What is the clinical evidence from this study?

Study design: Observational. Population: persistent atrial fibrillation with heart failure with preserved ejection fraction (n=50). Intervention: wearable single-lead ECG monitoring with AI-based prediction model. Primary outcome: prediction of 6–12-month change in NT-proBNP levels as a marker of heart failure progression (directional accuracy 0.82).

What does this research mean for the field?

Wearable ECG–based AI modeling predicts NT-proBNP changes with a root mean squared error of 1,667.04, achieving a directional accuracy of 0.82 in patients with persistent atrial fibrillation. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The study aims to evaluate the feasibility of predicting heart failure progression using features from wearable ECG in patients with persistent atrial fibrillation.

February 8, 2026Open Access

AI-based prediction of heart failure progression in persistent atrial fibrillation using wearable electrocardiography: a brief research report

Key Result

AI-based wearable ECG monitoring predicted 6–12-month changes in NT-proBNP with a directional accuracy of 82% in patients with persistent atrial fibrillation and heart failure.

Key Points

The study aims to evaluate the feasibility of predicting heart failure progression using features from wearable ECG in patients with persistent atrial fibrillation.
Fifty patients with persistent atrial fibrillation monitored with single-lead ECG for 3–7 days.
Combined heart rate variability and RR-interval features with baseline clinical metrics.
Developed a deep learning model using LSTM and attention mechanisms to predict changes in NT-proBNP.
Assessed model performance with RMSE, MAE, and directional accuracy of NT-proBNP changes.
Best performance obtained with combined clinical metrics and HRV features (RMSE 1,667.04; MAE 950.52).
Directional classification of NT-proBNP changes reached an accuracy of 0.82.
ECG-only models performed similarly to multimodal models.

Study Design

Type

Observational (n=50)

Multicenter

Structured PICO

Does an AI model using wearable ECG-derived features predict longitudinal changes in NT-proBNP in patients with persistent atrial fibrillation?

Population

50 adults (aged ≥20 years) with persistent atrial fibrillation, median age 70.0, 56% male, from a single center in the Republic of Korea. Exclusions: acute decompensated HF, pacemaker rhythm, incomplete follow-up, excessive motion artifact, and <48 h of analyzable ECG.

Intervention

Continuous single-lead wearable ECG monitoring for 3-7 days to extract heart rate variability (HRV) and RR-interval features, combined with baseline clinical metrics, analyzed using a context-aware deep learning model (LSTM) to predict NT-proBNP changes.

Outcome

Predicted change in NT-proBNP (directional change: increase or decrease) at 6-12 months follow-up.surrogate

Wearable ECG-derived digital biomarkers combined with AI can feasibly predict longitudinal trends in NT-proBNP in patients with persistent atrial fibrillation, offering a potential non-invasive approach for heart failure monitoring.

Main Result

Effect estimate: directional accuracy 0.82

Limitations

Small sample size (n=50) from a single center limits generalizability.
No external validation of the AI model was performed.
NT-proBNP measurements were obtained at routine clinical visits with heterogeneous sampling intervals.
The study used NT-proBNP trajectory as a surrogate endpoint, not clinical heart failure events.
High inter-patient variability in baseline NT-proBNP introduced large absolute prediction errors.
Results are exploratory and hypothesis-generating given limited sample and event data.

Abstract

Background Persistent atrial fibrillation (AF) frequently coexists with heart failure (HF), yet HF monitoring remains limited by the need for repeated blood-based biomarkers such as N-terminal pro-brain natriuretic peptide (NT-proBNP). Advances in wearable electrocardiography (ECG) and artificial intelligence (AI) now allow continuous extraction of digital physiologic signatures that may reflect hemodynamic stress. Objective To evaluate the feasibility of predicting HF progression using wearable ECG–derived features in patients with persistent AF. Methods Fifty patients with persistent AF underwent 3–7 days of single-lead ECG monitoring. Heart rate variability (HRV) and RR-interval features from 30 min windows were combined with baseline clinical metrics. A context-aware deep learning model using long short-term memory (LSTM) and attention mechanisms was trained to predict 6–12-month NT-proBNP changes. Model performance was assessed using root mean squared error (RMSE), mean absolute error (MAE), and the accuracy of directional NT-proBNP change. Results The best performance was achieved when clinical metrics, RR features, and long-term HRV summaries were combined (RMSE 1,667.04; MAE 950.52). Directional classification of NT-proBNP trajectories achieved an accuracy of 0.82. ECG-only models performed comparably to multimodal models. Conclusion Wearable ECG–based AI modeling is feasible for predicting trends in HF biomarkers in persistent AF. These results provide early evidence that ECG-derived digital biomarkers may offer a scalable, non-invasive approach for longitudinal HF monitoring.

Bookmark

View Full Paper