The burden of subclinical atrial fibrillation (AF) didn’t predict stroke or other embolic events, or the impact of direct oral anticoagulation, a NOAH-AFNET 6 substudy showed.

For combined risk for stroke, systemic embolism, or cardiovascular death, there was no significant relationship with an AF burden of 1% or less (P = .313) or more than 1% (P = .440), nor with the treatment impact of edoxaban vs placebo (P = .254 for interaction). Findings were similar when looking at ischemic stroke or systemic embolism alone as well.

“That means that patient characteristics beyond AF burden are required for decision making on the anticoagulation therapy in patients with device-detected atrial fibrillation,” Ulrich Schotten, MD, PhD, of Maastricht University in Netherlands, said during a presentation of the results at the Heart Rhythm Society (HRS) 2026 meeting in Chicago.

In clinically diagnosed AF, “there is a quantitative relation between the time spent in AF and the risk for stroke,” Schotten said. 

The key question is: “What is a meaningful AF burden relevant for therapeutic decisions?” said HRS session study discussant Elaine Wan, MD, of New York-Presbyterian/Columbia University Medical Center in New York City.

One major factor might have been the low AF burden in the NOAH-AFNET 6 trial, Schotten said. Just 0.4% of the 1693 patients who had data to calculate AF burden had the arrhythmia, and 21.5% of those had a burden greater than 1%.

Previously reported main results from the trial, which enrolled 2536 patients at risk for stroke with AF detected by implantable cardiac devices but no clinical diagnosis, found no significant benefit from edoxaban in preventing stroke, systemic embolism, or cardiovascular death. However, there was a higher risk for adverse events with edoxaban than seen with placebo. The ARTESIA trial subsequently showed that apixaban reduced risk for stroke or systemic embolism in a similar population, again with higher bleeding risk.

Together, the two trials suggested some effect is likely but is less than in clinically diagnosed AF. 

Substudies have tried to tease out patient characteristics that should sway clinical decision-making on anticoagulation. They have suggested greater likelihood of benefit from anticoagulation therapy in patients who had a CHA2DS2-VASC score of more than 4, prior stroke or vascular disease, but not longer or more frequent subclinical AF episodes. 

“What we learn overall here from these investigations is that if you look into a population with a very low burden, then naturally it is other factors that at least have the same contribution, if not even a stronger contribution, to the actual stroke risk,” Schotten said. 

Thus, it might not be surprising that a relatively higher ranking on that small scale of differences in AF burden was not associated with the treatment effect, he added. “And this may be relevant not only for device-detected AF. It may also turn out that in the postablation scenario that…this might also be relevant.”

The substudy also held value in pioneering methodology to estimate AF burden. 

The researchers used AI-based extraction of data from PDFs of 11,964 clinical routine pacemaker reports (an average of four per patient) and found that large language models could extract AF burden with 99.0% accuracy. And using mode switch as a surrogate for AF burden — after finding good agreement between the two — also had 98.6% accuracy compared with manual curation and visual inspection of the reports.

“Large language model-assisted extraction of AF burden from routine device reports can support individualized risk stratification beyond binary AF classifications,” Schotten said.

However, this method only successfully extracted AF burden from 48% of pacemaker reports, Wan said. She raised the question of how to increase the percentage of successful extraction.

“I think all of us now at HRS want to see how AI and large language models can integrate into our workflow flow,” Wan said. 

The study was partially funded by the German Ministry of Education and Research and Daiichi Sankyo Europe, and further supported by EU IHI network AF B STEP. Schotten disclosed relationships with the Dutch Heart Foundation, EU, Leduca Foundation, Roche, EP Solutions, Stanford University, Universita italiana svizzera, YourRhythmics BV, and Roche Diagnostics. Wan disclosed relationships with Boston Scientific and Abbott. 

Crystal Phend is an award-winning medical journalist with decades of experience reporting on clinical research and healthcare developments across specialties. When not walking the halls at a medical conference, she can be found at a keyboard in upstate New York.