Artificial Intelligence on the Twelve-Lead Electrocardiogram: What It Detects, What It Cannot, and Who Is Accountable

Abstract

A deep-learning model trained on the standard twelve-lead electrocardiogram (ECG) can now detect a left ventricular
ejection fraction (EF) of 35% or below directly from the waveform, at an area under the receiver-operating-characteristic
curve (AUC) of 0.93 in an independent test cohort of 52,870 patients, with sensitivity 86.3% and specificity 85.7%. In a
subsequent bias assessment, that accuracy held up across every racial and ethnic subgroup tested (AUC 0.930-0.961), even
though a separate network could distinguish Black from white patients from ECG features alone -- a reassuring finding on a
question the field was right to ask. When deployed in a pragmatic, cluster-randomised trial across 45 primary-care clinics (the
EAGLE trial), the algorithm increased low-EF diagnosis rates from 1.6% to 2.1% overall, and from 14.5% to 19.5% among
patients it flagged as high-risk -- a rare instance of an ECG-based artificial intelligence (AI) tool demonstrating a real-world
clinical benefit under randomised conditions rather than only strong retrospective accuracy. Nearly 90 FDA-authorised
AI/ML cardiovascular devices now exist, out of 882 across all specialties. This review sets out that evidence and closes with
the complication its subtitle names directly: the legal and regulatory frameworks that determine who is accountable when an
AI-ECG interpretation is wrong have not kept pace with the technology's clinical deployment, and current liability law
concentrates responsibility on the treating physician largely regardless of whether the algorithm's error was reasonably
detectable at the bedside.