In recent years, the use of artificial intelligence (AI) for the diagnosis, prognosis, and evaluation of risk factors and the response to treatment of different pathologies has shown significant advancements. This has had a positive impact on health care systems, improving the efficiency of medical handling and reducing the number of adverse events [1,2].

Cardiac electrophysiology has emerged as one of the fastest-growing subspecialties within cardiology over the last few decades [3,4], which is evidenced by numerous technological advancements aimed at detecting, predicting, and managing cardiac arrhythmias [5,6]. Among those, the prediction of cardiac arrhythmias is especially relevant for patients with atrial fibrillation (AF), who are 5 times more likely to experience thromboembolic events than patients without AF [7].

AI training of information systems based on enormous data quantities from clinical reports and complementary tests from millions of patients not only enables episode prediction in patients diagnosed with AF but also allows for the diagnosis of asymptomatic patients, allowing for early intervention and reducing the number of clinical events [8]. However, cardiac implantable devices can also be leveraged to train AI models for the detection of cardiac arrhythmias. They are particularly well suited for machine learning (ML)–based arrhythmia prediction because they provide continuous, high-fidelity rhythm monitoring over long periods, enabling precise quantification of atrial arrhythmia burden and enhancing diagnostic remote monitoring workflow and device-based therapies [9], with the additional advantage of being already implanted on the patient.

Although several authors have published many AI- and ML-based models for AF prediction in different contexts with good results [10-12], these models were primarily trained and evaluated for predictions over extended periods (often up to 1 year), which provided them with a great amount of input data. However, algorithms supporting short-term, dynamic AF prediction in patients with implantable cardiac devices are scarce. Short-term arrhythmia prediction is uniquely challenging because it requires anticipating nonlinear state transitions driven by heterogeneous and rapidly evolving physiological signals. While long-term prediction is valuable for estimating the risk of AF development in the general population—supporting primary prevention and population-level care planning—dynamic short-term prediction may be particularly impactful in patients who have already been diagnosed with AF. In this context, continuously collected data from implantable cardiac devices enable high-resolution, longitudinal assessment of cardiac rhythm, creating an opportunity for near-term forecasting of AF burden progression. Such device-based prediction could support timely optimization of medical therapy, generation of real-time alerts, and prevention of downstream adverse outcomes. Moreover, in patients with newly diagnosed AF, device-derived dynamic prediction may facilitate closer disease monitoring, guide medication escalation, and enable proactive risk factor modification [10,13].

For that reason, this study aimed to evaluate the use of AI on data extracted from remote pacemaker monitoring for short-term prediction of the onset or increase of AF episodes. Our hypothesis was that data obtained from remote pacemakers could be used to train AI models to predict the onset of atrial arrhythmias with a relatively high sensitivity and specificity without the need for additional datasets.

Data from 314 patients enrolled in the IA Pacing study between May 2022 and May 2024 were analyzed. A total of 65,243 data sequences from implantable devices were collected, of which 55,532 (85.1%) were used to train the algorithm, and the remaining 9711 (14.9%) were used to analyze the sensitivity and specificity of the generated predictions.

AI implementation is transforming health care in cardiology, especially in the context of heart rhythm–related disorders. We developed an initial algorithm with information obtained from the remote monitoring of pacemakers that predicted atrial arrhythmia onset within the 14 days following the 31-day monitoring period with a 66.4% sensitivity and 77.4% specificity. This study is, to the best of our knowledge, the first to use data exclusively from remote pacemaker monitoring to dynamically predict arrhythmia onset in patients with atrial arrhythmia in the short term.

Given how important arrhythmia detection is for the identification of patients at risk of having a stroke or systemic embolism, AI algorithms have lately been used for detection, prediction, and treatment tailoring of patients with AF [8]. Focusing on AF episode prediction, several authors have achieved significant advancements in developing AF-predicting algorithms in varying periods from initial measured data. Kao et al [14] developed a model able to predict new arrhythmic episodes in the span of a year with a 98.7% specificity using a database of clinical data gathered for 3 years from older patients with no electrocardiographic data. Similarly, Tiwari et al [15] predicted 6-month AF onset with 75% sensitivity and 85% specificity by developing a model using 200 clinical features from more than 2 million clinical records. Another algorithm worth mentioning developed from clinical records is the one from the Future Innovations in Novel Detection of Atrial Fibrillation study by Nadarajah et al [16] that predicted AF onset within 6 months with a sensitivity and specificity of 78.1% and 73.1%, respectively. Clinical data have also been used to train ML models to detect risk factors for developing AF in asymptomatic patients, as described by Hill et al [10]. Another author used data from 100,000 electrocardiograms (ECGs) taken from 40,000 patients in a mean of 2 years of follow-up to predict the time before a new arrhythmic episode, obtaining similar results to medical record–based predictions [11]. In our study, unlike the studies cited before, no medical record was used to train the system. The algorithm compared daily data from the remote monitoring of an implantable device, similarly to the recently published study by Gregoire et al [13] that trained an ML model using Holter recordings to predict short-term AF episodes. We consider this a strength of our study as only data obtained from the remote monitoring of the implantable device were needed to estimate the risk of experiencing an arrhythmic episode, thus eliminating the need for additional information from medical records.

