Objective: To determine whether a two-lead ECG (wECG), acquired using a self-applicable, wrist-worn device, provides diagnostic value in acute myocardial infarction (AMI) detection. Methods: Seventy-one patients with AMI, 26 with other cardiovascular diseases (CVDs), and 54 healthy controls were enrolled in the study. To acquire lead I, one electrode of the wrist-worn device was touched with the index finger of the opposite hand, while another lead was acquired by touching the electrode on the strap to a specific part of the body. The diagnostic value of the wECG is evaluated either directly or through synthesis of the standard 12-lead ECG using an echo state network. Results: Using the standard ECG, an AMI detector based on a convolutional neural network yields a sensitivity of 0.86 and a specificity of 0.71; the corresponding figures for blinded cardiologist diagnosis are 0.77/0.88. Using the synthesized ECG, the V3 electrode touch site yields the best performance, with a sensitivity/specificity of 0.84/0.71 for AMI detection and 0.67/0.86 for cardiologist diagnosis. Compared to the synthesized ECG across all touch sites, use of the wECG results in a substantially lower sensitivity, reduced by 0.12–0.13, while the specificity is comparable or slightly higher. Implementation of a two-stage approach, consisting of detector-based screening followed by cardiologist review of positive cases, results in correct decisions in 63% of the AMI cases using either the synthesized ECG or the wECG. Conclusion: The two-lead ECG acquired with a wrist-worn device offers diagnostic value and enables screening for AMI. Clinical and Translational Impact Statement: This work enables AMI detection using a two-lead ECG acquired with a wrist-worn device, either directly or through synthesis of the 12-lead ECG, supporting cardiovascular diagnostics in remote and resource-limited environments. (Category: Clinical Research).

While circadian phenotypes of paroxysmal atrial fibrillation (AF) have been observed in patient cohorts, differences in parameter definition and methodological limitations have left the existence and clinical relevance of distinct AF circadian phenotypes unclear. We hypothesize that paroxysmal AF comprises multiple circadian phenotypes (‘chronophenotypes’), each associated with different survival outcomes and population characteristics. We analyzed 24 h Holter recordings and clinical data from 58 995 examinations collected in 20 primary care facilities in Israel. AF episodes were detected using ArNet2 , a deep learning model for AF detection trained on over 51 000 h of Holter recordings. Unsupervised hierarchical clustering was then applied to identify different chronophenotypes of AF based on the short-term AF burden. Each chronophenotype was further characterized by demographic, clinical, and treatment differences, including survival outcomes. The analysis resulted in the following distinct chronophenotypes: Nocturnal-to-Morning (12 AM–10 AM), Evening-to-Early Morning (3 PM-4 AM), and Daytime (9 AM-7 PM). These chronophenotypes were associated with different AF burden and clinical outcomes. The Evening-to-Early Morning chronophenotype was associated with a higher AF burden ( $p \lt $ 0.001) than the other paroxysmal AF chronophenotypes. However, the Nocturnal-to-Morning and Daytime chronophenotypes were associated with a higher risk of mortality; Daytime was also associated with a higher risk of heart failure. The study introduces a new approach to discovering chronophenotypes from Holter recordings. The results support the existence of distinct AF chronophenotypes, which are associated with significantly different average AF burden ( $p \lt $ 0.001) and clinical outcomes. The open-source implementation is publicly available at https://github.com/aim-lab/AFtoolkit .

Background: A wrist-worn wearable device for acquiring limb and chest ECG leads (wECG) may constitute a promising approach to detection of acute myocardial infarction (AMI). However, it remains to be demonstrated whether the information conveyed by the wECG is sufficient for AMI detection. Objective: To explore explainable machine learning models for detecting AMI using the wECG. Methods: Two types of machine learning models are explored: a convolutional neural network (CNN) using the raw ECG as input and a gradient-boosting decision tree (GBDT) using clinically informative features. 123 participants were included, divided into patients with AMI, patients with other cardiovascular diseases, and healthy individuals. A wristworn device equipped with three biopotential electrodes was used to acquire two ECG leads with a single touch: limb lead I and another lead involving a specific body site, i.e., either the V3 or V5 electrode positions, or the abdomen. Results: The best performance on the test dataset is obtained using models that incorporate all four leads. The CNN model performs slightly better than the GBDT model, with a sensitivity of 0.77 and specificity of 0.75 compared to 0.77 and 0.72, respectively. When distinguishing between AMI and healthy participants, the specificity increases to 0.94 for the CNN model and 0.90 for the GBDT model. Feature importance analysis shows that the GBDT model primarily relies on the J point, while the CNN model primarily relies on the QRS complex. Conclusions: wECG-based AMI detection shows considerable promise in out-of-hospital settings. However, caution is needed as CNN explanations rarely agree with the ECG intervals typically analyzed in clinical practice

