Reclassifying Ablation Energy Risk Profiles in Atrial Fibrillation: A Narrative-Adjudicated Comparative Analysis of the MAUDE Database

Background: Pulsed field ablation (PFA) has been rapidly adopted for atrial fibrillation (AF) on the premise of greater safety than thermal ablation. Real-world signals of hemolysis, coronary spasm, and neurologic events have since emerged, and prior Manufacturer and User Facility Device Experience (MAUDE) analyses relied on term matching without prospective signal detection. Methods: We queried MAUDE for AF ablation catheter reports received January 2016 through June 2026 and assigned each to PFA, radiofrequency ablation (RFA), or cryoballoon ablation. We adjudicated injury and death narratives into predefined complication families and severity tiers with a large language model, checked against a larger model and by physician review. We calculated reporting odds ratios (ROR) and Bayesian shrinkage estimators compared against an empirical null, applied interrupted time series and prospective sequential detection, and repeated estimates on a 2026 hold-out. Results: Among 44,623 reports, 30.1% were PFA, 50.9% RFA, and 19.0% cryoballoon, with 91.8% model agreement on event occurrence (? = 0.83). PFA carried an ROR of 527 (95% CI, 33, 8,509) for hemolysis, 2.83 (2.55, 3.16) for arrhythmia, 2.83 (2.30, 3.48) for coronary, and 2.39 (2.10, 2.73) for cerebrovascular events, against 0.024 (0.008, 0.075) for esophageal and 0.34 (0.32, 0.37) for pericardial events. The cerebrovascular excess was largest for VARIPULSE, but without it the class proportional reporting ratio was 1.36 (1.13, 1.65), concordant with the 1.56 measured in EMBOL-AF. In the hold-out, the cerebrovascular, arrhythmia and hemolysis signals replicated, and coronary events kept their direction without meeting the signal rule. Prospective detection flagged cerebrovascular events 13 months before AVANT GUARD published its neurologic events. Conclusions: PFA showed disproportionate reporting of hemolysis, cerebrovascular, coronary, and arrhythmic events across sequential testing. Machine learning methods allowed for flagging these signals months before confirmatory trials, demonstrating a scalable tool for early detection of procedural risk.