Review of ECG Telemonitoring Techniques

Using Neural Network-Based Feature

Extraction
Arshad Abhineet Mishra Abhishek Kumar
(2200290210022) (2200290210007) (2200290210008)
arshad.2226en1010@kiet.edu abhineet.2226en1003@kiet.edu abhishek.2226en1032@kiet.edu
Kiet Group of Institution Kiet Group of Institution Kiet Group of Institution
Department Of Electrical and Department of Electrical and Department of Electrical and
Electronics Engineering Electronics Engineering Electronics Engineering

Supervisor
Dr. Snigdha Chaturvedi

Kiet group of institution

Department of Electrical and
Electronics Engineering

Abstract- Electrocardiography (ECG) is a

vital, non-invasive method for assessing
heart function and diagnosing cardiac issues. wireless sensor networks for ECG
Recent research has significantly advanced telemonitoring has proven effective for
heart disease prediction using ECG data remote monitoring of heart rate. An
through machine learning and deep learning advanced model combining CNNs and RNNs
approaches. Several studies developed has shown high sensitivity and specificity in
accurate ECG-AI models using 12- and detecting atrial fibrillation. Machine
single-lead ECG data, achieving high learning applied to electronic health records
accuracy in predicting fatal coronary heart has improved pre-emotive cardiovascular
disease (FCHD) risk when incorporating disease risk assessment. Overall, these
demographic and clinical data. studies highlight the potential of combining
Benchmarking revealed that convolutional wearable wireless sensors with neural
neural networks (CNNs), particularly ResNet networks for real-time, remote cardiac health
and inception architectures, are highly management
effective for ECG analysis. Various
keywords-ECG(electrocardiogram),
innovative methods, such as decision tree-
MLP(multi-layer-perceptron),
based random forest classifiers, wavelet
KAN(Kolmogorov-Arnold networks), Neural
transform, and independent component
Networks.
analysis, have shown improvements in heart
disease and arrhythmia classification. INTRODUCTION
Transfer learning has also proven beneficial,
as demonstrated by a CNN model for Summary Introduction: Wearable Wireless Sensor
heartbeat classification applied across Networks for ECG Telemonitoring with Neural
Network Feature Extraction Atrial Fibrillation (AF)
different ECG conditions. Additionally,
and cardiovascular disease (CVD) are major health
explainable AI techniques have ensured
concerns with significant economic burdens [1, 7].
model transparency and reliability in
Early detection and intervention are crucial for
predicting heart disease. The use of IoT and improving patient outcomes [7]. Traditional
methods for ECG analysis often require lengthy ECG in diagnosing cardiovascular diseases and
hospital stays or clinic visits, limiting their proposes a taxonomy of monitoring systems along
accessibility [6, 10].Wearable wireless sensor with an architectural model. The integration of
networks (WWSNs) offer a promising solution for advanced technologies such as AI and IoT is
continuous and remote ECG telemonitoring [4, 10]. essential for enhancing the effectiveness of
These networks consist of wearable sensors that healthcare monitoring. The study investigates
capture ECG data and transmit it wirelessly to a various machine learning algorithms used in IoT
central processing unit [4]. Machine learning, cloud-based monitoring systems and emphasizes
particularly deep learning, has emerged as a the significance of visualization methods for
powerful tool for analyzing ECG signals [6, 9]. detecting anomalies. Additionally, it explores
Deep learning models can automatically extract advancements in wireless technologies and
features from the ECG data, eliminating the need discusses the advantages of real-time monitoring
for manual feature engineering and potentially and wearable ECG systems. The paper also
improving accuracy [6, 9].This research addresses key challenges such as motion artifacts
investigates the application of a WWSN for ECG and signal noise, while emphasizing how wearable
telemonitoring, with a focus on utilizing a neural sensors and intelligent healthcare solutions are
network for feature extraction from the ECG signal. driving significant innovations in ECG monitoring.
This approach has the potential to improve At last, [2] This paper also concludes that more
accessibility, convenience, and potentially advancement in ECG monitoring can be done by
diagnostic accuracy for ECG analysis, aiding in adding more sensors that can track the patient’s
early detection and management of heart movements and can predict any abnormality in real
conditions. time. [2]This paper also puts some emphasis on
exploring the field of robotics and healthcare
LITERATURE REVIEW automation for more precise diagnosis and
treatment of patients.
[ [1]This research paper discusses the performance
of models trained on the PTB-XL dataset and tested In this paper [3] Kolmogorov-Arnold Networks
on the ICBEB2018 dataset for ECG analysis. It (KANs) are proposed as promising alternatives to
highlights the effectiveness of transfer learning multi-layer perceptrons (MLPs), which are
