AI-Based Detection of Myocardial Infarction

through Electrocardiogram Signals: A Review

Yasmine Shabaan Naglaa Fathy Tamer Abdelkader

Faculty of Computer and Information Faculty of Computer and Information Faculty of Computer and Information
Sciences Sciences Sciences
Ain Shams University Ain Shams University Ain Shams University
Cairo, Egypt Cairo, Egypt Cairo, Egypt
[email protected] [email protected] [email protected]

Rasha M. Ismail
Faculty of Computer and Information
Sciences
Ain Shams University
Cairo, Egypt
[email protected]

Abstract— Heart attack, medically termed Myocardial emergence of pathological Q waves, and occasionally the
Infarction (MI), happens when the heart muscle sustains occurrence of new conduction disturbances [3]. Most of
damage due to insufficient blood flow. MI ranks as the foremost published review articles regarding myocardial infarction
contributor to death among middle-aged and elderly individuals (MI) detection predominantly focus on traditional machine
on a global scale. AI-based approaches have the potential to learning methods [4] [5]. Furthermore, a review article
automatically diagnose MI by leveraging Electrocardiogram published in [6] pinpoints the utilization of both Machine
(ECG) signals. In this study, a comprehensive review is Learning (ML) and Deep Learning (DL) techniques in the
conducted to thoroughly evaluate Machine Learning (ML) and context of detecting myocardial infarction (MI) by analyzing
Deep Learning (DL) models, in identifying myocardial
ECG signals.
infarction (MI) through the analysis of ECG signals. The
manual extraction of features and the selection of ECG signals The objective of this research is to present a thorough
are necessitated by traditional machine learning approaches, overview of the methods proposed for analyzing the
whereas these tasks are automated by deep learning models. Electrocardiogram (ECG) signal to automate the detection of
Remarkably, Deep CNN (DCNNs) have demonstrated Myocardial Infarction (MI). We have examined an extensive
outstanding classification capabilities in the diagnosis of MI, collection of recent technical papers published within the last
leading to their increasing prominence in recent times. three years, focusing on the detection of Myocardial
Infarction. The essential elements of the proposed methods,
Keywords—Machine learning, Deep learning, Myocardial
Infarction, Heart Disorders, ECG.
encompassing preprocessing, feature extraction, and
classification approaches, are examined in-depth to provide a
I. INTRODUCTION comprehensive understanding. The next section in this paper
provides a description of Electrocardiogram (ECG) signals for
In recent times, cardiovascular disease (CVD) has MI Diagnosis Section 3 covers AI techniques for MI
emerged as a prominent global cause of mortality. Taking into Diagnosis Finally section 4 concludes the findings.
account the specific classifications of cardiovascular diseases
(CVDs), myocardial infarction (MI), commonly known as a II. ELECTROCARDIOGRAPHY FOR MI DIAGNOSIS
heart attack, is identified as posing a significantly elevated risk This section starts with a brief description about
of illness and death. It is responsible for approximately 15
electrocardiograms, followed by a summary of the datasets
million fatalities annually [1].
employed for automated Myocardial Infarction (MI)
Myocardial infarction (MI), occurs when a section of the heart detection.
