An AI read 16 years of heart records. Its maker says it hasn't been shown to help yet

Dyania Health and Nvidia announced a collaboration on Nov. 13 to run Dyania's Synapsis AI clinical reasoning engine on hospital-owned Nvidia infrastructure, at the scale of a health system's entire patient population 4

4

1

1

. The companies say the goal is to let a health system learn from the full histories of patients it has already treated 4

4

1

1

. That, they say, points to earlier identification of care needs 4

4

1

1

.

Synapsis reads the whole chart, covering written clinical records, imaging and pathology findings and lab results 2

2

3

3

4

4

. Its answers come with evidence traceable to their source in the chart 2

2

3

3

4

4

. Clinicians review the output and make every decision 2

2

3

3

. It runs on Nvidia accelerated computing inside the hospital's firewall, so records are processed onsite and, the companies say, never leave 2

2

3

3

4

4

1

1

. The stack is Nvidia's end to end: Dyania trains its models with Nvidia NeMo and optimizes them with TensorRT 2

2

4

4

1

1

.

Much of what matters clinically sits in narrative notes that resist analysis at scale, which has made manual chart review a bottleneck for registries and outcomes studies 4

4

1

1

. The field is crowded, too. Google Cloud launched Gemini for healthcare the same week, and Bunkerhill raised $55 million for hospital AI agents 3

3

. Proven gains for patients remain hard to find across hospital AI, and doctors have objected when claims run ahead of proof 3

3

.

The valve study

The collaboration's first work is a Cleveland Clinic study of 7,812 people given aortic valve replacements between 2005 and 2021, covering up to 16 years of follow-up care 4

4

1

1

. Most of that follow-up record sits in narrative text and ultrasound readings that registries struggle to capture 4

4

1

1

. Synapsis answered 88 clinical questions per patient, covering valve characteristics, co-existing conditions, lab trends, medications and later clinical events 3

3

4

4

1

1

. From those answers it built a model of valve failure and reintervention risk at three, five and ten years after surgery 3

3

4

4

1

1

.

In the model's findings, the kind of valve implanted mattered most for how quickly it deteriorated, and each follow-up ultrasound sharpens its outlook 4

4

1

1

. Nvidia said the same chart review done by hand would have taken clinicians years 2

2

3

3

. The study was built around the question valve patients actually ask: how long will my replacement valve last, and will I need another operation 4

4

1

1

.

What the first study shows

In Schlosser's telling, the collaboration is a watershed moment: for the first time, she says, a health system can reason across every patient's full record at once rather than one chart at a time 4

4

. Dyania says the same approach could extend to finding clinical-trial-eligible patients, closing care gaps in heart and cancer care, and filling registries 2

2

4

4

. The companies also plan joint case studies and are evaluating Nvidia's Confidential Computing and Blackwell Ultra GPUs for near-real-time clinical insight 4

4

1

1

.

The study demonstrates that the tool can assemble a dataset and forecast risk 3

3

. Whether patients actually do better is untested 3

3

. Cleveland Clinic, the study site, is also an investor in Dyania, an interest it disclosed last year 3

3

. Outside validation, Dyania says, is still to come 3

3

.