Using AI to aid in heart attack diagnosis in the emergency room does not improve patient outcomes, research shows

The use of artificial intelligence (AI) to support clinical decisions in the detection and treatment of myocardial infarction (MI; heart attack) in patients presenting to the emergency department with suspected heart disease does not improve cardiovascular outcomes. However, this AI-based clinical decision support is safe and promotes the acceptance of evidence-based care, according to new research presented in a hotline session at this year's ESC Congress 2024 in London, UK (30 August – 2 September).

“A key promise of AI in healthcare is that it will help healthcare professionals diagnose patients more quickly and accurately and quantify prognosis more objectively, ultimately enabling them to start appropriate treatment earlier to optimise patient outcomes,” said lead author Kristina Lambrakis from the Victorian Heart Hospital, Monash Health, the Victorian Heart Institute, Monash University and the College of Medicine and Public Health at Flinders University Australia.

She added: “Our large cluster-randomised trial of patients presenting to emergency departments across South Australia with suspected cardiac disease did not improve clinical outcomes, but highlighted the ability of real-time AI to influence clinical decisions and practice towards evidence-based care. Improving clinical outcomes will likely require greater adoption of AI insights and greater integration of AI insights into clinical workflows.”

The widespread use of high-sensitivity troponin (hs-cTn) testing, a biomarker detected in the blood that indicates myocardial injury, has improved the ability to identify individuals with even very mild myocardial injury. However, these more sensitive tests have also made troponin results more difficult to interpret, as elevated results can now have a variety of causes other than type 1 MI (heart attack due to coronary plaque rupture), with misinterpretation potentially leading to unnecessary investigations, risks, and costs.

In addition, these complex diagnostic decisions often have to be made by emergency department physicians who do not have advanced cardiology training and who work with patients who are difficult to distinguish from one another, making it even more difficult for them to distinguish between the different types of myocardial injury when interpreting hs-cTn results.

To help physicians make the correct diagnosis, the current 2018 Fourth Universal Definition of Myocardial Infarction (UDMI) guideline aims to aid the interpretation of hs-cTn by outlining MI and myocardial injury classifications that require specific treatments.

The 4th UDMI defines myocardial injury as an elevation of troponin with at least one value above the upper reference value of the 99th percentile. The difference in troponin values ​​between repeated sampling results defines whether the injury is acute or chronic. Myocardial injury can be caused by numerous situations, including sepsis, kidney disease, heart failure, and strenuous exercise.

To diagnose a myocardial infarction (MI), the patient must have acute myocardial injury as well as signs of acute myocardial ischemia—a lack of oxygen to the heart muscle detected by an ECG, symptoms such as pain in the chest, arms, or jaw, or evidence from imaging studies.

It is not known whether clinical use of the 4th UDMI classification system has resulted in improved patient outcomes. The use of AI algorithms based on the 4th UDMI to support clinical decisions could help in the interpretation of hs-cTn and improve outcomes.

To address this important gap in the evidence, the RAPIDx AI trial enrolled 14,131 patients (aged 18 years and over) presenting to the emergency department between 15 April and 31 December 2023 and receiving hs-cTn testing in six urban and six rural emergency departments across South Australia. Six hospitals were randomly assigned to the control group (ie, unchanged standard of practice) and six to the intervention group (ie, implementation of AI-based clinical decision support).

A total of 3,029 patients (mean age 74.5 years, 58% female) with myocardial injury (elevated troponin based on hs-cTn) and suspected cardiac cause were included in the primary intention-to-treat analysis (1,461 patients presenting to control hospitals and 1,568 to hospitals with AI decision support).

The main outcome measures were cardiovascular (CV) death, new/recurrent myocardial infarction, and unplanned cardiovascular readmissions within 6 months. The safety endpoint was death from any cause or myocardial infarction within 30 days in patients discharged directly from the emergency department.

In the intention-to-treat analysis, the researchers found no differences in outcomes between groups. 408/1568 participants (26.0%) in the intervention group and 385/1461 (26.4%) in the control group experienced cardiovascular death, myocardial infarction, or unplanned cardiovascular readmission within 6 months.

Likewise, in the entire cohort, the primary endpoint was achieved in 680 (9.4%) intervention group participants and 718 (10.4%) control participants.

Importantly, in patients in whom AI-guided decision support did not diagnose type 1 myocardial infarction (i.e., in whom coronary angiography was not found to be useful), the likelihood of undergoing invasive coronary angiography was 47% lower in the intervention group than in the usual care group (66/1,265; 5% vs. 112/1,186; 9.4%).

In addition, patients classified as having a type 1 heart attack by AI-based decision support were more likely to be prescribed statins (82% [248/303] compared to 68% [187/275]); they are more likely to receive antiplatelet therapy (56 % [248/303] compared to 44% [120/275]); and they are more likely to be prescribed a mineralocorticoid inhibitor (56 %[78/303] compared to 44% [48/275]) compared to the group receiving usual care.

Among patients discharged directly from the emergency department, those treated in hospitals with decision support were no more likely to die or have a heart attack within 30 days than those in hospitals receiving usual care (0.86% [40/4664] compared to 1.1% [48/4350] Non-inferiority p

“We found no increase in early hazards or negative impacts on emergency department discharge decisions, demonstrating the safety of AI-based clinical decision support,” said lead author Professor Derek Chew, from the Victorian Heart Hospital, Monash Health, Victorian Heart Institute, Monash University and College of Medicine and Public Health, Flinders University, Australia.

“Our next steps include exploring approaches to increase the trust and acceptance of AI-based clinical decision support in the clinical community, investigating new models of care into which such AI-based decision support tools could be integrated to increase the effectiveness and efficiency of the healthcare system, and evaluating AI-based decision support for other acute cardiac conditions where early detection is a key challenge and driver for optimizing outcomes.”