muscle doesn't receive enough blood flow, leading to damage
or death of the affected tissue. The delivery of oxygenated A. Basics of Electrocardiogram (ECG)
blood to the cardiac muscle is the responsibility of the In contemporary healthcare, electrocardiography stands as
coronary arteries. When there is an obstruction in these the prevailing and extensively employed approach for
arteries, it can lead to a diminished blood flow and cause a monitoring vital signs. Through Electrocardiogram (ECG)
specific segment of the heart muscle to experience a lack of recordings, the body's surface electric potential is captured,
oxygen, resulting in its death. The impairment or demise of mirroring the heart's electrical signal propagation [7]. An
cardiac muscle tissue leads to an alteration in the regular ECG signal offers a diverse array of extractable features,
cardiac conduction system, giving rise to potentially serving various purposes across distinct end-goals. The ECG
Arrhythmias that pose a risk to life [2]. waveform comprises essential elements such as T wave, P
The primary approach employed to detect myocardial wave, and QRS complex, with the QRS complex holding
ischemia and MI involves the use of a 12-lead prominence. Detecting the QRS complex is crucial across
electrocardiogram (ECG). There is a wide range of ECG numerous clinical scenarios, facilitating the measurement and
indicators that suggest ischemia and MI, including ST- diagnosis of various heart irregularities [8]. Fig.1 shows an
elevation and depression, abnormalities in T-waves, the example of an ECG pattern in normal heartbeat.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 Fig. 1. ECG pattern in normal heartbeat

B. Public ECG Datasets

Electrocardiogram signals (ECG) are vital for identifying
cardiovascular conditions without the need for invasive
Fig. 3. Summary of the ECG signal in PTB [11]
procedures. Therefore, researchers can gain valuable insights
from readily accessible, top-tier ECG datasets. Below, some PTB-XL Dataset [12] [13] :
of the well-known public ECG datasets are presented. These
datasets have been extensively utilized for research related to This dataset is an extension of the PTB Diagnostic ECG
myocardial infarction diagnosis. dataset, which was created by the PTB for research purposes.
The collection contains 21,837 clinical 12-lead ECG
PTB Diagnostic ECG Dataset [9] [10]: recordings, each lasting for 10 seconds. The dataset originates
from 18,885 patients, with a gender distribution of 52% male
It is the abbreviation of "Physikalisch-Technische and 48% female. The ages of the patients span the entire
Bundesanstalt", denoting the national metrology institute of spectrum from 0 to 95 years, with a median age of 62 and an
Germany. Within this dataset, there are a total of 549 records interquartile range of 22. The dataset's significance stems
stemming from 290 different individuals. Each individual can from its inclusive assembly of diverse co-occurring medical
be identified through one to five distinct records. These conditions, as well as a substantial representation of healthy
records encompass the simultaneous measurement of 15 control samples.
signals, which encompass the conventional 12 leads (i, ii, iii,
avr, avl, avf, v1, v2, v3, v4, v5, v6) along with the 3 Frank Limitations of the Dataset: The dataset exhibits class
lead ECGs (vx, vy, vz). A significant amount of ECG imbalance, meaning that certain classes of ECG signals are
recordings is available, encompassing both records from represented more frequently than others. This can pose
individuals in good health and those afflicted by diverse challenges when training machine learning models, as it may
cardiac conditions. result in biased performance towards the majority classes.
Limitations of the Dataset: The size of the sample is small. Fig.4 shows sample ECG signals inside PTB-XL dataset and
This limited sample size may impact the statistical power of Fig.5 shows the description of these signals.
studies and may restrict the ability to investigate rare or less
common cardiac conditions. Fig.2 shows sample ECG signals
inside PTB dataset and Fig.3 shows the description of these
signals.

Fig. 4. ECG report of a patient in PTB-XL [11]

Fig. 2. ECG report of a patient in PTB [11]

 III. AI TECHNIQUES FOR MI DIAGNOSIS
This section explains the recent techniques employed in the
automated detection of MI.
A. Machine learning-based publications on MI diagnosis
Machine learning techniques utilized for myocardial
infarction (MI) classification encompass support vector
machine (SVM) [16], PNN [17], k-nearest neighbor (KNN)
[18], and decision tree (DT) [19]. In this section, we delve into
an examination of the research papers within this domain.
In [17] The study utilized Lead II ECG signals extracted from
the PTB database, encompassing data from 148 individuals
with myocardial infarction (MI) and 52 individuals with a
normal cardiac profile. To obtain relevant features from the
Fig. 5. Summary of the ECG signal in PTB-XL [11] signals, a non-linear feature extraction process was applied.
Subsequently, a feature selection procedure was executed to
MIT-BIH Arrhythmia Dataset [14] [15]: identify optimal features, enhancing analysis speed.
Significant features were determined using Student's t-test and
This dataset was created by the Massachusetts Institute of ranked based on their t-values. These features were then input
Technology (MIT) and the Beth Israel Hospital. It contains a into Decision Tree (DT), K-Nearest Neighbor (KNN),
diverse set of ECG recordings captured from various sources, Probabilistic Neural Network (PNN), and Support Vector
including both healthy individuals and patients with different Machine (SVM) classifiers. The SVM classifier, using a 10-
types of arrhythmias. It includes 48 long-term ECG fold Cross-Validation (CV), yielded the highest sensitivity,
recordings, with each record lasting for approximately 30 specificity, and accuracy values of 98.89%, 93.81%, and
97.96%, respectively.
minutes. The data is digitally sampled at a rate of 360 samples
per second for each channel, with an 11-bit resolution In [19] A gradient boosted trees model was constructed for
spanning a range of 10mV. This sampling process captures myocardial infarction (MI) prediction. Patient data utilized for
details regarding the timing and nature of each heartbeat. training the model were collected between 2011 and 2015,
originating from a prominent academic medical center in the
Limitations of the Dataset: The recordings in the MIT-BIH Western United States. The machine learning algorithm
dataset were collected in the 1970s and 1980s. As a result, the utilized structured data encompassing demographics, present
dataset may not accurately represent the current clinical illness history, medical background, vital signs, and lab
landscape, including advancements in medical technology, results. The model achieved noteworthy performance metrics,
changes in diagnostic criteria, or shifts in patient including sensitivity of 0.87, specificity of 0.70, and an Area
characteristics. Fig.6 shows sample ECG signals inside MIT- Under the Receiver Operating Characteristic curve (AUROC)
BIH dataset and Fig.7 shows the description of these signals. of 0.87. Notably, it outperformed established tools such as
TIMI and GRACE.
In [20] A novel method based on Random Forst was proposed
to detect myocardial infarction (MI) leveraging 12-lead ECG
signals from PTB database. The accuracy of the proposed
method was 99.71% for the intra-patient scheme and 85.82%
for the inter-patient scheme. Specifically, the inter-patient
paradigm divides the training and testing datasets based on
individual patients, ensuring no overlap between the two.
Conversely, in the intra-patient paradigm, beats from the same
patient might appear in both the training and testing sets.

Fig. 6. ECG report of a patient in MIT-BIH [11] B. Deep learning-based publications on MI diagnosis
Learning from large datasets becomes feasible for DL
because of its intricate multi-layered structure. As a result,
DL models like DCNN [3], Long Short-Term Memory
(LSTM) [21], CNN-VGG Net [22], and DenseNet [23] can
be employed for disease classification, showcasing superior
performance in signal processing and classification compared
to ML methods.

In [22] , two innovative networks, namely VGG-MI1 and

VGG-MI2 were developed for retraining the existing VGG-
Net, specifically for the purpose of diagnosing myocardial
infarction (MI). This study utilized Lead II ECG signals with
a duration of two seconds obtained from the PTB database.
These signals comprised of 21,092 normal ECG beats and
Fig. 7. Summary of the ECG signal in MIT-HIB [11]
80,364 ECG beats associated with myocardial infarction. In [8], authors proposed a new neural network architecture
Each pair of beats represented a 128 × 128 grayscale image. structured as a collection of subnetworks termed "modules."
