Forecasting Imminent Atrial Fibrillation in Long-Term Electrocardiogram Recordings

Abstract

Background:

Despite the morbidity associated with acute atrial fibrillation (AF), no models currently exist to forecast its imminent onset. We sought to evaluate the ability of deep learning to forecast the imminent onset of AF with sufficient lead time, which has important implications for inpatient care.

Methods:

We utilized the Physiobank Long-Term AF Database, which contains 24-h, labeled ECG recordings from patients with a history of AF. AF episodes were defined as ≥5 minutes of sustained AF. Three deep learning models incorporating convolutional and transformer layers were created for forecasting, with two models focusing on the predictive nature of sinus rhythm segments and AF epochs separately preceding an AF episode, and one model utilizing all preceding waveform as input. Cross-validated performance was evaluated using area under time-dependent receiver operating characteristic curves (AUC(t)) at 7.5-, 15-, 30-, and 60-minute lead times, precision-recall curves, and imminent AF risk trajectories.

Results:

There were 367 AF episodes from 84 ECG recordings. All models showed average risk trajectory divergence of those with an AF episode from those without ~15 minutes before the episode. Highest AUC was associated with the sinus rhythm model [AUC=0.74; 7.5-minute lead time], though the model using all preceding waveform data had similar performance and higher AUCs at longer lead times.

Conclusions:

In this proof-of-concept study, we demonstrated the potential utility of neural networks to forecast the onset of AF in long-term ECG recordings with a clinically relevant lead time. External validation in larger cohorts is required before deploying these models clinically.

Introduction

Machine learning methods can inform clinical decision-making in various domains by improving diagnosis, predicting outcomes, and outperforming current prognostic indicators[1-3]. A comparatively new area of interest in this domain is predictive analytics, with models being created that demonstrate the ability to forecast clinically-meaningful events in real-time[4-6]. These algorithms demonstrate the potential for predictive analytics to make clinical tools that inform medical decision-making and improve clinical outcomes.

Atrial fibrillation (AF) is both the most common cardiac arrhythmia and associated with significant morbidity and mortality[7,8]. Considerable work has utilized machine learning and advanced analytic techniques to identify patients with subclinical AF during sinus rhythm[9-11]. While these models help ascertain those with or at risk of AF in hopes of mitigating its long-term sequelae, they cannot help in cases of acute AF, which is defined as AF in the setting of acute medical care or illness. Acute AF is well-known to occur with alarming frequency during the administration of medical therapies such as dialysis and immunotherapy, as well as in medical and surgical intensive care units. Its occurrence is associated with suboptimal medical therapy delivery, increased risk of cardiorespiratory decompensation, longer lengths of stay, mortality, and increased risk of AF recurrence[12]. Despite acute AF’s impact on patient outcomes, no models currently exist to forecast AF’s imminent onset.

In this context, we devised a proof-of-concept study that applied deep learning to waveform data from long-term ECG recordings of patients with a history of AF to forecast the onset of AF episodes. We hypothesized waveform signatures would predict imminent AF both in the sinus rhythm segments preceding an AF episode and the short epochs of AF before a longer AF episode. We sought to evaluate whether either of these rhythm signatures could predict AF’s onset independently, or if integrating data from both sinus rhythm segments and AF epochs would enhance our ability to forecast significant AF episodes accurately.

Materials & Methods

Data

The data for our study was extracted from the publicly-available Physionet Long-Term AF Database[13]. The database comprises 84 long-term, 2-channel ECG recordings approximately 24-25 hours in duration. These recordings were collected at Northwestern University, and subsequently digitized at 128 Hz with 12-bit resolution over a 20 mV range. Verified rhythm as well as individual beat annotations are available in the database. While demographic information on the entire cohort is not publicly available, the cohort is partially described[14].

For our purposes, the data was preprocessed using Python programming language (version 3.9, Python Software Foundation, Wilmington, DE, USA). The waveform signal amplitude was standardized utilizing segment interquartile range.

Outcome of Interest

The pre-labeled AF annotations from the Physiobank Long-Term AF Database were utilized. Due to clinical relevance[15,16], we defined an AF episode of interest as any AF annotation ≥5 minutes long. Accurate forecasting of such an episode was defined as the presence of at least part of an episode occurring before the end of identified lead time. This outcome was studied at several lead times (7.5-minutes, 15-minutes, 30-minutes, 60-minutes) to examine the maximal amount of lead time the forecasting model could afford physicians if enacted in clinical practice.

