Dyania Health is partnering with Nvidia to run its Synapsis AI clinical reasoning platform inside hospitals’ IT systems, enabling large-scale analysis without patient data leaving their networks.
The startup's Synapsis AI reads a patient's full medical record, including clinical notes, imaging and operative reports, labs and other largely unstructured information, and answers clinical questions while pointing back to supporting evidence in the chart. Dyania Health goes beyond summarizing medical records; it uses AI to convert messy, unstructured clinical records into usable datasets and insights.
Nvidia provides the computing infrastructure needed to perform that analysis across potentially entire patient populations behind the hospital firewall, without transferring patient records outside the health system, according to the company.
Dyania brings the clinical reasoning technology and implementation expertise, while Nvidia brings the accelerated computing platform and the software that lets it run securely and efficiently on-premises, Eirini Schlosser, founder and CEO of Dyania Health, told Fierce Healthcare.
Through this collaboration with Nvidia, Dyania's Synapsis AI could help health systems turn large volumes of unstructured clinical data into actionable insights at a scale that would be impractical with manual chart review, while keeping patient data inside the hospital’s own infrastructure, Schlosser said.
Health systems have vast amounts of valuable clinical information locked in unstructured medical records, but manually reviewing those records makes it difficult to use that data at scale for research, registries and predictive models. Synapsis AI aims to remove that bottleneck by analyzing patients’ records longitudinally, determining what happened and when, and identifying the clinical evidence supporting those conclusions, according to the company, which launched out of stealth two years ago.
Frontier models are becoming extraordinarily capable, but the architectures, training data sets, as well as logic are not built to perform clinical reasoning in a real-world healthcare system environment, Schlosser noted.
Dyania's Synapsis AI combines specialized models, a clinician-annotated clinical knowledge base, evidence tracing and hospital-specific implementation. Dyania says it has invested more than 25,000 physician hours annotating medical cases so that Synapsis AI can perform clinical reasoning. This specialized approach is better suited to complex clinical reasoning than a single general-purpose AI model, Schlosser asserted.
"Every patient's story is spread across years of notes, reports and results, and for the first time we can reason across all of it at once, the way a specialist does, for every patient in a health system rather than one chart at a time,” Schlosser said.
Synapsis AI is built on the Nvidia AI enterprise software stack end to end. Dyania trains its proprietary clinical models with the Nvidia NeMo framework for model training and reinforcement learning. Inside the health system, Synapsis AI runs on Nvidia TensorRT and TensorRT-LLM on the hospital's own accelerated computing infrastructure, so inference stays inside the hospital's private network.
"Synapsis AI sits within the health system firewall and is small enough to run inference on local clusters of GPU computing. It does not need to 'speak to the internet' or interact with cloud or frontier models. This provides immense benefits to the healthcare systems for both the comparable cost of tokenomics they would be paying for the frontier models in the cloud as well as for cybersecurity and data privacy of sensitive patient data," Schlosser said. "Hospitals can apply that capability to their patient histories under their own data governance, without sending those records to an external AI service."
Hospitals can use their own historical patient data to determine which treatments work for which patients, identify care gaps and ultimately recommend the next appropriate action, according to Schlosser.
Under the collaboration, the companies will work to bring Synapsis AI on Nvidia infrastructure to health systems in the U.S. and internationally, publish joint case studies and integrate Synapsis AI with Nvidia's healthcare software stack. The companies are exploring additional Nvidia technologies to secure AI processing, handle multiple clinical workloads at once and scale computing capacity, with the goal of moving from batch processing toward near-real-time clinical insights at the point of care.
"Health systems need AI they can trust, integrate into their workflows and operate efficiently. That means connecting information across clinical systems, validating performance for the patients they serve, protecting sensitive data and making results useful within the tools clinicians already use," Chelsea Sumner, Nvidia's U.S. healthcare AI startups ecosystem lead, told Fierce Healthcare.
"Dyania’s approach addresses part of that challenge by running inside the hospital’s own environment and linking findings to evidence in the patient record. Scaling also requires clinicians, technology teams and hospital leaders to work together from the start, with clear responsibility for reviewing results and monitoring performance. Success should be measured by whether AI helps teams find patients who need attention, make better-informed decisions and spend more time delivering care," Sumner said.
Putting Synapsis AI to work at Cleveland Clinic
Cleveland Clinic used Dyania Health’s Synapsis AI, running on Nvidia infrastructure, to analyze up to 16 years of records from 7,812 aortic valve replacement patients. The study addressed a central concern for patients: How long will the replacement valve last and will another procedure be needed?
Researchers reviewed nearly 8,000 patient records between 2005 and 2021, mostly narrative notes and echocardiogram reports that registries can struggle to capture.
"Understanding how a replacement heart valve performs means following a patient across years of care, connecting the original operative report with subsequent progress notes, hospitalizations, echocardiograms, other medical conditions and later interventions. Much of that information is buried in narrative records that researchers would otherwise need to review manually," Schlosser noted.
The AI extracted information from narrative clinical records and answered 88 questions per patient about their valves, health conditions, and subsequent care. Each answer linked to supporting evidence in the chart, he noted.
Researchers used that dataset to build a predictive model of valve deterioration and the risk of another intervention. The model estimates valve deterioration and repeat-procedure risk at three, five and 10 years post-procedure, with estimates that update as new medical information about the patient becomes available.
"Historically, this would have required human manual abstraction, and for 7,812 patients it would be humanly impossible to read and abstract features on an ongoing basis at the rate at which medical records are changing. Additionally, humans reading a medical record can often miss information if they skim through it," Schlosser said.
The tool is used to help inform care but requires external validation from clinicians.
"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," said Lars G. Svensson, M.D., Ph.D., chief of the Heart, Vascular and Thoracic Institute at Cleveland Clinic, in a statement.
The health system began working with Dyania Health a year ago to use its AI platform for clinical trial recruitment.
Synapsis AI made it possible for the first time to comprehensively analyze nearly two decades of patient data, Schlosser asserted. "What was even more groundbreaking was the ability to go from raw data in unstructured form to predictive risk models in one fell swoop. This enables precision care at scale and empowers physicians with stronger sets of information backing their clinical decisions about any given patient, including questions like 'when should the patient have their next echo?' or 'when will the patient likely need a reintervention?'" Schlosser said.
The Cleveland Clinic and Dyania Health collaboration reflects Nvidia's broader focus to provide healthcare and life sciences innovators the computing platform and AI tools to advance research and help care teams better understand patients’ needs, Sumner said.
The heart valve study shows how Synapsis AI could help health systems use their own patient data to determine which treatments work best for which patients and improve predictive models over time, ultimately supporting more personalized care recommendations, according to Dyania Health.
"Nvidia's infrastructure enables us to handle more work concurrently and use resources more efficiently. That could allow hospitals to examine larger populations, ask more detailed questions and revisit analyses more frequently. The longer-term opportunity is to move from periodic batches of records toward insights available closer to the moment a clinician needs them," Schlosser said.
"We can do that within the health system’s own environment, keeping patient records behind its firewall. We are also evaluating capabilities for distributed inference, workload isolation and protection of data during computation. Computing power alone does not make a clinical conclusion correct; it allows a carefully developed clinical system to work at greater scale and speed," he said.
Sumner asserts that healthcare organizations are just scratching the surface on the potential for AI to uncover clinically useful information buried in longitudinal patient records.
"Connecting a physician’s notes with earlier test results and treatment history could help identify patients who need follow-up, match people with clinical trials and give researchers a clearer understanding of how treatments perform over time," she said. "The Cleveland Clinic heart valve study is a concrete example. AI helped researchers assemble a detailed view of patients’ care to study how replacement valves deteriorate. The opportunity is to turn that knowledge into more personalized monitoring and better-informed care decisions. Clinicians need to be able to review the supporting evidence, validate the findings and decide what they mean for each patient."