To improve precision, data augmentation techniques were Each distinct "module" processes varying types of input data,
employed, thereby enhancing the trained model with a wide encompassing raw signals, extracted QRS complex entropies
range of intricate and diverse ECG content. VGG-MI1 QRS complexes, and raw signal entropy. The scope of the
achieved a remarkable accuracy of 99.02%, accompanied by research involves disease classifications across 2, 5, and 20
a sensitivity of 98.76% and specificity of 99.17%. Similarly, classes, utilizing 12-lead ECG signals sourced from the PTB-
VGG-MI2 demonstrated an overall accuracy, sensitivity, and XL dataset. This network achieved average accuracy of 90.0%
specificity of 99.22%, 99.15%, and 99.49%, respectively. ± 0.4% in the context of 2-class classifications. For 5-class
The main limitations of this research encompass the fact that categorizations, the accuracy stood at 76.2% ± 1.8%, while for
the proposed algorithm is more computationally expensive 20-class classifications, it reached 68.5% ± 1.3%. The
when compared to machine learning algorithms. techniques utilized in this research for computing entropy-
Furthermore, there is a need to conduct additional testing with based features and detecting R-waves have restricted
an extended dataset utilizing the proposed model. practicality because of their computational complexity.

In [21] approaches for myocardial infarction identification In [25] ECG data were used from the PTB-XL dataset that
from ECG signals were presented, incorporating a hybrid comprises 5 significant classes of 71 types of heart diseases.
CNN-LSTM method, CNN, and ensemble methodologies. These classes include hypertrophy, myocardial infarction,
This study employed a dataset of 123,998 single-lead ECG conduction disturbance, ST/T change and normal ECG. The
beats collected from the MIT-BIH arrhythmia DB and the experiment involved training two distinct neural networks –
PTBDB to construct the model. The issue of data imbalance, one designed for Few-Shot Learning (FSL) and another
specifically class imbalance, was successfully tackled by employing Deep Convolutional Neural Networks. These
incorporating the SMOTE + Tomek link sampling technique. networks' resultant vectors were subsequently compared. The
Remarkably high classification accuracies of 99.88%, research aimed to assess the efficacy of FSL (Few-Shot
99.82%, and 99.89% were achieved utilizing the hybrid CNN- Learning) for proximity-based classification of ECG signals.
LSTM, CNN, and ensemble techniques, respectively. The The FSL network demonstrated superior accuracy in the
main limitation of this research is the absence of cross- classification of healthy and sick patients, achieving a range
validation testing for the model. of 93.2% to 89.2%, in contrast to the softmax-based
classification network which attained an accuracy range of
In [24] PTB-XL dataset was employed to detect myocardial 90.5% to 89.2%. Additionally, the proposed FSL network
infarction (MI), focusing on a superclass categorized into the outperformed the softmax-based counterparts in categorizing
normal class (NORM) and four diagnostic classes: ST/T the five distinct disease classes, boasting an accuracy of 80.2%
change (STTC), MI, conduction disturbance (CD), and to 77.9%, as opposed to the range of 77.1% to 75.1% achieved
hypertrophy (HYP). The study exclusively utilized single- by the latter.
labeled data. The researchers explored two distinct scenarios.
In the initial scenario, they integrated the DCNN model for In [26], authors introduced EvoMBN, an evolving neural
both feature extraction and classification, establishing an network designed for myocardial infarction (MI) diagnosis.
average accuracy of 98.90%. In the second scenario, the The network employs a genetic algorithm (GA) to
proposed CNN served as a feature extractor, with the feature autonomously determine the best architectures for Multi-
set derived from the flatten layer's output. This arrangement branch Networks (MBN). Additionally, an innovative Lead
led to average accuracy rates of 99.20% when these features Squeeze and Excitation (LSE) block is developed to
were fed into the SVM classifier. compactly gather features from each of the branch networks.
