Atrial activity enhancement by Wiener filtering using an artificial neural network

A Review on the State of the Art in Atrial Fibrillation Detection Enabled by Machine Learning

Atrial Fibrillation (AF) the most commonly occurring type of cardiac arrhythmia is one of the main causes of morbidity and mortality worldwide. The timely diagnosis of AF is an equally important and challenging task because of its asymptomatic and episodic nature. In this paper, state-of-the-art ECG data-based machine learning models and signal processing techniques applied for auto diagnosis of AF are reviewed. Moreover, key biomarkers of AF on ECG and the common methods and equipment used for the collection of ECG data are discussed. Besides that, the modern wearable and implantable ECG sensing technologies used for gathering AF data are presented briefly. In the end, key challenges associated with the development of auto diagnosis solutions of AF are also highlighted. This is the first review paper of its kind that comprehensively presents a discussion on all these aspects related to AF auto-diagnosis in one place. It is observed that there is a dire need for low energy and low cost but accurate auto diagnosis solutions for the proactive management of AF.

Deep learning for cardiovascular medicine: a practical primer

Deep learning (DL) is a branch of machine learning (ML) showing increasing promise in medicine, to assist in data classification, novel disease phenotyping and complex decision making. Deep learning is a form of ML typically implemented via multi-layered neural networks. Deep learning has accelerated by recent advances in computer hardware and algorithms and is increasingly applied in e-commerce, finance, and voice and image recognition to learn and classify complex datasets. The current medical literature shows both strengths and limitations of DL. Strengths of DL include its ability to automate medical image interpretation, enhance clinical decision-making, identify novel phenotypes, and select better treatment pathways in complex diseases. Deep learning may be well-suited to cardiovascular medicine in which haemodynamic and electrophysiological indices are increasingly captured on a continuous basis by wearable devices as well as image segmentation in cardiac imaging. However, DL also has significant weaknesses including difficulties in interpreting its models (the 'black-box' criticism), its need for extensive adjudicated ('labelled') data in training, lack of standardization in design, lack of data-efficiency in training, limited applicability to clinical trials, and other factors. Thus, the optimal clinical application of DL requires careful formulation of solvable problems, selection of most appropriate DL algorithms and data, and balanced interpretation of results. This review synthesizes the current state of DL for cardiovascular clinicians and investigators, and provides technical context to appreciate the promise, pitfalls, near-term challenges, and opportunities for this exciting new area.

Ventricular activity morphological characterization: ectopic beats removal in long term atrial fibrillation recordings

Ectopic beats are early heart beats remarkably different to the normal beat morphology that provoke serious disturbances in electrocardiographic analysis. These beats are very common in atrial fibrillation (AF), causing important residua when ventricular activity has to be removed for atrial activity (AA) analysis. In this work, a method is proposed to cancel out ectopics by discriminating between normal and abnormal beats, with an accuracy higher than 99%, through QRS morphological delineation and characterization. The most similar ectopics to the one under cancellation are clustered to provide a very precise cancellation template. Simulated and real AF recordings were used to validate the method. A new index, able to estimate the presence of ventricular residue after ectopics cancellation, was defined. Results by using the 2, 4, 6, …, 30 most similar ectopics to the one under study yielded optimal cancellation for templates composed of 10 beats. Furthermore, these beats were very likely located close to the ectopic under cancellation, which could facilitate the algorithm implementation. As conclusion, the proposed method is an effective way to remove ectopics from long term AF recordings and get them ready for the application of any QRST cancellation technique able to extract the AA in optimal conditions. Moreover, it could also detect, characterize and remove ectopics in any other type of non-AF recordings.

Surface electrocardiogram reconstruction from intracardiac electrograms using a dynamic time delay artificial neural network

This study proposes a method to facilitate the remote follow up of patients suffering from cardiac pathologies and treated with an implantable device, by synthesizing a 12-lead surface ECG from the intracardiac electrograms (EGM) recorded by the device. Two methods (direct and indirect), based on dynamic time-delay artificial neural networks (TDNNs) are proposed and compared with classical linear approaches. The direct method aims to estimate 12 different transfer functions between the EGM and each surface ECG signal. The indirect method is based on a preliminary orthogonalization phase of the available EGM and ECG signals, and the application of the TDNN between these orthogonalized signals, using only three transfer functions. These methods are evaluated on a dataset issued from 15 patients. Correlation coefficients calculated between the synthesized and the real ECG show that the proposed TDNN methods represent an efficient way to synthesize 12-lead ECG, from two or four EGM and perform better than the linear ones. We also evaluate the results as a function of the EGM configuration. Results are also supported by the comparison of extracted features and a qualitative analysis performed by a cardiologist.

