Electrocardiogram Signal Classification Using Manual/Automated Features and Machine Learning Algorithms: A Large Cohort Study on Students Aging from 6 to 18 Years Old

doi:10.21203/rs.3.rs-1551010/v1

PDF

Article

Electrocardiogram Signal Classification Using Manual/Automated Features and Machine Learning Algorithms: A Large Cohort Study on Students Aging from 6 to 18 Years Old

https://doi.org/10.21203/rs.3.rs-1551010/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Cardiac rhythm abnormalities are one of the main cardiovascular diseases leading to morbidity and mortalities. Electrocardiogram (ECG), since its introduction, has been extensively used for the detection of rhythm disturbances The differentiation of abnormal ECG from normal one is the cornerstone of public health programs. The interpretation of ECGs by physicians is a time-consuming process that is not free of faults. In this study we sought to determine the accuracy of different machine learning models in distinguishing abnormal ECGs from normal ones in a large sample of in-house data sets of children’s who were examined using resting ECG machine during a community-based study. Altogether, 10745 ECGs were recorded for student aging from 6 to 18 years old. Manual and automatic ECG features were extracted for each participant. Features were normalized using Z-score normalization and went through the student’s t-test and chi-squared test to measure their relevance. All p-values were corrected with Benjamini-Hochberg (BH) method and were reported as q-value. We applied Boruta algorithm for feature selection and then implemented 8 different classifiers, including eXtreme Gradient Boosting (XGB), Support Vector Machine (SVM), Quadratic Discriminant Analysis (QDA); Random Forest (RF), Decision Tree (DT), K-nearest neighbors (KNN), Logistic Regression (LR), and Stack Learning (SL) for classification purpose. The dataset was split into training (80%) and test (20%) partitions. Hyper-parameter optimization of classifiers was performed on selected features by 10-fold cross validation and random search. Optimal models were made on training dataset and performance of the models was evaluated on the test data (unseen data) by 1000 bootstrap. We also reported mean ± SD and 95% confidence interval (CI) for each model. The performance of models was evaluated by sensitivity (SEN), specificity (SPE), AUC, and accuracy (ACC). In univariate analysis, the highest performance feature was heart rate and RR interval in the manual dataset with an AUC of 0.72, followed by heart rate in automated dataset with an AUC of 0.71. We also found QT in both of datasets had an AUC of 0.68. The Boruta method selected 15 and 16 features in the manual and the automated datasets, respectively. The best models in the manual dataset were RF and QDA model with AUC, ACC, SEN, SPE equal to 0.93,0.98, 0.69, 0.99 and 0.90, 0.95, 0.75, 0.96, respectively. In the automated dataset, QDA (AUC: 0.89, ACC: 0.92, SEN: 0.71, SPE: 0.93), and SL (AUC: 0.89, ACC: 0.96, SEN: 0.61, SPE: 0.99) reached best performances. This study demonstrated that the manual measurement of ECG features had better performance than the automated measurement but automated measurement in some model had promising result in discriminating normal and abnormal cases.

Electrocardiogram

Pediatric

Machine Learning

Classification

Manual/Automated Features

Cardiac rhythm abnormalities are one of the main cardiovascular diseases leading to morbidity and mortalities. Although these abnormalities are less common among children than adults, still a significant number of children struggle with them ^1–3. Electrocardiogram (ECG), since its introduction, has been extensively used for the detection of rhythm disturbances ^{4, 5}. ECG’s signals are created during depolarization and repolarization phases and can be used for interpreting rhythm disturbances ⁶. Different rhythms disturbances during the ECG recording, yield variations in the morphological characteristics of the ECG traces, such as QRS, and etc. ⁷. Thus, by detecting abnormal patterns on the ECG signals, we will be able to distinguish individuals at the risk of life-threatening arrhythmias during childhood or adolescence⁸. Accordingly, the differentiation of abnormal ECG from normal one is the cornerstone of public health programs.

The interpretation of ECGs by physicians is a time-consuming process that is not free of faults ^9–11. Hence, enormous endeavors have been applied to find practical tools for automated ECG feature extraction and to a further extend, differentiation of abnormal from normal ECGs ^12–16. Automatic feature extraction methods include morphological ECG feature ¹⁷, Wavelet Transform ¹⁸, Daubechies Wavelets transform ¹⁹, Discrete Wavelet Transform ²⁰, Slope Vector Waveform ²¹, Karhunen-Lòeve transform ²², Hermitian Basis ²³ and hybrid feature extraction method ¹³. Modular ECG Analysis System (MEANS) ²⁴, introduced in 1972, is a commercial method for analysis and interpretation of ECG. This method is now implemented in ECG devices and extracts feature automatically and real time from ECG signals.

