How do artificial intelligence tools compare to doctors for predicting heart attack biomarkers?

Artificial intelligence (AI) tools are being developed to help predict heart attack biomarkers, such as troponin levels, and to diagnose myocardial infarction (MI). Current evidence suggests that AI models can perform as well as or slightly better than traditional methods used by doctors, but most studies have not been tested in real-world clinical settings. This means AI is promising but not yet ready to replace physician judgment.

What the research says

A 2021 study compared an artificial neural network (ANN) to standard guideline-recommended algorithms and logistic regression for predicting acute MI using high-sensitivity cardiac troponin T levels. The ANN and logistic regression had similar accuracy (about 95% area under the curve), but the ANN reduced the number of patients in the 'intermediate risk' group by about 9%, meaning fewer patients needed additional testing ⁹. This suggests AI can help refine risk classification.

A systematic review of 120 studies on AI/ML methods for MI biomarker prediction found that most studies (74%) focused on prediction or prognosis, and the most common techniques were logistic regression (63%) and random forest (58%). However, only 37% of studies used independent external validation, and calibration and decision-curve analyses were rarely reported ⁴. This means many AI models may not perform as well when applied to new patient populations.

Another study tested an AI system called Auto-MACE to automatically adjudicate major adverse cardiovascular events (including MI) in a large trial. Auto-MACE agreed with a physician committee in 97% of deaths, 89% of potential MIs, and 88% of strokes when it was confident. However, it could only give a confident answer for 46% of potential MIs ⁶. So AI can be accurate but often needs human backup.

Overall, AI tools show promise for predicting heart attack biomarkers and may reduce uncertainty, but the evidence is limited by a lack of real-world testing and inconsistent reporting of model performance ⁴⁶⁹.

What to ask your doctor

• Are AI-based tools used at this hospital to help diagnose heart attacks?
• How do AI predictions compare with standard troponin testing in your experience?
• Should I be concerned about AI missing a heart attack diagnosis?
• What happens if the AI result is uncertain — does a doctor still review my case?
• Are there any ongoing studies at this hospital testing AI for heart attack detection?