Ushering in a New Era in Cardiology Research and Clinical Care

Artificial intelligence (AI) has already begun to transform cardiovascular medicine. Mass General Brigham's Daniel M. Huck, MD, and Antonis A. Armoundas, PhD, are at the forefront of efforts to understand and harness the power of AI, natural language processing (NLP), and other technologies to enhance medical research and clinical care in this field and beyond.

Extracting and Characterizing Cardiovascular Imaging Data With NLP

Coronary computed tomography angiography (CTA) is playing a rapidly growing role in evaluating and managing cardiovascular disease. The resulting vast volumes of cardiovascular imaging data have the potential to be used for improving patient care, advancing medical research and other efforts. But first, the data must be accurately extracted and characterized—an extremely labor-intensive pursuit.

Dr. Huck, a cardiologist and cardiovascular imaging specialist with the Cardiovascular Imaging and Prevention Group and the Division of Cardiovascular Medicine at Brigham and Women's Hospital, is part of a team of Mass General Brigham physician-scientists exploring the use of NLP to make data extraction and characterization more efficient.

"NLP is an AI technique that can process a large amount of semi-structured and unstructured textual data and package it into a more structured format that can then be used in medical research," Dr. Huck explains. "That's what we did in this project."

Dr. Huck is the corresponding author of a paper published in the Journal of Cardiovascular Computed Tomography that details the team's work. He and his colleagues used the open-source Canary NLP system (developed by Brigham endocrinologist and study co-author Alexander Turchin, MD) to design three NLP modules to phenotype coronary CTA text reports from the Brigham and Massachusetts General Hospital.

"Basically, the NLP system takes the semi-structured coronary CTA reports and translates what clinicians wrote at the time of interpretation into specific structured data that we can then use in our enlarged cardiovascular outcome studies," Dr. Huck says.

The investigators had over 30,000 coronary CTA reports at their disposal from the Mass General Brigham Coronary CTA Registry, one of the nation's largest coronary CTA registries. That immense amount of data would be impossible to interpret via traditional methods like human chart review, but NLP is up to the task.

First, a team must validate the NLP algorithm being employed. This involves reviewing a selection of reports to confirm the algorithm is successfully converting the semi-structured data into a more structured, research-friendly format. "After doing that on the testing set and a validation set, a total of 520 reports, we learned that we had over 95% accuracy for the degree of stenosis and presence of plaque," Dr. Huck says.

Following validation of the three modules, they were deployed on the more than 30,000 coronary CTA reports representing 28,175 Mass General Brigham patients from 2003 to 2021. A preliminary analysis found that 65% of the studies had one or more vessels with reported plaque or stenosis.

"New techniques like this are revolutionizing cardiovascular outcomes research," Dr. Huck says. "This approach is also scalable. We did 30,000 reports, but you can foresee this becoming millions of reports because it's just a matter of computing power and time."

AI INFORM Trial Looking for Evidence of Coronary Artery Calcification

Dr. Huck adds that once the accuracy of the NLP algorithm is validated, it could be applied to systemwide, state-level, or national data in different ways. He contends that this advancement could have a major impact on the future of medical research worldwide.

"Many countries with national health care are already using these approaches to digest the large amounts of data in their medical systems," he says. "In the United States, we have so many different data sources, and they don't always communicate with each other. We're going to need to develop approaches to integrate data from within the country and from other countries so that we can better understand how to improve cardiovascular health."

To that end, Dr. Huck and the Brigham are collaborating with other institutions on a study examining the use of AI to detect incidental coronary artery calcium on chest CT scans.

The AI INFORM multicenter trial will leverage AI technology developed by Nanox.AI to look for evidence of coronary artery calcification (CAC) in chest CTs that have already been done for non-cardiac reasons, such as screening for lung nodules. A patient's provider will be notified if the AI identifies the presence of coronary plaque.

"We know patients with CAC are at much higher risk for heart attacks and other cardiovascular events in the future," Dr. Huck says. "But this type of data often goes unutilized for various reasons, such as the lack of expertise among providers and the time constraints faced by providers today. We're using this information to inform providers and see if it will affect the prescription of preventive therapies for their patients."

AI INFORM is scheduled to begin in the second half of 2024.

How Can AI Improve Outcomes in Heart Disease?

Antonis A. Armoundas, PhD, a principal investigator with the Cardiovascular Research Center at Mass General, chaired two American Heart Association (AHA) committees that recently released scientific statements on technological innovations shaping the future of cardiovascular care.

The first AHA statement, published in Circulation, concerns the use of AI in improving outcomes in heart disease. It identifies best practices, gaps, and challenges in using AI tools, algorithms, and systems across imaging, electrocardiography, in-hospital monitoring, implantable and wearable technologies, genetics, and electronic health records.

"The motivation for this paper was to address the rapid expansion of AI in cardiovascular medicine and the huge investments by academia, industry, and government in a plethora of algorithms that do things that are of uncertain value," Dr. Armoundas explains. "I say 'uncertain value' because, despite the hype, very few algorithms have been able to be used clinically, and at scale."

Dr. Armoundas and his colleagues acknowledge that AI- and machine learning-based digital tools can improve disease screening, clarify what makes individual patients healthy, and facilitate the development of precision treatments for complex diseases. However, they argue, these tools have not yet been proven prospectively to justify widespread use and enhance care.

The authors call for "[r]obust prospective clinical validation in large diverse populations that minimizes bias" and maximizes the generalizablity of findings in order to avoid perpetuating existing healthcare inequalities.

"The majority of studies that have employed AI have been performed in retrospectively created cohorts of patients" Dr. Armoundas notes. "Also, they haven't included diverse patient groups. Being able to make accurate predictions with some known level of certainty will require more demographic diversity among patients, whether that's ethnicity or including individuals who live in remote and underserved areas. Inclusion, in turn, will inspire trust and lead to wider clinical acceptance and adoption."

A Call for Data Interoperability to Enrich Clinical Decision-Making

The second AHA statement, also from the American Heart Association, was published in Circulation: Genomic and Precision Medicine. Its focus involves data interoperability among an ever-growing number of wearable devices that individuals use to track health and illness. These technologies range from smartwatches and fitness trackers that collect health data primarily for recreational purposes to medical-grade devices that continuously collect more robust data.

"The medicine of the future will involve integration of natural history data, which includes genetics, genomics, and other health or non-health related data dating from our birth until today," Dr. Armoundas says. "But, it will also involve data acquired through many devices that have a monitoring, diagnostic, or perhaps therapeutic role."

In the paper, the authors outline the need to integrate and process data from these various devices into clinical workflows to enrich clinical decision-making. This is a complex technical endeavor that also must account for disparate policies across countries regarding data protection, handling, and sharing as well as how devices and systems communicate with each other.

"All of this information," Dr. Armoundas says, "must also be integrated into EHRs in forms that are recognizable for further processing by AI algorithms so it can be leveraged for clinical care and quality improvement."

Dr. Armoundas believes that AI is poised to transform medicine dramatically in the months and years to come. Keeping in mind the adverse outcomes that could result, he stresses the need to proceed with caution. But he is also excited about AI's potential to enhance patient care in and out of the hospital.

"For example, AI can be used to reduce incidents of false alarms in the ICU or to improve various processes around inpatient care," he says. "We also want to use AI and the increasing amounts of data that we can access to bring patients back to the hospital on an as-needed basis and at the appropriate time—not too early or too late—to prevent patient disease deterioration from developing. That will eventually help us not only to optimize patient care but also to reduce costs."