Artificial Intelligence Successfully Adjudicates Clinical Events in CV Trials

An artificial-intelligence (AI)-based model called Auto-MACE can adjudicate major adverse events in clinical trials, especially CV death and stroke, on par with expert physicians, according to new data.

The authors say their findings, presented this week at the American Heart Association 2025 Scientific Sessions and simultaneously published online in JACC, suggest that this technology could be used to streamline processes and save research dollars down the line.

“By decreasing the volume of cases requiring human review, AI can lessen a major driver of adjudication cost and timeline delays,” Pablo M. Marti-Castellote, PhD (Brigham and Women’s Hospital, Boston, MA), and colleagues write. “Applying a consistent AI-based adjudication model across all events within a trial and even between trials may improve reproducibility compared to adjudication by numerous reviewers with heterogeneous experience.

“AI has the potential to not only replicate human adjudication but actually improve upon its consistency and efficiency, paving the way for better clinical trials in the future,” they add.

Alexandra Popma, MD (Cardiovascular Research Foundation, New York, NY), who commented on the findings for TCTMD, said this “fantastic” study represents a formal step in the process of bringing AI into the fold of clinical trials.

“The challenge is how we are going to translate this into a product and an output that is going to be acceptable by regulatory agencies,” she said. The open question remains: “How are we going to do it in a way that can be ethical [and] that meets all the standards for transparency and traceability?”

Auto-MACE Findings

Researchers trained the Auto-MACE language model to adjudicate cardiovascular death based on five large cardiovascular clinical trials (INVESTED, DELIVER, PARAGON-HF, PRO2TECT, and INNO2VATE), nonfatal MI on data from PARAGON-HF, and stroke on data from PARAGON-HF, PRO2TECT, and INNO2VATE.

Among 5,661 PARADISE-MI participants with MI complicated by systolic dysfunction or pulmonary congestion, Auto-MACE confidently adjudicated 69% of deaths, 46% of potential MIs, and 81% of potential strokes. The model was in agreement with clinical event committee (CEC) adjudication in 97%, 89%, and 88% of these events, respectively.

Both Auto-MACE adjudication (HR 0.91; 95% CI 0.78-1.07) and CEC adjudication (HR 0.90; 95% CI 0.77-1.05) resulted in similar estimated lowering of composite MACE with sacubitril/valsartan versus ramipril.

“For CV death, Auto-MACE errors were rare and occurred primarily due to presentations involving a combination of CV issues but also infection, such as sepsis following lower-extremity revascularization or unwitnessed death at home in the setting of suspected infection,” Marti-Castellote and colleagues write.

Errors identifying MACE “were driven by inability to extract troponin data from tables and checkbox forms, as well as misinterpreting a previous MI (an inclusion criterion in PARADISE-MI) as a new MI event,” they continue. “For stroke, most errors were cases in which the model adjudicated stroke but the CEC found no event; oftentimes, the model misinterpreted previous stroke or evidence of previous stroke on brain imaging as a new stroke event.”

How are we going to do it in a way that can be ethical [and] that meets all the standards for transparency and traceability? Alexandra Popma

Looking to phase III pivotal trials of this technology, the authors say “the most pragmatic path is hybrid deployment together with careful human CEC oversight. Early dialogue with regulatory agencies will be crucial to ensure acceptance of AI-generated endpoint data.”

Popma said she understands how certain people or institutions might be hesitant about this technology changing the workflow of clinical trials, especially as processes have remained similar over the past several decades. But AI, which can be used in almost every facet of clinical trials from planning to execution and documentation, “really addresses a lot of the bottlenecks that we have,” she noted. “Everybody complains about the cost of clinical trials. Everybody complains about how consuming and how much a burden of the process is.”

Refinements in data security, multilanguage recognition, and upstream process changes will be figured out with time, according to Popma. “I'm not afraid of this,” she said. “I think that is a fantastic challenge. We need to address it. We need to come up with solutions.”