An echo state neural network for QRST cancellation during atrial fibrillation

A novel method for QRST cancellation during atrial fibrillation (AF) is introduced for use in recordings with two or more leads. The method is based on an echo state neural network which estimates the time-varying, nonlinear transfer function between two leads, one lead with atrial activity and another lead without, for the purpose of canceling ventricular activity. The network has different sets of weights that define the input, hidden, and output layers, of which only the output set is adapted for every new sample to be processed. The performance is evaluated on ECG signals, with simulated f-waves added, by determining the root mean square error between the true f-wave signal and the estimated signal, as well as by evaluating the dominant AF frequency. When compared to average beat subtraction (ABS), being the most widely used method for QRST cancellation, the performance is found to be significantly better with an error reduction factor of 0.24-0.43, depending on f-wave amplitude. The estimates of dominant AF frequency are considerably more accurate for all f-wave amplitudes than the AF estimates based on ABS. The novel method is particularly well suited for implementation in mobile health systems where monitoring of AF during extended time periods is of interest.

Feature extraction in time-frequency signal analysis by means of matched wavelets as a feature generator

The goal of presented work was to compare the usage of standard basic wave let function like e.g. bio-orthogonal or dbn with the optimized wavelet created to the best match analyzing ECG signals in the context of P-wave and atrial fibrillation detection. A library of clinical expert evaluated typical atrial fibrillation evolutions was created as a database for optimal matched wavelet construction. Whole data set consisting of 40 cases with long term ECG recording s were divided into learning and verifying set for the multilayer perceptron neural network used as a classifier structure. Compared with other wavelet filters, the matched wavelet was able to improve classifier performance for a given ECG signals in terms of the Sensitivity and Specificity measures.

Time-frequency characterization of atrial fibrillation from surface ECG based on Hilbert-Huang transform

In this paper, based on Hilbert-Huang transform (HHT), we develop a new non-invasive time-frequency analysis method to characterize the dynamic behaviour of atrial fibrillation (AF) from surface ECG. We first extract f waves from single-lead ECG records of AF patients using PCA analysis. To capture the non-stationary behaviours of AF signals at different time scales, we use HHT to find the Hilbert spectrum and instantaneous frequency (IF) distribution of residual signals from principal component analysis. Two important feature variables, namely mean IF (mIF) and index of frequency stability over time (IS), are derived from the IF distribution, and in combination will be able to effectively discriminate two different AF types: self-terminating and non-terminating termination. The proposed AF signal decomposition and analysis method will help us efficiently differentiate individual AF patients, advance our understanding of AF mechanisms, and provide useful guidelines for improving administration of AF patients, especially paroxysmal AF.

Rules extraction in SVM and neural network classifiers of atrial fibrillation patients with matched wavelets as a feature generator

Presented paper describes a system of biomedical signal classifiers with preliminary feature extraction stage based on matched wavelets analysis, where two structures of classifier using Neural Networks (NN) and Support Vector Machine (SVM) are applied. As a pilot study the rules extraction algorithm applied for two of mentioned machine learning approaches (NN & SVM) was used. This was made to extract and transform the representation of knowledge gathered in Black Box parameters during classifier learning phase to be better and natural understandable for human user/expert. Proposed system was tested on the set of ECG signals of 20 atrial fibrillation (AF) and 20 control group (CG) patients, divided into learning and verifying subsets, taken from MIT-BiH database. Obtained results showed, that the ability of generalization of created system, expressed as a measure of sensitivity and specificity increased, due to extracting and selectively choosing only the most representative features for analyzed AF detection problem. Classification results achieved by means of constructed matched wavelet, created for given AF detection features were better than indicators obtained for standard wavelet basic functions used in ECG time-frequency decomposition.

Adaptive singular value cancelation of ventricular activity in single-lead atrial fibrillation electrocardiograms

