Dyania Health announced a November 13 collaboration with Nvidia to expand Synapsis AI record analysis on hospitals' own infrastructure.

The big change

  • Synapsis links findings to chart evidence for clinicians to review. The important distinction is between locating information and deciding what it means: a suggested finding is not a final medical decision.
  • Keeping processing within an institution does not automatically make an application safe. A hospital still needs to decide who may access records, monitor use and check whether the system performs adequately for its intended task.
  • Earlier trial screening produced potential candidates, followed by clinical review and enrollment. That sequence matters: identifying a possible match does not mean a patient qualifies, enrolls or benefits from treatment.

From records to review

In a Fierce Healthcare interview, Dyania describes specialized models and a clinician-annotated knowledge base. It says physicians spent more than 25,000 hours annotating cases. This is the company's account of development, not an independent comparison proving superiority over other models.

The interview also describes hospital-specific implementation. Nvidia's Chelsea Sumner says clinicians, technology teams and hospital leaders need clear responsibility for reviewing results and monitoring performance. Those requirements help explain why buying computing capacity is only part of introducing an application into care.

A reviewer needs to know what question an output answers and which evidence supports it. An accurate summary of an incomplete record can still leave out information needed for a decision. The useful workflow therefore includes a way to question a finding and resolve missing or contradictory information, rather than assuming that a fluent answer is sufficient.

What the earlier trial study measured

In March, Cleveland Clinic reported prescreening for a phase 3 trial in transthyretin amyloid cardiomyopathy, a rare heart condition. Synapsis reviewed 1,476 patients in one week and found 46 potential matches. AI-assisted and clinician review identified 30 trial matches; seven enrolled before the site reached its enrollment goal.

The clinic reported 96.2% accuracy across 7,700 trial-specific questions in nine domains. This measures assessed answers, not treatment benefit or performance at another hospital. Clinical validation remained essential. The clinic disclosed its investment in Dyania and potential financial benefit from sales.

The clinic says the system combined structured medical-record data with analysis of notes and lab reports, giving reasons for inclusion or exclusion so research coordinators could check eligibility. Recruitment still involves eligibility review and a person's decision to participate; enrollment is a different outcome from improved health.

A different heart-valve question

Nvidia's account describes retrospective records from more than 7,800 valve-replacement patients treated in 2005–2021. Researchers used the resulting data to model deterioration and repeat-intervention risk at three, five and ten years. The account gives no prediction-performance figures or demonstrated patient benefit.

Historical model development does not establish how a tool performs when guiding future care. Before relying on a prediction, a hospital needs evidence that it works for the people and decisions concerned. BIG CHANGE did not test the software or independently verify deployment or security claims.