An AI Predicted This Patient’s Heart Attack Six Hours Before It Happened. The Cardiologist Didn’t Believe It at First.

In medicine, there’s a certain type of situation that doesn’t get discussed enough: when a machine flags something that a doctor didn’t see and the doctor has to make a real-time decision about whether to trust it. It’s not as dramatic as medicine on television. Compared to that, it is quieter. There’s a notification. A number is not within the anticipated range. And a clinician must balance that warning against everything their own training instructs them to look for, having been trained through years of education and pattern recognition developed from thousands of patient encounters. They take action on it occasionally. They don’t all the time.

AI Heart Attack Prediction — Key Facts & Reference

That moment is coming more often and with greater stakes in cardiology in particular. An AI model that can predict sudden cardiac death in patients with hypertrophic cardiomyopathy, a common inherited heart disease that affects approximately one in every 200 to 500 people worldwide, with 89 percent accuracy overall and 93 percent accuracy for patients between the ages of 40 and 60, was described by Johns Hopkins University researchers in July 2025 in Nature Cardiovascular Research. Hypertrophic cardiomyopathy is one of the main causes of sudden cardiac death in that age group. The comparison point is unsettling because the current clinical guidelines used by physicians in the US and Europe to determine which patients are at the highest risk are only roughly half as accurate. The work’s principal investigator, Natalia Trayanova, a professor of biomedical engineering at Johns Hopkins, put it bluntly: “not much better than throwing dice.”

The system, known as MAARS (Multimodal AI for Ventricular Arrhythmia Risk Stratification), examines a patient’s complete medical record in addition to contrast-enhanced cardiac MRI images. What sets this apart from previous methods is the MRI component. The development of fibrosis, or scarring, throughout the heart muscle in patients with hypertrophic cardiomyopathy increases the risk of a deadly arrhythmia. Physicians could view the raw MRI images, but visual review could not consistently interpret the specific patterns of scarring in a way that linked to mortality risk. MAARS discovered the signal that human readers were missing by using those images to train deep learning models. “We are able to extract this hidden information in the images that is not usually accounted for,” Trayanova said. Instead of just giving cardiologists a binary decision, the AI could also explain why a patient was deemed high risk, providing them with a foundation for customizing treatment.

The consequences go far beyond a single type of illness. A different study, the PECTUS-AI study, which was published in the European Heart Journal in September 2025, examined the effects of applying AI to coronary artery optical coherence tomography (OCT) imaging in patients who had already survived a heart attack. Which of the arteries’ remaining unstable plaques are most likely to burst and result in another incident? Manual OCT image review is slow, inconsistent between readers, and infrequently done over the whole vessel. The AI automatically processed complete pullbacks, mapping ten different tissue types, such as calcium, thrombus, and the media layer of the arterial wall. The two-year rate of death, non-fatal heart attack, or unscheduled revascularization was 12.3% when the AI detected high-risk thin-cap plaques anywhere throughout the imaged vessel segment. When no such plaques were discovered, the rate was 2.4 percent, with a negative predictive value of 97.6 percent and a hazard ratio of 5.50. Practically speaking, the AI was incorrect about the prognosis only 2.4% of the time when it stated that the patient’s arteries were free of high-risk plaques. That’s the kind of figure that begins to influence decisions about discharge.

Then there is the category of AI tools known as earlier-warning systems, which are made to detect risk weeks or even years before a cardiac event occurs. Researchers have shown that AI analysis of standard ECGs can 70% accurately predict dangerous arrhythmias up to two weeks in advance. An alternative method uses artificial intelligence (AI) to examine the fat tissue surrounding coronary arteries on CT scans, searching for inflammatory patterns that are associated with the risk of a heart attack. This method is said to be able to identify high-risk patients up to ten years prior to the event. Additionally, wearable sensor patches and machine learning algorithms that analyze continuous physiological data have been used at various centers to predict heart failure rehospitalization approximately one week in advance. In a pilot study at VA cardiac care facilities in Salt Lake City and Gainesville, 95% of clinicians responded within 24 hours after receiving AI-generated alerts.

The research recognizes the existence of the trust issue. Clinicians’ reactions to AI alerts were found to be highly influenced by their perceptions of the reliability of the data, their professional autonomy, and the cognitive burden of responding to automated notifications, according to a study looking at the integration of AI analytics into heart failure clinics. In the trial, 26% of AI alerts led to clinical action. This indicates that 74% did not—not necessarily because the alerts were inaccurate, but rather because the system for fostering clinician trust in the forecasts had not kept up with the advancements in technology.

As this field advances, there’s a sense that cardiology is getting close to a real turning point—not between human and machine judgment, but between reactive and predictive medicine. According to the conventional approach, a heart attack serves as the catalyst for intervention. The new one views the heart attack as a failure of past chances. The current AI tools are made to identify those opportunities. A different question that may ultimately determine how many of those opportunities are actually taken advantage of is how quickly clinicians, hospitals, and healthcare systems learn to use them.