Automatic Detection of Cardiovascular Abnormalities in ECG Images:

CNN and MobileNet

Aditi Jambotkar

, Regina Fernandes

, Shreya Arun Miskin

, Prema T. Akkasaligar

and Rajashri Khanai

Department of Computer Science and Engineering, KLE Tech. University’s Dr. MSSCET, Belagavi, India

Keywords:

ECG Images, CNN, MobileNet, Cardiovascular Diseases.

Abstract:

Cardiovascular diseases are becoming a leading cause of death worldwide. Detection of irregular heart activi-

ties like arrhythmia and heart attacks are critical for timely treatment. Automation in detecting cardiovascular

abnormalities is essential for providing timely diagnosis, especially in resource-limited settings where trained

medical professionals may be scarce. The paper aims to detect cardiovascular abnormality in ECG images

automatically using deep learning techniques. It uses a Convolutional Neural Network(CNN) and MobileNet

for efﬁcient and lightweight processing. The MobileNet model outperforms the CNN model, demonstrating

superior accuracy, precision and recall. The results show the potential of deep learning models in enhancing

the accuracy and automation of cardiovascular abnormality detection through ECG analysis. By automating

ECG interpretation, it enables early detection of abnormalities, reduces diagnostic delays, and improves pa-

tient care, particularly in resource-constrained settings.

1 INTRODUCTION

Electrocardiogram (ECG) imaging is an essential di-

agnostic tool used in cardiology to measure heart ac-

tivity. It monitors the electrical signals of the heart

and is basically used for detecting various abnormal-

ities in heart activities like arrhythmia and myocar-

dial infractions. Abnormalities in ECG images often

indicate severe cardiovascular conditions that require

immediate medical attention. Delayed diagnosis can

lead to critical consequences for patients. Automated

abnormality detection plays a crucial role in address-

ing this issue by enabling quick and accurate analysis

of ECG data, facilitating early and effective treatment.

Cardiovascular diseases are the leading cause of

death globally, taking an estimated 17.9 million lives

each year. According to (World Health Organiza-

tion, 2015), in 2000, around 14 million people died

from cardiovascular diseases globally, while in 2019,

it reached close to 18 million. The emerged need for

improved healthcare systems is more important than

https://orcid.org/0009-0002-9983-615X

https://orcid.org/0009-0002-4838-3824

https://orcid.org/0009-0001-6206-2309

https://orcid.org/0000-0002-2214-9389

https://orcid.org/0000-0002-5080-722X

ever, particularly as cardiovascular diseases remain to

be the leading cause of death worldwide. ECG imag-

ing is at forefront of heart care monitoring due to its

ability to capture the critical data of the heart.

(Agarwal et al., 2024) presents a novel approach

for detecting abnormalities in ECG images by uti-

lizing a MobileNet based CNN autoencoder. The

lightweight architecture is designed for efﬁcient pro-

cessing, making it ideal for real-time applications and

devices with limited computational resources. The

autoencoder learns compact representations of ECG

images during the encoding phase and reconstructs

them during decoding, allowing the system to iden-

tify abnormalities based on discrepancies between the

original and reconstructed images. By leveraging Mo-

bileNet’s efﬁciency and the autoencoder’s capability

to highlight subtle deviations, the method achieves

high diagnostic accuracy. The study demonstrates the

potential of this approach for improving anomaly de-

tection in ECG images, ensuring reliability and scala-

bility across different datasets and clinical scenarios.

The MobileNet50 CNN autoencoder method for ECG

anomaly detection has several potential drawbacks. It

faces challenge with overﬁtting on limited or biased

datasets, reducing generalizability across diverse pop-

ulations. Handling noisy ECG signals, common in

real-world scenarios, also degrade performance. The

754

Jambotkar, A., Fernandes, R., Miskin, S. A., Akkasaligar, P. T. and Khanai, R.

Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet.

DOI: 10.5220/0013601600004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 754-761

ISBN: 978-989-758-763-4

method’s reliance on a complex deep learning archi-

tecture introduces latency and impacting real-time ap-

plications. Additionally, without thorough compar-

ison to simpler techniques or clinical validation, its

real-world efﬁcacy remains uncertain. Addressing

these issues with robust datasets enhance its practical

applicability. The deep-learning models for MI

detection and QRS complex detection face many chal-

lenges. Deep learning techniques used like 1D CNNs

and hybrids of CNN-LSTM are limited by high com-

putational resources and data dependency. Image-

based methods and low-quality signal processing ma-

jorly have difﬁculties with noise reduction, feature ex-

traction etc. The models perform better in some en-

vironments but struggle with over-ﬁtting which im-

pacts the real-world applications. Enhanced machine

learning and deep learning algorithms need thorough

validation in clinical settings.

