Dyania Health

The model is the product. Synapsis AI is a medical language model the company pre-trains and fine-tunes to read full medical records and answer specific clinical questions (is this patient eligible, does this patient have this finding), used for trial prescreening through Patient Finder, observational studies, registries and quality measurement. The company's physicians review results before delivery, but that review exists to check the model's output; there is no workflow tool or data product sold without it.

People check every result; how missed patients are found is not described. The company's physicians review true and false positives and correct remaining hallucinations before results reach the customer, and in the published Cleveland Clinic work the clinic's research staff confirmed each match before patients were approached. The system answers defined clinical questions rather than writing summaries.

Company research on confidence estimation for multi-step model tasks addresses when the model should flag uncertainty, but no threshold used in the product is published, and nothing states how eligible patients the model never surfaces are detected.

The model's training is described; its versions and update practice are not. The company describes a proprietary medical language model pre-trained on a medical knowledge library of more than 120 billion characters, then fine-tuned on more than 25,000 physician-annotated cases (about 15,000 physician hours), reporting more than 95 percent accuracy against trained human reviewers without a published method.

Company research on confidence estimation in multi-step model tasks is published separately. Whether the model starts from an outside base model, how versions are managed, and how customers learn of an update to an installed system are not published.

The chain is structurally short: the company's own model, installed inside the customer. Synapsis AI is described as a proprietary model trained by the company and run behind the health system's firewall, with records stated never to leave, so no outside model provider or cloud host sits in the data path on the company's description. What remains is the company itself: its physicians review results before delivery, and its staff maintain the installation. Whether any base model or third-party component is inside, and what the company retains from reviews, are not stated.

Published work with a major academic health system, with the full paper not open access. Martyn and colleagues (Journal of Cardiac Failure, 2026; Cleveland Clinic with Dyania staff) evaluated the system for a phase 3 transthyretin amyloid cardiomyopathy trial; the company reports 1,476 patients reviewed in a week, 30 complete or high partial matches in two weeks, 35 percent more diverse patients than manual review, and 97 percent of those found missed by prior screening, but the article's text and method are behind the publisher's paywall.

The Cleveland Clinic polycythemia vera study presented at ASH 2025 is described by the clinic: 904 patients assessed in a week, 22 eligible, all confirmed, against 9 prescreened over a year. Cleveland Clinic moved to an enterprise-wide rollout after a year in cancer and cardiovascular disease.

The architecture keeps records in place; the training data's origin and the vendor's own access are not described. The company installs the software behind the health system's firewall and says patient records never leave its computing environment. Two things sit alongside that. The model was fine-tuned on more than 25,000 physician-annotated cases, and nothing says whose records those were, under what permission, or whether they came from current customers.

And the company's physicians review and correct results before delivery, so its staff see patient-level output; who can see what, from where (the company has offices in New Jersey and Athens), with what logging, is not stated. The system screens whole health system populations daily, before anyone has consented to a study. Nothing is published on retention of outputs.

Compliance is claimed and no position is taken. The company says its in-firewall architecture ensures compliance with the rules a health system must meet, naming HIPAA, HITRUST and GDPR (HITRUST is a voluntary certification, not a regulation), and elsewhere that it exceeds HIPAA and GDPR. Its privacy policy covers the website and the Greek and US companies' marketing contacts, not patient data in the service. No business associate agreement, role statement or contracting entity is published, though its physicians review patient-level results.

Controls are asserted with nothing independent behind them. The company's security case is architectural: installation behind the health system's firewall in a closed environment with local processing. It names HIPAA, HITRUST and GDPR as rules its architecture ensures compliance with; that is not a claim to hold HITRUST certification, and no SOC 2 report, ISO/IEC 27001 certification, HITRUST certification or penetration test is named, and there is no trust center. The architecture shifts the review toward the installed software and the company's access, which no attestation covers.

No device claim is made, and the product is scoped to chart review and abstraction for research, registries and quality measurement. The system identifies candidates and extracts information already documented in the record for a person to confirm; it does not diagnose or direct treatment. For trial prescreening the governing rules are human subjects protections: reading records before consent needs IRB approval with an authorization or waiver, held by the site or the company. Registry and quality abstraction carries accuracy obligations to the receiving program.

Equity is claimed as an outcome, with no groups or counts behind it. The company reports that in the first weeks of a Cleveland Clinic transthyretin amyloid cardiomyopathy trial, Patient Finder identified 35 percent more diverse patients than manual chart review, and its leadership states the aim of surfacing eligible patients from underrepresented groups.

Which groups, how many patients and what baseline go unstated, and the Journal of Cardiac Failure paper behind the study is not openly available, though it may hold the counts. No governance framework, evaluation method or monitoring commitment for the model is published.

An outside error figure with its denominator, which the company cites. Cleveland Clinic investigators led by Aaron Gerds reported at the 2025 American Society of Hematology meeting, and described in a Cleveland Clinic article the company links from its home page, that Synapsis AI assessed 904 patients with polycythemia vera against a phase 3 trial's 7 inclusion and 20 exclusion criteria within a week and identified 22 as eligible, all 22 confirmed by the clinic's research staff (100 percent positive predictive value), against 9 prescreened over 12 months by the usual process.

That measures false positives, not missed patients: how many eligible patients the model did not surface is not reported. The company also says remaining hallucinations are caught and corrected by its physicians before results are delivered. No warranty, indemnity or remediation commitment is published.

The model reads the record in production at a named health system; the EHR product and any write-back are not named. At Cleveland Clinic the system ran inside the health system's environment across its hospitals and outpatient centers, re-reading changing records daily and screening millions of records, and the company handles ingestion and structuring of EMR data itself. Output goes to reviewers rather than back into the chart, and no EHR vendor, integration standard or marketplace listing is named.

The deployment keeps data inside the customer; the edges are not described. The company installs Synapsis AI behind the health system's firewall in a closed environment, processes locally, and says patient records never leave the health system's computing environment. It also performs the deployment and all maintenance, and its physicians review results before delivery, which means company staff reach into the environment or its output. Remote access for support, its logging, outbound telemetry and how model updates arrive are not stated.

No price is published, and the cost frame is discoverable. The company says health systems with about 500,000 patients spend more than $5 million a year on staff doing manual chart abstraction, and describes a one-time deployment inside the health system followed by the company's ongoing maintenance and physician review. Buyers include health systems, physicians, pharmaceutical sponsors and researchers.

Coverage is named and validated for part of it. The company describes Synapsis AI as usable across disease areas, with oncology, cardiology and autoimmune conditions named, and across trial prescreening, observational studies, registry population and reporting, therapeutic gap identification and quality measurement, in North America and Europe. Published results concentrate on trial prescreening at Cleveland Clinic in transthyretin amyloid cardiomyopathy and polycythemia vera. Beyond that, breadth is described rather than demonstrated.