Forecasting atrial fibrillation five minutes ahead: a personalized neural network reads wearable ECG (arXiv, 2026)

Posted on 9 June 2026 on arXiv, this preprint from a team at Seoul National University trains a neural network to forecast, from a 60-second single-lead electrocardiogram (ECG) segment, whether an atrial fibrillation episode will begin within the next 5 minutes. By fine-tuning the model on each patient's first 24 hours of recording, the AUROC — the ability to separate a positive case from a negative one — climbs from about 0.61 to 0.71 on the training cohort, and from 0.59 to 0.69 on an independent Korean cohort. The idea is appealing (anticipate the event rather than detect it), the code is open, and the demonstration spans three real ambulatory-ECG databases. But performance stays modest, the decision thresholds are calibrated on each patient's own test data, the European validation cohort shrinks to 6 patients, and two of the authors are affiliated with the company that makes one of the devices studied.

The context

Atrial fibrillation is the most common sustained heart-rhythm disorder: the heart's atria contract chaotically instead of beating regularly. It multiplies stroke risk fivefold and weighs heavily in heart failure. Its paroxysmal form is especially treacherous: episodes come and go, sometimes lasting only minutes, often symptom-free, which makes them hard to catch during a one-off exam.

Wearable devices — smartwatches, ECG patches, long-duration holters — changed the game by recording rhythm continuously for days. Most existing algorithms detect fibrillation once it has begun. This paper's ambition is different: to forecast it a few minutes before onset. The clinical stake is concrete. If one could reliably announce an imminent episode, it would enable a so-called "pill-in-the-pocket" strategy (taking an antiarrhythmic on demand, just before the episode) or an alert in intensive care, where rapid-ventricular-response fibrillation can be dangerous. Prior work on this prediction remains scarce, often limited to a single dataset and hard to reproduce.

The method

The work, authored by Jangwon Suh, Soonil Kwon, Eue-Keun Choi, Wonjong Rhee and colleagues (Seoul National University, CHA University, and the company SEERS Technology), builds on three single-lead ambulatory ECG databases: ICENTIA11K (11,000 patients, CardioSTAT patch, 250 Hz), IRIDIA-AF (152 patients, holter, 200 Hz) and MobiCARE (379 patients, the eponymous device made by SEERS, 256 Hz, private data from Seoul National University Hospital).

The task is framed as a binary classification. The recording is cut into 60-second segments; a segment is labeled "pre-fibrillation" if it immediately precedes (within 5 minutes) the onset of an episode, and "non-fibrillation" if it lies at least 20 minutes from any episode. The model must say which type each window is — in other words, predict the imminence of an episode within a 5-minute horizon.

The network is an 18-layer 1D ResNet — a convolutional neural network (CNN), a family specialized in pattern recognition, here adapted to the one-dimensional ECG signal; "ResNet" denotes a residual-connection architecture that eases training of deep networks. The exact number of parameters is not stated. The core contribution is personalization: a "global" model is first trained on the large ICENTIA11K database, then fine-tuned (lightly re-trained) on each patient's first ~24 hours of recording, before being evaluated on the rest of their trace. Each patient thus gets their own calibrated model. Performance is measured by the AUROC and the AUPRC (area under the precision-recall curve, more demanding when classes are imbalanced). Decision thresholds are set per patient at the point that maximizes Youden's index — an important detail for what follows.

The results

Personalization helps, measurably but modestly. On the internal cohort ICENTIA11K (30 patients), the AUROC rises from 0.614 to 0.711 (p = 0.046) and the AUPRC from 0.546 to 0.649. On the external Korean cohort MobiCARE (31 patients), the AUROC climbs from 0.585 to 0.686 (p = 0.018), with a clear gain in sensitivity (0.668 to 0.775) and F1 score (0.538 to 0.628). On the European cohort IRIDIA-AF, the AUROC goes from 0.633 to 0.715, but without reaching statistical significance (p = 0.195) — there are only 6 patients. The authors also show that around a dozen episodes per patient must accumulate before personalization becomes useful, and that the model focuses on electrophysiologically plausible precursors (atrial premature beats, short supraventricular tachycardias), corroborated by a rise in heart rate just before the episode.

Clinical translation: these numbers must be read with caution. An AUROC around 0.70 means that, drawing a pre-fibrillation window and a normal window at random, the model correctly ranks the former as higher-risk about 70 times out of 100 — better than chance (50/100), far from a decisive tool. At the optimal threshold on MobiCARE, a sensitivity of 77.5% and a specificity of 63.5% imply that, out of 100 windows truly preceding an episode, about 78 would be flagged and 22 missed; but above all, out of 100 normal windows, nearly 37 would trigger a false alarm. And the evaluation is run on an enriched set (about one at-risk window for two normal ones): in real continuous monitoring, where the vast majority of minutes are fibrillation-free, the share of false alarms among all alarms would be far higher. The authors acknowledge it: precision must improve "to minimize false alarms and unnecessary medication use."