The network's performance was evaluated through a series of
In [23] DenseNet and CNN architectures were developed to five-fold inter-patient cross-validation experiments conducted
address the classification of individuals as either healthy on MI detection and localization using the PTB diagnostic
subjects or patients categorized into 10 distinct classes database. Furthermore, the architectural knowledge gained
determined by the myocardial involvement location in cases from the PTB database is successfully applied to the PTB-XL
of myocardial infarction (MI). This study involved detecting database without any modifications. Impressively, the model
R peaks within the individual lead signals of 12-lead ECGs attains an accuracy rate of 90.80% on the PTB-XL database
sourced from the PTB dataset. Extracted beats were then fed and 97.11% on the PTB database.
into both the CNN and DenseNet models for analysis. The
CNN and DenseNet models displayed impressive In [3], a new deep ensemble architecture named Deep Multi-
classification results, with average accuracy rates of 98.5% Scale Convolutional neural network Ensemble (DMSCE) is
and 98.9%, respectively. Although robust classification proposed. It integrates diverse scale-dependent DCNN (SD-
capabilities are provided by both models, there is a preference DCNN) expert classifiers. This innovative approach aims to
for the DenseNet model. This preference arises from its capture scale-specific disease characteristics within input
advantages, including reduced computational complexity and ECG data, generating diagnostic decisions. The DMSCE
improved feature reusability. To gain insight into the decision- framework comprises three key elements: the scale-
making processes of these models, the Grad-CAM technique dependent DCNN expert classifiers, the DCNN gating
was subsequently employed to analyze their respective network, and a strategically designed ensemble fusion
outputs. The main limitation of this research is the imbalance mechanism. The effectiveness of the DMSCE architecture
present in the dataset used for the study. was assessed using two publicly accessible ECG datasets:
CinC-training2017, a single-lead ECG dataset, and PTBXL- learning approaches, unlike conventional machine learning
2020, a 12-lead ECG dataset. Data augmentation methods are techniques. It is also noticed that machine learning techniques
utilized to address the problem of data imbalance within the achieved high accuracy values using datasets with smaller
training datasets. Remarkably, the model achieved an size. Yet, deep learning techniques outperform when dealing
accuracy rate of 85.65% on the PTB-XL database. with large-scale datasets. A summary of AI-based techniques
for automated MI diagnosis is presented in Table 1 and Table
In [27], a new model is presented, introducing a multilead 2.
branch that combines a residual network with squeeze and
excitation networks and a bidirectional long short-term TABLE I. A SUMMARY OF ML TECHNIQUES USED FOR MI
DIAGNOSIS
memory (LSTM) model, referred to as MLB-ResNet-SENet-
BL. This model aims to tackle the challenge of insufficiently Author Methods No. Dataset Classification
explored spatial-temporal relationships within and between (year used k-fold Results
published) CV
ECG leads. The study also includes an evaluation using Sridhar et Decision 10- PTB DB KNN
fivefold cross-validation for myocardial infarction (MI) al. [2020] Tree (DT) fold Acc: 0.973
detection and localization, relying on the PTB and PTB-XL KNN CV DT
datasets. For MI detection, the proposed MLB-ResNet- Probabilistic Acc: 0.955
Neural SVM
SENet-BL model achieved remarkable performance with an Network Acc: 0.979
accuracy of 99.94%, sensitivity of 99.99%, specificity of (PNN) PNN
99.80%, and an F1-score of 99.96%. When it comes to MI SVM Acc: 0.904
localization, the model also exhibited highly favorable test Wang et al. Random 10- PTB DB For the inter-
[2020] Forst (RF) fold patient scheme:
results, boasting an accuracy of 99.69%, sensitivity of CV ACC: 85.82%
99.58%, specificity of 99.94%, and an F1-score of 99.61%. For the intra-
Notably, the main limitations of this research revolve around patient scheme:
the challenge of data imbalance between MI patients and ACC :99.71%,
Kayikcioglu Weighted 10- Long-Term ACC: 94.23%
healthy control subjects, which has an impact on performance et al. [2020] k-nearest fold ST DB,
improvements. Additionally, the Grad-CAM-based neighbor CV European
interpretability method still lacks complete trust among (KNN) ST-T DB
medical professionals, indicating a need for further and
MIT-BIH
enhancements. Arrhythmia
DB.
