Page 68 - AIH-1-4

Artificial Intelligence in Health ML models for heartbeat classification

and smart healthcare systems. An EKG is a graphical expertise and recording quality of clinicians, yielding
representation of the heart’s electrical activity over time, high diagnostic accuracy but remaining prone to human
obtained by affixing electrodes on the skin surface. The errors. ML models, trained on large annotated datasets
5
human body generates various biomedical signals, and and enhanced with preprocessing techniques such as
electrocardiography sensors record the heart’s electrical Fourier transform (FT) and Gaussian noise injection, can
activity by generating three-lead EKGs for heart monitoring achieve similar or even superior accuracy at 90 – 99% in
and surface electromyography for muscle contractions. detecting cardiac abnormalities. 13
1
ECGs are crucial for diagnosing cardiovascular conditions This study integrates the use of the FT and Gaussian
and aid in the early detection of heart attacks and treatment noise injection techniques to enhance the capability of ML
to prevent them. However, identifying and classifying models trained and tested on the datasets provided by the
2
arrhythmias are challenging due to the requirement of University of Chinese Academy of Sciences, which will
extensive data analysis, potential human errors, and be presented in subsequent sections. The FT technique is
signal variability. Furthermore, developing an automated essential for preprocessing heartbeat signals in ML models,
detection system is difficult due to data complexity and converting time-domain signals into frequency-domain
noise; however, ECGs are indispensable for diagnosing ones to filter noise and improve signal quality, thereby
heart diseases. Anomalous deviations from a typical enhancing model performance to a level equal to or
ECG pattern are observed in various cardiac disorders, greater than that obtained with the 12-lead ECG analysis.
including irregular heart rhythms such as atrial fibrillation Unlike state-of-the-art algorithms such as convolutional
and ventricular tachycardia; diminished blood flow neural networks (CNNs), which filter noise implicitly, the
3
through the coronary arteries, which is observed in cases FT technique explicitly reduces computational demands
such as myocardial ischemia and myocardial infarction; and improves training efficiency. Traditional ML models
and disruptions in the electrolyte balance such as in benefit greatly from noise reduction, leading to improved
hypokalemia and hyperkalemia. 4 classification accuracy. In addition, Gaussian noise
Since the 20 century, ECG analysis has been vital for injection adds controlled noise during training, enhancing
th
diagnosing cardiovascular diseases by monitoring the model robustness and generalization capability, which
heart’s electrical activity. The conventional 12-lead ECG, is particularly useful for heartbeat classification, thereby
recorded by affixing electrodes to the chest and limbs, improving performance on unseen data and reducing the
is crucial for detecting arrhythmias. Early diagnosis is risk of overfitting. In practical applications, these methods
essential for effective treatment, sometimes requiring enhance accuracy and reliability, making them suitable for
over 24 h of continuous monitoring. Advances in the applications in clinical settings and resource-constrained
5
digital industry have improved devices, data acquisition environments such as wearable devices and mobile health
techniques, and computer-assisted diagnostics. Although applications.
ECGs are widely used by cardiologists to monitor heart This study aims to classify heartbeats using data
health, a manual analysis of the associated signals is from heartbeat sequences recorded by patient heartbeat
time-consuming and error-prone. Therefore, an accurate signal sensors. The key contributions of this study are
diagnosis of cardiovascular conditions is crucial as these summarized as follows:
conditions contribute to approximately one-third of global • Introduce the innovative application of the FT technique
mortality. The prevalence of irregular heart rhythms to preprocess ECG signals by converting time-domain
highlights the need for precise and cost-effective methods data into frequency-domain representations, allowing
to diagnose arrhythmic heartbeats. 6 for the effective noise reduction and enhancement of

The classification of ECG signals is challenging due signal quality.
to individual variability and the lack of standardization • Employ Gaussian noise injection as a novel approach
in feature extraction, often leading to low diagnostic to simulate real-world noise conditions by training
accuracy. To address the limitations of manual datasets to be more robust against various types
7-9
ECG signal analysis, numerous studies have applied of noise and artifacts commonly observed in ECG
machine learning (ML) techniques for accurate anomaly signals.
detection. 10-12 Traditional methods for this analysis involve • Conduct a comparative analysis of various ML models
signal preprocessing, handcrafted feature extraction, and by assessing their performance on the preprocessed
the use of ML and deep learning algorithms. However, ECG data. Emphasize the influence of the FT and
deep learning requires extensive datasets due to the large Gaussian noise integration techniques on key metrics
number of the associated parameters. The traditional such as accuracy, precision, recall, and F1 score for
12-lead ECG analysis in clinical settings depends on the each algorithm.

Volume 1 Issue 4 (2024) 62 doi: 10.36922/aih.3543