The Enrollment Gap

Clinical trial enrollment is in a state of systemic underperformance. According to data from the Tufts Center for the Study of Drug Development, approximately 80% of clinical trials fail to meet their original enrollment timelines.¹ A CenterWatch study of several hundred investigative sites found that upwards of 90% of clinical trials fail to meet timelines as indicated in the study contract,² and that 11% of research sites fail to enroll a single patient in a given trial.³

The consequences fall on multiple stakeholders. For pharmaceutical sponsors, each day of delay in a late-stage oncology trial can cost between $600,000 and $8 million in foregone revenue.⁴ Across the industry, slow enrollment is estimated to consume $50 billion annually in direct costs, extended CRO contracts and the downstream value destruction of delayed launches.¹

A cross-sectional study published in JAMA Surgery, examining 2,542 randomized clinical trials registered on ClinicalTrials.gov between 2010 and 2014, found that approximately 1 in 5 trials were completed within their planned timeframe, with a median delay of 12.2 months. Less than half of trials met their prespecified enrollment target upon completion, with the median shortfall accounting for 31% of the planned sample size.⁵

For academic medical centers, the impact is less visible but equally consequential

• Sponsored research funding that was budgeted for a trial does not materialize if enrollment targets are not met, directly affecting research revenue.
• Investigators whose studies fail to enroll become less competitive for future grant funding and industry partnerships.
• Patients who could have accessed novel therapies through a clinical trial are never identified or approached.
• Institutional reputation in competitive research networks is shaped, in part, by a site's track record of delivering enrolled patients on schedule.

“

These are not the failures of individual investigators or research teams. They reflect a structural problem in how eligible patients are identified, one that technology has not yet adequately solved.

The conventional explanation for poor enrollment is that there are not enough eligible patients. In most cases, this is inaccurate. The more precise explanation is that eligible patients exist within health system EHRs but cannot be systematically identified using current tools.

2.1The unstructured data problem

Up to 80% of clinically relevant information in electronic health records resides in unstructured data: physician notes, discharge summaries, radiology reports, operative notes and free-text clinical narratives.⁶ This is not a marginal share of the clinical record. It is the majority of it. And critically, it is the portion most likely to contain the nuanced clinical detail, such as disease severity, prior treatment response and comorbidity specifics, that determines trial eligibility.

Research published in the Journal of the American Medical Informatics Association demonstrated that structured EHR data achieved an average recall of 51.7% in identifying clinical concepts relevant to patient eligibility, while analysis of unstructured notes achieved 95.5% recall for the same task.⁷ In practical terms, a recruitment process that relies only on structured EHR fields will miss approximately half of the patients who actually qualify.

A study published in medRxiv analyzing over 19 million EHR notes from more than 500,000 patients found that NLP analysis of unstructured notes contributed a 29.5% relative increase in patients identified as smokers beyond what structured data alone captured, and a 19.3% relative increase for obesity.⁸ For eligibility criteria that depend on detailed clinical history, as most trial protocols do, the gap between structured-only and full-record analysis is material.

2.2The protocol complexity problem

Clinical trial inclusion and exclusion criteria are written in clinical language: free text, medical terminology and context-dependent conditions that cannot be directly translated into structured database queries. A coordinator screening patients for a complex oncology or rare disease trial may be working against 30 to 50 individual criteria, each requiring a different type of data from a different part of the patient record.

Tufts CSDD studies have consistently shown a positive association between protocol complexity, longer timelines, poorer patient recruitment and retention rates, and a higher incidence of protocol amendments.² More than 40% of trials amend their protocol before the first subject visit, delaying trials by an average of four months.⁹

2.3The coordinator burden problem

The current standard for patient identification in most research operations is manual chart review by clinical research coordinators. A time-and-motion study of clinical research coordinators in a pediatric emergency department found that patient screening consumed 31.6% of coordinator time, the single largest category of their work.¹⁰

An information extraction study examining heart failure patient screening found that an experienced research coordinator required two weeks, 80 hours, to manually pre-screen 198 patient records against a single trial's eligibility criteria, ultimately identifying 40 potentially eligible patients for further review.¹¹ An automated AI system completed the same pre-screening task in approximately 20 minutes.¹¹

A commentary published by the Association of Clinical Research Professionals (ACRP) in 2026 noted that research staff currently spend hours on pre-screening and must often manually evaluate large pools of potential participants for trial eligibility, diverting time and energy away from meaningful patient interaction.¹² At an institution running multiple concurrent studies, this burden is multiplicative, and it compounds directly into enrollment underperformance.

The enrollment gap is not closed by recruiting harder. It is closed by finding patients more accurately and earlier, using the data that already exists within the institution's EHR. This requires a fundamentally different approach to how patient records are structured, queried and connected to active research protocols.

3.1From episodic to longitudinal: the core distinction

Most EHR search tools operate on episodic data: what is recorded in structured fields during a specific encounter. A longitudinal patient record, by contrast, synthesizes the full clinical history of an individual patient across all encounters, data types and time periods. Structured fields, unstructured notes, lab values, imaging reports, medication history and historical visits, in one place.

