Transforming Chart Review in Clinical Trials: From Manual Effort to AI Efficiency
September 17, 2026
An experienced research coordinator spent 80 hours, or two full work weeks, reviewing 198 patient records to screen for a heart failure trial. Twelve patients were eligible. [3]
That’s the kind of arithmetic behind clinical trial enrollment. This inefficiency slows down clinical trials to the effect of missing enrollment timelines, driving coordinators to leave the field, and running out the clock so that patients who would have qualified aren’t screened in time to participate.
The Problem: Clinical Trials Are Expensive, and Unstructured Data Is a Major Bottleneck
The Time Cost Per Patient Is Steep
Pre-screening a single patient for a Phase III clinical trial takes on average 50 minutes of manual chart review by a trained coordinator. Because up to 88% of patients screened ultimately do not meet eligibility criteria, it takes an average of more than 7 hours of coordinator time to identify a single eligible patient for a Phase III trial. [1,2] In fields like oncology and neurology, where eligibility criteria can exceed 60 conditions per protocol, the time compounds quickly.
Manual Screening Is a Rate-Limiting Step in Trial Enrollment
Industry data on clinical trials shows that more than 80% of trials fail to meet their original enrollment timelines. [5] Enrollment challenges alone account for 30% of Phase III trial terminations, simply due to the inability to find and enroll enough patients fast enough. [2]
A University of Pennsylvania clinical trial designed specifically to study this problem described manual eligibility screening as a rate-limiting step in the clinical research enterprise, largely because eligibility data is buried in unstructured notes and there is insufficient coordinator bandwidth to surface it. [6]
Electronic health records have made much of clinical data easier to access and manage. But the richest, most clinically meaningful information about a patient still lives in free-text notes.
The Downstream Cost: Capacity First, Then Revenue
The first thing a site loses to slow pre-screening is capacity; the coordinator spends hours on charts that will never produce an enrollment and cannot take on additional protocols because they’re out of hours. The financial consequences follow from there. Every day a trial runs behind its enrollment timeline is a day of:
- Delayed per-patient payments from sponsors.
- Extended fixed infrastructure costs, including salaries, regulatory fees, site overhead, and clinical trial management software licensing.
- Milestone payment delays, since many sponsor contracts pay on enrollment milestones rather than elapsed time.
- Wasted research coordinator time.
The Five Pain Points in Trial Chart Review
Pain Point 1: Eligibility Criteria Are Complex and Growing
The median number of eligibility criteria per clinical trial has surged by 58% over the last two decades, from 31 criteria in 2001-2005 to 49 in 2016-2020. [2] In oncology, neurology, and infectious disease, criteria counts regularly exceed 60. Each criterion requires the coordinator to search a different part of the chart, cross-reference different note types, and make a clinical judgment call about how the documentation maps to the protocol language.
Pain Point 2: Unstructured Notes Are Not Easily Searchable by Standard EHR Queries
The most clinically meaningful information is typically in free text. Labs, vitals, and coded diagnoses are structured and can be queried. But the information about why a patient was treated, how they responded, what the pathology showed, and what the clinician’s impression was all lives in the clinical notes.
This problem is especially acute for eligibility determinations, which frequently hinge on exactly this kind of nuanced clinical detail: tumor histology, prior treatment response, symptom burden, and comorbidity characterization. The core challenge is that the nuance of a patient’s history and the richest clinical detail are largely absent from structured fields.
Pain Point 3: Inconsistency and Error Across Coordinators
Manual abstraction is not only slow but also variable. Different coordinators read the same note differently. The same coordinator may read a note differently on a Monday morning than on a Friday afternoon after reviewing 30 charts.
Studies routinely require dual abstraction and inter-rater reliability checks to manage this variability, which doubles the labor burden without doubling enrollment. The Penn AI-augmented abstraction trial required coordinators to achieve 80% agreement on a practice set before being allowed to abstract study charts. In addition to being inefficient, this was an acknowledgment that uncontrolled variability in chart abstraction is a recognized data quality risk in clinical research. [6]
Pain Point 4: Research Coordinator Burnout and Turnover
Turnover among clinical research coordinators is high, and the cost is not simply a replacement salary. It includes recruitment and marketing, education and development, lost productivity during orientation and training, and the emotional toll on the coordinators who stay. Each departing coordinator takes protocol knowledge, calibration, and institutional memory with them. Each new hire requires months of training before reaching full productivity.
Pain Point 5: Manual Abstraction Scales Linearly
The math of manual abstraction does not scale. Adding patients to a study means proportionally more chart review hours. Adding studies means proportionally more coordinators. The only way to grow a research program without proportional growth in headcount is to reduce the per-patient labor burden of chart review.
The Technique Is Already Validated, Now Adoption Is the Gap
Applying language models to chart abstraction is not a novel proposition, and the published results are encouraging. Cleveland Clinic researchers have reported using large language models to extract unstructured text data for research chart review. [4] A real-time automated screening system at Cincinnati Children’s Hospital Medical Center demonstrated that automated eligibility screening can run against live emergency department records. [7]
What has held adoption back at most institutions is not whether the technique works. It is governance, setup cost, and the requirement for programmer involvement in every study.
Introducing Brim
Brim is an AI-guided chart abstraction platform developed at Vanderbilt University Medical Center (VUMC) by Dr. Daniel Fabbri, PhD, a faculty member in biomedical informatics. Dr. Fabbri observed the chart abstraction problem firsthand across hundreds of research projects and other workflows, and set out to build a tool that used AI to make chart abstraction fast, accurate, and secure. Brim’s core technology is based on a project at Vanderbilt, which has received $3.9 million in total funding from the Advanced Research Projects Agency for Health (ARPA-H).