Artificial intelligence has revolutionized different fields of medical research since past two decades, and has stamped its effect on a big range of applications, from developing computer aided models for analysis of medical images ^25–28 and gene expressions ²⁹, to ECG signal analysis ³⁰.

Recently, Sahoo et al. ³⁰ systematically reviewed the present literature on the machine learning approaches for detection of different arrhythmias in ECG signals. Their survey shed light on the high demand for computer-aided decision-making systems for automated and real-time cardiac arrhythmias detection. They also pointed out that based on the literature, most of the developed models have been trained and validated on the ECG signals from MIT-BIH database, thus, they encouraged the scientific community to investigate on developing new databases for further validation of the whole flow. Moreover, there’s a lack of investigation on comparing different machine learning algorithms on similar circumstances, toward identification of the optimum modeling. In several studies, the accuracy of ECG classification has been varied from 83.3–99.66% that those were affected by the type of machine-learning (ML) models ^{31, 32}.

In this study we sought to determine the accuracy of several ML models in distinguishing abnormal ECGs from normal ones in a large sample of in-house data sets of children’s who were examined using resting ECG machine during a community-based study. In addition, we tried to compare the accuracy of the ML-based classification with the manual measurement of ECG features and automated measurement of ECG features.

Data collection

In this community-based study, a large-scale group of randomly selected students in Tehran urban area schools, Iran, underwent the electrocardiographic examinations using a resting machine (CP 150 model, Welch Allyn, USA) and a standard 12-lead ECG. The paper speed was 25 mm/s and a scale of 10 mm/mV standardization was obtained by a unique recorder. All ECGs were recorded while students were at rest and in a supine position. ECGs were re-taken if noise prevented accurate interpretation. A total of 10745 ECGs were collected. The evaluation of traces for identifying normal versus abnormal ECGs was performed based on the international guidelines used in physicians’ daily practice ^33–38. This study was conducted in accordance with the Declaration of Helsinki. The study protocol was evaluated and approved by the Ethics Committee of the National Committee for Ethics in Biomedical Research (Ethics Committee ID: IR.NIMAD.REC.1396.159). Participants' information were anonymized and informed consent was obtained from students' parents.

ECG features were measured manually by experienced nurses who had completed an educational course for evaluating ECG parameters. Visual measurements were performed with a scale rounding to a quarter of small box of ECG traces. Inter and intra-reader variabilities were calculated for nurses after repeated measurements on 30 ECGs. The readers scaled the QRS complex duration as well as PR, RR, and QT intervals. The evaluation of features was performed in three consecutive beats and the average of three beats was used in this study. For patients with sinus arrhythmia, these values are measured in two separate leads. The QT interval was calculated using the threshold method in lead II, so that the end of the T-wave intersecting with the isoelectric line was considered as the duration of QT interval. Finally, it was corrected for heart rate (QTc) based on the Bazett’s formula ³⁹.

In addition to manual measurements, we used the automated measurement of ECG features by the electrocardiogram machine. The automated features were also implemented to compare them with the manually measured features.

Finally, all ECG traces were assessed by a pediatric cardiologist who was certified with a fellowship of electrophysiology so as to identify normal and abnormal ECGs. All ECGs with abnormalities were again examined by an electrophysiologist. The ECG findings considered as abnormal ECGs included sinus tachycardia, sinus bradycardia, premature ventricular contraction, premature atrial contraction, non-conducted premature atrial contraction, prolonged PR interval, prolonged QTc, and right bundle branch block. The details of each group of patients are summarized in Table 1, and Fig. 1 shows major ECG abnormalities detected in study dataset.

Table 1

Number of cases for each type
ECG type	Number
Normal	10043
Sinus tachycardia	446
Sinus bradycardia	145
Premature atrial contraction (PAC)	43
Premature ventricular contraction (PVC)	33
Prolonged QTc	17
Prolonged PR	10
Right bundle branch block (RBBB)	8

Data Pre-processing

After preparing manual and automated datasets provided by the investigators of epidemiological study, all missed data (see Table 2) in both manual and automated datasets were imputed with the means of each feature in normal and abnormal groups. The ECG features in both manual and automated datasets as well as baseline characteristics of students are summarized in Table 2

Table 2

Characteristics of patients and ECG features in Train and Test dataset
Variables	Missed data	Train	Test
Categorical	No.(%)	No.(%)	No.(%)
Outcome Normal		8034 (93.5%)	2009 (93.5%)
Abnormal		562 (6.5%)	140 (6.5%)
Gender Male		4119 (48%)	1005 (47%)
Female		4477 (52%)	1144 (53%)
Axis_Manual Normal		7989 (93%)	2007 (93%)
LAD		64 (0.74%)	20 (0.93%)
RAD		533 (6.2%)	118 (5.5%)
Indeterminate		10 (0.12%)	4 (0.19%)
Continous		mean ± SD	mean ± SD
Age		12 ± 3.1	12 ± 3.1
RR_Manual		680 ± 120	680 ± 120
HeartRate_Manual		91 ± 16	91 ± 15
PR_Manual		130 ± 19	130 ± 18
QRS_Manual		79 ± 12	78 ± 11
QT_Manual		330 ± 24	330 ± 24
QtcBazet_Manual		400 ± 31	400 ± 30
LeadI_Manual	71 (0.66%)	4.1 ± 3.5	4.2 ± 3.4
LeadAVF_Manual	71 (0.66%)	7.8 ± 4.4	7.7 ± 4.4
WEIGHT	139 (1.29%)	49 ± 19	49 ± 19
HEIGHT	141 (1.31%)	150 ± 17	150 ± 17
WC	360 (3.35%)	70 ± 11	70 ± 11
RHSBP	142 (1.32%)	110 ± 12	110 ± 12
RHDBP	142 (1.32%)	73 ± 7.5	73 ± 7.4
WCHR		0.47 ± 0.059	0.46 ± 0.059
BMI		21 ± 4.9	21 ± 4.9
PR2_Auto		130 ± 18	130 ± 17
QRS_Auto		82 ± 13	82 ± 12
QT_Auto		340 ± 26	340 ± 25
QTc_Auto		420 ± 26	420 ± 22
P_axis_Auto	145 (1.35%)	46 ± 24	47 ± 21
QRS_axis_Auto	113 (1.05%)	61 ± 31	61 ± 30
T_axis_Auto	5 (0.05%)	40 ± 22	40 ± 18
HearRate_Auto		92 ± 16	93 ± 16
*Auto: Automated, LAD: left axis deviation; RAD: right axis deviation

Univariate analysis

In order to investigate the correlations between features, Spearman correlation test was performed on ECG features. The spearman correlation coefficient (SCC) over 0.90 was selected as threshold for omitting collinear features. Features were normalized using Z-score normalization and went through the student’s t-test and chi-squared test to measure their relevance. All p-values were corrected with Benjamini-Hochberg (BH) method and were reported as q-value. P-value and q-value < 0.05 were considered statistically significant. After construction of receiver operating characteristics (ROC) curve, the area under the curve (AUC) was calculated for selected features.

Multivariate analysis

Feature Selection

Given the big size of the dataset and the low number of ECG features, the feature selection was based on using all relevant features. To deal with this situation, we applied Boruta algorithm. This algorithm is classified as a wrapper feature selection method. At the first step it creates shadow features by randomly shuffling copy of all features and using random forest (RF) classification algorithm which is followed by computing Z-score base on loss accuracy. The maximum Z-score among shadow attributes is set as criteria for selecting features. This process was repeated to determine all important features or to reach RF limit runs ⁴⁰.

Machine Learning Classifiers

In this study, we implemented 8 different classifiers, including eXtreme Gradient Boosting (XGB), Support Vector Machine (SVM), Quadratic Discriminant Analysis (QDA); Random Forest (RF), Decision Tree (DT), K-nearest neighbors (KNN), Logistic Regression (LR), and Stack Learning (SL). The SL classifier model used base learners as prediction and super learner which fit on this prediction (instead of features). In this study we used XGB, SVM, QDA and RF as base learners, based on experiment, and generalized linear model as super learner based on Kabir et al study ⁴¹. Table 3 show hyper-parameter ranges for classifiers optimization.

Table 3

hyper-parameter ranges for classifiers optimization
Classifier	Hyper-parameter	Range
XGB	eta	0.025, 0.05, 0.1, 0.3
	max_depth	2–10, step = 1
	nrounds	50-1000, step = 50
	colsample_bytree	0.4, 0.6, 0.8, 1.0
	subsample	0.5, 0.75, 1.0
	gamma	0, 0.05, 0.1, 0.5, 0.7, 0.9, 1.0
	min_child_weight	1, 2, 3
SVM	cost	0.1–10, step = 0.1
SVM	gamma	0.1–10, step = 0.1
DT	minsplit	5–20, step = 1
DT	minbucket	3–10, step = 1
KNN	k	1–12, step = 1
RF	ntree	50-1000, step = 50
	mtry	1–10, step = 1
	nodesize	1–20, step = 1
QDA	method	"moment","mle","t"
QDA	nu	3–10, step = 1
LR	-	-
SL	-	-
*XGB: eXtreme Gradient Boosting, Support Vector Machine, QDA: Quadratic
Discriminant Analysis, RF: Random Forest, DT: Decision Tree, KNN:
K-nearest neighbors, LR: Logistic Regression, and SL: Stack Learning.

