The latest science coming out of ACC.25, held March 29-31 in Chicago, IL, illuminated new approaches in how artificial intelligence (AI) can be used to improve the efficiency of cardiovascular care, expand the reach into underserved communities and more.

Studies examined the latest trends in application such as AI assistance for risk stratification in the emergency department (ED), clinical trial enrollment, ultrasound capture and echocardiogram analysis.

Deep Learning ECG Model For ED Risk Stratification

A late-breaking clinical trial presented Saturday and simultaneously published in the European Heart Journal delved into the use of a deep learning electrocardiogram (ECG) model in the ED for risk stratification on coronary revascularization need.

This study was the first of its kind to be designed for broad applicability in a general EP population. Antonius Büscher, MD, et al., used data from 144,691 ED visits (median age 60 years, 53% women, 0.6% revascularization) from a single U.S. center to train a convolutional neural network model in categorizing risk into three groups: low, immediate and high.

The model was then tested on an additional cohort of 35,995 visits and compared against clinician ECG interpretation and cardiac troponin T (TnT), where it performed better than clinicians and similarly to TnT testing.

In an external validation performed on a data set of 18,673 ED visits in Europe (median age 55 years, 49% women, 1.5% revascularization, 1% with type 1 myocardial infarction [MI]), the model outperformed both clinician interpretation and cardiac TnT testing, achieving an area under the receiver operating characteristic curve of 0.91 (95% CI, 0.91-0.91).

In determining both revascularization and type 1 MI, the model continued to outperform clinicians and was competitive with high-sensitivity TnT with a higher specificity but lower sensitivity, indicating what investigators called "complementary diagnostic information."

"A common goal with these types of models is to make the most use of the data that we have, and to complement physicians’ clinical reasoning by picking up some nuances in the data that might be obscure to the human eye," Büscher said. "I think we are still at the beginning of this field, but in the not too far future, these types of models will be used as routine diagnostic tests."

Improving Access to Clinical Trial Participation

Another late-breaking clinical trial, presented on Sunday, shared how implementation of digital health technologies can help reach patients blocked from clinical trial participation due to either lack of involvement by community physicians and infrastructure or challenges in participant recruitment and retention.

With the Health360x™ Registry, a clinical research platform service that integrates social determinants of health data and electronic health records with AI and machine learning, barriers at the patient, provider and practice levels were targeted to improve clinical trial recruitment and retention in predominantly African American, Latinx and rural communities.

The registry currently has 11,374 participants (median age 68 years, 85% Black, 8% White, 5% Hispanic/Latinx, 0.5% Asian, 59% women) enrolled in seven studies – five sponsored by the National Institutes of Health, one sponsored by industry, and one quality improvement project – with 100% screening success.

Elizabeth O. Ofili, MD, MPH, FACC, et al., said the key to success was in "meeting sites where they are," and measuring site compliance, H360x.ai Chatbot usage and patient satisfaction.

Future steps will focus on the automation of AI and machine learning support for trial management, including site initiation and training to reduce burden on participating practices and improve site activation and recruitment.

AI Assistance in Imaging

Additionally, two Monday presentations looked at AI use to aid novice sonographers in obtaining echocardiogram views and cardiologists in reporting severe aortic stenosis (AS), respectively.

The first was an international, multicenter, prospective noninferiority trial, which evaluated whether nurses with no prior ultrasound experience could obtain diagnostic-level acquisitions or 10 echocardiographic views with assistance from the AI software HeartFocus.

Trained on a total of 1,483 patients and 1,204,113 ultrasound images, HeartFocus has a combination of one deep-learning algorithm for view detection, 10 deep-learning algorithms for live guidance and two recording algorithms. The program indicates where to hold the probe while auto-recording results.

Seeking to determine whether novice ultrasound exams were high-quality enough to analyze left ventricle size and function, right ventricle size, and presence of nontrivial pericardial effusion, the trial included seven experts and eight novice nurses conducting sequential exams on 240 patients (48.8% women, 70.4% with cardiac abnormalities, 18.3% with implanted cardiac devices). Five outside expert cardiologists then reviewed the 480 images.

No difference in quality was found between novice and expert exams, and novices performed high quality exams starting with their first exam. With these findings, Caroline Ong, MD, FACC, et al., said HeartFocus is a "safe and reliable" integration into standard practice.

The second study explored AI assistance in cardiologist reporting of AS with the EchoSolv-AS model, now cleared by the U.S. Food and Drug Administration.

David A. Playford, MBBS, PhD, FACC, et al., used AI to detect AS in 200 echocardiograms and compared results to clinician findings. Secondary endpoints included reading time, referral rates, cardiologist concordance and cardiologist sensitivity.

The AI model was 100% accurate at detecting severe AS, while clinicians misdiagnosed severe AS 6-54% of the time, with an overall error rate of 20.6%. Most misdiagnoses were of low-gradient severe AS.

Despite the 100% accuracy of the model, no cardiologist consistently accepted AI recommendations. The researchers describe these results as a signal that lack of trust will be a major barrier in AI integration moving forward.

Keywords: ACC Annual Scientific Session, ACC25, Artificial Intelligence