The challenges observed in the problem space in-

clude data availability and format, where the open-

access datasets for ECG images are limited. ECG

data is usually recorded on paper in clinical settings

which makes it difﬁcult to digitize, store and ana-

lyze the data. ECG images usually consists of text-

annotations, grid lines and other background elements

that often interfere with the image extraction. The

other challenges include model performance and val-

idation, where achieving high sensitivity is crucial as

false negatives in MI detection can have serious health

consequences. These challenges highlight the impor-

tance of using reliable image processing and machine

learning methods to accurately identify heart diseases

using ECG images.

The present study introduces an advanced abnor-

mality detection model that utilizes CNNs for accu-

rate and reliable classiﬁcation. The model is specif-

ically designed to handle ECG image data, ensuring

that it undergoes thorough preprocessing to improve

input quality. By leveraging the preprocessing steps,

the model enhances the consistency and quality of the

ECG image data. Furthermore, the CNN-based archi-

tecture excels in extracting meaningful and relevant

features from the processed ECG images, enabling ef-

ﬁcient classiﬁcation of abnormalities. This approach

aims to address challenges in medical diagnostics by

providing a robust and precise solution for ECG anal-

ysis.

The paper is organized as follows. Section 2 pro-

vides a brief review of the literature survey on the

recent works. Section 3 contains the problem state-

ment, background and provides proposed methodol-

ogy. Section 4 discusses the implementation details

along with results and discussions. Finally, the paper

concludes in Section 5.

2 RELATED WORKS

The detection of ECG peaks has seen exceptional

progress with the application of advanced image

processing techniques and deep learning algorithms.

This survey explores a wide range of methods pro-

posed by researchers for accurate and efﬁcient detec-

tion of ECG features which enables the detection of

arrhythmia and other cardiac conditions.

(Sane et al., 2021) developed a computerized

method for detecting Myocardial Infarction(MI) by

using a dataset containing 12-lead ECG images. They

included a two-step approach which involved im-

age processing to extract ECG signals from the im-

ages, followed by a one-dimensional CNN to classify

MI. The model is validated on the PTB diagnostic

database. This method avoids the need for manual

computation and handcrafted features. The advantage

of this model is that it is adaptable to various datasets

and provides real-time assessment, which makes it

well-suited for critical health conditions. However,the

limitation lies in the challenge of extracting signals

from ECG images and the lack of open access to ECG

image datasets.

(Zhou et al., 2020) introduced a novel deep-

learning algorithm for the real-time prediction of R-

peaks in ECG signals using a combined model of

CNN and LSTM network. The designed strategy is to

predict the next R-peak by computing the variability

of previous ECG intervals which indicate the poten-

tial future health problems like depression, anxiety,

asthma and sudden infant death. It also proves to be

a strong indicator for the onset of myocardial infrac-

tions. The model is validated on MIT-BIH arrhyth-

mia dataset and the combined model outperforms the

standalone models like CNN or LSTM by combin-

ing their strengths. CNN has been used for ﬁltering

noise and extracting visual pattern, LSTM for han-

dling temporal dynamics in ECG signals. The advan-

tage of the model includes its ability to perform real-

time monitoring. However, the accessibility of high-

quality ECG data and computational resources are the

key-issues.

(Yuen et al., 2019) developed a innovative CNN-

LSTM model for identifying QRS complexes in noisy

ECG signals. In this approach, CNN is used to extract

features, LSTM identiﬁes the QRS complex timings,

and a multi-layer perceptron makes the ﬁnal predic-

tions. The model performs well in scenarios where

the training and testing datasets have data from dif-

ferent patients. It is tested using the MIT-BIH dataset

and evaluated based on metrics like precision, recall,

and F1 score. Advantages of this approach include

adaptability to noisy signals which makes it suitable

Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet

755

to generalize on unseen patient data.

