AI Tool Adding Speed and Diversity to Cleveland Clinic Trials

ATTR-CM has historically been characterized as a rare disease but is now more often viewed as an unrecognized cause of heart failure in older adults, he says. Disease prevalence has gone up since 2014, when significant advances in nuclear scans moved the technology from research into standard clinical practice. A heart biopsy was previously required to arrive at a diagnosis. The first targeted treatment for the condition wasn’t approved until 2019. 

Thinking Like a Doctor 

“Without the involvement of EMR-based tools to help find patients, [trial] recruitment is somewhat of an idiosyncratic process,” Martyn says. “If you happen to be seeing the patient and you’re involved with the study, then you would think about enrolling. So, generally patients are enrolled from the hubs where research is centralized, and where investigators are actively seeing these patients.” 

This is true for diseases like cardiac amyloidosis, as well as other types of cardiomyopathies causing heart failure, he adds. While patients often get referred to high-volume centers for specialized care, and thus sought as trial sites by study sponsors, “that doesn’t mean those patients don’t exist elsewhere.”  

Systematic efforts in the past to look more broadly for participants have been limited to sweeps of structured data in EMR systems, notably ICD codes and labs. The problem with this, according to Martyn, is that the information isn’t very specific. Patients might have been coded for amyloidosis based on a clinical suspicion that was later disproven, or for a non-cardiac rather than a cardiac type of amyloidosis. 

Roughly 80% of EMR data lives in unstructured form within clinical notes, and imaging reports, he says, pointing to the benefits of LLMs. The only other way to extract the information is via manual review, a time-consuming process that has been the cornerstone for determining clinical trial eligibility for years. In the absence of automation tools, research coordinator screening and eligibility determinations also constituted a sizable proportion of clinical trial budgets.  

Training of Synapsis AI involved not only the use of medical corpus—a massive dataset of text curated from medical science literature, journals, clinical notes, and textbooks—to teach the platform to understand complex medical terminology and clinical reasoning. It also had in-house physicians manually annotating thousands of medical charts accessed in a HIPAA-compliant fashion “so the algorithm could learn off those conclusions,” Martyn says, which is its key differentiator from frontier LLMs like Claude and GPT-4 and GPT-5.  

“What I particularly like about the research we did was that it ... happened in real-world EMRs” rather than using synthetic or simulated data, he says. Much of the focus of AI has been on a novelty such as the ability to crack the United States Medical Licensing Examination board exam rather than comprehensively processing health record data. In the real world, physicians assimilate multiple data types to arrive at clinical conclusions, says Martyn. Accordingly, in what could soon be the new normal at Cleveland Clinic, the broad range of EMR notes could be processed by Synapsis AI to process and make a clinical determination, such as whether a patient has symptomatic heart failure. 

The determination isn’t based purely on symptoms referenced in clinical notes, Martyn stresses. The AI system can, for example, note that a patient has a heart valve problem as indicated in the echo report, an elevated N-terminal pro-brain natriuretic peptide value associated with a heart failure exacerbation, shortness of breath and leg swelling per the clinical notes, fluid in the lungs suggested by an X-ray, and an admission report coded as a heart failure hospitalization where IV diuretics were administered six months ago. “All those pieces of information come together to make a conclusion that it is very likely this patient currently is in a heart failure exacerbation,” he says. 

Considering just lab values or a prior history of hospitalization captured in claims or coding data, “you could misclassify reasons for hospitalization,” says Martyn. This more holistic, clinician-oriented approach is what gives Cleveland Clinic physicians confidence in the conclusions reached by Synapsis AI. “We saw multiple data types being pulled together in a logical way to make a singular conclusion.” 

Possibilities and Limitations 

Like other large academic centers, Cleveland Clinic is rapidly growing the amount of its on-premises computing power to accommodate AI solutions. The processing time relates to how much computing power is devoted to a task and speaks to the need for pre-filtering data, says Martyn, both to ensure data quality and keep large-scale deployment feasible. 

Moving forward, it will be fascinating to see how Synapsis AI approaches rejecting patients for other studies who are truly ineligible, he adds. If it goes after obvious cutoffs that are easily verifiable, such as age or lab-based renal function, an initial cut might be made using rule-based querying of structured data in the EMR, reserving the use of AI for the more sophisticated inclusion/exclusion criteria where it’s needed. 

Martyn says he is personally excited about the automation of chart reviews, recalling the long hours he spent during his medical school, residency, and fellowship days poring through patient files trying to construct data tables in Excel in the pursuit of retrospective research. When searching for specific patient populations for studies, this laborious chore was historically circumvented by simply looking at ICD-10 codes and hoping that they were a good approximation of the truth. 

“But then you’re ... getting less quality data due to prioritizing expediency over accuracy,” Martyn points out. “I don’t think we’re going to have to make that tradeoff to the same degree anymore.”