Atrial Fibrillation Detection Using ECG Recordings Based on Genetic Optimization ^†

Abstract

Recently, mobile healthcare is emerging using portable and wearable devices. This article uses short-single-lead ECG signals to develop an automatic AF detection system. Heart rate variability (HRV) and frequency analysis are used for feature extraction. The innovative contribution is to develop a Genetic Optimization Algorithm for detecting atrial fibrillation in short ECG recordings. The validation of the results is carried out with the publicly available dataset which comprises short ECG recordings by the support vector machine algorithm. The accuracy varies from 94.2 to 96.5 for N versus A classification under noise levels ranging up to 30 dB. A maximum accuracy of 82.7% is obtained for N versus A versus O.

1. Introduction

Atrial fibrillation (AF) is a highly prevalent arrhythmia of the heart. It is quite difficult to find unless a specific arrhythmia episode occurs during the investigation. There is an urgent need for the automatic detection of AF, as it can result in cardiac strokes and death if diagnosis and treatment are delayed. One common method of identifying AF is thought to be the study of ECG data. ECG data have been analyzed taking into account noise, atrial fibrillation (A), other arrhythmias (O), and normal rhythm (N). A result that is understandable for physicians can be produced by employing various algorithms to extract the valuable elements from the raw data. In linked analysis, several approaches are used, motivated either by information analytics or medical data. Physicians employ physiologically interpretable AF signals by taking aberrant P wave and QRS shape measurements into account. This research uses publicly available, open datasets to validate the suggested technique. The suggested approach starts with pre-processing the ECG signal, extracts features, uses a genetic algorithm (GA) to optimize the features, and then uses a support vector machine (SVM) classifier to categorize the data. The suggested technique outperforms the current algorithm with an SNR of dB, achieving improved accuracy.

2. Materials and Methods

2.1. Preprocessing

The preprocessing step includes noise reduction, ECG signal segmentation, and the transformation of non-uniform pulses into uniform pulses. This work uses data from the public domain, including 8528 ECG records. Among the data that were collected, 60% showed normal arrhythmia (N), 28% showed Other arrythmia (O), 9% showed atrial fibrillation and the remaining 3% showed noise. The ECG recordings have been cleared up of noise using filters. The ECG pulses are split into pulses with the use of the R peaks. The resulting beats yield non-uniform samples. In this work, the MIT-BIH open-source dataset [1] is used for experimentation and analysis.

2.2. Spectral and HRV Feature Extraction

The ECG’s identified QRS complex generate RR signals. The Heart Rate Variability (HRV) signal, which offers details on differences among successive heartbeats, has been used to obtain the Congestive Heart Failure (CHF)-related parameter from the ECG. HRV can be examined in terms of frequency or time. Since the time domain parameters of HRV are derived using well-used statistical techniques, they are the simplest. The initial stage in obtaining HRV-based features is R peak detection. Because of its better performance, a well-known real-time QRS detection algorithm, originally developed by Pan and Tompkins (P&T), has been used [2]. The oscillations that occur between successive heartbeat intervals are known as HRV. It is thought that the vagal and autonomic nervous systems both influence HRV. Only rudimentary information regarding cardiac activity is available in the HRV-based features. The technique for QRS detection is used to identify these R peaks. The formulas for calculating RR interval sequences, the Catalecticant matrix descriptor, Complex Correlation Measure and signal-to-noise ratio were discussed in detail in [3].

2.3. Genetic Algorithm (GA)-Optimized Feature Selection

In order to reduce the feature set and enable an extremely simplistic and efficient model that works effectively in classification, feature selection is a crucial step in the process. In this work, the best features for classifying the ECG signals into normal, atrial fibrillation and other rhythms are chosen using the Genetic Optimization Algorithm (GA) [4]. A selection technique based on tournaments has been employed to apply to the GA. The number of characteristics has been employed to determine the chromosome’s size. Finding a chromosome which conveys an array of traits with the maximum classification accuracy is the goal of using GA [5]. First, a random population is generated to represent various locations inside the searching region. The fitness function has been utilized to evaluate each member of the initial population. Reproduction, crossover, and mutation are crucial procedures in GA that create a subsequent iteration from the current one. GA maintains high fitness chromosomes for subsequent generation and destroys low fitness chromosomes through repetitions till an optimal genome has been obtained.

According to the findings of the research, a meta-heuristic-based strategy outperforms a statistical method in the classification of ECG AF. A feature selection procedure is used to avoid overfitting. Genetic Optimization Algorithm (GA) and maximum-relevance–min-redundancy (MRMR) are integrated. One potential method for reducing feature counts is the use of GA. Additional meta-heuristic optimization methods, such evolutionary strategies and simulated annealing, are also suitable for this use case. But GA [6] is a robust and stochastic method based on natural evolutionary concepts that has been shown to eventually lead to a semi-optimal solution for a wide range of challenging situations. In order to lay out the potential response as a statistical vector, the initial step of the GA method is to determine its encoding. After that, the code attempts to dynamically create a population. Also, it is utilized for feature distribution and reduction in numerous applications.

