Dyania Health Runs Synapsis AI on Hospital-Owned NVIDIA GPUs

By Grace Kim

Dyania Health and NVIDIA said on November 13, 2026 that Synapsis AI now runs on hospital-owned NVIDIA accelerated computing, with an initial Cleveland Clinic study covering 7,812 aortic-valve replacement patients. The design keeps patient records inside health-system networks by running inference on on-premises GPU servers instead of shipping data to an external environment.

Dyania Health describes Synapsis AI as a clinical reasoning engine. It reads years of notes, operative reports, imaging reports and lab results, then answers clinical questions with evidence linked back to the chart. Its outputs are clinical signals that show care teams which patients need attention and why. Clinicians review the evidence and make every decision themselves.

What does the on-premises stack actually do?

The deployment uses NVIDIA AI Enterprise across the software stack. Dyania trains its proprietary clinical models with NeMo, then optimizes inference on hospital GPU servers with TensorRT and TensorRT-LLM. Inference stays within the hospital's private network.

NVIDIA said the system remains assistive: it surfaces evidence from medical records that clinicians must confirm and act on. That distinction matters, because the deployment automates chart review and signal detection — not the clinical decisions that follow.

The companies are targeting a manual bottleneck in hospital research and care management. Important details often sit in narrative text rather than structured fields, which limits how many charts staff can review for registries, outcomes studies and predictive models. Synapsis AI applies longitudinal review across an entire health system's population, with potential uses including care-gap detection, observational research and clinical-trial matching. In plain terms: the software reads what humans don't have time to read.

What did the Cleveland Clinic study find?

The study examined patients who received aortic valve replacements between 2005 and 2021, reviewing up to 16 years of care per patient. Synapsis AI answered 88 clinical questions for each patient, covering valve characteristics, co-existing conditions, lab trends, medications, follow-up ultrasound measurements and subsequent clinical events. Much of that history lived in narrative notes and echocardiogram reports.

NVIDIA said the platform completed chart review across the full cohort — work that would have taken clinicians years to perform manually. The resulting dataset trained a predictive machine-learning model that estimates valve-failure and reintervention risk at three, five and 10 years, and revises those estimates with each new ultrasound.

Dyania reported that implanted valve type was the strongest determinant of how quickly a valve deteriorated, while age and other medical conditions influenced whether and when a second procedure occurred. The analysis also showed that follow-up information, such as post-surgery ultrasounds, can substantially raise predictive accuracy around future interventions.

"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, chief of Cleveland Clinic's Heart, Vascular & Thoracic Institute.

One caveat: the study demonstrates dataset construction and risk prediction, not improved patient outcomes. Dyania says the tool is intended to inform care and requires external clinical validation before broader clinical reliance.

What comes next?

The companies plan to pursue health-system deployments in the United States and internationally, publish joint case studies and integrate Synapsis AI with NVIDIA's healthcare software stack. Technologies under evaluation include:

• NVIDIA Confidential Computing for training and inference
• NVIDIA Dynamo for distributed inference serving
• Multi-Instance GPU for partitioning GPUs across concurrent clinical workloads

The companies are also evaluating scaling to multiple Blackwell Ultra GPUs as processing moves from batch reviews toward near-real-time clinical insight at the point of care. These remain evaluation goals, not completed deployments.

For health systems, the practical question is whether externally validated models can turn population-scale record review into reliable clinical action while clinicians keep control of care decisions.