Model Development and Evaluation

First, the dataset was split into training (80%) and test (20%) partitions (stratified by target). Then, the relevant feature selection was performed in training dataset using the Boruta classification algorithm. Hyper-parameter optimization of classifiers was performed on selected features by 10-fold cross validation and random search. Optimal models were made on training dataset and performance of the models was evaluated on the test data (unseen data) by 1000 bootstrap. We also reported mean ± SD and 95% confidence interval (CI) for each model.

For each model, we calculated a confusion matrix containing true positive (TP), true negative (TN), false positive (FP), and false negative (FN). Then, the performance of models was evaluated by sensitivity (SEN), specificity (SPE), AUC, and accuracy (ACC). Comparison of the performances of manual and automated datasets and also comparisons of the performances of different machine learning algorithms was performed by Wilcoxon rank test while p-values were adjusted with BH method. Univariate and multivariate analysis were implemented on R 4.0.

Univariate Outcomes

The inter- and intra-observer variabilities for ECG intervals were good to excellent among four readers (intraclass correlation coefficient greater than 0.75 for all values). The SCC of ECG features is summarized in Fig. 2. RR interval and heart rate in the manual dataset had SCC of -1. Also, body mass index (BMI), had correlation values of 0.87, and 0.86, with weight and waist circumference (WC), respectively. Weight and height were also correlated (SCC = 0.85). As shown in Fig. 2, heart rate in automated dataset had SCC values of -0.84, 0.84, and − 0.82 with RR, heart rate (manual dataset) and QT (automated dataset), respectively. Figure 3 shows univariate analysis for each feature. Based on the results, most of features have significant p and q values (< 0.05) for discriminating normal and abnormal cases except QRS_Manual, lead AVF_manual, QTc and QRS among automated dataset, and WC to height ratio and gender (chi-squared test) among clinical data. The highest performance feature was heart rate and RR interval in the manual dataset with an AUC of 0.72, followed by heart rate in automated dataset with an AUC of 0.71. We also found QT in both of datasets had an AUC of 0.68.

Multivariate Outcomes

The Boruta method selected 15 and 16 features in the manual and the automated datasets, respectively. Figure 4 shows importance of features in both datasets. Among the manual dataset, lead AVF, lead I and gender, and among automated dataset only gender were found to be as unimportant features and others remain as important to construct models. Following the results of univariate analysis, stating that RR and heart rate in the manual dataset are completely correlated, RR feature was removed for model construction, owing to heart rate’s higher importance (results shown in Fig. 4 shows heart rate have better feature importance than RR).

Table 4 shows mean ± SD and 95% CI of the performance of different models, trained on both manual and automated feature sets, reported by averaging the output of 1000 bootstraps on the test datasets. The best models in the manual dataset were RF and QDA model with AUC, ACC, SEN, SPE equal to 0.93,0.98, 0.69, 0.99 and 0.90, 0.95, 0.75, 0.96, respectively. In the automated dataset, QDA (AUC: 0.89, ACC: 0.92, SEN: 0.71, SPE: 0.93), and SL (AUC: 0.89, ACC: 0.96, SEN: 0.61, SPE: 0.99) reached best performances.

