similarly Labels array has 162 rows and 1 column of
character vectors corresponding to the data. The first 96
Deep neural network such as craftnet can also ne used for rows are ECG data. i.e Data (162x65536 matrix) with
classification of various types of arrythmia disease. It also corresponding labels in ECG data. Labels of 'ARR' are
modified copies of the two ECG recordings in the 48 data
files contained in the MIT-BIH Arrhythmia Database. efficients. MODPWT is superior over DWT
Labels of 'CHF' are modified copies of the two ECG because it aligns the decomposed wavelet and
recordings in the 15 data files contained in The BIDMC scaling coefficient at each decomposition levels
Congestive Heart Failure Database. The final 36 rows of with the original time series [15]. It is applied for
ECG Data with the labels of 'NSR' are modified copies of finding significant features at each frequency
the two ECG recordings in the 18 data files contained in points.
the MIT-BIH Normal Sinus Rhythm Database. The  Multifractal wavelet leader estimates of the
variations in ECG recordings for various cases of cardiac second cumulant of the scaling exponents and
rhythm is shown in figure 2 the range of Holder exponents, or singularity
spectrum time Window = 8192

These extracted features are then subjected to

classification. The signals are to be classified to three
classes such as Abnormal Arrhythmia defined as ARR,
Congestive Heart Failure defined as CHF and Normal
Sinus Rhythm defined as NSR. The Holder exponent can
detect the occurrence of discontinuity in a dynamic signal.
In addition, it can also find the time of occurrence of
discontinuity. Figure 3 depicts the range of holder
exponents in the singularity spectra obtained for the first-
time window and wavelet variancefor various types of
rhythms is shown in figure 4.

Fig. 3. Box plot for holder exponent range for various type of rhythms

Fig. 2. Signal samples showing heart disorders.