This distinction is material for clinical trial matching. Many eligibility criteria require evidence of a condition or prior treatment over a defined historical period, not just what was coded in the last encounter. A patient who received a specific therapy 18 months ago, documented only in a clinical note, is invisible to structured-field search. They are not invisible to longitudinal patient intelligence.

3.2How Prompt Opinion works

Prompt Opinion is an AI-powered clinical intelligence platform built natively on FHIR (Fast Healthcare Interoperability Resources) standards. Developed in close collaboration with health system partners, the platform ingests both structured and unstructured EHR data across a broad range of clinical data types, including clinical notes, laboratory results, radiology reports, pathology, medications, procedures, encounters and genomics, and makes this data continuously queryable against active research protocols.

The platform connects to health system EHR environments, including Epic, through SMART on FHIR APIs: the same standardized, vendor-approved integration framework that major EHR systems already expose for third-party applications. Implementation works within existing data infrastructure rather than requiring custom integrations, parallel data pipelines or direct database access.

For health IT and security teams evaluating the platform, Prompt Opinion operates inside a certified environment: HITRUST e1, SOC 2, HIPAA and ONC Certified Health IT. These are the core security and compliance frameworks that health system security and vendor assessment teams require, and they are available for review as part of any institutional security assessment.

The clinical trial matching workflow

1. 1

   Protocol ingestion

   Inclusion and exclusion criteria are extracted from trial protocols and translated from clinical language into structured, queryable logic.

2. 2

   Longitudinal record construction

   The platform builds a continuously updated, FHIR-native patient record by ingesting structured EHR fields, clinical notes, lab reports, imaging and historical encounter data, normalizing terminology across ICD, SNOMED, LOINC and RxNorm and reconciling events across encounters and time.

3. 3

   AI eligibility matching

   Each patient's longitudinal record is matched against the extracted protocol criteria, with the AI scoring and ranking patients by likelihood of eligibility.

4. 4

   Missing data identification

   For patients who are close to eligible but have a data gap, for example a required lab value not yet on record, the system flags them separately so coordinators can decide whether the gap can be resolved, rather than dismissing a potentially eligible patient.

5. 5

   Coordinator dashboard

   Coordinators receive a prioritized, actionable worklist rather than a chart review queue: new matches, candidates needing a specific lab review, patients ready for outreach and cases requiring clinical review. The coordinator's role shifts from searching to acting. Each candidate arrives with the matching rationale surfaced and the next step pre-determined.

3.3A note on implementation and burden

Any new technology introduced into a clinical research environment must be evaluated not only on its potential benefit but on the burden it places on already-stretched teams. Prompt Opinion is designed with this constraint as a first principle, not an afterthought.

For academic medical centers, the consequences of systematic enrollment underperformance fall across two dimensions that are deeply connected but rarely addressed together: the operational burden on research teams, and the revenue yield from the institution's sponsored research portfolio. Improving enrollment performance addresses both simultaneously.

4.1Research operations

Systematic enrollment underperformance creates a specific set of operational and strategic pressures for research leadership:

• Studies that miss enrollment targets require timeline extensions, additional site activations or protocol amendments, each of which consumes research operations resources without proportional benefit.
• Coordinators allocated to studies that enroll poorly are effectively under-utilized relative to their cost, while coordinators on high-demand studies are over-stretched.
• Investigator confidence in the institution's ability to deliver enrollment affects which studies investigators choose to pursue and whether external sponsors select the institution as a site.
• Research networks and multi-site collaborations increasingly track per-site enrollment performance. Poor performance affects future inclusion.

AI-assisted patient matching addresses each of these pressures by improving the yield of the existing research operations infrastructure: more enrolled patients per coordinator hour, more studies meeting their enrollment targets, and a stronger institutional track record for future sponsor engagement.

4.2Research revenue

The financial case for improving clinical trial enrollment at an academic medical center operates on two dimensions: revenue generation and cost efficiency.

A validation conducted with a major academic medical center demonstrated strong clinical interest and feasibility, with institutional security review processes, standard for health systems of this scale, identified as the primary pathway to full deployment. This validation confirms both the technical capability of the platform and the appetite among research leadership for this category of solution.

The immediate application of AI-powered longitudinal patient intelligence is clinical trial matching: identifying eligible participants faster, with less coordinator burden, across a broader share of the EHR population. But the infrastructure required to do this well, a continuously updated, FHIR-native, AI-queryable patient record, creates the foundation for a broader set of capabilities that academic medical centers will increasingly need.

The path from today's enrollment gap to tomorrow's research intelligence infrastructure runs through a single decision: making the patient data that already exists inside the EHR continuously accessible, structured and queryable for research purposes. Prompt Opinion is built to make that transition as low-friction as possible for institutions ready to take it.

The clinical trial enrollment problem is well understood. Its persistence is not a reflection of insufficient effort. It is a reflection of a structural gap between the data that exists in health system EHRs and the tools available to translate that data into research participation.

For academic medical centers, closing this gap is not merely an operational improvement. It is a strategic imperative: for research revenue, for investigator retention, for institutional competitiveness in sponsored research, and for fulfilling the fundamental mission of translating clinical knowledge into patient benefit.

“

The patients who qualify for your active studies are already in your EHR. The question is whether your current infrastructure can find them.