(Cai and Hu, 2020) introduced two deep learn-

ing models for QRS complex detection in ECG sig-

nals: a CNN and a hybrid Convolutional Recurrent

Neural Network(CRNN). The CNN is fundamentally

conposed of convolutional blocks and Sqeeze-and-

Exitation networks, while the CRN combines both

convolutional and recurrent layers to improve feature

extraction and temporal dependency learning. The

model is evaluated on four open access ECG datasets.

Advantages of this model involves its noise resistance

and high accuracy, however it faces issues such as

over-ﬁtting, computational complexity which limits

their effectiveness in real-time applications.

(Zahid et al., 2022) created a reliable system to

detect R-peaks in low-quality Holter ECG signals. It

uses a 1D CNN combined with a veriﬁcation model

to reduce false alarms. The approach includes a

encoder-decoder structure that generates a 1D seg-

mentation map for precise localization of R-peaks

from ECG inputs. The model is tested on China Phys-

iological Signal Challenge(CPSC) and MIT-BIH Ar-

rhythmia database, achieving a excellent performance

on F1 score on CPSC database while showing better

results on MIT-BIH database. The advantages of this

approach lies in the adaptability to low-quality signals

and its generalization across different datasets, which

makes it suitable for real-time applications. How-

ever,the issue of over-ﬁtting remains as a challenge,

particularly in variable ECG environments.

(Das, 2024) proposed a comprehensive system

for blood pressure prediction using three machine

learning algorithms: Support Vector Classiﬁer (SVC),

Random Forest Classiﬁer, and Naive Bayes Classi-

ﬁer. The approach is tested on a dataset consisting

of 3000 image samples, of which 2005 represented

cases of high blood pressure and 495 represented nor-

mal blood pressure. The strength of the model lies in

the synergistic use of multiple algorithms, which sig-

niﬁcantly enhances the accuracy and reliability of pre-

dictions by leveraging the strengths of each method.

SVC effectively handles the separation of complex

data, Random Forest contributes robustness through

ensemble learning, and Naive Bayes adds simplicity

and speed to the classiﬁcation process. One major

drawback is the increased computational complexity,

which arises from training multiple algorithms and

integrating their outputs. Efﬁciently handling large

datasets remains challenging due to high memory us-

age and processing time.

(Yang et al., 2021) proposed a hybrid deep learn-

ing model for non-invasive, cuff-less blood pressure

estimation. The study focuses on using raw ECG im-

ages directly as input for deep learning models. Two

types of experiments are caried out. In the ﬁrst, phys-

ical characteristics and features from the ECG signals

are extracted and used with traditional ML techniques

like SVR, LASSO, Ridge regression, KNN, Multi-

ple Linear Regression, and AdaBoost. In the second,

DL models such as CNNs, LSTMs, and fully con-

nected networks are applied for testing and analysis

of model. The advantage of the model is its ability

to automatically extract features from raw PPG and

ECG signals which reduces the complexity of use for

users. However, a disadvantage is the model’s re-

liance on insufﬁcient data for optimal performance.

Table 1 illustrates recent 2024 literature surveys for

cardiovascular diseases detection in ECG Images.

In conclusion, the literature survey provides a

comprehensive overview of research on abnormalities

in ECG images, highlighting a wide range of stud-

ies that identify key trends and systematic methodolo-

gies. By addressing the existing gaps in this domain,

the study aims to contribute signiﬁcantly to the ad-

vancement of accurate and efﬁcient detection of car-

diovascular abnormalities.

3 PROPOSED METHODOLOGY

The study aims on detecting abnormalities in ECG

images by utilizing advanced and enhanced deep

learning techniques to identify irregular heart activ-

ities such as arrhythmia and MI. Consider a patient

experiencing chest pain. An ECG is conducted, and

the image is analyzed for any signs of MI or other

irregular heart conditions. Currently, the analysis is

performed manually by trained professionals. How-

ever this can lead to delays and many inconsisten-

cies. An automated model quickly detects whether

the ECG contains any abnormalities which takes less

usage of time and treatment decisions can be taken on

time.

The objectives are as follows :

• To preprocess the data by cleaning it to improve

the image quality of ECG images for better per-

formance of the model.

• To design and implement robust deep learning

models for accurate detection and classiﬁcation of

abnormalities in ECG signals.

• To assess the developed model’s performance by

using the evaluation metrics.