from a larger dataset to a smaller one, with the fundamental to deep learning due to their ability to
xresnet1d101 model showing the best performance. approximate nonlinear functions as supported by
The results emphasize the importance of deep the universal approximation theorem [3]. Despite
learning algorithms, specifically convolutional their widespread use, MLPs have significant
neural networks, in ECG analysis. Models trained drawbacks, such as high parameter consumption
on PTB-XL and fine-tuned on ICBEB2018 and low interpretability, particularly in transformer
outperform those trained from scratch on models [3]. KANs, inspired by the Kolmogorov-
ICBEB2018. The xresnet1d101 model Arnold representation theorem, differ from MLPs
demonstrates robustness and adaptability across by utilizing learnable activation functions on edges
different datasets. Deep learning algorithms, (weights) instead of fixed functions on nodes
especially CNNs, perform well in ECG analysis (neurons), replacing linear weight matrices with
tasks. Large, diverse datasets like PTB-XL enable learnable 1D functions parameterized as splines,
comprehensive benchmarking and evaluation of which reduces computational complexity [3]. This
algorithms.[ [1] Although pretraining and training hybrid structure leverages the strengths of both
from scratch on a large dataset do not significantly splines and MLPs, addressing the curse of
deviate from the results, interpretability and dimensionality faced by splines and the lower
uncertainty quantification are crucial for accuracy of MLPs in low dimensions. Empirically,
developing clinically impactful decision support KANs demonstrate superior performance in
systems for ECG interpretation. function fitting and small-scale AI and science
tasks, achieving comparable or better accuracy than
[ [2]This paper provides a comprehensive review of larger MLPs and exhibiting faster neural scaling
ECG monitoring systems, covering their laws [3]. Applications of KANs include solving
architecture, operational processes, challenges, and physical equations, partial differential equations
emerging trends. It highlights the critical role of (PDEs), and integration into machine learning
models such as transformers, resulting in the research highlights the potential scalability of
innovative models like "transformers" [3]. The the framework by suggesting future enhancements,
interpretability of KANs makes them valuable for such as integrating additional sensor nodes to
scientific collaboration, aiding in the rediscovery of monitor additional vital signs and optimizing WSN
mathematical and physical laws. Despite limited parameters for improved efficiency and
mathematical understanding, KANs show promise adaptability. In conclusion, this research paper
for enhancing accuracy, efficiency, and provides a thorough analysis of wearable WSN and
interpretability in various AI and scientific IoT integration for ECG telemonitoring,
applications, positioning them as viable alternatives showcasing advancements in remote healthcare
to MLPs in the evolving landscape of deep learning monitoring systems. By bridging the gap between
[3]. Hence, this research is very recent, and there continuous health data acquisition and timely
has been nearly no research done on ECG feature medical intervention, the study contributes to
extraction using Kolmogorov-Arnold networks. personalized healthcare delivery and improved
Further research can be conducted to explore the patient outcomes. Future directions include refining
application for arrhythmia detection and other the system architecture and expanding its
cardiovascular diseases application to include a wider range of
physiological parameters, further advancing remote
[4]This research paper investigates the integration health monitoring capabilities in clinical
of wearable wireless sensor networks (WSN) with practice..as mentioned above further research can
the Internet of Things (IoT) for ECG be conducted using [3]KAN networks instead of
telemonitoring, as explored by Amina El Attaoui et multi-layer perceptrons for further improvement in
al. [5]. It addresses a critical healthcare need for models and accuracy
efficient remote monitoring solutions, particularly
beneficial in managing chronic diseases and
improving the quality of life for elderly [5]The article explores the early prediction of heart
populations. The focus on real-time ECG signal disease through supervised learning, specifically
monitoring underscores the importance of utilizing stochastic gradient boosting. It introduces
continuous health data acquisition in the early the HDP-DTRF methodology, which leverages a
detection and proactive management of cardiac decision tree-based random forest classifier to
abnormalities. The proposed system architecture enhance prediction accuracy. This approach aims to
consists of three main components: ECG sensing improve the precision of heart disease predictions
nodes for signal acquisition, Gateway nodes using advanced machine-learning techniques. The
equipped with Multilayer Perceptron Neural DTRF classifier, an ensemble learning model, is
Networks (MLPNN) for noise reduction and particularly effective in managing noisy data and
