Artificial intelligence arrived in cardiology before most of us were asked whether we wanted it. It is already reading electrocardiograms in emergency departments, quantifying strain on echocardiographic loops, scoring coronary plaque on CT, and stratifying risk from registry data that no individual clinician could hold in their head. The question is no longer whether it belongs in practice. It is which parts of it have earned the right to influence a decision, and which are a logistic regression with a marketing budget.
Where the evidence is genuinely strong
Three areas have moved past proof of concept. The first is ECG interpretation. Convolutional networks trained on millions of tracings now detect left ventricular systolic dysfunction from a twelve-lead ECG with an area under the curve that a cardiologist reading the same tracing cannot match, because the signal they are using is not one the human eye was ever taught to look for. The same architecture has been applied to occult atrial fibrillation and to structural disease.
The second is echocardiographic quantification. Automated border detection and strain analysis reduce the interobserver variability that has always been echocardiography’s weakest point. When two sonographers report ejection fractions eight points apart, the problem is not the machine.
The third is triage of imaging volume. Algorithms that flag the studies most likely to be abnormal do not have to be better than a radiologist to be useful; they only have to reorder the worklist.
Where it is weaker than it looks
Almost every published model performs worse outside the population that trained it. A risk score built on a North American registry applied to a South Asian cohort will misestimate, because the underlying prevalence, the age of onset and the phenotype are different. Rheumatic valve disease is common in the wards I work on and effectively absent from most training sets. This is not a technical detail; it is the single most important question to ask of any tool offered to you.
Prospective, randomised evidence that an algorithm changes an outcome, rather than a metric, remains scarce. An AUC of 0.94 in a retrospective cohort tells you the model separates two groups in data that has already happened. It does not tell you that acting on it earlier helps anyone.
What the doctor remains accountable for
When you countersign an automated report, you own it. That is not a philosophical position, it is a medicolegal one, and it will not be softened by the observation that the software was confident. Three practical rules follow.
Know the training population. Ask what the model was built on and whether your patient would have been eligible. Know the failure mode. Every model fails in a characteristic direction; a system trained mostly on sinus rhythm will behave unpredictably in atrial fibrillation. Read the tracing anyway. The moment automated interpretation becomes the primary read rather than a second opinion, the skill that lets you catch its error begins to decay.
Where this is going
The near-term gains are unglamorous and real: ambient documentation that returns minutes to a consultation, worklist triage, automated quantification that removes variability, and retrieval tools that surface the relevant guideline rather than requiring you to remember which society updated what. The far-term claims — autonomous diagnosis, treatment selection without a clinician — remain claims.
The clinicians who will use these tools best are the ones who read an ECG well without them. If you want to strengthen that foundation, start with the systematic ECG reading sequence and the echocardiography normal values, or take a structured programme from the Academy course catalogue.
Further reading across the Academy: ECG interpretation, echocardiography, interventional cardiology, and the full eBook library.