B. Preprocessing

The initial step involved in signal pre-processing is

scaling. Then scaling was done on the raw files as defined
in the Physionet data base. The data were re-sampled to a
uniform sample rate of 128 hertz. Certain ECG files
contained two records. Those files were segregatedto two
separate records. The data length was truncated into a
common length which is of 65536 samples. These signal
after pre-processing is subjected to feature extraction.
Fig. 4. Box plot for wavelet variance for various type of rhythms
C. Feature Extraction
D. Classification
Feature extraction involves extraction of signal features
from the ECG signal. The following are the parameters SVM is a type of supervisory algorithm that can be
assigned for extracting ECG signal features. applied for both regression and classification problems.
SVM classifier creates a decision boundary or hyperplane
 Autoregressive model (AR) coefficients of order
for classification using support vectors. The number of
4
classification is decided based on the number of support
 Shannon entropy (SE) values for the maximal
vectors formulated. SVM classifier has shown significant
overlap discrete wavelet packet transform
classification rate in many applications, especially in
(MODPWT) is considered at level 4. MODPWT
medical diagnosis applications [16, 17]. SVMs are
performs a series of filtering operation where
valuable tools in the field of medical diagnosis,
variation in the signal can be transformed to co-
particularly in arrhythmia classification. Their ability to
effectively classify arrhythmias based on extracted study made with this tool aid in development of many
features from ECG data helps healthcare professionals such algorithms for the diagnosis of Arrhythmia. The
make accurate and timely decisions, leading to better future work involves further enhancement of this tool to
patient care and outcomes. classify all class of Arrhythmia and also to develop as a
user-friendly device.
IV. RESULTS AND DISCUSSION
REFERENCES
In this work the extracted features under various classes
are trained against SVM classifier. The performance of [1] R. Li, S. Ji, S. Shen, P. Li, X. Wang, T. Xie, et al., "Arrhythmia
the algorithm is evaluated using statistical parameters Multiple Categories Recognition based on PCA-KNN Clustering
Model", 2019 8th International Symposium on Next Generation
such as precision, recall and F1 score. The overall Electronics (ISNE), pp. 1-3, 2019.
accuracy achieved with this tool is 97.95 percent. The [2] Binson V.A., Subramoniam, M., and Mathew, L. “MOS Based
results obtained against these statistical parameters for Sensor Array System for the Detection of Human Breath Volatile
various classes are shown in Table 1. Organic Compounds”. Ann Romanian Soc Cell Biol, vol. 25,
pp.2069-2081, Apr 2021.
The performance metrics obtained using Support Vector [3] W. Sun, N. Zeng and Y. He, "Morphological Arrhythmia
Automated Diagnosis Method Using Gray-Level Co-Occurrence
Machine (SVM) for the classification of different heart Matrix Enhanced Convolutional Neural Network", IEEE Access,
conditions are as follows: For Abnormal Arrhythmia, the vol. 7, pp. 67123-67129, 2019.
SVM achieved a precision of 91%, indicating that 91% of [4] Binson V. A., Subramoniam M., Ragesh G. K., and Kumar A.
the predicted cases of abnormal arrhythmia were indeed "Early detection of lung cancer through breath analysis using
correct, while the recall of 98% signifies that 98% of adaboost ensemble learning method." proc. ACCESS, IEEE,
Ernakulam, India 2021. pp. 183-187.
actual abnormal arrhythmia cases were correctly
[5] Binson V.A., Subramoniam, M., and Mathew, L. “Detection of
identified. The F1 Score, a harmonized measure of COPD and Lung Cancer with electronic nose using ensemble
precision and recall, is 94%, suggesting a balance learning methods”. Clin Chim Acta, vol. 523, pp. 231-8, Dec 2021.
between precision and recall. For Congestive Heart [6] Binson V. A., Thomas S.,. Ragesh G. K., and Kumar A. "Non-
Failure, the SVM achieved a precision of 91%, indicating invasive Diagnosis of COPD with E-nose Using XGBoost
Algorithm." proc. ACCESS, IEEE, Ernakulam, India 2021. pp.
91% of the predicted cases were accurate, but with a 297-301.
lower recall of 70%, signifying that 70% of actual [7] Elif IZCI and Mehmet Akif OZDEMIR, Arrhythmia Detection on
congestive heart failure cases were correctly identified. ECG Signals by Using Empirical Mode Decomposition, IEEE,
2018.
The F1 Score for Congestive Heart Failure is 80%. In the [8] S. Y. Şen and N. Ozkurt, "ecg Arrhythmia Classification By Using
case of Normal Sinus Rhythm, the SVM yielded a high Convolutional Neural Network And Spectrogram", 2019
precision of 98%, meaning that 98% of the predicted Innovations in Intelligent Systems and Applications Conference
(ASYU), pp. 1-6, 2019.
cases were true positives, while the recall of 95%
[9] S. Ji, R. Li, S. Shen, B. Li, B. Zhou and Z. Wang, "Heartbeat
indicates that 95% of actual normal sinus rhythm cases classification based on multifeature combination and stacking-
were identified. The F1 Score for Normal Sinus Rhythm DWKNN algorithm", J. Healthcare Eng., vol. 2021, pp. 1-14, Jan.
is 96%, reflecting the model's overall effectiveness in 2021.
classifying this specific condition. These metrics provide [10] Ullah H, Heyat M B, Akhtar S F, Muaad A, Islam S, Abbas Z, et
al., "An End-to-End Cardiac Arrhythmia Recognition Method with
valuable insights into the SVM's performance in an Effective DenseNet Model on Imbalanced Datasets Using ECG
accurately distinguishing between these different heart Signal", Computational Intelligence and Neuroscience, pp. 1-23,
conditions, with precision emphasizing the model's 2022.
reliability in making positive predictions, recall [11] S. Ilbeigipour, A. Albadvi and E. Akhondzadeh Noughabi, "Real-
highlighting its ability to capture true cases, and the F1 Time Heart Arrhythmia Detection Using Apache Spark Structured
Streaming", J Healthc Eng, vol. 2021, 2021.
Score representing a balanced assessment of both
[12] Bae T.W., Lee S.H, and Kwon K.K, “An adaptive median filter
precision and recall. based on sampling rate for R-peak detection and major-arrhythmia
analysis”. Sensors, vol. 20, pp.6144, Feb 2020.
Table 1. Performance indices for various type of Rhythm [13] Amin Ullah et al., "A Hybrid Deep CNN Model for Abnormal
Arrhythmia Detection Based on Cardiac ECG Signal", Sensors,
Performance Index vol. 21.3, pp. 951, May 2021.
Type of
Rhythm [14] Çınar A and Tuncer S A, “Classification of normal sinus rhythm,
Precision (%) Recall (%) F1 Score (%) abnormal arrhythmia and congestive heart failure ECG signals
using LSTM and hybrid CNN-SVM deep neural networks”,
ARR 91 98 94 Comput Methods Biomech Biomed Engin. vol. 24, pp. 203-214,
Feb 2021.
CHF 91 70 80 [15] S. Gurumoorthy, N. B. Muppalaneni and G. S. Kumari, "Eeg
signal denoising using haar transform and maximal overlap
NSR 98 95 96 discrete wavelet transform (modwt) for the finding of epilepsy" in
Epilepsy-Update on Classification Etiologies Instrumental
Diagnosis and Treatment., IntechOpen, 2020
[16] Binson, V. A., M. Subramoniam, Youhan Sunny, and Luke
Mathew. "Prediction of pulmonary diseases with electronic nose
V. CONCLUSION using SVM and XGBoost." IEEE Sens J, Aug 2021
[17] VA, Binson, M. Subramoniam, and Luke Mathew. "Noninvasive
The primary tool developed for classification of detection of COPD and Lung Cancer through breath analysis using
Arrhythmia has been discussed. This tool is developed MOS Sensor array based e-nose. Expert Rev. Mol. Diagn, vol. 21.
using wavelet transform features and SVM classifier. This no. 11, pp.1223-1233, Sep 2021.
tool achieved a classification accuracy of 98 %. The pilot