Background: With the availability of a wrist-worn device capable of acquiring two ECG leads with a single touch, synthesis of the 12-lead ECG may be accomplished to facilitate clinical interpretation. Objective: This study proposes an echo state network (ESN) for synthesizing the 12-lead ECG from two leads simultaneously acquired using a wrist-worn device. Methods: The wrist-worn device, equipped with three electrodes, was used to acquire two ECG leads from 51 healthy participants, 29 patients with acute myocardial infarction, and 12 patients with other cardiovascular diseases. The person-specific synthesis is based on the ESN, a recurrent neural network, trained on a single resting, standard 12-lead ECG through a highly efficient training process. To explore the importance of different electrode touch sites, the participants were instructed to touch sites on the body corresponding to the electrode positions for acquiring the precordial leads V3 and V5, as well as the abdomen. Results: Using the ESN, the lowest RMS error between the standard and the synthesized ECGs is obtained for leads I and V1, irrespective of participant group and touch site from which the two-lead ECG was acquired. The ESN outperformed a linear regression-based transformation matrix, especially for the precordial leads where the RMS error was up to three times higher than that of the ESN. Conclusion: ESN-based synthesis of the 12-lead ECG based on a two-lead ECG holds promise as a valuable tool for screening abnormalities in the ECG.

Introduction: The presence of fibrillatory waves (f-waves) is important in the diagnosis of atrial fibrillation (AF), which has motivated the development of methods for f-wave extraction. We propose a novel approach to benchmarking methods designed for single-lead ECG analysis, building on the hypothesis that better-performing AF classification using features computed from the extracted f-waves implies better-performing extraction. The approach is well-suited for processing large Holter data sets annotated with respect to the presence of AF. Methods: Three data sets with a total of 300 two- or three-lead Holter recordings, performed in the USA, Israel and Japan, were used as well as a simulated single-lead data set. Four existing extraction methods based on either average beat subtraction or principal component analysis (PCA) were evaluated. A random forest classifier was used for window-based AF classification. Performance was measured by the area under the receiver operating characteristic (AUROC). Results: The best performance was found for PCA-based extraction, resulting in AUROCs in the ranges 0.77--0.83, 0.62--0.78, and 0.87--0.89 for the data sets from USA, Israel, and Japan, respectively, when analyzed across leads; the AUROC of the simulated single-lead, noisy data set was 0.98. Conclusions: This study provides a novel approach to evaluating the performance of f-wave extraction methods, offering the advantage of not using ground truth f-waves for evaluation, thus being able to leverage real data sets for evaluation. The code is open source (following publication).

Background and Objective: Growing evidence shows that certain acute exposures, especially alcohol, may trigger episodes of paroxysmal atrial fibrillation (AF). However, there is a lack of methods for assessing the relation between triggers and AF episodes in individual patients. The present paper proposes an approach to identifying AF triggers based on the assumption that the post-trigger AF burden is larger than the pre-trigger AF burden during the analysis time interval.Method: For the purpose of identification, a measure of relational strength between pre- and post-trigger burden is introduced, accounting for the cumulative effect of the triggers contained in the observation interval. The proposed approach is explored for different types of AF episode pattern, generated using the alternating, bivariate Hawkes model, whose conditional intensity function is designed to account for the effect of alcohol. In total, 7200 different AF patterns were generated for different numbers of AF triggers and alcohol units.Results: The simulation study demonstrates that, depending on the pattern type, the relational strength increases 3–6 times with alcohol consumption in comparison with no consumption.Conclusions: The proposed approach to identifying triggers in individual patients with paroxysmal AF should facilitate the implementation of longitudinal studies for the objective assessment of trigger effect on AF occurrence.