feature extraction, and an IoT platform enabling mitigating overfitting issues. Its adaptability and
seamless data transmission and remote access. This parameter-free nature significantly contribute to its
architecture not only provides immediate access to robust performance. By integrating multiple
ECG data for healthcare professionals but also decision trees, the DTRF model generates more
empowers patients with self-monitoring accurate classification outcomes. The proposed
capabilities, potentially reducing hospital visits and HDP-DTRF approach demonstrates notable
enhancing healthcare efficiency. To validate the improvements in precision, recall, F1-score, and
effectiveness of MLPNN in accurately extracting overall accuracy compared to traditional methods,
R-peaks and calculating heart rates from raw ECG marking a significant advancement in the early
signals, the study leverages datasets from the MIT- detection of heart disease. This HDP-DTRF
BIH database [5]. Experimental results demonstrate approach resulted in 86% of precision, 86% of
robust performance in various environments, recall, 85% of F1-score, and 96% of accuracy on
including indoor and outdoor settings, with ECG publicly available real-world datasets, which are
sensing nodes operating reliably on battery power higher than traditional methods.
over extended periods. Technical evaluations
include rigorous assessments of system parameters
such as autocorrelation, end-to-end delay, and [6]The study presents an innovative deep-learning
coverage, confirming the system's reliability and model designed for the real-time detection of Atrial
feasibility for real-world deployment. Furthermore, Fibrillation (AF) in ECG recordings. This model
integrates Convolutional Neural Networks (CNNs) medicine, enabling individualized risk assessments
and Recurrent Neural Networks (RNNs) to and proactive interventions. This approach not only
effectively extract features from RR intervals. improves the accuracy of CVD predictions but also
Notably, the model demonstrates superior aligns with modern healthcare's movement towards
performance compared to existing methodologies, more personalized and preventive care strategies.
achieving high sensitivity and specificity across
three different databases. The preprocessing phase
involves converting raw ECG data into RR [8]This study aimed to create precise ECG-AI
intervals, followed by segmentation. The model models for predicting the risk of fatal coronary
architecture includes convolutional layers for heart disease (FCHD) and assessing time-
feature extraction and Long Short-Term Memory dependent risk factors. By analyzing ECG data
(LSTM) layers for capturing temporal from over 50,000 patients, the researchers found
dependencies, complemented by a pooling layer for that the most effective model combined
downsampling. This combination allows the model demographics, clinical data, and ECG-AI,
to process 24-hour ECG recordings in less than a achieving high accuracy in predicting FCHD risk.
second, showcasing its computational efficiency. Both the 12-lead and 1-lead ECG-AI models
Overall, the proposed model not only surpasses showed strong predictive performance. The top
state-of-the-art methods in AF detection but also model, known as C12-ECG-AI-Cox, performed
ensures rapid and reliable analysis, making it a exceptionally well by integrating demographic
significant advancement in the field of information, clinical data, and ECG-AI, indicating
cardiovascular health monitoring. [6] This algorithm its potential for clinical use to identify individuals
was tested on the unseen datasets to examine its at high risk. Detailed subgroup analyses
robustness in detecting AF (Atrial Fibrillation) for demonstrated that the model's ability to predict
new recordings which resulted in 98.96% and FCHD risk was consistent across various genders,
86.04% for specificity and sensitivity, respectively. races/ethnicities, and patients with existing
coronary artery disease. This suggests that the
model is robust and widely applicable.
[7] This research article explores the early Additionally, certain age groups showed even
prediction of cardiovascular disease (CVD) through higher accuracy in predicting cardiac events,
the application of machine learning algorithms to highlighting the effectiveness of ECG-based
health records. It underscores the significance of models for specific age demographics in risk
incorporating genetic predispositions, lifestyle assessment. The research also emphasizes the
choices, and clinical markers into these models to critical role of ECG-based predictions in clinical
achieve precise risk assessments. The study practice. These models can identify high-risk
emphasizes how such integration can advance individuals, allowing for early interventions and
preventive healthcare strategies and facilitate customized preventive strategies. The findings
personalized interventions by leveraging electronic indicate that ECG-AI models could significantly
health records (EHRs). Integrating diverse data contribute to personalized medicine, aiding
sources, including genetic information, lifestyle clinicians in making more informed decisions and
factors, and clinical markers, into machine learning improving patient outcomes. In summary,