In [28], authors proposed two hybrid models, namely CNN Lin et al. k-nearest 10- PTB DB ACC: 99.57%
LSTM and CNN BILSTM, for the automated detection of [2020] neighbor fold
myocardial infarction (MI). These models are designed for (KNN) CV
Panchavati Gradient 3-fold Electronic ACC: 87.00%
multiclass classification, categorizing outputs into MI, et al. [2021] boosted tree CV health
Healthy, and Other Cardiovascular Diseases (CVD), and a model records
comparison of their performance was conducted. (systolic/dias
Additionally, the study focused on identifying the most tolic blood
pressure, etc)
relevant ECG leads for efficient computational resource
utilization during model development and training. Notably, Sulthana et Multi-linear 5- At a hospital, ACC: 94.03%
this research stands out as the first to apply multiclass MI al. [2022] regression+ fold a clinical
detection to the PTB XL dataset, as opposed to the smaller CV dataset was
PTB dataset, and it employed data balancing techniques to Probabilistic collected
PCA + from the
mitigate biases. Remarkably, the system demonstrated the cardiology
ability to self-learn without manual feature extraction, RBF based department.
leading to quicker training times. The performance SVMs
evaluation of the models indicated an overall accuracy of
89% for CNN LSTM and 91% for the CNN BILSTM model.
However, a notable drawback is that these models are
considered black-box models, lacking the capability to TABLE II. A SUMMARY OF DL TECHNIQUES USED FOR MI DIAGNOSIS
provide explanations that support their prediction results. As Author (year Methods No. Dataset Classification
a potential avenue for future research, it would be beneficial published) used k- Results
to enhance this work by incorporating explainable machine fold
learning models to address this limitation. CV
Alghamdi et CNN-VGG 10- PTB VGG-MI1
al. [2020] Net fold DB Acc: 99.02% with
C. Discussion VGG-MI1 CV augmentation
VGG-MI2 Acc: 97.57% without
A notable benefit of deep learning methods lies in their ability augmentation
to process unprocessed ECG signals directly, bypassing the VGG-MI2
need for preprocessing or segmentation. This capability helps Acc: 99.22% with
augmentation
mitigate errors introduced in various processing stages, Acc: 98.07%
notably during preprocessing and segmentation. without
Furthermore, hand-crafted features are not required by deep augmentation
Rai et al. hybrid CNN- NC MITDB Hybrid CNN-LSTM uncover that deep learning models attain the highest accuracy
[2021] LSTM and Acc: 99.88 % in diagnosing MI and exhibit compatibility with large
PTBDB
CNN datasets. Frequently, deep learning models function like
Acc: 99.82% opaque systems, creating challenges in comprehending the
CNN rationale behind their predictions. In medical contexts,
Śmigiel et al. DNN 5- PTB- On 2 Classes interpretability is crucial for clinicians to trust and make
[2021] fold XL DB Acc: 90.0%
CV On 5 Classes informed decisions based on the model's outputs.
Acc:76.2% ± 1.8% Consequently, many researchers are embracing DL
On 20 Classes: techniques, although they come with their own limitations
Acc: 68.5% ± 1.3% requiring future improvements.
Liu et al. Multi-branch 5- PTB On PTB:
[2021] Network with fold DB Acc: 97.11% REFERENCES
GA CV
PTB- On PTB-XL: [1] W. Liu, J. Ji, S. Chang, H. Wang, J. He, and Q.
XL DB Acc: 90.80% Huang, “Evombn: Evolving multi-branch networks
Hammad et al. End-to-end 10- PTB- CNN
[2022] CNN model fold XL DB Acc:98.9
on myocardial infarction diagnosis using 12-lead
for each class CV electrocardiograms,” Biosensors (Basel), vol. 12, no.
1, Jan. 2022.
The features [2] H. M. Rai and K. Chatterjee, “Hybrid CNN-LSTM
are extracted
deep learning model and ensemble technique for
using CNN,
followed by SVM classifier automatic detection of myocardial infarction using
classification Acc: 99.20% big ECG data,” Applied Intelligence, vol. 52, pp.
with an SVM 5366–5384, 2021.
classifier.