Quality of ECG images is assumed that the ECG

images provided are of acceptable resolution and

sufﬁcient quality for processing is present. The ECG

images are assumed to have consistent dimensions

INCOFT 2025 - International Conference on Futuristic Technology

756

Table 1: Literature Surveys for Cardiovascular Diseases Detection.

Year Paper Title Dataset Approach Accuracy Precision Recall F-score Gaps identi-

ﬁed

2024 Detection of

Cardiovascular

Disease Using

ECG images

(Jessy et al.,

2024)

ECG dataset col-

lected from hospitals

SVM, K-NN, DT,

RF, Naive Bayes

98.23% 98.31% 97.50% 97.90% Limited gen-

eralizability;

focus on

classiﬁcation

only

2024 Interpreting

Deviant Heart

Patterns: Mo-

bileNet CNN

Autoencoder

(Agarwal et al.,

2024)

ECG images dataset

with anomalies

MobileNet (CNN) 76.93% 55.23% 70.00% 61.63% Low accuracy

and not effec-

tive on larger

ECG datasets

2022 Pan-

Tompkins++:

A Robust Ap-

proach to Detect

R peaks in ECG

Signals (Imtiaz

and Khan, 2022)

MIT-BIH Arrhyth-

mia, European ST-T,

PhysioNET PTB,

Atrial Fibrillation

dataset

CNN, CRNN 97.49% 95.89% 96.00% 95.94% High process-

ing demands

2022 Energy Efﬁcient

Compression Al-

gorithm of ECG

Signals (Fathi

et al., 2022)

Real-time ECG im-

ages dataset

Krawtchouk and

AALO algorithm

92.00% 91.00% 90.00% 90.50% Lack of

anomaly

detection

capabilities

2022 Intelligent Sys-

tem for ECG

Classiﬁcation

Using CNN

(Hammad and

Abdulbaqi, 2022)

Publicly available

ECG dataset

1D CNN 90.00% 88.00% 87.00% 87.50% Limited in

handling

diverse ECG

signal varia-

tions

Figure 1: Proposed Methodology for Cardiovascular Abnormality Detection.

after pre-processing, making them compatible with a

CNN model. It is assumed that a sufﬁcient amount of

labeled ECG image data is provided for testing and

training.

The proposed solution utilizes deep learning

models, speciﬁcally CNN and MobileNet, to au-

tomate the detection of ECG signal abnormalities,

offering a fast, accurate, and reliable diagnostic tool

Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet

757

for cardiovascular diseases. The methodology begins

by preprocessing ECG signals to ensure consistency

across ECG samples. CNN extracts critical temporal

and spectral features, such as P and T waves and QRS

complexes, which are vital for identifying cardiac

conditions. Its convolutional layers capture complex

patterns, pooling layers reduce dimensionality while

preserving essential information. The fully connected

layers classify signals into normal or abnormal cate-

gories. Complementing this, MobileNet a lightweight

architecture, is ﬁne-tuned on the ECG dataset for

efﬁcient real-time analysis, using depthwise separa-

ble convolutions to optimize feature extraction with

minimal computational overhead. A global average

pooling layer and customized dense layers adapt

MobileNet for ECG classiﬁcation. Both models are

trained on the same dataset with early stopping to

avoid overﬁtting and ensure optimal performance.

Comprehensive evaluation metrics, including ac-

curacy, precision, recall, F1-score, and confusion

matrix, are used to assess their effectiveness. The

framework uses CNN’s robustness and MobileNet’s

efﬁciency to deliver a scalable and accessible solution

for ECG analysis.

The schematic diagram illustrating the proposed

methodology is depicted in Figure 1. Initially, ECG

images are drawn from the dataset consists of la-

beled samples of ECG signals. Preprocessing steps

such as resizing the images to a standard dimension

and normalizing pixel intensity values are performed

to ensure data consistency and quality. These steps

aim to enhance the clarity of the ECG images, which

is crucial for accurate feature extraction. The pre-

processed ECG images are then trained using CNN

and MobileNet models. The CNN architecture in-

corporates convolutional layers that use ﬁlters to ex-

tract local features, pooling layers to reduce spa-

tial dimensions, and fully connected layers to pre-

dict the class labels by analyzing the acquired fea-

tures. Activation function, such as ReLU, introduce

non-linearity, enabling the network to learn complex

patterns and relationships in ECG images. CNNs ex-

cel in automatically extracting hierarchical represen-

tations of features from ECG images, which facili-

tates effective classiﬁcation of normal and abnormal

ECG. Similarly, the MobileNet model is employed to

efﬁciently analyze ECG images with reduced com-

putational complexity. It uses depth-wise separable

convolutions to optimize the feature extraction pro-

cess while maintaining high classiﬁcation accuracy.