Table 4

Confidence interval 95% and Mean ± SD performance of different models by AUC, ACC, SEN, SPE metrics.
Type	AUC	ACC	SEN	SPE
Auto_DT	0.8–0.8	0.96–0.96	0.62–0.62	0.99–0.99
Auto_DT	0.8 ± 0.021	0.96 ± 0.0043	0.62 ± 0.042	0.99 ± 0.0027
Auto_KNN	0.84–0.84	0.94–0.94	0.21–0.22	0.99-1
Auto_KNN	0.84 ± 0.021	0.94 ± 0.0049	0.21 ± 0.036	0.99 ± 0.0015
Auto_LR	0.72–0.72	0.94–0.94	0.055–0.057	0.99-1
Auto_LR	0.72 ± 0.035	0.94 ± 0.0051	0.056 ± 0.019	0.99 ± 0.00047
Auto_QDA	0.89–0.89	0.92–0.92	0.71–0.72	0.93–0.93
Auto_QDA	0.89 ± 0.019	0.92 ± 0.006	0.71 ± 0.039	0.93 ± 0.0055
Auto_RF	0.89–0.89	0.96–0.96	0.58–0.59	0.99–0.99
Auto_RF	0.89 ± 0.021	0.96 ± 0.0043	0.59 ± 0.043	0.99 ± 0.0026
Auto_SL	0.89–0.89	0.96–0.96	0.6–0.61	0.99–0.99
Auto_SL	0.89 ± 0.021	0.96 ± 0.0042	0.61 ± 0.042	0.99 ± 0.0026
Auto_SVM	0.89–0.89	0.95–0.95	0.41–0.42	0.99–0.99
Auto_SVM	0.89 ± 0.021	0.95 ± 0.0045	0.41 ± 0.042	0.99 ± 0.0022
Auto_XGB	0.89–0.89	0.96–0.96	0.51–0.52	0.99–0.99
Auto_XGB	0.89 ± 0.02	0.96 ± 0.0044	0.51 ± 0.044	0.99 ± 0.0024
Manual_DT	0.86–0.86	0.97–0.97	0.69–0.69	0.99–0.99
Manual_DT	0.86 ± 0.019	0.97 ± 0.0033	0.69 ± 0.038	0.99 ± 0.0017
Manual_KNN	0.88–0.88	0.96–0.96	0.46–0.46	0.99-1
Manual_KNN	0.88 ± 0.02	0.96 ± 0.0041	0.46 ± 0.041	0.99 ± 0.0015
Manual_LR	0.89–0.9	0.97–0.97	0.6–0.61	0.99–0.99
Manual_LR	0.9 ± 0.019	0.97 ± 0.0037	0.61 ± 0.042	0.99 ± 0.0017
Manual_QDA	0.9–0.9	0.95–0.95	0.75–0.75	0.96–0.96
Manual_QDA	0.9 ± 0.018	0.95 ± 0.0046	0.75 ± 0.037	0.96 ± 0.0041
Manual_RF	0.93–0.93	0.98–0.98	0.69–0.69	0.99-1
Manual_RF	0.93 ± 0.015	0.98 ± 0.0033	0.69 ± 0.039	0.99 ± 0.0016
Manual_SL	0.91–0.92	0.98–0.98	0.7–0.71	0.99-1
Manual_SL	0.91 ± 0.018	0.98 ± 0.0033	0.71 ± 0.038	0.99 ± 0.0016
Manual_SVM	0.9–0.9	0.97–0.97	0.63–0.63	0.99-1
Manual_SVM	0.9 ± 0.019	0.97 ± 0.0036	0.63 ± 0.042	0.99 ± 0.0016
Manual_XGB	0.91–0.92	0.97–0.97	0.69–0.69	0.99–0.99
Manual_XGB	0.92 ± 0.017	0.97 ± 0.0034	0.69 ± 0.039	0.99 ± 0.0019
*Auto: automated, AUC: area under the curve, ACC: accuracy, SEN: sensitivity, SPE: specificity, XGB: eXtreme Gradient Boosting, Support Vector Machine, QDA: Quadratic Discriminant Analysis, RF: Random Forest, DT: Decision Tree, KNN: K-nearest neighbors, LR: Logistic Regression, and SL: Stack Learning.

Figure 5 shows box plots for different models in both datasets, manual versus automated. Wilcoxon p-value indicated significant difference between the manual and the automated datasets. In all models and metrics, the manual dataset had better performance than automated datasets (except KNN and LR with SPE metric). Figure 6 demonstrates the results of comparison of the performance of the models (with different metrics) using Wilcoxon test. As shown in Fig. 6 The performances of the models were significantly different except a few cases.

In this study we used a relatively large-scale dataset containing more than ten thousand ECG traces obtained during a community-based study among school-aged population in Iran. Based on the ML-approached models used in this study, the manual measurements of ECG features were better than automated measurements for discriminating normal from abnormal ECGs. Among ML models trained with manual features RF model with AUC, ACC, SEN, SPE equal to 0.93,0.98, 0.69, 0.99 and QDA with AUC, ACC, SEN, SPE equal to 0.90, 0.95, 0.75, 0.96 had the highest performances, while using the automated features QDA (AUC: 0.89, ACC: 0.92, SEN: 0.71, SPE: 0.93), and SL (AUC: 0.89, ACC: 0.96, SEN: 0.61, SPE: 0.99) reached best performances.

