Arrhythmias & electrophysiology

Local Conformal Auditing Exposes and Restores Reliability of AI-ECG Atrial Fibrillation Detection Under External Deployment Shift

Heart RhythmPublished four-class 12-lead rhythm classifier trained on PTB-XL (Germany; 7,831 ECGs, 7,164 patients) was evaluated unchanged on Chinese (Chapman-Shaoxing; n=2,289) and US (Georgia; n=1,990) cohorts

Authors' conclusion

AI-ECG uncertainty can fail silently despite high AUC. A local conformal audit detects when uncertainty estimates are untrustworthy and recalibrates the uncertainty layer without changing the classifier, converting trust into a measurable, subgroup-aware, task-specific guarantee.

Study context

Population
12-lead ECGs from a German training source (PTB-XL) and external Chinese (Chapman-Shaoxing) and US (Georgia) cohortsECGs per cohort · PTB-XL (Germany; training, ECGs) 7,831 · PTB-XL (Germany; patients) 7,164 · Chinese (Chapman-Shaoxing) 2,289 · US (Georgia) 1,990
Target condition
atrial fibrillation (AF)
Methodological detailsexternal deployment shift, 90%-target split-conformal coverage, local labeled audit without retraining
  • External deployment shift
  • 90%-target split-conformal coverage
  • Local labeled audit without retraining

Performance

As reported in the abstract

Test or modelAtrial fibrillation (AF) AUC
China (Chapman-Shaoxing)0.998 (95% CI 0.997-0.999)
US (Georgia)0.936 (0.919-0.952)

Atrial fibrillation (AF) AUC

China (Chapman-Shaoxing)0.998 (95% CI 0.997-0.999)
US (Georgia)0.936 (0.919-0.952)

Further findings

OutcomeChinese menChinese womenDifference (95% CI)
Coverage in Chinese cohort by sex (subgroup)Coverage lower in men0.7070.8220.115 (0.081–0.148)

Expected calibration error (ECE)

ECE, source
0.051
ECE, external cohorts
0.197 and 0.237

Source-calibrated conformal coverage

Coverage, source
0.901
Coverage, external cohorts
0.758 and 0.722
  • Both P<0.001

Label-free shift corrections

Label-free corrections failed

Coverage after local recalibration

Mean coverage with 50 labeled ECGs
0.90
95% simulation interval, 50 ECGs
0.82-0.97
95% simulation interval, 200 ECGs
0.86-0.94

AF auto-clear with local risk control

Held-out ECGs certified for auto-clearing, China
68%
US cohort: abstained

Original publication

Local Conformal Auditing Exposes and Restores Reliability of AI-ECG Atrial Fibrillation Detection Under External Deployment Shift. Heart Rhythm. Published 4 November 2026. doi:10.1016/j.hrthm.2026.10.064

Composed by AI from the published abstract and checked against it. How Briefs are made →

Get them every week

New studies on this topic in My Week and in your Monday e-mail.

We e-mail you a sign-in link and a code. No password, no spam.