Key result
An optimized LLM accurately estimates AF burden from clinical notes and identifies patients likely needing ablation.
Why the study?
AF guidelines increasingly emphasize burden, but prolonged ambulatory ECG recordings are not always practical, raising the question of whether LLMs can estimate AF burden from EHR notes without ECGs.
Can a large language model accurately estimate atrial fibrillation burden and infer disease progression from electronic health record clinical notes without using ECGs?
Observational (n=3,494)
No
Can a large language model accurately estimate atrial fibrillation burden and infer disease progression from electronic health record clinical notes without using ECGs?
An optimized large language model can accurately estimate atrial fibrillation burden from clinical notes and identify patients at higher risk for requiring ablation or cardioversion, serving as a potential low-cost surrogate for ambulatory ECGs.
Answers are grounded in this paper and cite their sources.
May support LLM-derived AF burden from notes as ECG surrogate; hypothesis-generating, requires prospective validation before clinical use.
Background Guidelines for atrial fibrillation (AF) increasingly stress the importance of burden, yet prolonged ambulatory ECG recordings may not always be practical. We hypothesized that large language models (LLMs) could estimate patients’ AF burden from the electronic health record (EHR) in a large population without using ECGs, and thus enable the study of AF progression and outcomes on an unprecedented scale. Purpose To develop a method for estimating quantitative AF burden from full-text clinical notes using optimized prompts in a HIPAA-compliant GPT-4o validated against contemporaneous ambulatory recordings in AF patients. We further evaluated whether such LLM-inferred AF burden corresponded to cardioversion or AF ablation in a large population of patients with AF. Methods From the Stanford AF registry, we randomly created a partition of 3494 patients with AF clinical notes spanning at least one year (Fig A). In a development cohort of 36 patients (132 notes, 1464 ± 549 words, 33.3% female), we optimized the LLM prompt on ambulatory ECG recordings and matched AF notes to identify burden (Fig B). In a validation cohort, this optimal LLM prompt was tested in a set of clinical notes paired with ambulatory ECG recordings. We applied this LLM on a hold-out test cohort of 3,415 patients (40.20% female, N=70,621 notes, 1054 ± 746 words), in whom we compared inferred AF burden trends to ablation and cardioversion procedures. Results In the validation cohort, the LLM achieved an F1 score of 0.84 for 0-33% burden and 0.71 for 67-100%. In the hold-out test cohort, 59.12% of notes had low AF burden (0-33%), 5.29% notes indicated medium burden (34-66%), and 35.58% of notes indicated high AF burden (67-100%). Patients with an increasing LLM-inferred AF burden between their first and most recent visits were more likely to undergo ablation or cardioversion (19.04%) than others (10.32%; p<0.05) (Fig C). Conclusions An optimized LLM system may serve as an effective, low cost surrogate for ambulatory ECG recordings to determine AF burden trends in patients. If further refined, this approach can be used broadly to study AF burdens and outcomes on an unprecedented scale.A. Data splits; B. Function of LLM. LLM can predict AF progression & outcome
No takes yet. Share an insight, caveat, or question.
Brennan et al. (2025) conducted an observational in Atrial fibrillation (n=3,494). Increasing LLM-inferred AF burden vs. No increase in LLM-inferred AF burden was evaluated on Ablation or cardioversion procedures (p=<0.05). Patients with an increasing LLM-inferred atrial fibrillation burden from clinical notes were more likely to undergo ablation or cardioversion than those without an increase (19.04% vs 10.32%; p<0.05).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: