TrialsgrambetaRCTs are beautiful

O'Sullivan LLM (cardiology)

2026

A Large Language Model for Complex Cardiology Care

In 9 general cardiologists assessing complex suspected genetic cardiomyopathy, an LLM assistant alongside the raw investigations improved blinded subspecialist-rated quality of triage, diagnosis and management compared with cardiologists working unassisted.

Result

Blinded subspecialist preference across a ten-domain rubric of triage, diagnosis and management quality
Absolute difference
14.0 points
NNT
8

Absolute difference and NNT are derived from the event rates.

Also reported: Clinically significant errors 13.1% vs 24.3% (P=0.033); missing content 17.8% vs 37.4% (P=0.0021).

Original publication
Subtopic
Physician decision-making
Design
RCT, cardiologists randomised, retrospective real-world case data, blinded subspecialist rating
Relevance
Important: one of several pillars.
Source
O'Sullivan JW, et al. Nat Med. 2026;32(2):616-623.doi 10.1038/s41591-025-04190-9

Discussion & critique

Extends the AMIE programme from simulated primary care (AMIE) into a subspecialty where expertise is scarce, with the LLM reading raw diagnostic data rather than text alone. Only nine cardiologists, retrospective cases from a single subspecialist practice, preference-based rubric scoring by three raters, and no patient outcomes: a signal for augmentation, not evidence of benefit.