The proper analysis and characterization of atrial fibrillation (AF) from surface electrocardiographic (ECG) recordings requires to cancel out the ventricular activity (VA), which is composed of the QRS complex and the T wave. Historically, for single-lead ECGs, the averaged beat subtraction (ABS) has been the most widely used technique. However, this method is very sensitive to QRST wave variations and, moreover, high-quality cancelation templates may be difficult to obtain when only short length and single-lead recordings are available. In order to overcome these limitations, a new QRST cancelation method based on adaptive singular value cancelation (ASVC) applied to each single beat is proposed. In addition, an exhaustive study about the optimal set of complexes for better cancelation of every beat is also presented for the first time. The whole study has been carried out with both simulated and real AF signals. For simulated AF, the cancelation performance was evaluated making use of a cross-correlation index and the normalized mean square error (nmse) between the estimated and the original atrial activity (AA). For real AF signals, two additional new parameters were proposed. First, the ventricular residue (VR) index estimated the presence of ventricular activity in the extracted AA. Second, the similarity (S) evaluated how the algorithm preserved the AA segments out of the QRST interval. Results indicated that for simulated AF signals, mean correlation, nmse, VR and S values were 0.945 +/- 0.024, 0.332 +/- 0.073, 1.552 +/- 0.386 and 0.986 +/- 0.012, respectively, for the ASVC method and 0.866 +/- 0.042, 0.424 +/- 0.120, 2.161 +/- 0.564 and 0.922 +/- 0.051 for ABS. In the case of real signals, the mean VR and S values were 1.725 +/- 0.826 and 0.983 +/- 0.038, respectively, for ASVC and 3.159 +/- 1.097 and 0.951 +/- 0.049 for ABS. Thus, ASVC provides a more accurate beat-to-beat ventricular QRST representation than traditional techniques. As a consequence, VA cancelation is optimized and the AA can be extracted more precisely. Finally, the study has proven that optimal VA cancelation is achieved when a number between 20 and 30 complexes is selected following a correlation-based strategy.

P wave detector with PP rhythm tracking: evaluation in different arrhythmia contexts

Automatic detection of atrial activity (P waves) in an electrocardiogram (ECG) is a crucial task to diagnose the presence of arrhythmias. The P wave is difficult to detect and most of the approaches in the literature have been evaluated on normal sinus rhythms and rarely considered arrhythmia contexts other than atrial flutter and fibrillation. A novel knowledge-based P wave detector algorithm is presented. It is self-adaptive to the patient and able to deal with certain arrhythmias by tracking the PP rhythm. The detector has been tested on 12 records of the MIT-BIH arrhythmia database containing several ventricular and supra-ventricular arrhythmias. On the overall records, the detector demonstrates Se = 96.60% and Pr = 95.46%; for the normal sinus rhythm, it reaches Se = 97.76% and Pr = 96.80% and, in the case of Mobitz type II, it demonstrates Se = 72.79% and Pr = 99.51%. It also shows good performance for trigeminy and bigeminy, and outperforms some more sophisticated techniques. Although the results emphasize the difficulty of P wave detection in difficult arrhythmias (supra and ventricular tachycardias), it shows that domain knowledge can efficiently support signal processing techniques.

Noninvasive ECG as a Tool for Predicting Termination of Paroxysmal Atrial Fibrillation

Atrial fibrillation (AF) is the most common cardiac arrhythmia and entails an increased risk of thromboembolic events. Prediction of the termination of an AF episode, based on noninvasive techniques, can benefit patients, doctors and health systems. The method described in this paper is based on two-lead surface electrocardiograms (ECGs): 1-min ECG recordings of AF episodes including N-type (not terminating within an hour after the end of the record), S-type (terminating 1 min after the end of the record) and T-type (terminating immediately after the end of the record). These records are organised into three learning sets (N, S and T) and two test sets (A and B). Starting from these ECGs, the atrial and ventricular activities were separated using beat classification and class averaged beat subtraction, followed by the evaluation of seven parameters representing atrial or ventricular activity. Stepwise discriminant analysis selected the set including dominant atrial frequency (DAF, index of atrial activity) and average HR (HRmean, index of ventricular activity) as optimal for discrimination between N/T-type episodes. The linear classifier, estimated on the 20 cases of the N and T learning sets, provided a performance of 90% on the 30 cases of a test set for the N/T-type discrimination. The same classifier led to correct classification in 89% of the 46 cases for N/S-type discrimination. The method has shown good results and seems to be suitable for clinical application, although a larger dataset would be very useful for improvement and validation of the algorithms and the development of an earlier predictor of paroxysmal AF spontaneous termination time.