Several studies used automatic feature extraction from ECG and applied ML algorithms to distinguish normal from abnormal patients. Moridani et al.¹⁴ used automated algorithms to differentiate normal and abnormal ECG signals. They extracted heart rate variability (HRV) features by determining RR interval and achieved AUC of 89.3% and 94.7% for MLP and SVM classifiers, respectively. Jadhav et al. ¹⁵ in a similar study used feature elimination and ensembles learning for ECG arrhythmia diagnosis. They reported ACC: 91.1%, SEN: 85% and SPE: 96% with 15 and 20 features, respectively. In another study, Venkatesan et al. ¹⁶ used HRV feature extraction method, along SVM classifier and achieved accuracy of 96% for discrimination of normal from abnormal ECGs. Singh et al. ¹³ used dual tree complex wavelet transform (DTCWT) with linear discriminate analysis (LDA) as a hybrid feature extraction. Among their models back propagation neural network (BPNN), SVM and KNN reached highest accuracy of 99.7%. In this study we discriminate normal from abnormal ECGs in students aging from 6 to 18 years old, using manual and automated ECG feature and ML algorithms. Also, we used SL as ensample learning in our study and achieved AUC of 89% and ACC of 96% in automated dataset.

Ribeiro et al. ⁴² applied deep neural network on more than two million 12-lead ECG for automatic diagnosis of six abnormalities from ECGs. Mean of F1 measure was 80% in test dataset. They used 1-degree AV block (1dAVb), left bundle branch block (LBBB), atrial fibrillation (AF) abnormalities which are rare in the teenagers and also used binary classification for each abnormality. In our study we used PVC, PAC, Prolonged QTc and Prolonged PR instead of 1dAVb, LBBB and AF and combined abnormalities to perform a binary classification.

Since the ECGs in this study were taken from an epidemiological study, the number of abnormal cases was much lower than normal and the dataset was unbalanced. To overcome this problem, various metrics such as AUC, accuracy, sensitivity and specificity have been used and also bootstrapping was performed on test dataset. Another limitation of this study is the use of an single algorithm in our device (MEANS interpretation algorithm ²⁴) for automatic feature extraction. Using different automated algorithms to extract ECG features may improve model performance because extracting manual features is time consuming and non-reproducible. In future studies, extraction of various features from ECG signal as well as the use of deep learning methods could be performed.

In univariate analysis, heart rate and QT in manual and automated dataset had the highest performance in discriminating normal and abnormal cases in population of school age children. In multivariate analysis, QDA and SL in manual and automated dataset had the best result. This study demonstrated that the manual measurement of ECG features had better performance than the automated measurement but automated measurement in some model had promising result in discriminating normal and abnormal cases.

Acknowledgements

The study grant was mainly provided by the National Institute for Medical Research Development (NIMAD) under the identification number of 962126. The other portion of funding was provided by Rajaie Cardiovascular Medical and Research Center, Tehran, Iran. The SHED LIGHT investigators would like to thank who helped us during the electrocardiographic evaluations, particularly Ms Maryam Arvanesh, Mrs Akram Iranshahi, Mrs Akram Tabatabaee, and Mrs Maryam Alibakhshi.

Contributions

M.O., S.H. and M.K. conceptualized, administered project, acquired funding, designed the study. G.H. and I.S. designed and developed model and performed statistical analysis and figure generation. G.H. and Y.R. drafted the manuscript. M.K., Y.R., I.S., S.K. and M.A. interpreted the results, and critically reviewed. I.S., M.A. and S.K edited the manuscript. Y.R., G.H., B.K.B., I.S., S.K. and M.A. acquired and prepared data. N.S., A.T. and M.K. labeled data. All authors have read and approved the final manuscript.

Data Availability

The data used in this study are available from the corresponding author upon request.

Competing interests

The authors declare no competing interests.