This makes it particularly suitable for real-time appli-

cations in clinical settings. Finally, the trained mod-

els are evaluated using performance metrics such as

accuracy, precision and recall to validate their effec-

tiveness. The evaluated models are then used to detect

abnormalities in new ECG images, classifying them

into normal or abnormal categories. The results are

visualized to provide clinicians with interpretable in-

sights, aiding in decision-making for cardiac health

monitoring and diagnosis.

4 RESULTS AND DISCUSSION

The experiments are conducted on a machine with

an Intel(R) core(TM) i3-7020U processor operating

at 2.30 Ghz equipped with 4 GB RAM and running

on Windows 10. The implementation is carried out

using Python, with Tensorﬂow and keras serving as

the primary libraries for developing the deep learn-

ing models. Additional libraries such as NumPy and

Pandas are used for efﬁcient data manipulation and

analysis.

Figure 2: Sample Images: (a) Myocardial Infraction, (b)

Abnormal Heartbeats, (c) History of MI, (d) Normal Heart-

beats.

The dataset used in this study is sourced from

mendeley data and comprises a collection of 12-lead

ECG images, categorized into various classes such

as normal, myocardial infarction, abnormal heartbeat,

and history of myocardial infarction. The Figure 2

shows the sample images. The proposed method is fo-

cused on the targeted analysis of ECG images, specif-

ically concentrating on identifying abnormalities in-

dicative of cardiac conditions. The entire dataset

consists of 1,377 ECG images, divided into train-

ing and testing subsets. The training dataset com-

prises 929 samples, used to train the model by extract-

ing and learning features, essential for distinguish-

ing normal from abnormal ECG images. The testing

dataset includes 448 samples, reserved for evaluating

the model’s performance and generalization capabil-

ity. The dataset ensures a balanced representation of

normal and abnormal ECG images to support effec-

INCOFT 2025 - International Conference on Futuristic Technology

758

tive classiﬁcation and robust model performance. In

the study, a comprehensive approach is applied for

the detection and classiﬁcations of ECG peaks, aim-

ing to distinguish between normal and abnormal pat-

terns. The approach focuses on automating the ECG

images, which is critical in diagnosing cardiac condi-

tions. Two deep learning models, a custom CNN and

MobileNet architectures are implemented to perform

the classiﬁcation task.

The ﬁrst step in this study involves designing and

training a custom CNN to classify the ECG images

into four predeﬁned categories. The architecture con-

sisted of an input layer accepting images of size(100,

100, 3) followed by three CNN ﬁlter sizes of 32, 64

and 128. These layers are accompanied by ReLU ac-

tivation functions and max-pooling layers to reduce

the spatial dimensions of the feature maps. The ﬁ-

nal layers include dense layer with 128 neurons us-

ing ReLU activation function. The output layer con-

sists of four neurons employing the softmax function

for multi-class classiﬁcation, where the four neurons

indicate myocardial infarction, abnormal heartbeats,

history of myocardial infarction, and normal heart-

beats. The model is trained over 20 epochs. The per-

formance of the CNN model is remarkable achieving

a test accuracy of 90.98% and a minimal test loss of

0.029. It is observed that, as epochs increase, accu-

racy increases and loss decreases.

The next step employs MobileNet, a lightweight

and efﬁcient CNN architecture, for image classiﬁ-

cation. The approach is particularly effective for

medical imaging tasks due to its computational

efﬁciency and high accuracy. The base MobileNet

model is pre-trained on ImageNet data but initialized

with custom weights in this implementation. The

fully connected top layers are removed to allow

ﬁne-tuning for type speciﬁc task. A custom classiﬁer

is added, which includes a global average pooling

layer followed by two dense layers, one with 128

neurons using ReLU activation and the other with

softmax activation to output probabilities for multiple

classes. The MobileNet model is compiled using the

Adam optimizer with a learning rate of 0.0001. The

categorical cross entropy loss function is employed

as the dataset is multi-class in nature. Accuracy

is chosen as the evaluation metric. To prevent

overﬁtting and ensure early convergence, an early

stopping callback is applied monitoring validation

loss and restoring the best weights if no improvement

is observed after given consecutive epochs. Training

the model occurred over a minimum of 50 epochs,

although the early stopping criterion terminated

training early upon validation loss stabilization. The

training results showed a progressive improvement

in accuracy and reduction in loss over epochs. The

ﬁnal evaluation gives test accuracy of approximately

91.73% and a test loss of 0.445. Figure 3, shows the

accuracy of the MobileNet model and it is observed

that, as epochs increase, accuracy also increases.

