TrialsgrambetaRCTs are beautiful

DeepRhythmAI

2025

Artificial Intelligence for Direct-to-Physician Reporting of Ambulatory Electrocardiography

In 14,606 ambulatory ECG recordings annotated beat by beat, AI interpretation reported directly to the physician improved sensitivity for critical arrhythmias against cardiologist consensus panels compared with certified ECG technicians.

Result

Efficacy: Sensitivity for critical arrhythmias versus cardiologist consensus-panel annotationrelative risk of a missed diagnosis 14.1 (95% CI 10.4-19.0)

Below 1 favours certified ECG technicians, above 1 favours AI interpretation reported directly to the physician.

Event rates

The estimate refers to a different reading of the endpoint than the rates.

Also reported: False positives 12 vs 5 per 1,000 patient-days.

Original publication
Subtopic
Atrial fibrillation
Design
Non-randomised diagnostic-accuracy comparison against 17 cardiologist consensus panels
Relevance
Supporting: context rather than must-know.
Source
Johnson LS, et al. Nat Med. 2025;31(3):925-931.doi 10.1038/s41591-025-03516-x

Discussion & critique

The largest head-to-head of AI against human technicians for ambulatory ECG, supporting AI-first reporting with physician review. It is an accuracy study, not an outcome trial: the reference standard was a cardiologist-adjudicated random sample of 5,235 events rather than every recording, the comparison was not randomised, and the price of the very high sensitivity was a higher false-positive burden that physicians must absorb.

Discussed in podcasts