This prospective observational study enrolled 50 patients with persistent AF who underwent wearable ECG monitoring between 2023 and 2024 at St. Vincent's Hospital, The Catholic University of Korea. Adults aged 20 years and older were eligible. Exclusion criteria included acute decompensated HF, pacemaker rhythm, incomplete follow-up data, excessive motion artifact (>20% unusable signal), and <48 h of analyzable ECG. Follow-up NT-proBNP measurements were obtained during routine clinical visits, resulting in a sampling interval of 6–12 months. All participants provided written informed consent. The study was approved by the institutional review board (VC23OISI0082) and conducted in accordance with the principles outlined in the Declaration of Helsinki.

Participants wore a single-lead wearable ECG device (MobiCARE, Seers Technology Co., Korea) continuously for 3 to 7 days. The ECG was segmented into 30 min windows to derive sequential HRV features (HRV30 min). Long-term HRV summaries and RR-interval–based features (HRV72 h and RR) were calculated from the entire recording period. Baseline clinical metrics (CM) were extracted from electronic medical records. Feature distributions were assessed using Q–Q plots and the Shapiro–Wilk test. All features were standardized using z-score scaling and Yeo–Johnson transformation. A 30 min window length was used to balance temporal resolution and noise reduction in persistent AF. Feature groups are summarized in Table 1.

ECG preprocessing was performed using device-generated R-peak annotations together with noise-labeled intervals. R-peaks occurring within noise-marked segments were removed, and no additional ectopic-beat filtering was applied, consistent with the physiology of persistent AF. For each 30 min window, RR intervals were computed from successive R-peak timestamps. Windows with fewer than two valid RR intervals were treated as invalid, and all HRV metrics were coded as −1. Time-domain HRV metrics were computed directly from the cleaned RR series. At the same time, frequency-domain features were derived by converting RR intervals to seconds, applying linear interpolation to obtain an evenly sampled signal, and estimating power spectra using Welch's method. Windows with insufficient RR data for spectral estimation likewise had frequency-domain HRV metrics assigned a value of −1.

Although conventional short-term HRV measures have limited physiological interpretability in persistent AF because beat-to-beat irregularity obscures sinus-node–mediated autonomic modulation, a prior study has shown that long-window HRV and RR-variability–derived metrics can still reflect global autonomic tone and hemodynamic burden in AF. Based on this evidence, we used 30 min aggregated windows as surrogate markers of autonomic dynamics, as longer windows reduce stochastic irregularity and capture lower-frequency physiologic patterns relevant to AF-related remodeling (11).

To ensure signal reliability, artifact-related noise was mitigated by excluding recordings with more than 20% unusable segments, removing physiologically implausible RR intervals, and applying window-level aggregation, which collectively improves robustness in AF-derived HRV estimation.

A hybrid deep learning model was developed to integrate time-varying ECG-derived features with static clinical information. For each patient, HRV features, RR-interval dynamics, and interaction terms were aggregated into a multivariate time series that reflects autonomic and beat-to-beat variability over the monitoring period. Clinical metrics, which are inherently non-temporal, were replicated across all 30 min windows and concatenated with the sequential features to create a unified model input. This allowed the network to learn temporal autonomic patterns while simultaneously leveraging stable patient-specific characteristics. Baseline NT-proBNP and meanRMSSD were also incorporated as context vectors within the attention mechanism to modulate temporal weighting in a patient-dependent manner.

The model architecture consisted of a three-layer long short-term memory (LSTM) network with 64, 32, and 16 hidden units, followed by a context-aware attention module and a two-layer regression head. This design enabled the model to capture both short-term fluctuations and longer-range patterns in autonomic variability. Training was performed using the Adam optimizer (learning rate 0.001), a batch size of 8, and mean squared error as the loss function, with early stopping applied based on training performance (patience = 10). To minimize overfitting in this small-sample setting, several regularization strategies were used, including L2 weight decay (0.001) and dropout (0.5).

Additional details regarding the model architecture, attention mechanisms, input construction, and statistical feature analysis are provided in the Supplementary Methods.

Model performance was assessed using leave-one-out cross-validation (LOOCV) at the patient level, ensuring that no information from the held-out patient was used in model training. For each iteration, the model generated a continuous prediction of the change in NT-proBNP, which was subsequently categorized as an increase or a decrease. The primary endpoint was the predicted change in NT-proBNP. Performance metrics included root mean squared error (RMSE), mean absolute error (MAE), and directional accuracy.

Among the 50 patients included in the analysis, the median age was 70.0 [60.25–78.75] years, and 56% were male. Baseline clinical characteristics are shown in Table 2.

The best regression performance was achieved when combining clinical metrics, RR-interval features, and long-term HRV summaries (HRV72h), yielding an RMSE of 1,667.04 and an MAE of 950.52. When regression outputs were converted into a trend label (increase vs. decrease/stable), the model achieved directional accuracy of 0.82 (Supplementary Table S1). Predicted and observed NT-proBNP trajectories for individual patients are shown in Figure 1 and Supplementary Table S2. Figure 2 illustrates the architecture of the context-aware LSTM model.

Models using ECG-only features performed comparably to those incorporating multimodal clinical variables.

As shown in Supplementary Table S3, patients with increasing NT-proBNP experienced more HF-related hospitalizations (9 vs. 5), and additional events such as cardioversion and stroke. However, these differences did not reach statistical significance due to the limited number of clinical events.