Cleveland Clinic uses AI to review records of more than 7,800 heart-valve patients

Years of medical records became material for heart-valve risk modeling in a Cleveland Clinic study described by Nvidia on November 13. Dyania Health’s Synapsis AI reviewed records for more than 7,800 patients. The work helped build predictions of valve deterioration and the need for another procedure—not evidence that using those predictions improves care.

From scattered records to a research dataset

The study covered patients who underwent aortic valve replacement between 2005 and 2021. Rather than examine only information recorded around surgery, Synapsis worked across their medical histories, including echocardiograms added afterward. Those later records supplied information for modeling what happened to implanted valves over the years following the procedure.

The underlying task is extracting usable information from records that are not neatly organized into data fields. Nvidia describes patient histories spread across physicians’ notes, operative reports, imaging and pathology reports, and discharge summaries. Synapsis is designed to connect findings across notes, reports and laboratory results, even when those findings are separated by years.

For this study, the system answered 88 questions about each patient, The Next Web reported. Its answers linked back to information in the medical record. That traceability is part of the platform’s design: it surfaces clinical signals with supporting source material, rather than presenting care teams with an answer detached from the chart.

The extracted information supported a predictive model covering three, five and ten years after surgery. Its targets were valve deterioration and reintervention risk—the chance of needing another procedure. Nvidia says completing the cohort’s chart review manually would have taken clinicians years; that is the supplier’s estimate of the work avoided.

Using technology to bring together and analyze information that can help us predict the duration of implanted heart valves is a promising step towards enhanced outcomes and more personalized care for patients

Dr. Lars Svensson, chief of Cleveland Clinic’s Heart, Vascular & Thoracic Institute

Hospital servers, with clinicians still responsible

The system’s computing setup keeps the review inside the health system. Synapsis runs on hospital-owned Nvidia GPU servers behind the institution’s firewall, according to Nvidia. Patient records are processed onsite rather than sent outside the hospital. The Next Web reported that the Cleveland Clinic work ran on the hospital’s own servers.

Dyania uses Nvidia’s NeMo software libraries to build and train its clinical models. It then uses TensorRT and TensorRT-LLM to make those models run faster and more efficiently on hospital GPU servers. Nvidia therefore supplies both the computing hardware and software tools used to prepare the platform for local operation.

That automation does not transfer clinical authority to the software. Nvidia describes Synapsis as an assistive tool: clinicians must confirm the evidence it surfaces and decide what action to take. Its role is to bring relevant information forward for review, while medical decisions remain with the care team.

Prediction is not yet a patient benefit

Svensson’s description emphasizes the prospect of more personalized care. The demonstrated work is narrower: creating a dataset and using it for risk modeling. The Next Web distinguished those results from proof of improved care and reported that Dyania says outside validation is still needed. The model’s clinical value remains unproven.

There is also a financial relationship relevant to interpreting the hospital’s role. Cleveland Clinic disclosed in a March 3 announcement that it had invested in Dyania and could benefit financially from sales of the technology. The hospital is therefore not only a research user of the system, but also an investor in its supplier.

For now, the study’s concrete contribution is to research: turning years of follow-up records into data for estimating implanted-valve risk. Whether those estimates lead to better care is a separate question.