Model Development

All patients in the dataset were included in the analysis. Data were split into 3-minute consecutive windows of time with a 30-second stride. The input to our model was three minutes of waveform signal from two electrocardiogram channels (3 min*60 sec/min*128 Hz*2 channels). We utilized a deep network architecture consisting of three 1-D convolutional layers, three transformer layers (one encoder layer, two decoder layers) and three fully-connected linear layers. Specifics regarding architecture can be visualized in Figure 1. A domain generalization-regularization term, maximum mean discrepancy, was added before the fully connected layers to avoid overfitting patient-specific patterns and improve generalizability. ReLU was the chosen activation function for all layers. Output varied based on model iteration and were either 1) the parameters (γ, β) of a Weibull Distribution characterizing the probability of onset of AF using a survival analysis approach or 2) the probability of the input segment being AF.

Figure 1:

Basic neural network architecture for the forecasting models.

Three models were created to assess whether signatures capable of forecasting AF were contained in sinus rhythm segments and/or AF epochs immediately preceding an AF event. For the model examining sinus rhythm signatures, windows of pure sinus rhythm where time-to-event (TTE)<1 hour were utilized and trained based on the exact TTE with the loss function being −log P(TTE=tte \ x). Any window where the time to an AF event is >1 hour was censored and downsampled by 90% prior to model training. With this model, the chosen output was Weibull Distribution parameters (γ, β) characterizing the probability of onset of AF using a survival analysis approach. The associated cumulative distribution function determines the probability of an AF event occurring in a specified lead time. For each lead time (7.5-, 15-, 30-, and 60-minutes), the associated probabilities were assessed across different classification thresholds (0 to 1) to create time-dependent receiver operating characteristic (ROC) curves. Further description of time-dependent ROC curve creation is in Supplemental Description 1.

For the second model, we analyzed whether forecasting signals existed in the short epochs of AF before a sustained, 5-minute or longer episode using a novel AF detector-turned-forecaster approach. Nearly the same architecture was used, but we instead trained the model with labeled segments of sinus rhythm and AF, and the output was the probability of the input window being AF. These probabilities were then utilized to indirectly create a novel forecasting mechanism by examining how well the probability of a segment being AF performed as a surrogate for the likelihood of imminent AF at different lead times. This was examined at an array probability thresholds (0 to 1), allowing for time-dependent ROC curve creation.

Lastly, all windows with TTE<1 hour were utilized to create a forecaster to see if the combined information of signals from both sinus rhythm and short AF segments was superior to either rhythm type alone. The model construction and output mirrored that of the first iteration on the sinus rhythm segments, except we now included windows of all rhythm types as input.

Model Testing/Evaluation

All models were trained using 3-fold cross-validation. In addition to evaluating models using areas under the time-dependent ROC curves (AUC(t)) at varying lead times, we also produced precision-recall curves for each model to assess performance further. We additionally examined the true positive rate (TPR), or the model’s ability to correctly forecast AF’s onset at the specified lead time, at a false positive rate (FPR) of 0.25.

Furthermore, we created imminent risk trajectories to compare the risk scores for those with an AF episode ≥5 minutes occurring in 1 hour compared to those without. The trajectory of those with imminent AF was constructed by gathering the risk scores from windows during the hour before AF, plotting the scores over time, and fitting a locally estimated scatterplot smoothing (LOESS) model[17] to represent the average risk trajectory. To obtain a risk trajectory for those without imminent AF, we utilized 60-minute segments where patients did not have an impending AF episode, computed the corresponding risk scores, then smoothed using LOESS regression.

Lastly, we qualitatively examined the risk trajectories of individual patients up to 2-hours before an AF episode depending on availability of preceding data. All analyses used Python programming language (version 3.9, Python Software Foundation, Wilmington, DE, USA).

Results:

Data Summary

There were 367 AF episodes in total from 84 long-term ECG recordings. The cumulative amount of time in AF per patient ranged from 40 seconds to 25 hours and 44 minutes. Ten patients had persistent AF for their entire recordings. Looking at AF episodes ≤5 minutes, we saw an increase in the percentage of time in AF overall leading up to our more sustained episodes of interest.