[3] E. Prabhakararao and S. Dandapat, “Multi-Scale
Jahmunah et DenseNet 10 - PTB DenseNet Convolutional Neural Network Ensemble for Multi-
al. [2022] fold DB Acc:98.9% Class Arrhythmia Classification,” IEEE J Biomed
CNN CV CNN
Acc:98.5 Health Inform, vol. 26, no. 8, pp. 3802–3812, Aug.
Pałczyʼnski et Few-Shot 5- PTB- Few-Shot Learning 2022.
al. [2022] Learning fold XL DB Acc:93.2% [4] A. Chakraborty, S. Chatterjee, K. Majumder, R. N.
CV DNN Shaw, and A. Ghosh, “A Comparative Study of
DNN Acc:90.5 %
Myocardial Infarction Detection from ECG Data
Prabhakararao Deep multi- 5- PTB- On PTB-XL :
et al. [2022] scale fold XL ACC: 85.65 Using Machine Learning,” in Lecture Notes in
convolutional CV Networks and Systems, Springer Science and
neural CinC- On CinC- Business Media Deutschland GmbH, 2022, pp. 257–
network 2017 Training2017 :
267.
ensemble Precision : 88.4
(DMSCE) Recall :88.2 [5] S. Ansari et al., “A Review of Automated Methods
F1:88.3 for Detection of Myocardial Ischemia and Infarction
Han et al. residual 5- PTB MI detection Using Electrocardiogram and Electronic Health
[2023] network with fold DB and Acc:99.94% Records,” IEEE Rev Biomed Eng, vol. 10, pp. 264–
squeeze and CV Se: 99.99%
excitation PTB- Sp: 99.80%
298, 2017.
networks and XL DB F1-score: 99.96% [6] J. Hassannataj Joloudari et al., “Application of
a bidirectional MI localization artificial intelligence techniques for automated
long short- Acc: 99.69% detection of myocardial infarction: a review,”
term memory Se: 99.58%
(LSTM) Sp: 99.94%
Physiol Meas, vol. 43, no. 8, p. 08TR01, Aug. 2022.
(MLB- F1-score: 99.61% [7] E. Merdjanovska and A. Rashkovska,
ResNet- “Comprehensive survey of computational ECG
SENet-BL) analysis: Databases, methods and applications,”
Hasbullah et CNN LSTM NC PTB CNN LSTM Expert Systems with Applications, vol. 203. Elsevier
al. [2023] XL Acc: 89%
Ltd, Oct. 01, 2022.
CNN CNN BILSTM [8] S. Śmigiel, K. Pałczyński, and D. Ledziński, “Deep
BILSTM Acc: 91% Learning Techniques in the Classification of ECG
Signals Using R-Peak Detection Based on the PTB-
XL Dataset,” Sensors, vol. 21, no. 24, p. 8174, Dec.
2021.
[9] “https://physionet.org/content/ptbdb/1.0.0/,ptbdb”.
IV. CONCLUSION
[10] R. Bousseljot, D. Kreiseler, and A. Schnabel,
The detection of myocardial infarction (MI) is greatly aided “Nutzung der EKG-Signaldatenbank CARDIODAT
by ECG signals, but manual analysis is both time-consuming der PTB über das Internet,” Biomedizinische
and subject to bias. Machine learning (ML) relies on Technik/Biomedical Engineering, pp. 317–318, Jul.
handcrafted ECG features, while deep learning (DL) excels 1995.
in automatic feature extraction. Our study focuses on [11] PhysioBank and PhysioToolkit, “Physionet:
evaluating current methods of myocardial infarction components of a new research resource for complex
detection utilizing both ML and DL models. The findings
 physiologic signals,” Circulation, vol. 101, pp. e215– methods for urban healthcare in smart cities,”
e220, 2000. Multimed Tools Appl, Mar. 2020.