Chow GV, Marine JE, Fleg JL. Epidemiology of arrhythmias and conduction disorders in older adults. Clinics in geriatric medicine 2012;28:539–553.
Sekar RP. Epidemiology of arrhythmias in children. Indian pacing and electrophysiology journal 2008;8:S8.
Benjamin EJ, Blaha MJ, Chiuve SE, et al. Heart Disease and Stroke Statistics—2017 Update: A Report From the American Heart Association. Circulation 2017;135:e146-e603.
Chiu SN, Wang JK, Wu MH, Chang CW, Chen CA, Lin MT, Wu ET, Hua YC, Lue HC. Cardiac conduction disturbance detected in a pediatric population. The Journal of pediatrics Jan 2008;152:85–89.
Sox HC, Jr., Garber AM, Littenberg B. The resting electrocardiogram as a screening test. A clinical analysis. Annals of internal medicine Sep 15 1989;111:489–502.
Hu YH, Palreddy S, Tompkins WJ. A patient-adaptable ECG beat classifier using a mixture of experts approach. IEEE transactions on bio-medical engineering Sep 1997;44:891–900.
Moraes JCTB, Seixas MO, Vilani FN, Costa EV. A real time QRS complex classification method using Mahalanobis distance. Paper presented at: Computers in Cardiology; 22–25 Sept. 2002, 2002.
Gonzalez Corcia MC, Sieira J, Pappaert G, de Asmundis C, Chierchia GB, La Meir M, Sarkozy A, Brugada P. Implantable Cardioverter-Defibrillators in Children and Adolescents With Brugada Syndrome. Journal of the American College of Cardiology Jan 16 2018;71:148–157.
Clifford GD, Liu C, Moody B, Li-wei HL, Silva I, Li Q, Johnson A, Mark RG. AF classification from a short single lead ECG recording: The PhysioNet/computing in cardiology challenge 2017. Paper presented at: 2017 Computing in Cardiology (CinC)2017.
Van de Loo A, Arendts W, Hohnloser SH. Variability of QT dispersion measurements in the surface electrocardiogram in patients with acute myocardial infarction and in normal subjects. The American journal of cardiology 1994;74:1113–1118.
Massel D. Observer variability in ECG interpretation for thrombolysis eligibility: experience and context matter. Journal of thrombosis and thrombolysis 2003;15:131–140.
Christov I, Bortolan G. Ranking of pattern recognition parameters for premature ventricular contractions classification by neural networks. Physiological measurement Oct 2004;25:1281–1290.
Singh R, Rajpal N, Mehta R. Abnormality detection in ECG using hybrid feature extraction approach. Paper presented at: 2018 First International Conference on Secure Cyber Computing and Communication (ICSCCC)2018.
Moridani M, Zadeh MA, Mazraeh ZS. An efficient automated algorithm for distinguishing normal and abnormal ECG signal. IRBM 2019;40:332–340.
Jadhav S, Nalbalwar S, Ghatol A. Feature elimination based random subspace ensembles learning for ECG arrhythmia diagnosis. Soft Computing 2014;18:579–587.
Venkatesan C, Karthigaikumar P, Paul A, Satheeskumaran S, Kumar R. ECG signal preprocessing and SVM classifier-based abnormality detection in remote healthcare applications. IEEE Access 2018;6:9767–9773.
Tadejko P, Rakowski W. Mathematical morphology based ECG feature extraction for the purpose of heartbeat classification. Paper presented at: 6th International Conference on Computer Information Systems and Industrial Management Applications (CISIM'07)2007.
Zhao Q, Zhang L. ECG feature extraction and classification using wavelet transform and support vector machines. Paper presented at: 2005 International Conference on Neural Networks and Brain2005.
Mahmoodabadi S, Ahmadian A, Abolhasani M. ECG feature extraction using Daubechies wavelets. Paper presented at: Proceedings of the fifth IASTED International conference on Visualization, Imaging and Image Processing2005.
Lin H-Y, Liang S-Y, Ho Y-L, Lin Y-H, Ma H-P. Discrete-wavelet-transform-based noise removal and feature extraction for ECG signals. Irbm 2014;35:351–361.
Xu X, Liu Y. ECG QRS complex detection using slope vector waveform (SVW) algorithm. Paper presented at: The 26th Annual International Conference of the IEEE Engineering in Medicine and Biology Society2004.
Jager F. Feature extraction and shape representation of ambulatory electrocardiogram using the karhunen-lòeve transform. Electrotechnical Review 2002;69:83–89.
Ahmadian A, Karimifard S, Sadoughi H, Abdoli M. An efficient piecewise modeling of ECG signals based on Hermitian basis functions. Paper presented at: 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society2007.
Van Bemmel J, Kors J, Van Herpen G. Methodology of the modular ECG analysis system MEANS. Methods of information in medicine 1990;29:346–353.
Amini M, Nazari M, Shiri I, Hajianfar G, Deevband MR, Abdollahi H, Arabi H, Rahmim A, Zaidi H. Multi-level multi-modality (PET and CT) fusion radiomics: prognostic modeling for non-small cell lung carcinoma. Phys Med Biol Oct 14 2021;66.
Amini M, Hajianfar G, Hadadi Avval A, Nazari M, Deevband MR, Oveisi M, Shiri I, Zaidi H. Overall Survival Prognostic Modelling of Non-small Cell Lung Cancer Patients Using Positron Emission Tomography/Computed Tomography Harmonised Radiomics Features: The Quest for the Optimal Machine Learning Algorithm. Clin Oncol (R Coll Radiol) Feb 2022;34:114–127.
