NVIDIA shows how hospital-local AI supports chart review NVIDIA shows how hospital-local AI supports chart review

What does the software actually produce?

Dyania describes a system that connects information across notes, imaging reports and laboratory results. Its output is a set of signals linked to the underlying records, for clinical review. Dyania says Cleveland Clinic used it to support research on aortic valve replacement patients treated between 2005 and 2021. That work assembled information for predictions about deterioration and further intervention. The distinction matters: extracting a usable history is one task; deciding what to do for a patient is another.

Why keep the work inside the hospital?

In NVIDIA’s account, Synapsis operates on the health system’s own accelerated servers, behind its firewall. This puts computation near the records with processing kept onsite. For builders, that separates deployment location from the usefulness of an answer. A local system still needs a way to show the source passage and put its conclusion in front of a qualified reviewer. Locality describes where the work happens. Traceability describes how someone can inspect it.

What evidence predates this announcement?

Cleveland Clinic described a different Synapsis application on March 3: pre-screening for a rare-disease clinical trial. It reported that the system reviewed 1,476 patients in one week and identified 46 potential matches. Following clinician review, 30 patients were identified as trial matches, including 29 missed by traditional recruitment. These are separate steps with separate counts. A potential match enters the review queue before enrollment. The hospital reported seven enrollments over six days through AI-assisted screening, compared with ten over ninety days through traditional screening. The process included auditable reasons for eligibility decisions.

How would the review chain work in practice?

Consider a hypothetical screening task with three requirements and ten candidate records. The system returns thirty criterion-level answers, each accompanied by the relevant passage and its date. A reviewer checks whether the passage supports the answer, whether it describes the right time period, and whether information is missing. A correct extraction from an old note can still be unsuitable for a current eligibility decision. Recording both extraction errors and eligibility disagreements gives the team two different measurements. This hypothetical test design illustrates the review process. It illustrates why a useful summary and a final decision should remain separately inspectable.

How does screening fit into a longer deployment?

Cleveland Clinic’s October 26 account places its enterprise deployment in November 2025, after cardiology and oncology pilots. The hospital says participating study teams subsequently enrolled forty percent more patients, forty-seven percent faster on average. That describes an operational recruitment result, not the heart-valve study. The chronology also matters. NVIDIA’s November account is a new description of an existing clinical workflow, not evidence that hospital deployment began that week. For a builder, the relevant progression runs from a defined screening task to reviewed candidates and then recorded enrollment. Those stages give a deployment more informative checkpoints than a single claim that the model reads records.