FHIR-native AI agents and clinical trial enrollment: from weeks to hours

Clinical trial enrollment is one of medicine’s most expensive and persistent operational failures. Roughly 80 percent of clinical trials fail to meet their enrollment targets on time, and the average per-site screening timeline — from protocol development to first eligible patient identified — stretches into weeks or months for complex eligibility criteria. The consequence is not just delayed drug development; it is delayed access to potentially life-saving treatments for patients who are eligible but never identified.

The emerging application of FHIR-native AI eligibility screening agents addresses this bottleneck directly. These systems connect to health system EHR data through FHIR APIs, parse trial eligibility criteria in real time, and generate continuously updated lists of potentially eligible patients — reducing screening timelines from weeks to hours in early implementations. But making these systems work at scale requires infrastructure prerequisites, regulatory clarity, and institutional commitment that are not uniformly in place.

Why trial enrollment remains a bottleneck

The conventional enrollment process is astonishingly manual. A clinical research coordinator receives a protocol, manually reviews inclusion and exclusion criteria, queries the EHR using whatever search tools are available — which may be limited — and generates candidate lists that are then individually chart-reviewed. For trials with complex eligibility criteria, this process for a single site can take weeks. Patients meeting eligibility criteria are routinely missed because the manual process cannot comprehensively screen the full patient population.

The problem is compounded by the fact that eligibility criteria are written in natural language, often with nested conditions, exclusions based on laboratory values within specific time windows, and requirements that combine diagnosis codes, medication history, and procedure history in ways that do not map cleanly to simple database queries. The knowledge representation challenge alone is substantial.

Biomedical informatics groups have been working on automated eligibility screening for years using rule-based approaches and NLP. What changed is the combination of large language model capability for parsing complex natural language criteria, FHIR APIs that provide standardized programmatic access to EHR data, and the deployment of these tools as AI agents that can operate continuously rather than as one-off batch queries.

What FHIR-native eligibility screening agents do

A FHIR-native eligibility screening agent operates in several stages. First, it ingests the trial eligibility criteria — either from ClinicalTrials.gov via API or from a protocol document — and parses the criteria into a structured representation that can be evaluated against patient data. Modern LLM-based parsers handle the ambiguity and nested logic of natural language criteria substantially better than prior NLP approaches.

Second, the agent connects to the health system’s FHIR server, typically the FHIR R4 endpoint that EHR vendors are now required to support, and pulls the relevant clinical data elements: diagnosis codes, medication lists, laboratory results, procedure history, vital signs, demographic information. The USCDI (United States Core Data for Interoperability) data set defines the minimum required elements, and most major EHR vendors now support USCDI v3 through their FHIR APIs.

Third, the agent evaluates each patient’s record against the parsed eligibility criteria and generates a probability or confidence score for eligibility, along with structured explanations of which criteria are met, which are not, and which cannot be determined from available data.

The result is a continuously updated, systematically comprehensive eligible patient list that clinical research coordinators can use as the starting point for chart review, rather than performing the initial screening manually.

The NEJM AI pragmatic trial framework

Research published in NEJM AI has proposed a framework for evaluating these tools through pragmatic trial designs — comparing standard enrollment processes with AI-assisted enrollment on enrollment rate, screen failure rate, time-to-enrollment, and population diversity outcomes. This framework matters because the evidence base for FHIR-native screening agents is still developing, and health systems need a principled approach to evaluating performance in their specific context.

The pragmatic trial framework also addresses an important validity question: does comprehensive AI-assisted screening change the characteristics of enrolled populations in ways that affect trial generalizability? Early evidence suggests yes — AI screening tends to identify patients at smaller sites, from underrepresented demographic groups, and with atypical clinical presentations who would be missed by manual processes. This is a benefit from a population representativeness standpoint but requires attention from trial designers.

Interoperability prerequisites

Health systems looking to deploy or participate in FHIR-native eligibility screening programs need to assess their FHIR infrastructure honestly. The technical prerequisites are specific.

FHIR R4 server availability is the baseline requirement — and while CMS has mandated FHIR R4 APIs for payer-provider data exchange, health system FHIR server implementations vary considerably in completeness and query performance. A FHIR API that supports patient-level queries but cannot support population-level queries efficiently will be inadequate for systematic screening applications.

USCDI data completeness matters more than FHIR API availability. If laboratory results are missing from the structured FHIR data because they come from a reference lab that is not integrated into the EHR’s FHIR data model, the eligibility screening agent cannot evaluate laboratory-based criteria. Assessing data completeness against the eligibility criteria relevant to target therapeutic areas is a prerequisite step.

Patient consent and IRB authorization for using clinical data for trial matching must be in place. This varies by institution and by data use agreement structure. Some health systems have established blanket research authorization frameworks that cover AI-assisted screening; others require study-specific authorization.

Regulatory status and what health systems need to participate

The regulatory status of FHIR-native eligibility screening agents depends on how they are positioned. A tool that identifies potentially eligible patients for coordinator review — where the coordinator makes the eligibility determination — is generally regarded as research informatics infrastructure rather than a medical device requiring FDA clearance. A tool that makes autonomous eligibility determinations without human review enters more complex regulatory territory.

Most commercial implementations in the current market take the coordinator-review model, which is both the appropriate clinical safeguard and the clearer regulatory path. Health systems should confirm this design with any vendor they evaluate.

For health systems looking to participate as sites in trials using these tools, the practical requirements are: an assessed and deployment-ready FHIR R4 implementation, a data governance framework that authorizes the use, a clinical research team trained on AI-assisted screening workflows, and a sponsor or coordinating center that has done the implementation work to integrate the tool with the health system’s FHIR endpoint.

The technology is ready. The bottleneck to broader adoption is infrastructure and institutional readiness, not capability.