Paradigm Health

Language models do the eligibility reasoning, and the company published a direct comparison making that concrete: it evaluated its AI native matching against a traditional rule based platform relying on structured data and pre parsed fields across four cancer trials, reporting that in three of them the AI approach reduced screening volumes by 31 to 87 percent by surfacing candidates closer to full eligibility.

The architecture is described specifically as structured data prefilters enhanced with LLMs reasoning over full text clinical notes. Trial eligibility criteria live in prose, which is precisely why rules engines underperform here.

The design keeps a human at every consequential step and pairs software with staff rather than substituting for them. AI surfaces candidates with clear rationale and supporting clinical context for site review, continuously monitors eligibility and prioritizes candidates, while recruitment coordinators work alongside site teams on a shared platform to confirm eligibility and progress patients. Surfacing rationale rather than a bare match score is what makes the site review meaningful, and final eligibility determination and consent remain with the research team.

Unusually specific about architecture and its evolution. The company describes moving away from an earlier approach of fine tuned specialized models built use case by use case, each requiring separate training and validation, toward general purpose LLM evaluation, and reports needing roughly one tenth of the data to validate the newer approach. Its current design is documented as structured prefilters plus LLM reasoning over clinical notes. What is not published is model versioning, ongoing evaluation methodology, or performance by trial complexity.

The chain here has four parties and now runs continuously, which is the substance of this record. Since April 2026 the company has operated a collaboration with the United States drug regulator under which trial data is analysed and key safety and efficacy signals are reported to sponsors and to the regulator in near real time, with named pharmaceutical companies participating.

So the path runs from a provider's record, to this platform, to the sponsor, to a federal regulator, continuously rather than at scheduled milestones, and each hop is a different legal relationship needing its own basis. Two credits are due. The company states that transfer through the platform is traceable and auditable and minimises the movement of unnecessary datasets, which is a minimisation commitment rather than a general assurance.

And the architecture uses source linked case report forms with integrated source viewers, so a monitor validates a data point against the originating clinical record rather than a transcription, which is better science. That second feature is also the thing to examine most closely, because it grants sponsor or research organisation monitors a view into a provider's own record.

Establish what that view exposes, whether it is scoped to enrolled participants and to protocol required fields, whether access is logged and visible to the provider, and how it is revoked. Also unpublished: where de identification occurs if at all, what the regulator receives and in what form, and retention at each hop.

Strong on both scale and comparative rigor. Network reach is specific: 45 states, 166 provider organizations, 2,100 care locations including 123 community oncology practices and 43 health systems and academic centers of which 23 are NCI designated, which the company frames as putting 70 percent of the US cancer population within reach of a trial site.

Outcome claims include health systems accruing four times faster and a named practice reporting a 45 percent enrollment increase over two years. Most importantly, the head to head comparison against rule based matching was presented as an abstract at a 2025 ASCO symposium rather than asserted in marketing. Figures remain vendor reported.

Converted from Not Rated. The prior note asked exactly the right question, to map what patient level data crosses from provider to sponsor and at what point in the funnel. That question has since grown a third leg and become more urgent.

Since April 2026 the company has operated a collaboration with the United States drug regulator under which trial data is analysed and key safety and efficacy signals are reported to sponsors and to the regulator in near real time, with named pharmaceutical companies participating. So the flow is now provider record, to this company's platform, to the sponsor, to a federal regulator, continuously rather than at scheduled milestones. Each hop is a different legal relationship and each needs its own basis.

Two credits are due. The company states that transfer through the real time platform is traceable and auditable and minimises the movement of unnecessary datasets, which is a data minimisation commitment rather than a general assurance. And its architecture uses source linked electronic case report forms with integrated source viewers, so a monitor validates a data point against the originating clinical record rather than against a transcription.

That second feature is also the thing to examine most closely, because it grants sponsor or contract research organisation monitors a view into a provider's own record. Establish what that view exposes, whether it is scoped to enrolled participants and to the fields the protocol requires, whether access is logged and visible to the provider, and how it is revoked.

What is not published: where de identification occurs in the pipeline if it occurs at all, what the regulator receives and in what form, retention at each hop, and what happens to a provider's data when a study closes or the relationship ends.

Converted from Not Rated. Business associate status is structurally required and nothing is published, which is the same conclusion the prior note reached, but there are now three distinct legal questions rather than two and only one of them is a business associate question.

First, the platform integrates directly into provider electronic health records across a network the company describes as spanning thousands of care locations. Receiving identifiable records on behalf of a provider makes the company a business associate of each of them. No agreement availability statement, no role characterisation and no identification of the contracting entity was located.

Second, identifying trial eligible patients within routine care means reading records before anyone has consented to a study. The instrument for that is institutional review board approval with an authorisation or a waiver, held by someone, and the question of whether the vendor screens as the provider's agent or on its own account determines who that someone is. Nothing published addresses it.

Third, and this one is new and specific to this company. Under its regulator collaboration, trial data now moves to a federal agency continuously rather than at scheduled submission points. A participant's informed consent describes who receives their data and for what. Adding a real time channel to a regulator is a change in that description, not merely a change in timing. Establish whether consent documents at participating sites cover it, whether the transfer is of identifiable, coded or aggregated data, and under what authority the agency holds it.

A buyer should not treat a strong answer on any one of these as covering the other two. They answer to different authorities and fail in different ways.