models enhances the preemptive risk assessment combining demographics, clinical data, and ECG-
for CVD. Advanced machine learning technologies AI into predictive models represents a significant
play a pivotal role in refining these risk advancement in assessing cardiovascular risk. The
assessments and improving the predictive accuracy high accuracy and broad applicability of these
of CVD outcomes. Critical steps such as data models underscore their potential to enhance
preprocessing and feature selection are essential in preventive healthcare strategies and improve the
preparing EHRs for effective machine-learning early detection of fatal coronary heart disease.
applications. Various machine learning models,
including logistic regression, decision trees,
random forests, support vector machines, and
[9]This paper introduces a novel approach for
neural networks, have demonstrated efficacy in
arrhythmia classification by combining
predicting CVD. Utilizing EHRs for prediction
morphological and dynamic features extracted from
supports the paradigm shift towards precision
ECG signals. Morphological features are derived monitoring. The studies reviewed highlight the
using Wavelet Transform and Independent efficacy of deep learning models, particularly
Component Analysis, while dynamic features are convolutional neural networks (CNNs) and their
obtained from RR interval information. For variants, in various aspects of ECG interpretation,
classification, a Support Vector Machine (SVM) is including arrhythmia detection, atrial fibrillation
employed, categorizing the data into 15 distinct (AF) diagnosis, and risk prediction for
classes and achieving an impressive overall cardiovascular diseases (CVD) such as coronary
accuracy of 99.66% on a dataset of 85,945 heart disease (FCHD) [[1][6][8][10] ]..
heartbeats. The method is validated using the MIT-
BIH Arrhythmias Database, demonstrating superior The integration of large-scale datasets like PTB-XL
performance compared to previous methodologies. and MIT-BIH has enabled the development of
To further enhance accuracy, the study employs robust models through techniques such as transfer
two fusion methods to integrate results from two learning, enhancing the accuracy and generalization
ECG leads. This combination of morphological and of ECG-based AI systems. For instance, the
dynamic features, along with the use of fusion xresnet1d101 model demonstrated superior
techniques, significantly improves the robustness adaptability when fine-tuned on smaller datasets
and precision of arrhythmia classification. The like ICBEB2018, illustrating the effectiveness of
approach's high accuracy underscores its potential pretraining on comprehensive datasets for specific
for clinical application in reliable arrhythmia clinical applications [[1] ]..
detection.
Moreover, the emergence of innovative AI
[10]This paper introduces a method utilizing deep architectures like Kolmogorov-Arnold Networks
convolutional neural networks (CNNs) for (KANs) presents promising alternatives to
classifying heartbeats, specifically identifying five traditional MLPs, offering improved interpretability
different types of arrhythmias. Additionally, it and computational efficiency. While KANs are
explores a technique for transferring the knowledge relatively new in ECG feature extraction, their
gained from this task to the classification of potential for enhancing accuracy and scalability
myocardial infarction, achieving high accuracy in suggests a fertile area for future research and
both classifications. The study focuses on deep development [[3]]..
learning approaches for precise heartbeat
classification and effective knowledge transfer Further development in ECG telemonitoring
between different cardiac tasks. The method systems, leveraging wearable wireless sensor
demonstrates impressive accuracy in categorizing networks (WSNs) and IoT integration, holds
both arrhythmias and myocardial infarction. promise for advancing personalized healthcare
Preprocessing steps include cropping individual solutions. These systems not only facilitate real-
beats and applying zero padding to ensure uniform time monitoring but also empower patients with
length. The deep learning framework allows the continuous health data access, reducing
model to learn task-specific features, requiring a dependency on traditional clinical visits and
large dataset for effective training. The research enhancing proactive disease management [[4][5]]..
utilizes labeled ECG records from the PhysioNet
MIT-BIH Arrhythmia Database and the PTB Looking ahead, future research directions should
Diagnostic ECG Database, providing a robust focus on addressing current challenges such as
foundation for training and validation. The results noise reduction, signal artifacts, and the integration
highlight the effectiveness of deep learning in of multi-modal health data (genetic, lifestyle, and
improving the accuracy and reliability of cardiac clinical) into AI models. Additionally, exploring
condition classifications. explainable AI methodologies will be crucial for
gaining insights into model decisions, thereby
CONCLUSION fostering trust and adoption in clinical settings [[7]