[12] “https://physionet.org/content/ptb-xl/1.0.1/,ptb- [23] V. Jahmunah, E. Y. K. Ng, R.-S. Tan, S. L. Oh, and
xldb”. U. R. Acharya, “Explainable detection of myocardial
[13] Wagner et al., “PTB-XL, a large publicly available infarction using deep learning models with Grad-
electrocardiography dataset,” PhysioNet. CAM technique on ECG signals,” Comput Biol Med,
[14] “https://physionet.org/content/mitdb/1.0.0/,mitdb”. vol. 146, p. 105550, Jul. 2022.
[15] G. B. Moody and R. G. Mark, “The impact of the [24] M. Hammad et al., “Automated Detection of
MIT-BIH arrhythmia database.,” IEEE Eng Med Biol Myocardial Infarction and Heart Conduction
Mag, vol. 20, no. 3, pp. 45–50, 2001. Disorders Based on Feature Selection and a Deep
[16] A. R. Sulthana and A. K. Jaithunbi, “Varying Learning Model,” Sensors, vol. 22, no. 17, p. 6503,
combination of feature extraction and modified Aug. 2022.
support vector machines based prediction of [25] K. Pałczyński, S. Śmigiel, D. Ledziński, and S.
myocardial infarction,” Evolving Systems, vol. 13, Bujnowski, “Study of the Few-Shot Learning for
no. 6, pp. 777–794, Dec. 2022. ECG Classification Based on the PTB-XL Dataset,”
[17] C. Sridhar et al., “Accurate detection of myocardial Sensors, vol. 22, no. 3, p. 904, Jan. 2022.
infarction using non linear features with ECG [26] W. Liu, J. Ji, S. Chang, H. Wang, J. He, and Q.
signals,” J Ambient Intell Humaniz Comput, vol. 12, Huang, “EvoMBN: Evolving Multi-Branch
no. 3, pp. 3227–3244, Oct. 2020. Networks on Myocardial Infarction Diagnosis Using
[18] Z. Lin, Y. Gao, Y. Chen, Q. Ge, G. Mahara, and J. 12-Lead Electrocardiograms,” Biosensors (Basel),
Zhang, “Automated detection of myocardial vol. 12, no. 1, p. 15, Dec. 2021.
infarction using robust features extracted from 12- [27] C. Han, J. Sun, Y. Bian, W. Que, and L. Shi,
lead ECG,” Signal Image Video Process, vol. 14, no. “Automated Detection and Localization of
5, pp. 857–865, Jul. 2020. Myocardial Infarction With Interpretability Analysis
[19] S. Panchavati et al., “Retrospective validation of a Based on Deep Learning,” IEEE Trans Instrum
machine learning clinical decision support tool for Meas, vol. 72, 2023.
myocardial infarction risk stratification,” Healthc [28] S. Hasbullah, M. S. Mohd Zahid, and S. Mandala,
Technol Lett, vol. 8, no. 6, pp. 139–147, Aug. 2021. “Detection of Myocardial Infarction Using Hybrid
[20] Z. Wang, L. Qian, C. Han, and L. Shi, “Application Models of Convolutional Neural Network and
of multi-feature fusion and random forests to the Recurrent Neural Network,” BioMedInformatics,
automated detection of myocardial infarction,” Cogn vol. 3, no. 2, pp. 478–492, Jun. 2023.
Syst Res, vol. 59, pp. 15–26, Jan. 2020. [29] İ. Kayikcioglu, F. Akdeniz, C. Köse, and T.
[21] H. M. Rai and K. Chatterjee, “Hybrid CNN-LSTM Kayikcioglu, “Time-frequency approach to ECG
deep learning model and ensemble technique for classification of myocardial infarction,” Computers
automatic detection of myocardial infarction using and Electrical Engineering, vol. 84, Jun. 2020.
big ECG data,” Applied Intelligence, vol. 52, no. 5,
pp. 5366–5384, Aug. 2021.
[22] A. Alghamdi et al., “Detection of myocardial
infarction based on novel deep transfer learning