OBJECTIVE:The episode patterns of paroxysmal atrial fibrillation (AF) may carry important information on disease progression and complication risk. However, existing studies offer very little insight into to what extent a quantitative characterization of AF patterns can be trusted given the errors in AF detection and various types of shutdown, i.e., poor signal quality and non-wear. This study explores the performance of AF pattern characterizing parameters in the presence of such errors. METHODS:To evaluate the performance of the parameters AF aggregation and AF density, both previously proposed to characterize AF patterns, the two measures mean normalized difference and the intraclass correlation coefficient are used to describe agreement and reliability, respectively. The parameters are studied on two PhysioNet databases with annotated AF episodes, also accounting for shutdowns due to poor signal quality. RESULTS:The agreement is similar for both parameters when computed for detector-based and annotated patterns, which is 0.80 for AF aggregation and 0.85 for AF density. On the other hand, the reliability differs substantially, with 0.96 for AF aggregation but only 0.29 for AF density. This finding suggests that AF aggregation is considerably less sensitive to detection errors. The results from comparing three strategies to handle shutdowns vary considerably, with the strategy that disregards the shutdown from the annotated pattern showing the best agreement and reliability. CONCLUSIONS:Due to its better robustness to detection errors, AF aggregation should be preferred. To further improve performance, future research should put more emphasis on AF pattern characterization.

Objective: Non-sustained supraventricular tachycardia (nsSVT) is associated with a higher risk of developing atrial fibrillation (AF), and, therefore, detection of nsSVT can improve AF screening efficiency. However, the detection is challenged by the lower signal quality of ECGs recorded using handheld devices and the presence of ectopic beats which may mimic the rhythm characteristics of nsSVT. Methods: The present study introduces a new nsSVT detector for use in single-lead, 30-s ECGs, based on the assumption that beats in an nsSVT episode exhibits similar morphology, implying that episodes with beats of deviating morphology, either due to ectopic beats or noise/artifacts, are excluded. A support vector machine is used to classify successive 5-beat sequences in a sliding window with respect to similar morphology. Due to the lack of adequate training data, the classifier is trained using simulated ECGs with varying signal-to-noise ratio. In a subsequent step, a set of rhythm criteria is applied to similar beat sequences to ensure that episode duration and heart rate is acceptable. Results: The performance of the proposed detector is evaluated using the StrokeStop II database, resulting in sensitivity, specificity, and positive predictive value of 84.6%, 99.4%, and 18.5%, respectively. Conclusion: The results show that a significant reduction in expert review burden (factor of 6) can be achieved using the proposed detector. Clinical and Translational Impact: The reduction in the expert review burden shows that nsSVT detection in AF screening can be made considerably more efficiently.

Objective.Despite the growing interest in understanding the role of triggers of paroxysmal atrial fibrillation (AF), solutions beyond questionnaires to identify a broader range of triggers remain lacking. This study aims to investigate the relation between triggers detected in wearable-based physiological signals and the occurrence of AF episodes.Approach.Week-long physiological signals were collected during everyday activities from 35 patients with paroxysmal AF, employing an ECG patch attached to the chest and a photoplethysmogram (PPG)-based wrist-worn device. The signals acquired by the patch were used for detecting potential triggers due to physical exertion, psychophysiological stress, lying on the left side, and sleep disturbances. To assess the relation between detected triggers and the occurrence of AF episodes, a measure of relational strength is employed accounting for pre- and post-trigger AF burden. The usefulness of ECG- and PPG-based AF detectors in determining AF burden and assessing the relational strength is also analyzed.Main results.Physical exertion emerged as the trigger associated with the largest increase in relational strength for the largest number of patients (p < 0.01). On the other hand, no significant difference was observed for psychophysiological stress and sleep disorders. The relational strength of the detected AF exhibits a moderate correlation with the relational strength of annotated AF, withr = 0.66 for ECG-based AF detection andr = 0.62 for PPG-based AF detection.Conclusions.The findings indicate a patient-specific increase in relational strength for all four types of trigger.Significance.The proposed approach has the potential to facilitate the implementation of longitudinal studies and can serve as a less biased alternative to questionnaire-based AF trigger detection.

The tools for spectrally analyzing heart rate variability (HRV) has in recent years grown considerably, with emphasis on the handling of time-varying conditions and confounding factors. Time-frequency analysis holds since long an important position in HRV analysis, however, this technique cannot alone handle a mean heart rate or a respiratory frequency which vary over time. Overlapping frequency bands represents another critical condition which needs to be dealt with to produce accurate spectral measurements. The present survey offers a comprehensive account of techniques designed to handle such conditions and factors by providing a brief description of the main principles of the different methods. Several methods derive from a mathematical/statistical model, suggesting that the model can be used to simulate data used for performance evaluation. The inclusion of a respiratory signal, whether measured or derived, is another feature of many recent methods, e.g., used to guide the decomposition of the HRV signal so that signals related as well as unrelated to respiration can be analyzed. It is concluded that the development of new approaches to handling time-varying scenarios are warranted, as is benchmarking of performance evaluated in technical as well as in physiological/clinical terms.

