Short-term prediction of atrial fibrillation from ambulatory monitoring ECG using a deep neural network

Abstract

Background Atrial fibrillation (AF) is associated with significant morbidity but remains underdiagnosed. 24-hour ambulatory ECG is largely used as a tool to document AF but yield remains limited. We hypothesize a deep learning model can identify patients at risk of AF in the 2 weeks following a 24-hour ambulatory ECG with no documented AF. Methods We identified a training set of Holter recordings of 7 to 15 days duration, in which no AF could be found in the first 24 h. We trained a neural network to predict the presence or absence of AF in the 15 following days, using only the first 24 h of the recording. We evaluated the neural network on a testing set and an external dataset not used during algorithm development. Results In the testing data set, out of 9993 Holters with no AF on the first day, we found 361 (4%) recordings with AF within the 15 subsequent days of monitoring (5808, 218 (4%) respectively in the external dataset). The neural network could discriminate future AF with an area under the receiver operating curve, a sensitivity and specificity of 79·4%, 76% and 69% respectively (75·8%, 78% and 58% in the external dataset), and outperformed ECG features previously shown to be predictive of AF. Conclusion We show here the very first study of short-term AF prediction using 24-hour Holter monitoring. This could help identify patients who would benefit the most from longer recordings and proactively initiate treatment and AF mitigation strategies in high-risk patients.

Artificial intelligence for atrial fibrillation detection, prediction, and treatment: A systematic review of the last decade (2013–2023)

“…Similar approaches were employed by Tohyama et al (2023) and Tzou et al (2021), who proposed CNNs to predict postoperative AF. Singh et al (2022) proposed a framework called CardioNet that analyzed cardiac abnormalities by directly modeling physiological relationships from unstructured ECG reports using natural language processing. Khurshid et al (2022) developed a CNN model to predict long‐term AF risk using 12‐lead ECG data from multiple datasets.…”

Prediction of atrial fibrillation from at-home single-lead ECG signals without arrhythmias

“…The value of AF prediction from a 12-lead ECG without AF occurrences has recently been shown in a prospective study, with a specificity of 98% but a sensitivity of only 7.5% in the proposed operating point 14 . A longer rhythm observation has been performed with Holter monitors, showing the advantage of considering an interval longer than the 10-s, 12-lead ECG in order to improve AF prediction, with an AUC, a sensitivity, and specificity of 0.79, 76%, and 69% 15 . While previous studies have explored AF prediction using short 12-lead ECGs, our study complements these investigations by analyzing a single-lead ECG in a scenario with an extended duration of monitoring, which allowed us to also detect individuals with low AF burden that a brief 10-s 12-lead ECG snapshot might miss.…”

“…To alleviate the cost of non-targeted screening, risk stratification of individuals most likely to benefit from long-term monitoring would be of great clinical benefit 10 . One method of risk stratification can be enabled by future AF prediction from findings in a non-AF ECG signal, as shown with 10-s, in-clinic 12-lead ECGs in retrospective 11 – 13 and prospective 14 trials, or using single-lead 24-hour, non-AF ECG recordings to predict AF in the next 14 days 15 .…”

Can 24 h of ambulatory ECG be used to triage patients to extended monitoring?

“…A previous study has applied a machine learning approach to predict AF based on shorter ambulatory monitoring, which did not result in substantially better ROC statistics (Singh et al, 2022). An advantage of the present approach is its comparative simplicity, and the fact that the relation between ECG features and AF risk is described.…”