Sinus Rhythm Model

The model utilizing sinus rhythm segments with TTE< 1 hour as input had an AUC(t)>0.7 for lead times of 7.5- and 15-minutes (see Table 1). Model performance lowered with lead times of 30- and 60-minutes. With 15-minutes of lead time, our forecaster predicted onset of AF with a TPR of 56% at a FPR of 25%. Figure 2a demonstrates an acute increase in risk based on sinus rhythm signals in those who will develop sustained AF approximately 20 minutes before the episode. Figure 3a illustrates the associated time-dependent ROC and precision-recall curves for the model. At a recall of 80%, the precision using a 60-minute lead-time is ~55%.

Table 1:

Summary of model performance using area under the time-dependent ROC curve (AUC) and the true positive rate (TPR) at a false positive rate of 0.25.

Model Type	AUC (7.5 min)	TPR (7.5 min)	AUC (15 min)	TPR (15 min)	AUC (30 min)	TPR (30 min)	AUC (60 min)	TPR (60 min)
Input: Preceding Sinus Rhythm	0.74 +/− 0.06	0.59 +/− 0.08	0.71 +/− 0.05	0.56 +/− 0.08	0.65 +/− 0.04	0.50 +/− 0.05	0.61 +/− 0.03	0.41 +/− 0.04
Atrial Fibrillation Detector-as-Predictor	0.69 +/− 0.02	0.52 +/− 0.03	0.68 +/− 0.02	0.52 +/− 0.02	0.67 +/− 0.02	0.50 +/− 0.02	0.65 +/− 0.02	0.45 +/− 0.03
Input: All Preceding Waveform	0.73 +/− 0.02	0.60 +/− 0.03	0.72 +/− 0.02	0.58 +/− 0.02	0.72 +/− 0.02	0.58 +/− 0.02	0.71 +/− 0.02	0.57 +/− 0.03

^**Note: Values are accompanied by 95% confidence intervals.

Figure 2:

Imminent and stable risk trajectories when 2a) utilizing sinus rhythm segments where TTE < 1 hour, 2b) focusing on AF epochs preceding an AF episode and 2c) utilizing all waveform data where TTE< 1 hour as input.

Figure 3:

Receiver Operating Characteristic (ROC) curves and Precision-Recall Curves at varying lead times for 3a) the sinus rhythm model, 3b) the AF detector-as-a-predictor model, and 3c) the model utilizing all preceding waveform data.

Atrial Fibrillation Detector-as-a-Forecaster Model

The detector-as-a-forecaster model was fed with 6,740 positive labels for AF and 21,083 negatively labeled segments. The model had slightly lower AUC(t)s than the previous model, with its best-performing AUC(t) being 0.69 at a 7.5-minute lead time. Model performance did not lower significantly with increased lead time, with AUC(60 minutes) being 0.65. Figure 2b demonstrates an acute increase in risk in those who will develop sustained AF approximately 15 minutes before the episode using this model. Figure 3b illustrates the associated ROC curves and precision-recall curves, respectively, for different lead times for the model.

Combined Rhythm Model

Our model utilizing all waveform data with TTE<1 hour had similar performance compared to the previously best performing model, with the highest AUC(t) being 0.73 at a 7.5-minute lead time. Model performance at a 60-minute lead time was higher than both previous models, maintaining an AUC(t)>0.70 (see Table 1). Figure 2c demonstrates an acute increase in risk based on waveform signals in those who will develop sustained AF approximately 15 minutes before the episode similar to previous models, but also demonstrates a difference in the baseline risk between those who develop a sustained episode and those who do not. Figure 3c illustrates the associated ROC curves and precision-recall curves for different lead times for the model.

Individual Risk Trajectories

Despite the average risk trajectory exhibiting a distinctive increase in risk prior to an AF episode, there are significant variations in individual risk trajectories. See Supplemental Figure 1 for examples.

Discussion:

We evaluated the ability of deep learning to forecast the likelihood of imminent AF with sufficient lead time for potential clinical preventative measures. By analyzing both the preceding sinus rhythm segments in isolation before a sustained event and the preceding AF epochs using our novel detector-as-a-predictor approach, we have provided evidence that forecasting signals exist in both rhythms. While shorter lead times tend to have higher AUC(t)s, the model utilizing all rhythm data yields more stable AUC(t)s across lead times from 7.5 minutes to 60 minutes than the other two models. Our precision-recall curves suggest higher degrees of accuracy with longer lead times, though this is biased as longer lead times provide more leeway in predicting when AF is to occur. Despite our models’ performances on the group level, our individual risk trajectories show varying performance, highlighting that there are multiple different patterns that precede AF entry, which supports what we observe clinically. Further work delineating these sub-phenotypes would be prudent for a clinically applicable model.