Figure 4, shows the loss of the MobileNet model and

it shows that, as epochs increase, loss decreases.

The confusion matrix in Figure 8 shows the

performance of the model on the ECG dataset. It

is observed that the model accurately classiﬁes the

majority of normal and abnormal heartbeats, as well

as MI and patients with MI history. However, there

are a few misclassiﬁcations, with a small number of

abnormal heartbeats being misclassiﬁed as normal,

and vice versa. Overall, the loss is minimal, indi-

cating that the model performs well in detecting and

classifying ECG images. These highlight the capabil-

ity of the MobileNet model in effectively classifying

images. Compared to traditional CNN’s, MobileNet

offered a signiﬁcant advantage in terms of reduced

computation and memory requirements while main-

taining high accuracy. The efﬁciency is achieved by

employing depth wise separable convolutions which

reduces the number of parameters and computational

cost. The inclusion of dropout layers and global

average pooling further reduced overﬁtting while

enabling robust feature extraction. In summary, the

MobileNet-based models successfully demonstrated

high accuracy and efﬁciency, underlining its potential

for real-world applications in detecting abnormalities

in ECG images.

The Table 2 clearly highlights the superiority

of MobileNet over CNN in detecting ECG abnor-

malities. MobileNet achieves a higher accuracy of

91.73% compared to CNN’s 90%, and excels with an

F1-Score of 96% versus CNN’s 95%. MobileNet also

outperforms CNN in precision and recall, achieving

96% for both, compared to CNN’s 95% precision and

94% recall, further emphasizing its effectiveness in

minimizing false positives and negatives. When com-

pared to other existing methods, such as the approach

proposed by (Agarwal et al., 2024), which focuses

on ECG anomaly detection using MobileNet50 CNN

autoencoder, and the method by Sane et al. (Cai and

Hu, 2020), which centers on detecting myocardial

infarction from 12-lead ECG images, MobileNet

demonstrates superior performance. The methods in

(Agarwal et al., 2024) and (Cai and Hu, 2020) report

lower accuracies and F1-Scores, with precision and

recall metrics that fall short of MobileNet’s consistent

96% values. Overall, MobileNet outperforms CNN

and existing methods in accuracy, precision, recall,

and F1-Score, solidifying its position as a highly

Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet

759

effective model for detecting abnormalities in ECG

images.

Figure 3: Accuracy of MobileNet model.

Figure 4: Loss of MobileNet model.

5 CONCLUSIONS

The study utilizes advanced DL models: CNN and

MobileNet, for classiﬁcation of ECG images into four

classes to detect abnormalities. The CNN model, fea-

turing three convolutional layers to extract key pat-

Figure 5: Confusion Matrix of MobileNet.

Table 2: Quantitative Comparison of CNN and MobileNet.

Performance parameters (Sane

et al.,

2021)

(Agarwal

et al.,

2024)

