Paradigm Health

Language models do the eligibility reasoning in the core matching module, inside a business that sells people, a network and data capture as well. The company describes structured-data prefilters followed by large language models reasoning over full clinical notes, and reports in an ASCO Quality abstract that this cut screening volumes by 31 to 87 percent in three of four oncology trials against a rules-based approach.

Around it the company sells recruitment coordinators and eligibility specialists who work patients through enrollment, a research network enlarged by the December 2025 purchase of Flatiron Health's clinical research business, EHR-integrated eSource and data capture, feasibility and protocol services, and a sponsor offering for FDA's Real-Time Clinical Trials pilot. Those carry value without the models.

A person confirms every candidate; what catches a missed patient is not described. The AI continuously monitors eligibility and prioritizes candidates with rationale and supporting clinical context; the company's eligibility specialists and site teams review, confirm eligibility and follow up, re-screening as records change; final eligibility and consent rest with the research team. Nothing the models do reaches a patient without that review.

What is missing is the other side of the error: nothing states how eligible patients the models never surface are found, whether screens are sampled or back-tested against manual review, or what happens after a miss.

The approach is described; the models are not named for the current product. The company describes structured-data prefilters, where hard criteria are applied, followed by large language models reasoning over full clinical notes, pathology and imaging reports, with the documents behind each inclusion or exclusion shown to reviewers. In 2024 it published work with OpenAI evaluating GPT-4 for trial screening, which it found far better than a custom medical language model. Which models run in production now, whether they are fine-tuned, and how versions are managed and customers notified are not published.

The host is named; the production model layer is not. All environments run in AWS US-East-1. The company's 2024 case study with OpenAI evaluated GPT-4 for screening, but no page says which model provider serves the product today, and no sub-processor list is published. The data path is long: provider record, the company's platform, sponsors and contract research organizations, and in the Real-Time Clinical Trials pilot the FDA.

Named deployments with figures for the matching, and a peer reviewed study of the data capture platform the company acquired. Ochsner Health (47 hospitals, more than 370 care centers) reports a 41 percent increase in screening capacity, 3.6 times more patients actively followed and about 75 percent fewer patients needing manual review; an ASCO Quality abstract reports 31 to 87 percent lower screening volumes in three of four oncology trials against rules-based matching; unnamed sponsors report twice the planned enrollment rate and a 43 percent median enrollment lift across six trials.

None has a published method beyond the abstract. Ghanim and colleagues (npj Health Systems, 2026; Flatiron Health with Johnson and Johnson, before Paradigm bought the business) ran a prospective study at four community oncology sites: 82 patients in 11 months, 0.59 percent missing data, 91 percent of 4,568 data points eligible for automated EHR-to-EDC transfer, and fewer than 2 percent queried. That study tests eSource and data capture with rule-based patient identification, not the language model matching.

The security program is described; what happens to patient data at each hop is not. The platform reads provider records inside participating organizations, passes candidates and study data to sponsors, and in FDA's Real-Time Clinical Trials pilot sends pre-defined safety, efficacy and data quality signals to the Agency during a study. Source-linked case report forms let monitors check a value against the originating record.

The June 2026 privacy notice excludes trial participants and research, and nothing published covers where de-identification happens, what the FDA and sponsors receive and in what form, retention at each hop, whether patient data is used to develop or tune models, or what happens at study close. Nothing is published on safety engineering for the language models.

Business associate status is stated with an outside check behind it; no agreement is published. The security and privacy program overview says the company operates as a business associate, holds a third-party HIPAA security compliance attestation with no exceptions noted, has annual third-party compliance audits, and runs a data privacy program under a named data protection officer. The website privacy notice (June 2026) excludes trial participants.

Two questions remain separate from business associate status: screening records before anyone consents needs an IRB approval with an authorization or waiver, held by the site or the company, and sending trial signals to the FDA in near real time must be covered by participants' consent.

Recognized attestations are named on the company's own security page; the reports are not published. The security and privacy program overview says the company holds a third-party SOC 2 Type 2 attestation against the security, confidentiality and availability criteria with an unqualified opinion and no exceptions noted, a third-party HIPAA security compliance attestation, annual third-party audits, and a program built on NIST SP 800-53 revision 5 under a named chief information security officer. The period and scope of the SOC 2 report, and whether the component deployed inside providers' environments is covered, are not stated, and there is no downloadable report or trust portal.

No device claim is made, and the product is scoped as trial software. Trial matching identifies research candidates rather than informing diagnosis or treatment, so human subjects rules govern it: IRB oversight, informed consent and Good Clinical Practice, with electronic records rules for the eSource and data capture products.

The company has a Research Collaboration Agreement with the FDA, says it helped shape the Real-Time Clinical Trials initiative, and offers sponsors the technology to join that pilot, in which pre-defined trial signals reach the Agency during a study. That is a role in an FDA program about trial data, not a regulatory status for the product.

Equitable access is the stated mission and part of the design; no evaluation of the matching model across groups is published. Extending trials into community oncology practices addresses who gets offered a trial, and the npj Health Systems study of the eSource platform (2026) reports that 30.5 percent of its 82 participants identified as Black or Hispanic, meeting its diversity target early.

The company's writing on trial matching says a solution must show the documents behind each inclusion or exclusion. What is missing is measurement of the AI: whether candidates the models surface differ by race, age, sex or record completeness from those found by manual screening, when a model trained on how records are written decides who is seen.

Each candidate comes with its evidence and passes a person; nothing stands behind the output. The AI surfaces candidates with a rationale and the clinical documents behind it for site teams and the company's eligibility specialists to confirm, and final eligibility and consent stay with the research team.

The published figures are screening efficiency (31 to 87 percent fewer charts screened in three oncology trials; about 75 percent fewer patients needing manual review at Ochsner Health), not an error rate: no count of eligible patients the models missed is published. No warranty, indemnity or remediation commitment is published.

Record integration is running at named sites through standards; no named EHR integration is documented. The platform deploys inside participating provider organizations and reads structured and unstructured record data, with pipelines that normalize data through FHIR, HL7 and data warehouse integrations. The npj Health Systems study (2026) shows an EHR-to-electronic-data-capture connector in production at four community oncology sites, moving structured data in a median 1.7 days.

The Flatiron partnership commits to integrating the company's tools natively into Flatiron's OncoEMR, stated as future work. No EHR vendor listing or integration documentation shows a named, two-way connection.

Residency is committed for the platform; the in-provider component and the international network are open. The security and privacy program overview says all production and lower environments are hosted in Amazon Web Services' US-East-1 region inside a virtual private cloud, under AWS's shared responsibility model.

The platform also runs inside participating providers' record environments, and the company describes network sites in Japan and Israel; what the embedded component sends outbound, and how non-US participant data is handled and transferred, are not stated. The website privacy notice relies on standard contractual clauses for EEA transfers.

The two sided structure can be worked out, though no price is published. Sponsors and contract research organizations pay for patient identification, feasibility, study conduct and the FDA pilot offering across the company's research-ready network, while provider organizations receive the platform and recruitment support. Money flows from the sponsor side to a platform the provider hosts. The terms on the provider side, meaning what it pays and receives, what data obligations attach and how trials are chosen for its patients, are not published.

Coverage is named and validated for part of it. Oncology is the center: a research network of community oncology practices, health systems and academic cancer centers, enlarged by the Flatiron clinical research business, plus Ochsner Health across many specialties; the platform supports phase 1 to 4 and post-approval studies, with sites in Japan and Israel. Named results cover Ochsner and oncology trials; the peer reviewed study covers eSource in four community oncology practices.