An artificial intelligence (AI) system analyzed full electronic medical records (EMRs), accurately identifying patients eligible for clinical trials for a rare heart condition, and potentially streamlining one of the most labor-intensive steps in clinical research.
In the study, published in the Journal of Cardiac Failure, researchers deployed an AI-enabled platform within a large health system’s secure network to automate chart review for a phase 3 trial for transthyretin amyloid cardiomyopathy. The system combined structured data, such as diagnoses and prescriptions, with unstructured clinical notes to evaluate complex eligibility criteria across entire patient records.
Manual chart review remains a major bottleneck in clinical research because patient data are spread across hundreds of documents and data fields. By contrast, the AI system processed 1,476 patient records with amyloid-related diagnostic codes in six days, automatically assessing dozens of inclusion and exclusion criteria drawn from the trial protocol.
When evaluated against physician review, the large language model component answered trial-relevant questions with 96.2 percent accuracy, correctly resolving 7,409 of 7,700 criteria assessments.
The system also demonstrated strong performance in identifying potentially eligible patients. Among 46 candidates flagged as possible matches, 93 percent were categorized correctly before investigator review. After clinician validation, 30 patients were considered appropriate for recruitment.
Notably, of those 30 AI-identified patients, 37 percent were Black, compared with 7 percent of those identified manually. Additionally, while 93 percent of patients identified by traditional methods were already registered with a cardiologist, only 60 percent of the AI-identified patients were linked. These figures suggest that AI might expand trial participation among under-represented populations.
Equally important for clinical research workflows, the platform generated traceable explanations for each decision. Investigators could view supporting evidence drawn from both structured EMR fields and narrative clinical documentation, enabling rapid verification without extensive manual chart searching.
The researchers also assessed the system’s ability to exclude ineligible patients. In a physician review of 200 rejected cases, the AI’s decisions proved correct in 99 percent of instances, indicating strong reliability in filtering out unsuitable candidates.
Notably, most eligible patients identified by the AI system had not been detected through standard screening processes. Of the 30 eligible individuals identified during one week of AI-assisted review, nearly all had been missed previously, suggesting automated approaches may broaden recruitment and reduce bias in clinical trials.
Although further evaluation is needed to compare AI-assisted screening directly with conventional recruitment methods, the findings suggest that integrated AI systems could substantially reduce the workload associated with chart review while improving the efficiency of patient identification for complex clinical studies.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