CNN MobileNet

Accuracy 86.21% 76.93% 90% 91.73%

Precision 91.30% 55.23% 95% 96%

Recall 85% 70% 94% 96%

F1-Score 88.05% 61.63% 95% 96%

terns like P waves, QRS complexes, and T waves, is

followed by dense layers for classiﬁcation. Trained

over 20 epochs, it achieved a test accuracy of 90.98%,

highlighting its efﬁcacy in diagnosing cardiovascu-

lar conditions. MobileNet, a lightweight and pre-

trained DL architecture, is also employed for this

study. Known for its computational efﬁciency, Mo-

bileNet demonstrated robust performance by lever-

aging its depthwise separable convolutional layers to

extract features efﬁciently. Fine-tuned for the ECG

classiﬁcation task, MobileNet achieved a test accu-

racy of 91.73%, proving its adaptability and effective-

ness in handling medical image data. MobileNet’s

lightweight design and faster inference time make

it highly suitable for real-time applications, such as

portable ECG monitoring devices and telemedicine

platforms. Among the two approaches, MobileNet

emerges as the more practical solution for real-world

deployment due to its efﬁcient architecture and scal-

ability. However, the MobileNet model demonstrates

superior accuracy, showcasing its capability for de-

tailed and precise classiﬁcation, which is particularly

beneﬁcial in controlled or research settings. Future

advancements aim to improve ECG analysis by devel-

oping more accurate and efﬁcient tools for detecting

cardiac anomalies. These innovations will leverage

AI and real-time monitoring to enable early diagno-

sis and better outcomes. However, their applicability

may be limited to ECG-speciﬁc data, requiring com-

plementary tools for broader diagnostic needs.

INCOFT 2025 - International Conference on Futuristic Technology

760

REFERENCES

Agarwal, M., Gill, K. S., Malhotra, S., and Devliyal, S.

(2024). Decoding Abnormal Heart Patterns: ECG

Anomaly Detection using MobileNet50 CNN Autoen-

coder Method. In 2024 IEEE International Confer-

ence on Information Technology, Electronics and In-

telligent Communication Systems (ICITEICS), pages

1–4.

Cai, W. and Hu, D. (2020). Qrs Complex Detection using

novel Deep Learning Neural Networks. IEEE Access,

pages 97082–97089.

Das, S. (2024). Prediction of Blood Pressure using Machine

Learning. International Journal of Intelligent Systems

and Applications in Engineering, pages 579–586.

Fathi, I. S., Makhlouf, M. A. A., Osman, E., and Ahmed,

M. A. (2022). An Energy-Efﬁcient Compression Al-

gorithm of ECG Signals in remote Healthcare Moni-

toring Systems. IEEE Access, pages 39129–39144.

Hammad, A. H. and Abdulbaqi, A. S. (2022). Intelli-

gent System for Classiﬁcation ECG based on Con-

volutional Neural Networks (CNN). In 2022 8th In-

ternational Conference on Contemporary Information

Technology and Mathematics (ICCITM), pages 101–

106.

Imtiaz, M. N. and Khan, N. (2022). Pan-tompkins++: A

robust approach to detect R-Peaks in ECG Signals. In

2022 IEEE International Conference on Bioinformat-

ics and Biomedicine (BIBM), pages 2905–2912.

Jessy, K. J., Manikandan, G., Hemalatha, S., and Veronica,

V. (2024). Detection of Cardiovascular Disease using

ECG Images in Machine Learning and Deep Learn-

ing. International Journal of Scientiﬁc Research in

Science and Technology, pages 131–139.

Sane, R. K. S., Choudhary, P. S., Sharma, L. N., and Danda-

pat, P. S. (2021). Detection of Myocardial Infarction

from 12 lead ECG Images. In 2021 National Confer-

ence on Communications (NCC), pages 1–6.

World Health Organization (2015). Target 3.4: Re-

duce by one third Premature Mortality from Non-

Communicable Diseases through Prevention and

Treatment and Promote Mental Health and Well-

Being. Accessed: 2024-12-11.

Yang, S., Zhang, Y., Cho, S.-Y., Correia, R., and Morgan,

S. P. (2021). Non-invasive Cuff-Less Blood Pressure

Estimation using a hybrid Deep Learning Model. Op-

tical and Quantum Electronics, page 93.

Yuen, B., Dong, X., and Lu, T. (2019). Inter-patient CNN-

LSTM for QRS Complex Detection in noisy ECG Sig-

nals. IEEE Access, pages 169359–169370.

Zahid, M. U., Kiranyaz, S., Ince, T., Devecioglu, O. C.,

Chowdhury, M. E. H., Khandakar, A., Tahir, A., and

Gabbouj, M. (2022). Robust R-Peak Detection in low-

quality Holter ECGs using 1d Convolutional Neural

Network. IEEE Transactions on Biomedical Engi-

neering, pages 119–128.

Zhou, P., Schwerin, B., Lauder, B., and So, S. (2020). Deep

Learning for real-time ECG R-Peak Prediction. In

2020 14th International Conference on Signal Pro-

cessing and Communication Systems (ICSPCS), pages

1–7.

Automatic Detection of Cardiovascular Abnormalities in ECG Images: CNN and MobileNet

761