Converted from Not Rated. No SOC 2, HITRUST, ISO 27001 or equivalent attestation was located and there is no trust centre.

One thing needs stating carefully so it is neither over credited nor missed. The company has engineered infrastructure to a federal regulator's specification, working with that agency since early 2026 to define reporting and validation protocols and to establish software and data interoperability. That is genuine external technical engagement of a kind almost no vendor in this index has, and a buyer will hear it in a sales conversation. It is not a security attestation. The work described concerns reporting protocols, validation criteria and interoperability, not an examination of access management, change control, incident response or the vendor's own environment. Regulatory collaboration and security assurance are different things and the first does not evidence the second.

The absence carries more weight here than the grade alone conveys, because of what the platform touches. Software embedded in provider record systems across a network the company describes as thousands of care locations in three countries, moving participant level data continuously to sponsors, to a contract research organisation partner and to a federal agency. That is a concentration point, and a compromise would not be confined to one organisation's data.

Ask whether any attestation exists and what it is scoped to, what the embedded component transmits outbound, what access company staff hold to a deployed instance with what logging, and how sponsor tenancies are isolated from one another given that competing sponsors run studies across the same provider network.

No FDA pathway applies to trial matching software, which identifies candidates for research rather than informing diagnosis or treatment. The governing regime is human subjects research regulation, meaning IRB oversight, informed consent, and Good Clinical Practice, plus sponsor obligations under trial protocols. Worth noting for buyers that an algorithm influencing who gets offered a trial has equity implications that sit inside research ethics rather than device regulation.

Equitable access is the company's stated founding mission, and the network strategy is a substantive response: extending trial availability into 123 community oncology practices rather than concentrating in academic centers directly addresses the geographic and socioeconomic skew in trial participation. That is more than positioning.

What is missing is measurement: no published analysis of whether AI surfaced candidates differ demographically from those found by manual screening, which is the specific risk when a model trained on documentation patterns decides who gets seen.

This company documents an architectural change and what it cost to validate, which is a form of disclosure almost nobody in this index attempts. It describes moving away from fine tuned specialised models built use case by use case, each requiring separate training and validation, toward general purpose model evaluation, and reports needing roughly one tenth of the data to validate the newer approach.

Publishing that you replaced your own architecture is an admission that the earlier one had limits, and quantifying the validation burden tells a buyer something concrete about how the change alters the evidence position rather than just the engineering.

The current design is documented as structured prefilters plus reasoning over clinical notes, which is a real description: the prefilter stage is where hard eligibility criteria should be applied deterministically, and knowing it exists tells a reviewer which failures would be logic errors and which would be reading errors. Held at C because nothing measures the result.

No model versioning practice, ongoing evaluation methodology or performance by trial complexity was located, and trial complexity is the variable that matters, since a protocol with a handful of structured criteria and one with layered biomarker and prior therapy requirements are different problems. No warranty, indemnity or remediation commitment was found. Ask for performance stratified by protocol complexity, how versions are managed, and what the prefilter stage handles versus the reasoning stage.

Deep EHR integration is the precondition for the product and is described as such: the platform deploys inside participating provider organizations, integrating with site EHR systems to evaluate structured and unstructured clinical data including molecular and genomic testing results against protocol criteria, with embedded screening workflows in clinical workflow rather than a separate portal. Sustaining that across 166 provider organizations and 2,100 locations is a substantial integration footprint rather than a pilot.

Converted from Not Rated. The prior note flagged international expansion as raising cross border questions the published materials do not address. The expansion is now confirmed and the questions are unchanged.

The company describes its provider network as spanning the United States, Japan and Israel, across a large number of care locations. Japan and Israel each maintain their own data protection regimes with their own rules on transferring personal information abroad, and neither is answered by a United States compliance posture. A trial running across that network moves participant data between three jurisdictions.

Nothing was located on hosting location, region availability, tenancy separation between provider organisations or between sponsors, retention, or subprocessors. That last one matters more than usual here, because the company also partners with a global contract research organisation and now transmits to a federal regulator, so the set of parties touching the pipeline is larger than a two party vendor relationship.

One architectural point does emerge from the company's own description and it cuts both ways. The platform embeds data capture directly into the clinician's existing record system rather than requiring separate entry, which reduces duplication and keeps the source authoritative. It also means the company's software is operating inside the provider's environment, which raises exactly the questions an in boundary deployment always raises and which are not answered: what the component transmits outbound, what telemetry leaves, and what the vendor can reach during support.

Ask for the deployment topology per jurisdiction, and for the transfer mechanism relied on for each cross border flow.

No published pricing, but the commercial structure is disclosed in outline and is unusual enough to matter: sponsors and CROs pay for access to a research ready provider network, while provider organizations gain trial access and recruitment support. The company reports relationships with 75 percent of the top 20 largest biopharma sponsors. A provider should establish what it pays, what it receives, and what obligations attach, since the economics flow primarily from the sponsor side.

Oncology is the clear center of gravity, evidenced by the NCI designated center count, community oncology network, and cancer population reach framing, though the platform supports Phase I through IV trials generally and the company has launched a trial design service for interventional and pragmatic studies embedded in routine care. Settings span community practice through academic medical center, which is the meaningful breadth here, with announced international expansion beyond the US network.