In recent years, there has been significant interest in utilizing machine learning to advance care in those with AF. A wide array of approaches for detecting AF has been developed, ranging from utilizing sophisticated heart rate featurization with simpler classification schema such as random forest[18] to utilizing deep learning[19]. Given the amount of AF thought to be missed diagnostically due to clinicians not capturing a patient’s AF paroxysms, there have also been major strides in diagnosing patients with AF from their sinus rhythm ECG[9-11]. While there has long been interest in predicting acutely when AF will occur in real-time, there have unfortunately not been similar advances in the literature. In 2001, significant interest in this task was demonstrated by having predicting imminent AF be the subject of the yearly Physionet/Computing in Cardiology Challenge[20]. The challenge was set up as a classification task to distinguish between 30-minute segments before AF and segments far away from an AF episode. At that time, the best-performing model had an accuracy of under 80%. Since then, there has not been significant progress in acutely forecasting the onset of imminent AF.

While our study was performed using outpatient data due to its availability, it provides a foundation for larger, more generalizable models to be built to forecast AF. When thinking about the profound effect acute AF has on patients and the healthcare system, its footprint is often biggest in medical and surgical intensive care units[7]. Conditions such as post-cardiac surgery, sepsis, and need for dialysis all predispose to AF, and an AF episode can quickly cause detrimental consequences to their care. Many afflicted patients do not even have a prior history of AF[12]. Given that we found forecasting signals in sinus rhythm segments preceding AF, there is promise that this type of forecasting algorithm could be effective in cases of new-onset AF. An AF forecasting model would be a powerful tool for intensivists as it affords them the opportunity to potentially mitigate AF before its occurrence, thus improving clinical outcomes.

We presented our study as a proof-of-concept work as its limitations are necessary to consider and improve upon before the clinical adoption of any similar model. The size of the Physionet Long-Term AF Database is an inherent limitation, and the population itself is limited to people with a known history of AF. While we utilized cross-validation and added a domain-generalization regularization term to our architecture to create a more generalizable model, having a larger dataset that included patients without a history of AF would be prudent in the next iterations of model development. We also did not have access to demographic or clinical data, and it is well known that different cohorts have different risks of AF at baseline. Additionally, we grouped all atrial fibrillation events as a common process, though there are several sub-phenotypes of atrial fibrillation morphologies. Examining these sub-phenotypes in a larger dataset may show that they have different predictive performances based on these approaches. Lastly, we defined an AF episode as 5-minutes of consecutive atrial fibrillation based on what we deemed clinically significant[15,16], but there are likely a variety of endpoints one could consider, such as different percentages of AF burden over a predetermined period. How the endpoint is defined could potentially affect the accuracy of our forecasting algorithm.

Our team plans to pursue an intensive care adaptation of this approach utilizing stored demographic, clinical, and telemetry data from patients at the University of Pittsburgh Medical System. This will add clinical relevance and improve sample size compared to our previous model, making it an intuitive next step toward improving patient care.

Conclusions

In summary, our work demonstrated the ability to use deep learning to forecast imminent AF development. This proof-of-concept study provided evidence of signals capable of acutely forecasting AF in both sinus rhythm segments and short bursts of AF prior to a more sustained episode, suggesting that AF has the potential to be forecasted in different sub-phenotypes of patients about to enter AF. Further work in larger datasets would potentially allow for clarification of these sub-phenotypes and subsequently an improved ability to forecast AF on the individual level.

Supplementary Material

2

Supplemental Figure 1: Individual risk score trajectories for 4a) an individual who had increased bigeminy prior to their AF episode, 4b) an individual who had alternating sinus bradycardia with many short epochs of atrial fibrillation prior to their AF episode, 4c) an individual who had increased ectopy near the AF episode, and 4d) an individual who had normal sinus rhythm and short spurts of bigeminy, atrial fibrillation, and bradycardia prior to their AF episode.

• Atrial fibrillation has the potential to be forecasted in real-time

• Forecasting signatures are present in sinus rhythm preceding atrial fibrillation

• Risk trajectories show an increase in risk of an event ~15 minutes prior

• Different sub-phenotypes of rhythm patterns before atrial fibrillation exist