Brim is in production today at academic medical centers including UCSF, Vanderbilt, Johns Hopkins, and UNC, supporting clinical registries, academic research teams, and clinical trial sites. Brim routinely achieves agreement higher than 90% with significant time savings. In one Johns Hopkins example, Brim had a 96.7% mean accuracy across 7 registry variables in pancreatic cancer, comparing favorably with an ontology-driven NLP platform.
How Brim Works
Brim processes unstructured clinical notes using large language models (LLMs) to extract structured data, then presents those extractions to a human coordinator with the source text highlighted as evidence. The coordinator confirms, overrides, or refines the result. Over time, Brim improves from that feedback on a per-project basis, while keeping the patient data inside the institution.
The platform is designed explicitly for non-programmers. Once Brim is deployed at an institution, clinicians, researchers, and coordinators can define the variables they need extracted in plain language, with no data science, prompt engineering, or IT involvement required.
What This Looks Like for Trial Pre-Screening
Consider the heart failure screen example from the beginning of this article: 198 candidate charts, 80 hours of chart review, and twelve eventual enrollees. Here's what using Brim might look like for that trial:
- Define the criteria once. A coordinator writes each eligibility criterion as a variable in plain language using the same language they would use to explain the criterion to a new hire.
- Run the cohort, not the chart. Brim processes the notes for all 198 candidates against the full criteria set, rather than one coordinator working through one chart at a time.
- Review all of the charts, or only the most important ones. Brim enables full visibility into every value so that projects that review every data point still save significant time. Coordinators can also choose to review only the included patients, or only the high-confidence results, creating an even more efficient workflow.
- Confirm against the source. Every extracted value appears alongside the sentence in the note that supports it. The coordinator can Accept, Edit, or Remove the answer, and leave a note with their reasoning. This combines the speed and breadth of AI with the judgment of a human reviewer.
- Automatic refinement. Corrections improve extraction for the remaining batches of the same study, and criteria definitions are reusable across protocols with overlapping requirements.
Brim shifts the experience of chart review for clinical trials from poking around through the patient’s chart to reviewing and deciding based on conclusions.
Brim’s Five Core Advantages
Brim Addresses the Most Common Governance Concerns
The most common objection to AI tools in clinical research is data governance. Brim’s architecture is designed specifically to address it.
It’s Time to Stop Losing Patients to Paperwork
Manual chart abstraction is a structural bottleneck. It costs health systems millions in delayed trial revenue, drives coordinators out of the field, and means that patients who could benefit from clinical trials never get screened in time to participate.
The constraint is not coordinator effort or commitment. It is that many criteria require a manual search through free text, a cost that grows in lockstep with a program’s ambition. Brim transforms chart review from searching through the note to reviewing and validating a draft abstraction.
See it on your own protocol. The fastest way to evaluate Brim is against charts your team has already abstracted manually, so you can compare its output to a known answer key. If you’re interested, email enterprise@brimanalytics.com or book a demo.
References
[1] Penberthy LT, Dahman BA, Petkov VI, DeShazo JP. Effort required in eligibility screening for clinical trials. J Oncol Pract. 2012;8(6):365-370. doi:10.1200/JOP.2012.000646
[2] Callies A, Bodinier Q, Ravaud P, Davarpanah K. Real-world validation of a multimodal LLM-powered pipeline for high-accuracy clinical trial patient matching. Commun Med. 2025. doi:10.1038/s43856-025-01256-0. Preprint: arXiv:2503.15374
[3] Adupa AK, Garg RP, Corona-Cox J, Shah SJ, Jonnalagadda SR. An information extraction approach to prescreen heart failure patients for clinical trials. arXiv:1609.01594 [cs.CL]. Preprint posted September 6, 2016
[4] Ejaz A, Shu J, Dalton J, Deshpande A, Wong KK. P-1965. A smart way of chart review for research: utilizing large language models (LLM) for extraction of unstructured text data. Open Forum Infect Dis. 2026;13(suppl 1):ofaf695.2132. doi:10.1093/ofid/ofaf695.2132
[5] Wandile PM. Patient recruitment in clinical trials: areas of challenges and success, a practical aspect at the private research site. J Biosci Med. 2023;11(10):103-113. doi:10.4236/jbm.2023.1110010
[6] Parikh RB, Kolla L, Beothy EA, et al. Human-AI teaming to improve accuracy and efficiency of eligibility criteria prescreening for oncology trials: a randomized evaluation trial using retrospective electronic health records. Nat Commun. 2026;17:2306. doi:10.1038/s41467-026-68873-8
[7] Ni Y, Bermudez M, Kennebeck S, Liddy-Hicks S, Dexheimer J. A real-time automated patient screening system for clinical trials eligibility in an emergency department: design and evaluation. JMIR Med Inform. 2019;7(3):e14185. doi:10.2196/14185
[8] McPhaul T, Kreimeyer K, Baras A, Botsis T. Automated extraction of cancer registry data from pathology reports: comparing LLM-based and ontology-driven NLP platforms. medRxiv. Preprint posted March 23, 2026. doi:10.64898/2026.03.20.26348915
[9] Brim Analytics. Case study: reducing surgical rescheduling from weeks of manual review to hours with Brim. Accessed September 14, 2026