Deep Learning With Longitudinal Electrocardiograms Predicts Atrial Fibrillation Better Than a Single Electrocardiogram

BACKGROUND: Atrial fibrillation (AF) is often paroxysmal and asymptomatic, yet early identification could enable targeted surveillance. Deep learning can predict future AF from a single 12-lead electrocardiogram (ECG), but AF develops through progressive atrial remodeling. OBJECTIVES: We hypothesized that modeling each patient's longitudinal ECG history would improve 1-year AF prediction over a matched single-ECG model. METHODS: We developed a transformer-based model that accepts variable-length sequences of prior 12-lead ECGs with time-aware encoding of irregular intervals. We compared it with a single-ECG model sharing the same encoder, covariates, preprocessing, outcome definition, and splits. Models were developed at Weill Cornell Medicine/NewYork-Presbyterian and evaluated on MIMIC-IV-ECG without and with fine-tuning. The outcome was AF documented more than 1 and within 365 days after an index ECG. Discrimination used patient-cluster bootstrapping with 1 random index ECG per patient. RESULTS: Cohorts included 107,811 institutional and 55,307 MIMIC-IV-ECG patients. The fine-tuned longitudinal model with age and sex discriminated best (area under the receiver-operating characteristic curve: 0.85; C-index 0.84), exceeding the single-ECG model with age and sex (paired area under the receiver-operating characteristic curve: +0.007; 95% CI: 0.003-0.012; +0.014 with ≥1 prior ECG). Without retraining on MIMIC-IV-ECG, the ECG-only advantage was confined to patients with ≥1 prior ECG (+0.014; 95% CI: 0.004-0.024). After external-domain adaptation, the matched advantage was +0.007 (95% CI: 0.002-0.012). High-risk patients had lower AF-free survival (HR: 8.8; 95% CI: 7.8-10.0). CONCLUSIONS: A longitudinal deep learning model using serial 12-lead ECGs improved 1-year AF prediction over a matched single-ECG model. Longitudinal ECG modeling may provide a scalable approach to identify patients warranting enhanced AF surveillance.

Authors

Institutions

Publication Details

Journal
JACC Advances
Published
2026-11-17
DOI
https://doi.org/10.1016/j.jacadv.2026.103468
Primary Topic
ECG Monitoring and Analysis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deep Learning With Longitudinal Electrocardiograms Predicts Atrial Fibrillation Better Than a Single Electrocardiogram

Nusrat Binta Nizam, Mert Rory Sabuncu, Sunwoo Kwak, Fengbei Liu et al.
1 links
JACC Advances
ECG Monitoring and Analysis
article

Deep Learning With Longitudinal Electrocardiograms Predicts Atrial Fibrillation Better Than a Single Electrocardiogram

Nusrat Binta Nizam, Mert Rory Sabuncu, Sunwoo Kwak, Fengbei Liu, Peter M. Okin
article en

Abstract

BACKGROUND: Atrial fibrillation (AF) is often paroxysmal and asymptomatic, yet early identification could enable targeted surveillance. Deep learning can predict future AF from a single 12-lead electrocardiogram (ECG), but AF develops through progressive atrial remodeling. OBJECTIVES: We hypothesized that modeling each patient's longitudinal ECG history would improve 1-year AF prediction over a matched single-ECG model. METHODS: We developed a transformer-based model that accepts variable-length sequences of prior 12-lead ECGs with time-aware encoding of irregular intervals. We compared it with a single-ECG model sharing the same encoder, covariates, preprocessing, outcome definition, and splits. Models were developed at Weill Cornell Medicine/NewYork-Presbyterian and evaluated on MIMIC-IV-ECG without and with fine-tuning. The outcome was AF documented more than 1 and within 365 days after an index ECG. Discrimination used patient-cluster bootstrapping with 1 random index ECG per patient. RESULTS: Cohorts included 107,811 institutional and 55,307 MIMIC-IV-ECG patients. The fine-tuned longitudinal model with age and sex discriminated best (area under the receiver-operating characteristic curve: 0.85; C-index 0.84), exceeding the single-ECG model with age and sex (paired area under the receiver-operating characteristic curve: +0.007; 95% CI: 0.003-0.012; +0.014 with ≥1 prior ECG). Without retraining on MIMIC-IV-ECG, the ECG-only advantage was confined to patients with ≥1 prior ECG (+0.014; 95% CI: 0.004-0.024). After external-domain adaptation, the matched advantage was +0.007 (95% CI: 0.002-0.012). High-risk patients had lower AF-free survival (HR: 8.8; 95% CI: 7.8-10.0). CONCLUSIONS: A longitudinal deep learning model using serial 12-lead ECGs improved 1-year AF prediction over a matched single-ECG model. Longitudinal ECG modeling may provide a scalable approach to identify patients warranting enhanced AF surveillance.

JACC AdvancesVol. 5(12)
Cornell University (US), Weill Cornell Medicine (US), Cornell Tech (US)
NewYork-Presbyterian Hospital
Good health and well-being
ECG Monitoring and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon