AI Is Becoming a Cardiologist's Co-Pilot, But Who Teaches It Clinical Judgment?

Artificial intelligence is no longer waiting at the door of medicine. It is already here. In cardiology, AI can analyse ECGs, interpret images, identify patterns, estimate risk, and support clinical decisions. But as its capabilities grow, an important question remains: are we teaching AI medicine, or are we teaching it clinical judgment?

Knowing medicine is not the same as practising it. A cardiologist does more than recognise patterns. We interpret those patterns in context, weigh uncertainty, question unexpected findings, and decide what matters for the individual patient. AI can tell us what is statistically likely. The clinician must still decide what is clinically appropriate. This becomes particularly important as AI systems produce increasingly convincing answers. Fluency is not understanding, and confidence is not accuracy. A beautifully structured recommendation can still be wrong when the clinical context is misunderstood.

Consider a scenario many of us would recognise. An AI-enabled ECG algorithm flags a patient's tracing as showing a pattern strongly associated with a specific arrhythmia risk, and confidently recommends a corresponding treatment pathway. On the surface, the recommendation looks reasonable and well-supported. What the algorithm does not know is that this patient had a similar tracing five years ago that was fully investigated and explained by a benign, patient-specific conduction variant something documented only in a prior discharge summary the model never had access to.

It does not know if the patient stopped taking medication two days earlier because of a side effect, or if a family member mentioned a symptom in passing that reframes the whole picture. None of this contradicts the AI's output on the data it was given. It simply was not given the whole patient. A clinician who defers to the recommendation without asking what might be missing could set in motion further tests, treatment changes, or unnecessary anxiety for the patient, all built on an interpretation that was statistically sound but clinically incomplete.

This is not a flaw unique to any one tool. It is a structural feature of how these systems work: they reason from the data in front of them, not from the full arc of a patient's history, context, and lived circumstances. That gap is precisely where clinical judgment has to step in. Perhaps this is where medical education has something important to contribute.

We have spent decades teaching clinicians not only what to know, but how to reason through cases, uncertainty, reflection, discussion, and experience. If AI is going to become part of clinical practice, perhaps we should ask whether we are teaching it in a similar way.

Future physicians will need more than the ability to use AI. They will need the ability to interrogate it, and to build that interrogation into habit rather than treating it as an occasional exercise.

Before acting on an AI-generated recommendation, a clinician should be able to answer a short set of questions almost reflexively:

• What evidence supports this recommendation, and is that evidence relevant to this specific patient, or only to the population the model was trained on?
• What assumptions are being made: about data quality, about the completeness of the clinical picture, about a "typical" presentation that this patient may not fit?
• What information is missing that a human clinician would normally gather through history-taking, examination, or simply noticing something that does not add up?
• What happens if the recommendation is wrong: is this a low-stakes suggestion, or one that could set off a chain of consequential decisions?
• And perhaps the most important question: Would I make the same decision without the AI?

If the answer is no, we may not have created a co-pilot. We may have created dependence. Teaching this kind of structured scepticism should sit alongside teaching clinicians how to use these tools in the first place. Not as an afterthought, but as a core clinical skill in its own right.

I remain optimistic about AI in cardiology. But I believe its greatest value will not be in replacing clinical judgment. It will be in strengthening it. The future should not be autonomous medicine, but augmented medicine: technology that helps clinicians see more, process more, and think more clearly, while leaving responsibility where it belongs: with the clinician.

The goal should not be to create an AI that behaves like a cardiologist. It should be to create AI that helps cardiologists become better cardiologists. Because ultimately, the question is not how intelligent our machines become. It is whether we remain wise enough to know when to trust them.

About the Author

References

American Heart Association Scientific Statement, Use of Artificial Intelligence in Improving Outcomes in Heart Disease, Circulation, 2024. https://www.ahajournals.org/doi/10.1161/CIR.0000000000001201

American Heart Association Scientific Statement, Value Creation Through Artificial Intelligence and Cardiovascular Imaging, Circulation, January 2024. https://www.ahajournals.org/doi/10.1161/CIR.0000000000001202