A low-complexity algorithm for detection of atrial fibrillation using an ECG

Abstract

The computational cost of our algorithm is low as all features are derived from RR intervals and are processed by a single hidden layer neural network. This makes it potentially suitable for low-power devices.

Automated diagnosis of atrial fibrillation in 24-hour Holter recording based on deep learning:a study with randomized and real-world data validation

“…This study used the RR intervals as the input to the deep learning model for two reasons: 1) The irregularity of RR intervals is a main feature of the AF 21 ；2) Another feature of the AF is the appearance of fine or coarse fibrillatory waves and the absence of P-wave 22 , however, this feature can be easily confused by common perturbations in Holter recording 23 . Our experiments showed that our RR-interval-based method and the raw-ECG-based method achieved similar performance (appendix p 4-6), which indicates that the RR-interval-based features may be the major ones for detecting AF in the deep learning model.…”

ECG Monitoring Systems: Review, Architecture, Processes, and Key Challenges

“…Finding the optimal set of features that capture the true nature of the ECG signals remains a challenging task. Researches tried to apply multiple techniques to address the aforementioned issue, such as adopting Discrete Wavelet Transform and the Pan Tompkins Method to improve heartbeat abnormality classification from ECG signals in [93], or using time domain, frequency domain and distribution features for detection of atrial fibrillation (AF) in [46]. Keeping the good reliability of data and the quality of the signal are also challenges facing smartphone-integrated ECG monitoring systems; such a case was handled in [94] by enhancing the feature extraction process.…”

“…This massive diversity in ECG monitoring systems' contexts, technologies, computational schemes, and purposes makes it hard for researchers and professionals to design, classify, and analyze ECG monitoring systems. Some efforts attempted to provide a common understanding of ECG monitoring systems' processes [42][43][44][45][46][47], guiding the design of efficient monitoring systems. However, these studies lack comprehensiveness and completeness.…”

Application of Dense Neural Networks for Detection of Atrial Fibrillation and Ranking of Augmented ECG Feature Set

“…All included papers can be considered as a truthful basis of related works to this study; therefore, a detailed list of embedded classifiers and input features is further reviewed: Support vector machine (SVM) classifiers are input with 47 features from the statistical and morphological rhythm representation [ 24 ]; 33 features expressive to the signal power, spectrum, entropy, RR intervals and P-waves [ 25 ]; and 61 features from the time-frequency ECG representation, both average and variability of RR intervals, and the average beat morphology [ 26 ]; Linear and quadratic discriminant classifiers are input with a set of 122 RR-interval features from their time domain, frequency domain and distribution (histogram) representations [ 27 ] and 44 features measured by heart rate variability (HRV) analysis, average beat morphology analysis, and atrial activity analysis focused on the P-wave amplitude in the average beat and f-waves amplitude in TQ intervals [ 28 ]; Decision tree classifiers are input with 30 multi-level features, including measures of AF, morphology, RR intervals and similarity between beats [ 29 ]; morphological coefficients and HRV features calculated by ECG waveform fitting with a piecewise linear spline [ 30 ]; 400 hand-crafted features, reflecting the complex physiology of cardiac arrhythmias visible in single-channel ECG [ 31 ]; and 74 features of the R-peak amplitude, RR-interval statistics, PQRST statistics, ECG signal irregularity, entropy, noise content and four additional sparse coding features [ 32 ]; A multi-layer binary classification architecture is input with subsets of 77, 66 and 19 features selected from 188 dimensional feature pool containing time, frequency, morphological and statistical domain ECG features [ 33 ]; A multi-stage classifier, combining SVM, decision tree and threshold-based rules is quantifying both atrial and ventricular activity, estimated by local features (beat classification) and global features (rhythm, signal quality, similarity) [ 34 ]. Advanced multi-stage classifiers, combining decision trees and neural networks (NNs) include: a tree gradient boosting model and a recurrent long, short-term memory (LSTM) network as a global classifier that uses 42 summary ECG features of the full record and a sequence classifier that works on a beat-by-beat basis using individual features of each cardiac cycle [ 35 ]; a bagged tree ensemble with 43 input features based on QRS detection and PQRS morphology connected in parallel to a convolutional neural network (CNN) and a shallow NN for analysis of raw filtered ECG signals (8× envelograms, 1× band-pass) [ 36 ]; a nine-layer CNN for segmentation of P, QRS and T waves, inter-beat segments, noise and arrhythmic beats, additionally augmented by hand-crafted features, thus supplying a set of 181 features to eXtreme Gradient Boosting trees to classify the heart rhythm [ 37 ]; a densely connected ...…”

“…Linear and quadratic discriminant classifiers are input with a set of 122 RR-interval features from their time domain, frequency domain and distribution (histogram) representations [ 27 ] and 44 features measured by heart rate variability (HRV) analysis, average beat morphology analysis, and atrial activity analysis focused on the P-wave amplitude in the average beat and f-waves amplitude in TQ intervals [ 28 ];…”