Khodabakhshi Z, Amini M, Mostafaei S, Haddadi Avval A, Nazari M, Oveisi M, Shiri I, Zaidi H. Overall Survival Prediction in Renal Cell Carcinoma Patients Using Computed Tomography Radiomic and Clinical Information. J Digit Imaging Oct 2021;34:1086–1098.
Shiri I, Sadr A, Amini M, et al. Decentralized distributed multi-institutional PET image segmentation using a federated deep learning framework. Clin Nucl Med 2022;In press.
Shiri I, Amini M, Nazari M, Hajianfar G, Haddadi Avval A, Abdollahi H, Oveisi M, Arabi H, Rahmim A, Zaidi H. Impact of feature harmonization on radiogenomics analysis: Prediction of EGFR and KRAS mutations from non-small cell lung cancer PET/CT images. Comput Biol Med Mar 2022;142:105230.
Murat F, Yildirim O, Talo M, Baloglu UB, Demir Y, Acharya UR. Application of deep learning techniques for heartbeats detection using ECG signals-analysis and review. Comput Biol Med May 2020;120:103726.
Ye C, Coimbra MT, Kumar BV. Arrhythmia detection and classification using morphological and dynamic features of ECG signals. Paper presented at: 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology2010.
Tabassum T, Islam M. An approach of cardiac disease prediction by analyzing ECG signal. Paper presented at: 2016 3rd International Conference on Electrical Engineering and Information Communication Technology (ICEEICT)2016.
Sharma S, Drezner JA, Baggish A, et al. International Recommendations for Electrocardiographic Interpretation in Athletes. Journal of the American College of Cardiology 2017;69:1057–1075.
Rautaharju PM, Surawicz B, Gettes LS, et al. AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram: part IV: the ST segment, T and U waves, and the QT interval: a scientific statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: endorsed by the International Society for Computerized Electrocardiology. Circulation Mar 17 2009;119:e241-250.
Maron BJ, Friedman RA, Kligfield P, et al. Assessment of the 12-lead ECG as a screening test for detection of cardiovascular disease in healthy general populations of young people (12–25 Years of Age): a scientific statement from the American Heart Association and the American College of Cardiology. Circulation Oct 7 2014;130:1303–1334.
Kligfield P, Gettes LS, Bailey JJ, et al. Recommendations for the standardization and interpretation of the electrocardiogram: part I: the electrocardiogram and its technology a scientific statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society endorsed by the International Society for Computerized Electrocardiology. Journal of the American College of Cardiology Mar 13 2007;49:1109–1127.
Hancock EW, Deal BJ, Mirvis DM, et al. AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram: part V: electrocardiogram changes associated with cardiac chamber hypertrophy: a scientific statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: endorsed by the International Society for Computerized Electrocardiology. Circulation Mar 17 2009;119:e251-261.
Ackerman MJ, Priori SG, Willems S, et al. HRS/EHRA expert consensus statement on the state of genetic testing for the channelopathies and cardiomyopathies this document was developed as a partnership between the Heart Rhythm Society (HRS) and the European Heart Rhythm Association (EHRA). Heart rhythm Aug 2011;8:1308–1339.
Bazett HC. An analysis of the time relations of electrocardiograms. Heart 1920;7:353–370.
Kursa MB, Rudnicki WR. Feature Selection with the Boruta Package. J Stat Softw 2010-09-16 2010;36:13.
Kabir MF, Ludwig SA. Enhancing the Performance of Classification Using Super Learning. Data-Enabled Discov Appl 2019/01/17 2019;3:5.
Ribeiro AH, Ribeiro MH, Paixão GM, Oliveira DM, Gomes PR, Canazart JA, Ferreira MP, Andersson CR, Macfarlane PW, Meira Jr W. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nature communications 2020;11:1–9.

No competing interests reported.

PDF

Reviews received at journal
14 May, 2022
Reviewers agreed at journal
06 May, 2022
Reviewers invited by journal
02 May, 2022
Editor assigned by journal
02 May, 2022
Editor invited by journal
19 Apr, 2022
Submission checks completed at journal
19 Apr, 2022
First submitted to journal
12 Apr, 2022

You are reading this latest preprint version

Electrocardiogram Signal Classification Using Manual/Automated Features and Machine Learning Algorithms: A Large Cohort Study on Students Aging from 6 to 18 Years Old

Status:

Version 1

Abstract

Figures

Introduction

Methods

Data collection

Data Pre-processing

Univariate analysis

Multivariate analysis

Feature Selection

Machine Learning Classifiers

Model Development and Evaluation

Results

Univariate Outcomes

Multivariate Outcomes

Discussion

Conclusion

Declarations

References

Competing Interests

Status:

Version 1