In conclusion, this literature review has

[10]]..
underscored the significant advancements and
In summary, the convergence of AI with ECG
diverse applications of artificial intelligence (AI) in
analysis is poised to revolutionize cardiovascular
ECG analysis and cardiovascular health
 healthcare by enabling early detection, personalized
risk assessment, and targeted interventions.
Continued interdisciplinary collaboration and
technological innovation will drive further
advancements, ultimately improving patient
outcomes and healthcare efficiency on a global
scale.

[1] N. Strodthoff, P. Wagner, T. Schaeffter, and W. Samek, “Deep Learning for ECG Analysis: Benchmarks
and Insights from PTB-XL,” IEEE J Biomed Health Inform, vol. 25, no. 5, pp. 1519–1528, May 2021,
doi: 10.1109/JBHI.2020.3022989.

[2] M. A. Serhani, H. T. El Kassabi, H. Ismail, and A. N. Navaz, “ECG monitoring systems: Review,
architecture, processes, and key challenges,” Sensors (Switzerland), vol. 20, no. 6. MDPI AG, Mar. 02,
2020. doi: 10.3390/s20061796.

[3] Z. Liu et al., “KAN: Kolmogorov-Arnold Networks,” Apr. 2024, [Online]. Available:
http://arxiv.org/abs/2404.19756

[4] A. El Attaoui, M. Hazmi, A. Jilbab, and A. Bourouhou, “Wearable Wireless Sensors Network for ECG
Telemonitoring Using Neural Network for Features Extraction,” Wirel Pers Commun, vol. 111, no. 3,
pp. 1955–1976, Apr. 2020, doi: 10.1007/s11277-019-06967-x.

[5] A. P. Jawalkar et al., “Early prediction of heart disease with data analysis using supervised learning
with stochastic gradient boosting,” Journal of Engineering and Applied Science, vol. 70, no. 1, Dec.
2023, doi: 10.1186/s44147-023-00280-y.

[6] R. S. Andersen, A. Peimankar, and S. Puthusserypady, “A deep learning approach for real-time
detection of atrial fibrillation,” Expert Syst Appl, vol. 115, pp. 465–473, Jan. 2019, doi:
10.1016/j.eswa.2018.08.011.

[7] D. R. Deepa, V. B. Sadu, C. G. Prashant, and D. A. Sivasamy, “Early prediction of cardiovascular disease
using machine learning: Unveiling risk factors from health records,” AIP Adv, vol. 14, no. 3, Mar. 2024,
doi: 10.1063/5.0191990.

[8] O. Akbilgic, “Time-Dependent ECG-AI Prediction of Fatal Coronary Heart Disease”, doi:
10.1101/2023.10.11.23296910.

[9] IEEE Staff and IEEE Staff, 2010 32nd Annual International Conference of the IEEE Engineering in
Medicine and Biology Society.

[10] M. U. Sreeja and M. H. Supriya, “A Deep Convolutional Model for Heart Disease Prediction based on
ECG Data with Explainable AI,” WSEAS Transactions on Information Science and Applications, vol. 20,
pp. 254–264, 2023, doi: 10.37394/23209.2023.20.29.