What is good

Forecasting rather than detecting, on real and varied data. Tackling the 5-minute forecast of fibrillation, not its mere detection, across three ambulatory-ECG databases from three different devices, is an ambitious and useful choice. Testing the same model on cohorts acquired with distinct hardware and sampling rates is exactly the kind of generalization test the literature too often skips.

An honest, quantified deployment policy. Rather than touting personalization as a magic wand, the authors map its learning curve: they show that with too few episodes (3), it brings nothing, and propose a "data-aware" strategy — start with the global model, switch to the personalized one only once about a dozen episodes have accumulated. It is a rare and welcome stance that resists over-promising.

Open code, partly public data, interpretability. The code is released under an MIT license, and two of the three databases (ICENTIA11K, IRIDIA-AF) are publicly accessible. The authors support their predictions with Grad-CAM attribution maps showing that the network relies on recognized warning signs, and with a physiological analysis (rise in heart rate and in RMSSD variability before the episode) that makes the learned signal plausible rather than opaque.

What is less good

Thresholds calibrated on the patient's test data (failure mode: the misleading metric). Sensitivities, specificities and predictive values are computed at the optimal Youden threshold on each patient's own test set. The authors themselves note these figures are "descriptive" and not prospective performance estimates: in practice, a patient's ideal threshold is not known in advance. Add to this the class imbalance in the evaluation (1 at-risk window for 2 normal ones), far more favorable than the reality of continuous monitoring where fibrillation is rare — a classic base-rate effect that inflates the reported positive predictive value (0.58) well beyond what would be seen in the field.

Tiny, filtered validation cohorts (failure mode: selection and population bias). Of IRIDIA-AF's 167 recordings, only 6 patients were retained; 31 of 379 for MobiCARE; 30 for ICENTIA11K personalization. The inclusion criteria (fibrillation burden under 60%, a minimum number of episodes) effectively exclude patients with extreme profiles, and the European cohort is too small to conclude anything (non-significant result). The authors also note that IRIDIA-AF provides no patient identifiers, so individual uniqueness cannot be guaranteed. Generalization to a broad population therefore remains an open question.

Weak comparator, domain shift and conflicts of interest (failure modes: biased comparator and population bias). The only real comparator is the global model versus the personalized one; no simple baseline based on heart-rate variability (HRV) is implemented, though it would have made a natural clinical reference (the authors set it aside, citing the short window and noise). The model is trained on lead I and tested on lead II for MobiCARE, a non-trivial domain shift. Finally, two authors are employed by SEERS Technology, maker of the mobiCARE device, and the senior author is a shareholder and declares numerous industry ties — a conflict of interest to keep in mind, since one of the three validation databases is precisely that maker's. The study is retrospective, with no prospective trial or regulatory approval, and the model weights are not released.

What it changes

For the research community, the paper offers a reproducible baseline and a clear recipe (global model then personalization) for a still poorly charted problem, the short-term forecasting of fibrillation. The open code and the learning-curve analysis make it an honest starting point for whoever wants to do better — with architectures more recent than ResNet, and above all thresholds and an evaluation representative of real continuous monitoring.

For clinicians, nothing changes today: this is a preprint not yet peer-reviewed, with modest discrimination (AUROC ~0.70), non-transferable thresholds and no prospective validation. The idea of anticipating an episode to adjust treatment remains a credible horizon, but the false-alarm rate expected in real conditions prevents it, as it stands, from guiding a therapeutic decision.

For patients and the public, the message is measured. The promise of a watch or patch that "warns before the episode" is real in principle, but this work mainly shows the road ahead: moving from a statistically detectable signal to a tool reliable enough not to drown the user in false alarms. Anticipating is not yet preventing.

Going further

The full preprint is available on arXiv (DOI 10.48550/arXiv.2606.10900), posted 9 June 2026 under a CC BY-NC-ND 4.0 license, by a team from Seoul National University, CHA University and SEERS Technology. The code is released under an MIT license on GitHub. The ICENTIA11K (PhysioNet) and IRIDIA-AF (Zenodo) databases are publicly accessible; the MobiCARE database remains private. For the clinical framework of atrial fibrillation, the European Society of Cardiology (ESC) guidelines offer a reliable entry point.