Background: Slower adaptation of the QT interval to sudden changes in heart rate has been identified as a risk marker of ventricular arrhythmia. The gradual changes observed in exercise stress testing facilitates the estimation of the QT-RR adaptation time lag. Methods: The time lag estimation is based on the delay between the observed QT intervals and the QT intervals derived from the observed RR intervals using a memoryless transformation. Assuming that the two types of QT interval are corrupted with either Gaussian or Laplacian noise, the respective maximum likelihood time lag estimators are derived. Estimation performance is evaluated using an ECG simulator which models change in RR and QT intervals with a known time lag, muscle noise level, respiratory rate, and more. The accuracy of T-wave end delineation and the influence of the learning window positioning for model parameter estimation are also investigated. Results: Using simulated datasets, the results show that the proposed approach to estimation can be applied to any changes in heart rate trend as long as the frequency content of the trend is below a certain frequency. Moreover, using a proper position of the learning window for exercise so that data compensation reduces the effect of nonstationarity, a lower mean estimation error results for a wide range of time lags. Using a clinical dataset, the Laplacian-based estimator shows a better discrimination between patients grouped according to the risk of suffering from coronary artery disease. Conclusions : Using simulated ECGs, the performance evaluation of the proposed method shows that the estimated time lag agrees well with the true time lag.

Introduction: Deep learning models for detecting episodes of atrial fibrillation (AF) using rhythm information in long-term, ambulatory ECG recordings have shown high performance. However, the rhythm-based approach does not take advantage of the morphological information conveyed by the different ECG waveforms, particularly the f-waves. As a result, the performance of such models may be inherently limited. Methods: To address this limitation, we have developed a deep learning model, named RawECGNet, to detect episodes of AF and atrial flutter (AFl) using the raw, single-lead ECG. We compare the generalization performance of RawECGNet on two external data sets that account for distribution shifts in geography, ethnicity, and lead position. RawECGNet is further benchmarked against a state-of-the-art deep learning model, named ArNet2, which utilizes rhythm information as input. Results: Using RawECGNet, the results for the different leads in the external test sets in terms of the F1 score were 0.91--0.94 in RBDB and 0.93 in SHDB, compared to 0.89--0.91 in RBDB and 0.91 in SHDB for ArNet2. The results highlight RawECGNet as a high-performance, generalizable algorithm for detection of AF and AFl episodes, exploiting information on both rhythm and morphology.

A new model for the estimation of the QT-RR adaptation time lag using exercise ECG stress testing has been proposed, assuming a linear heart rate trend during the test. In this work, simulated ECGs have been generated based on heart rate patterns with oscillations at different frequencies to demonstrate that the assumption can be relaxed so that the QTRR adaptation time lag can be adequately estimated for any heart rate trend whose frequency content is below a certain frequency, which depends on the QT-RR time lag.

Slowed adaptation of the QT interval to sudden abrupt heart rate (HR) changes has been identified as a marker of ventricular arrhythmic risk. However, abrupt HR changes are difficult to induce in patients. Quantifying the QT adaptation time in gradual HR changes, as observed in ECGs recording during an exercise stress test, has been recently proposed. The time lag between the QT series and an instantaneous memoryless HR-dependent QT series along stress test was computed as QT memory. Here, this method was evaluated in a control scenario using simulated exercise stress test ECG signals presenting different QT adaptation times. The method robustness was studied by contaminating the ECGs with muscular noise (MN) signals with different Signal-to-Noise ratio (SNR) values, either synthetic or extracted from real recordings. We found that delineation of the T-wave end point in the first transformed lead from Periodic Component Analysis offers the best performance for low SNR. Moreover, we confirmed that the estimator provides an unbiased estimate of the QT memory introduced in the simulations for the studied range of SNR values (25 to 50 dB).