Analysis of esophageal atrial recordings through wavelet packets decomposition

In this paper the processing of esophageal atrial electrograms by means of wavelet packets (WP) decomposition is presented. WP is described as a flexible, signal-adaptive, tool, which can be easily tuned to enhance characteristics of esophageal signals. Two aspects are mainly investigated: (i) the possibility to obtain automatic, reliable detection of atrial activation in 24h Holter recordings and (ii) the development of an algorithm for discrimination between atrial flutter (AFLU) and atrial fibrillation (AF) episodes. WP decomposition was used as a framework for pre-processing the esophageal signal and to build a set of orthonormal sub-signals which can be selected and combined according to the signal processing task to be performed: (i) in the detection of atrial activation, sub-band signal characteristics were explored at different scales by using the modulus maxima criteria and (ii) in the discrimination between AFLU and AF the coarser approximation of the esophageal signal was studied by spectral analysis. A reliable detection of atrial activation was obtained (Sensitivity (SE): 99.08%; positive predictability (+P): 98.98%). In addition a quantitative index able to discriminate between AFLU (SE: 97.5%; +P: 98.7%) and AF (SE: 98.7%; +P: 97.5%) episodes was introduced.

Spatiotemporal Blind Source Separation Approach to Atrial Activity Estimation in Atrial Tachyarrhythmias

The analysis and characterization of atrial tachyarrhythmias requires, in a previous step, the extraction of the atrial activity (AA) free from ventricular activity and other artefacts. This contribution adopts the blind source separation (BSS) approach to AA estimation from multilead electrocardiograms (ECGs). Previously proposed BSS methods for AA extraction--e.g., independent component analysis (ICA)--exploit only the spatial diversity introduced by the multiple spatially-separated electrodes. However, AA typically shows certain degree of temporal correlation, with a narrowband spectrum featuring a main frequency peak around 3.5-9 Hz. Taking advantage of this observation, we put forward a novel two-step BSS-based technique which exploits both spatial and temporal information contained in the recorded ECG signals. The spatiotemporal BSS algorithm is validated on simulated and real ECGs from a significant number of atrial fibrillation (AF) and atrial flutter (AFL) episodes, and proves consistently superior to a spatial-only ICA method. In simulated ECGs, a new methodology for the synthetic generation of realistic AF episodes is proposed, which includes a judicious comparison between the known AA content and the estimated AA sources. Using this methodology, the ICA technique obtains correlation indexes of 0.751, whereas the proposed approach obtains a correlation of 0.830 and an error in the estimated signal reduced by a factor of 40%. In real ECG recordings, we propose to measure performance by the spectral concentration (SC) around the main frequency peak. The spatiotemporal algorithm outperforms the ICA method, obtaining a SC of 58.8% and 44.7%, respectively.

Atrial activity extraction for atrial fibrillation analysis using blind source separation

This contribution addresses the extraction of atrial activity (AA) from real electrocardiogram (ECG) recordings of atrial fibrillation (AF). We show the appropriateness of independent component analysis (ICA) to tackle this biomedical challenge when regarded as a blind source separation (BSS) problem. ICA is a statistical tool able to reconstruct the unobservable independent sources of bioelectric activity which generate, through instantaneous linear mixing, a measurable set of signals. The three key hypothesis that make ICA applicable in the present scenario are discussed and validated: 1) AA and ventricular activity (VA) are generated by sources of independent bioelectric activity; 2) AA and VA present non-Gaussian distributions; and 3) the generation of the surface ECG potentials from the cardioelectric sources can be regarded as a narrow-band linear propagation process. To empirically endorse these claims, an ICA algorithm is applied to recordings from seven patients with persistent AF. We demonstrate that the AA source can be identified using a kurtosis-based reordering of the separated signals followed by spectral analysis of the sub-Gaussian sources. In contrast to traditional methods, the proposed BSS-based approach is able to obtain a unified AA signal by exploiting the atrial information present in every ECG lead, which results in an increased robustness with respect to electrode selection and placement.

Sequential Characterization of Atrial Tachyarrhythmias Based on ECG Time-Frequency Analysis

A new method for characterization of atrial arrhythmias is presented which is based on the time-frequency distribution of an atrial electrocardiographic signal. A set of parameters are derived which describe fundamental frequency, amplitude, shape, and signal-to-noise ratio. The method uses frequency-shifting of an adaptively updated spectral profile, representing the shape of the atrial waveforms, in order to match each new spectrum of the distribution. The method tracks how well the spectral profile fits each spectrum as well as if a valid atrial signal is present. The results are based on the analysis of a learning database with signals from 40 subjects, of which 24 have atrial arrhythmias, and an evaluation database with 211 patients diagnosed with atrial fibrillation. It is shown that the method robustly estimates fibrillation frequency and amplitude and produces spectral profiles with narrower peaks and more discernible harmonics when compared to the conventional power spectrum. The results suggest that a rather strong correlation exist between atrial fibrillation frequency and f wave shape. The developed set of parameters may be used as a basis for automated classification of different atrial rhythms.