2.4. Support Vector Machine (SVM) Based Classification

SVM [7] is used to lower the measured risk and uncertainty range for the classification algorithm or prediction function. In order to properly classify high-dimensional data as well, the challenge is resolved through developing an exponential or polynomial decision rule [8]. To categorize the segments as either normal or atrial fibrillation, the retrieved features are used for classification. The support vector machine algorithm is the method utilized to do this. In total, 60% of the data is utilized for training and 40% is used for testing when the data are trained and tested. Performance metrics including accuracy, specificity, and sensitivity are assessed using these data.

3. Results and Discussion

The present section compares and validates the suggested GA-SVM-based AF detection system’s performance using a range of parameters. The goal of the proposed research is to develop an automated detection system that can recognize and categorize AF based on ECG data. The HRV spectral feature extraction, GA feature selection, and SVM classification techniques are used in this study to achieve this goal. This study conducts a thorough simulation evaluation to look at and contrast the overall AF detection performance and outcomes of the suggested model. Additionally, the MIT-BIH public dataset was employed in this study to assess and contrast the outcomes of the GA-SVM-based AF detection system.

The convergence curve for the GA optimization method with respect to different iteration counts is displayed in Figure 1. Usually, the fitness value of the optimization algorithm is used to validate and assess its performance. It is clear from the data that the GA requires fewer iterations to obtain the optimal fitness value.

The most popular metrics, including accuracy, sensitivity, and specificity, are computed in this study to assess the AF detection and classification performance of the GA-SVM classifier. These metrics are computed mathematically as indicated by the following equations: [9]

$$Sensitivity=\frac{TP}{TP+FN}$$

(1)

$$Specificity=\frac{TN}{TN+FP}$$

(2)

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

(3)

Based on the characteristics of accuracy, sensitivity, and specificity,

Table 1. Comparison of performance metrics at different SNR values with respect to all the features and top 10 features.

SNR	FEATURES	CLASSIFIER	N/A			O/A			N/O			N/O/A
			Acc	Sen	Spec	Acc	Sen	Spec	Acc	Sen	Spec	Acc	Sen	Spec
∞ dB	All	SVM	95.3	74.3	98.5	89.1	67.5	95.9	82.7	62.5	92.3	79.7	63.2	98.2
		GA-SVM	96.4	79.7	99.1	92.0	75.3	98.1	85.2	68.3	94.9	82.8	71.6	99.3
	Top 10	SVM	96.2	82.6	98.2	89.7	74.6	94.4	80.3	61.6	89.2	78.2	73.0	97.3
		GA-SVM	96.5	85.0	98.3	90.7	77.6	95.1	81.6	63.9	91.3	80.1	75.9	98.2
30 dB	All	SVM	95.8	76.3	98.8	90.0	72.6	95.5	82.7	64.9	91.2	80.2	68.6	98.0
		GA-SVM	96.4	78.4	99.1	91.4	72.1	97.8	85.0	67.4	94.3	82.5	68.4	98.8
	Top 10	SVM	96.4	83.6	98.3	89.6	75.0	94.2	80.0	59.7	89.7	78.0	73.5	97.2
		GA-SVM	96.3	84.2	98.2	90.7	76.5	95.1	81.2	61.9	90.3	80.5	75.2	98.1
20 dB	All	SVM	95.3	74.3	98.4	89.5	69.7	95.7	81.9	64.5	90.2	79.3	65.7	98.1
		GA-SVM	96.2	76.4	98.9	91.3	70.5	97.4	84.2	65.4	93.3	82.7	68.2	98.7
	Top 10	SVM	95.9	82.4	97.9	88.6	72.5	93.6	78.7	58.9	88.1	76.4	71.0	96.9
		GA-SVM	96.3	83.7	98.2	89.6	74.4	94.8	80.1	60.9	89.9	80.6	72.4	98.0
10 dB	All	SVM	94.8	66.1	99.1	87.6	60.1	96.3	81.2	56.2	93.1	77.9	54.5	98.5
		GA-SVM	95.0	69.0	98.9	90.1	64.1	97.3	83.6	59.8	92.1	82.0	58.7	98.6
	Top 10	SVM	93.8	64.9	98.2	85.9	58.2	94.6	78.5	52.8	90.7	75.1	51.2	97.7
		GA-SVM	94.2	67.0	98.1	87.0	61.2	94.4	79.9	55.9	89.3	79.5	54.4	98.0

compare the SVM and GA-SVM classification algorithms with respect to different SNR levels, such as ∞ dB, 10 dB, 20 dB, and 30 dB. where the cases such as Normal/Arrhythmia, Other/Arrhythmia, Normal/Other, and Normal/Other/Arrhythmia are included, respectively. When compared to the traditional SVM classification technique, the GA-SVM classification technique yields better prediction results, as indicated by the expected findings.

4. Conclusions

The novel contribution of this work is to create a highly effective and qualified methodology used for accurately detecting AF from the provided short-single-lead ECG signals. The HRV-based feature extraction methodology with a high level of robustness contributes to improved atrial fibrillation detection accurately with confidence. A genetic algorithm along with an SVM classifier are used for AF detection. It has superior accuracy and specificity, making this algorithm a compete prototype for preliminary screening. The algorithm is validated using a publicly available dataset consisting of short ECG recordings acquired under unstructured environments from portable devices.