Modifiable factors, such as alcohol or physical exertion, may trigger atrial fibrillation (AF) episodes. Identifying and eliminating these triggers can lead to effective strategies which reduce risk of AF recurrence. This study aims to evaluate pre- and post-trigger AF in long-term photo-plethysmogram (PPG) signals obtained during daily living from patients with paroxysmal AF. Thirty-seven patients were instructed to wear a wrist-worn device for a week. They were also asked to log suspected triggers using a smartphone app. Of these patients, 15 experienced AF episodes, resulting in an average AF burden of 0.15. The results indicate that longer post-trigger analysis time intervals resulted in better performance of PPG-based AF detection. The sensitivity was highest for the 16-h post-trigger interval (0.76) and lowest for the 4-h interval (0.43). In contrast, the specificity slightly decreased with an increasing longer post-trigger analysis time interval, being 0.98 and 0.95 for the 4-h and 16-h intervals, respectively. The PPG-based post-trigger AF burden was approximately half of that determined by the annotated ECG-based AF pattern. The study suggests that long-term PPG-based monitoring is a suitable alternative for detecting post-trigger AF instead of ECG-based. However, the accuracy of AF burden estimation using PPG-based technology still calls for improvement.

The present article proposes an ECG simulator that advances modeling of arrhythmias and noise by introducing time-varying signal characteristics. The simulator is built around a discrete-time Markov chain model for simulating atrial and ventricular arrhythmias of particular relevance when analyzing atrial fibrillation (AF). Each state is associated with statistical information on episode duration and heartbeat characteristics. Statistical, time-varying modeling of muscle noise, motion artifacts, and the influence of respiration is introduced to increase the complexity of simulated ECGs, making the simulator well suited for data augmentation in machine learning. Modeling of how the PQ and QT intervals depend on heart rate is also introduced. The realism of simulated ECGs is assessed by three experienced doctors, showing that simulated ECGs are difficult to distinguish from real ECGs. Simulator usefulness is illustrated in terms of AF detection performance when either simulated or real ECGs are used to train a neural network for signal quality control. The results show that both types of training lead to similar performance.

AbstractBackgroundAccess to long‐term ambulatory recording to detect atrial fibrillation (AF) is limited for economical and practical reasons. We aimed to determine whether 24 h ECG (24hECG) data can predict AF detection on extended cardiac monitoring.MethodsWe included all US patients from 2020, aged 17–100 years, who were monitored for 2–30 days using the PocketECG device (MEDICALgorithmics), without AF ≥30 s on the first day (n = 18,220, mean age 64.4 years, 42.4% male). The population was randomly split into equal training and testing datasets. A Lasso model was used to predict AF episodes ≥30 s occurring on days 2–30.ResultsThe final model included maximum heart rate, number of premature atrial complexes (PACs), fastest rate during PAC couplets and triplets, fastest rate during premature ventricular couplets and number of ventricular tachycardia runs ≥4 beats, and had good discrimination (ROC statistic 0.7497, 95% CI 0.7336–0.7659) in the testing dataset. Inclusion of age and sex did not improve discrimination. A model based only on age and sex had substantially poorer discrimination, ROC statistic 0.6542 (95% CI 0.6364–0.6720). The prevalence of observed AF in the testing dataset increased by quintile of predicted risk: 0.4% in Q1, 2.7% in Q2, 6.2% in Q3, 11.4% in Q4, and 15.9% in Q5. In Q1, the negative predictive value for AF was 99.6%.ConclusionBy using 24hECG data, long‐term monitoring for AF can safely be avoided in 20% of an unselected patient population whereas an overall risk of 9% in the remaining 80% of the population warrants repeated or extended monitoring.

Objective. This study proposes a novel technique for atrial fibrillatory waves (f-waves) extraction and investigates the performance of the proposed method comparing with different f-wave extraction methods. Approach. We propose a novel technique combining a periodic component analysis (PiCA) and echo state network (ESN) for f-waves extraction, denoted PiCA-ESN. PiCA-ESN benefits from the advantages of using both source separation and nonlinear adaptive filtering. PiCA-ESN is evaluated by comparing with other state-of-the-art approaches, which include template subtraction technique based on principal component analysis, spatiotemporal cancellation, nonlinear adaptive filtering using an echo state neural network, and a source separation technique based on PiCA. Quality assessment is performed on a recently published reference database including a large number of simulated ECG signals in atrial fibrillation (AF). The performance of the f-wave extraction methods is evaluated in terms of signal quality metrics (SNR, ΔSNR) and robustness of f-wave features. Main results. The proposed method offers the best signal quality performance, with a ΔSNR of approximately 22 dB across all 8 sets of the reference database, as well as the most robust extraction of f-wave features, with 75% of all estimates of dominant atrial frequency well below 1 Hz.