Page 51 - AIH-1-4

Artificial Intelligence in Health A fuzzy system for heartbeat classification

condition detection, a variable threshold is applied to the time-frequency plane to display the locations of different
adaptive neuro-fuzzy inference system (ANFIS) output, peaks, the TERMA algorithm designates specific areas of
and the parameters are adjusted accurately. The proposed interest to locate the desired peak. In the study, a DL-based
1
system has three parts. First, the input ECGs undergoes system is presented; using convolutional neural networks
preprocessing to eliminate noise. In the second step, feature (CNNs) for ECG classification with the PhysioNet MIT-
extraction prepares the inputs for the ANFIS, while heart BIH Arrhythmia database. The suggested system uses a
1
1
conditions are labeled using the subtractive clustering 1-D convolutional deep residual neural network (ResNet)
method to train the ANFIS. model, which uses the input heartbeats directly to extract
ANFIS is a binary classifier, so to classify six classes; six features. To handle the class imbalance in the training
separate ANFIS systems are required. This mechanism, as dataset and effectively classify five heartbeat types in the
the third stage, can be known as a multi-binary classifier. test dataset, the synthetic minority oversampling technique
2
Across this process, classification is achieved; however, to was employed. In addition, raw ECG recordings are
3
enhance accuracy, six thresholds are applied to the output classified using deep CNNs. However, these CNNs require
of each ANFIS. By tuning the threshold values, a variable- extensive annotated samples for effective training, which
threshold system is created that yields optimal results. The can be costly to obtain. To mitigate this issue, transfer
3
categorization of ECG arrhythmias is crucial for quick learning is utilized. Using the largest available collection
identification and diagnosis of cardiovascular disorders. of continuous raw ECG signals, the first CNNs were pre-
A more accurate diagnosis allows for timely and suitable trained. Next, the networks were refined for the most
interventions when needed. These diagnostic techniques common cardiac arrhythmia and atrial fibrillation using a
can be integrated into electronic devices used by individuals small data set. An artificial NN approach was presented for
4
with cardiac problems, and in the event of an emergency, the automatic identification and categorization of ECG.
appropriate alerts can be delivered to medical facilities, To thoroughly mine the hierarchical and time-sensitive
such as hospitals or physicians. In addition, hospitals features of ECG data, a dense heart rhythm network has
can regularly employ these methods to diagnose cardiac been developed that combines a 24-layer deep CNN and
problems early, reducing the need for human intervention. bidirectional long short-term memory. The original ECG
With the use of ECG classification approaches, individuals is filtered using a combination of wavelet transform and
with established cardiac diseases can be monitored median filtering to remove the influence of noise on the
continuously. As the disease progresses or as a patient signal. In addition, three different sizes of convolution
responds to treatment, changes in the ECG pattern over kernels (32, 64, and 128) are used to mine the detailed
time can help doctors make well-informed judgments features from the ECG signal.
about modifying treatment plans. With significant The symlet wavelet transform was presented to detect
accuracy and precision, the suggested method can make the QRS complex and reduce the error. Using values of RR
5
a significant contribution to this field by classifying six intervals, amplitude, and Hjorth parameters, some features
classes of heart diseases. Since biomedical signals are very of the ECG were extracted for heartbeat classification.
6
sensitive, even a small improvement in speed, accuracy, Techniques, including variation mode decomposition,
and precision can greatly impact individuals’ health and phase space reconstruction, euclidean distance, and
well-being. Related works are explored in Section 2, while Shannon energy envelope were employed to detect
Section 3 discusses the data and methods employed in this myocardial dysfunction. The Hilbert-Huang transform was
7
research. The results are discussed in Section 4, and the used for feature selection, which includes a set of essential
conclusion is provided in Section 5. features. The ANFIS employed Lyapunov exponents for
8
the ECG classification. A reliable beat classification was
9
2. Related works performed using the wavelet transform and principal
10
Researchers aim to achieve accurate and high-speed component analysis-independent component analysis.
classification while keeping computational costs low. Robust An extreme learning machine was applied to the MIT/
automated diagnostic approaches require preprocessing BIH database, and feature selection was performed using
of the ECG, enhancing the signal, extracting features, the variances of the wavelet transform and parameters
11
and classifying the data. Different techniques have been of the autoregressive model. Furthermore, the ECG
explored in the literature for detecting diseases through classification was done by an automatic, reliable, two-stage
ECG analysis. To enhance ECG analysis, an algorithm is hybrid hierarchical approach. The ANFIS, along with
12
put forth that makes use of the fractional Fourier transform a fuzzy rule-based model classifier, was used to identify
(FrFT) and two-event-related moving averages (TERMA) premature ventricular contraction (PVC) beats with a high
algorithms. While the FrFT rotates ECG signals in the readability-accuracy trade-off. 13

Volume 1 Issue 4 (2024) 45 doi: 10.36922/aih.3367