Researchers say they hope that using AI to determine results will reduce the high cost and complexity of conducting large-scale clinical trials.
New data shows that an artificial intelligence (AI)-based model called Auto-MACE can determine serious adverse events in clinical trials, particularly CV death and stroke, on par with expert physicians.
The authors describe the findings presented this week at the American Heart Association’s 2025 Scientific Sessions. published online Jack, This suggests that this technology could potentially be used to streamline processes and save research funds in the future.
“By reducing the volume of cases that require human review, AI can alleviate key drivers of adjudication costs and schedule delays,” write Pablo M. Martí Castellote, Ph.D., Brigham and Women’s Hospital in Boston, Massachusetts, and colleagues. “Applying an AI-based adjudication model that is consistent across all events within a trial, and even across trials, may improve reproducibility compared to adjudication by a large number of reviewers with disparate experience.
“AI has the potential to not only replicate human judgment, but actually improve its consistency and efficiency, paving the way for better clinical trials in the future,” they added.
Commenting on the TCTMD findings, Alexandra Popma, MD, PhD, Cardiovascular Research Foundation, New York, NY, said this “excellent” study represents a formal step in the process of incorporating AI into clinical trials.
“The challenge is how do we translate this into products and deliverables that are acceptable to regulators,” she said. The open question remains: “How do we do that in a way that can be ethical?” [and] Does it meet all the standards of transparency and traceability?”
Automatic MACE findings
Researchers trained the Auto-MACE language model to determine cardiovascular death based on five large cardiovascular clinical trials (Invested, deliver, Paragon-HF, protectand inno two bait), Nonfatal MI is based on PARAGON-HF data, and stroke is based on PARAGON-HF, PRO2TECT, and INNO2VATE data.
Of 5,661 items paradise sumi participants In cases with myocardial infarction complicated by systolic dysfunction or pulmonary congestion, Auto-MACE confidently determined 69% of the chances of death, 46% of the chances of myocardial infarction, and 81% of the chances of stroke. This model agreed with Clinical Events Committee (CEC) adjudication in 97%, 89%, and 88% of these events, respectively.
Similar estimated reductions in combined MACE with sacubitril/valsartan and ramipril occurred for both Auto-MACE determination (HR 0.91; 95% CI 0.78-1.07) and CEC determination (HR 0.90; 95% CI 0.77-1.05).
“For CV deaths, Auto-MACE errors were rare and occurred mainly due to a combination of CV issues as well as infections, such as sepsis after lower extremity revascularization and unwitnessed deaths at home in the setting of suspected infection,” Marti-Castellote et al. wrote.
The errors in identifying MACE “were caused by the inability to extract troponin data from tables and checkbox forms and from misinterpretation of previous MIs (inclusion criteria for PARADISE-MI) as new MI events,” the researchers continued. “For stroke, most errors were cases where the model determined it was a stroke, but the CEC did not detect the event. In many cases, the model incorrectly interpreted a previous stroke or evidence of a previous stroke on brain images as a new stroke event.”
how to do it in an ethical way [and] Does it meet all standards of transparency and traceability? alexandra popma
Looking ahead to the technology’s pivotal Phase III trials, the authors state that “the most viable path forward is a hybrid deployment with careful human oversight of CEC. Early dialogue with regulators is critical to ensure acceptance of AI-generated endpoint data.”
Popma said he understands how certain people and institutions may be hesitant about this technology changing clinical trial workflows, especially since the processes have remained similar over the past several decades. But AI, which can be used in nearly every aspect of clinical trials from planning to execution to documentation, “really addresses a lot of the bottlenecks that we have,” she said. “Everyone complains about the cost of clinical trials. Everyone complains about how long and burdensome the process is.”
Improvements in data security, multilingual awareness and upstream process changes will become apparent over time, Popma said. “I’m not afraid of this,” she said. “I think this is a great challenge. We need to address it. We need to come up with solutions.”
