In a cardiology unit in south London, a machine now reads every electrocardiogram before a human ever sees it. It flags the traces that matter, ranks them by urgency, and — increasingly — catches the subtle signatures of disease that the human eye was never trained to notice. The cardiologists have not been replaced. They have been reordered around the machine.
This is the quiet arrival of artificial intelligence in clinical cardiology: not a dramatic rupture, but a steady rearrangement of who looks at what, and when. The models at the centre of it were trained on datasets of a scale no individual clinician could ever absorb — tens of millions of annotated heartbeats, each one a small lesson in the difference between the ordinary and the ominous.
The Promise
The case for these systems is, on its face, overwhelming. In several large validation studies, algorithms have matched or exceeded specialist performance in detecting conditions like atrial fibrillation and left-ventricular dysfunction — sometimes from readings that appeared, to the human eye, entirely normal.
The appeal is not only accuracy but reach. An algorithm can, in principle, extend specialist-grade screening into places where a cardiologist has never set foot — rural clinics, pharmacies, the sensor on a wrist. Early pilots in Lagos and Nairobi have begun testing exactly this proposition.
"The question was never whether the model could read the heart. It was whether we could trust the conditions under which it learned."
The Data Problem
Every one of these systems is a reflection of the data that shaped it — and that data is far from neutral. Models trained overwhelmingly on one population can falter, quietly, when applied to another. A signature of disease learned in Boston may not generalise to Lagos, and the failure may be invisible until it is counted in outcomes.
- Training cohorts remain skewed toward high-income, well-instrumented health systems.
- Consent for the reuse of clinical data is frequently ambiguous or absent.
- The commercial value of these datasets is now considerable — and contested.
Which raises the question that follows every one of these tools into the clinic: who owns the heartbeat? The patient who produced it, the hospital that recorded it, or the company that turned ten million of them into a product?
Who Bears the Risk
When an algorithm misses a diagnosis, the lines of responsibility blur in ways medicine has not yet resolved. The clinician who deferred to the model, the institution that deployed it, and the developer who built it each hold a portion of the liability — and none holds all of it.
For now, the consensus among the researchers we spoke to is cautious optimism, heavily qualified. The technology is real, its gains are measurable, and its risks are not hypothetical. What remains unsettled is the governance around it — the standards, the consent, the accountability — and that, they argue, is now the more urgent frontier.
This article is part of The Node Journal's continuing series on artificial intelligence in clinical medicine. Reporting was supported by the Node Editorial Fund.