Although our algorithm does not compute the time before a new arrhythmic episode, it does enable the identification of patients at risk of experiencing episodes of supraventricular arrhythmias in the short term (14 days following the measurement), which would improve their medical care. Our model correctly predicted whether arrhythmias would increase in almost 75% of the cases, with 66.4% global sensitivity and 77.4% specificity. The previously mentioned study from Gregoire et al [13] predicted short-term paroxysmal AF episode onset with higher sensitivity and specificity than ours (86.6% vs 83.0%, respectively). However, it is worth mentioning that, while their approach relies on comprehensive ECG recordings from a Holter monitor, our analysis is based solely on limited, secondary data derived from pacemaker use (eg, daytime heart rate, respiratory rate, or apnea counts). Although these variables provide less information, they may still be highly informative as they are known to be correlated with arrhythmic episodes [17]. Additionally, it is noteworthy that our algorithm was trained using data obtained from a device already implanted in the patient, thereby eliminating the need for additional medical visits. A table comparing the main features of our study to those of the ones mentioned previously can be found in Multimedia Appendix 7.

In patients with no registered history of arrhythmia (232/314, 73.9%), who constitute the group most likely to benefit from early therapeutic interventions, the analysis showed higher values of sensitivity and specificity (69% and 80%, respectively). Algorithm performance was different according to baseline arrhythmia status (higher sensitivity in patients with arrhythmia in the last 31 days and higher specificity in patients with no episodes in the same period). We think this is due to the different data distribution among these 2 groups as sensitivity and specificity heavily depend on the prevalence of the episode of interest (in this case, arrhythmia).

The fact that our algorithm predicted the occurrence of arrhythmias with a similar performance to that of the model by Tiwari et al [15] (which predicted arrhythmias in a 6-month time window based on clinical records from 2 million younger patients) with relatively few data is a promising starting point to continue enhancing AI training and, therefore, the obtained results. Similarly to Tiwari et al [15], we deem it necessary to optimize AI models to increase their sensitivity and reduce false positives. One strength of our approach is that it requires only a small amount of data for training because the implantable devices continuously monitor cardiac activity. We also used a tool to standardize variable names across devices, which allowed us to combine datasets and increase the amount of data for training, which is another strong point of this study.

It is worth mentioning some limitations of the proposed model. First, the training did not consider demographic variables. Including these factors could increase the predictive capabilities and justify the use of AI over classic risk factor predictors [18]. Moreover, it has been observed that models encompassing more than one type of variable—such as clinical data and ECGs, echocardiograms, or magnetic resonance imaging, for instance—increase prediction sensitivity [19]; this should be taken into account when designing and training future AI algorithms. Second, although the number of prediction and reality sequences was high (more than 9000), data pairs were taken from a cohort of 314 patients, which could be considered low. The size of the cohort could then be increased to augment the training database and refine the output of the model, optimizing AI training and reducing the chance of overfitting the model [20]. It is also worth mentioning that our algorithm did not have external validation as it is a first-phase development, and testing the finalized model on independent data is necessary to draw more specific conclusions on its clinical impact. Additionally, the ability of cardiac devices to differentiate arrhythmia types should be considered when using the extracted data to train the model [21,22]. On this matter, our model did not distinguish among AF, atrial tachycardia, and atrial high-rate episodes, making it hard to fully characterize the clinical relevance of the algorithm. Further research is being conducted considering minimal times, ECG information, and heart rate to fully characterize arrhythmia type. Finally, we extracted the data from devices from the same vendor, and we cannot affirm that this model can be reproducible with other commercial brands. However, fine-tuning methods can be used to train the algorithm using information from other devices. Taking all the aforementioned factors into consideration, further research could focus on increasing the prediction period and on differentiating types of arrhythmias given the clinical implications of de novo episodes of AF and the need for anticoagulation therapy.

Finally, we would like to mention the importance of an intuitive web design to present the obtained data. It has been proven that the use of mobile apps in health care is determined by how user-friendly and intuitive they are and the amount of data that must be input by users (eg, health care professionals) for the apps to function (eg, patient demographics, clinical variables, or diagnostic data) as a higher data entry burden may negatively affect usability and adoption [23]. If an application quickly displays the prediction along with the patient’s arrhythmia risk percentage, it is more likely that health care professionals will use it.

In conclusion, this study provides preliminary evidence that changes in arrhythmic burden may be estimated with limited but measurable sensitivity and specificity using data derived from implantable cardiac devices. A graphical overview of the study and its main findings is provided in Multimedia Appendix 8 (Visual Abstract). The predictive performance of the model may improve with further development and validation in larger cohorts that include additional demographic and clinical variables. These elements, together with the user-friendly visualization of predictions in a simple application, suggest that this initial phase of the study may serve as a basis for future research. Further studies are warranted to determine the clinical utility of this approach for the monitoring and management of patients with implantable cardiac pacing devices.