Brim Field Notes Recap: How VUMC Cut Chart Review Time for Surgery Rescheduling
October 27, 2026
Before a child has surgery at Vanderbilt's pediatric ambulatory surgery center, a nurse practitioner has to confirm they are healthy enough to be there. That review used to start with a blank chart and 30 minutes of reading. Today it starts with Brim.
In the first session of Brim Field Notes, our conversation series on real AI chart abstraction work, Brim CEO Dan Fabbri talked with Beth Bass, MSN, APRN, FNP-BC, NE-BC, Advanced Practice Manager at Vanderbilt University Medical Center (VUMC). Beth led the project that brought Brim into clinical use at VUMC for identifying which patients needed to be rescheduled from the ASC to Main Campus. She was candid about what worked, what broke, and her advice for teams undergoing similar projects.
After eight months of blinded validation, Brim went from 62% to 99% accuracy against expert review. It has been live in the clinical workflow since March 2026 as a tool to help nurse practitioners make this call.
The problem: essential screening that doesn't scale
Monroe Carell Jr. Children's Hospital at Vanderbilt is busy. To keep the ORs at the main campus available for the most complex patients, VUMC opened a freestanding pediatric ambulatory surgery center (ASC) in Murfreesboro, about 40 minutes away.
The ASC only takes otherwise healthy children having routine procedures, under strict, CMS-required anesthesia eligibility criteria. The list covers nearly every body system: cardiac defects, severe sleep apnea, uncontrolled asthma, diabetes, bleeding disorders, personal or family history of anesthesia problems, airway-affecting syndromes, BMI, and more.
Verifying those criteria meant reading charts by hand. "A lot of this is buried," Beth said. A nurse practitioner might be three or four notes deep before finding the history that matters. The numbers added up fast:
- About 30 minutes per comprehensive chart review
- Up to 3 reviews per child: at scheduling, with the family before surgery, and on the day of surgery
- 188% growth in surgical volume from 2021 to 2025, to nearly 6,000 cases a year
And the same nurse practitioners doing those reviews also care for patients at the ASC on the day of surgery.
The Approach: Verification, not Discovery
Beth and Dan were introduced by the Chief of Pediatric Surgery. After their first meeting, Beth saw a fit: the ASC's red and green eligibility criteria were already written down as yes-or-no questions.
The team turned those criteria into Brim variables with precise definitions. Brim reads each scheduled patient's notes from Epic and uses the variables and rules to recommend whether the patient should be rescheduled to the main campus. Every True answer links to the source note, so a clinician can see exactly why Brim flagged it and confirm or override in a click.
From day one, two design choices prioritized patient safety:
- Clinicians are the gold standard. "Their decision is the decision," Beth said. Brim never decides eligibility. It moves the likely-ineligible patients to the top of the list.
- No review step goes away. Every patient still gets the full CMS-required chart review, a history with a parent or guardian, and a day-of-surgery check with anesthesia. If Brim misses something, those steps catch it.
The team also tuned Brim to err on the conservative side. Beth would rather see a flag for neurofibromatosis type 1, which the ASC can handle, than risk missing type 2, which it can't.
Before Brim, nurse practitioners reviewed charts in schedule order: the 7 o'clock patient, then the 8 o'clock, with no risk stratification. Now they open Brim first, clear the flagged patients, and are able to reschedule them weeks earlier. That frees the ASC slot for another patient and gives families more notice. Beth described the shift as moving from discovery to verification.
Results: Validating Before Moving into Clinical Use
Phase 1: Blinded Validation
Before Brim touched a real scheduling decision, the team ran it on 3,425 past cases over eight monthly cycles. The validation was blinded both ways: Brim never saw the nurse practitioners' reviews, and they never saw Brim's output. Beth then compared the two case by case, with the clinicians as ground truth.
The first cycle came in at 62% accuracy, mostly false positives. "We did not want to create more work for an APP by deploying a model that was 62% accurate," Beth said. The errors broke down into a few categories:
Accuracy climbed to 81%, then 96%, then 99%, and held at 98 to 99% through the last cycles.
Governance
Brim was the first AI clinical decision support tool VUMC's IRB had reviewed. Approval took about six months and two bodies: the IRB and VUMC's AI technology subcommittee. Reviewers wanted it "clear as day" that clinicians, not software, make the decisions. The team also documented how Brim fits FDA's definition of clinical decision support and put a conflict-of-interest plan in place for Dan's role on the research team.
Phase 2: Going Live
Since launching on March 30 2026, Brim has reviewed more than 3,000 patients in clinical use with:
- 96% average accuracy
- 97% specificity, so few unnecessary moves to the main campus
- 90% sensitivity, agreement with nurse practitioners on patients who truly need to move
Beth was candid that sensitivity swings week to week. Only a handful of patients each week are truly ineligible, so one miss moves the number a lot. Some of those misses are gray-area cases the chart can't capture, like how much support a child with autism needs, or a family history that only comes out when a nurse calls the parents. That is exactly why every review step stays in place.
The team also kept improving the workflow after go-live. For example, nurse practitioners found a weekly Monday batch too heavy, so uploads moved to daily.
Beth's playbook for leaders
Asked what she would tell a health system starting this next quarter, Beth offered five steps:
- Start from written criteria. Rules your clinicians already apply make the best variables. Turn each one into a true-or-false question.
- Validate blind against experts. Keep the AI and the clinicians from influencing each other, then compare case by case.
- Tune in small steps. Change one variable at a time and watch sensitivity closely.
- Bring governance in early. Prepare a workflow showing human oversight, a user guide, an FDA clinical decision support analysis, and any conflict-of-interest plan before you submit.
- Keep clinicians in control. "AI is a tool to use, but it is not the decision maker."
She added one more: deploying is not the last step. Someone has to own monitoring. At VUMC, nurse practitioners accept or override every output, and Beth reviews overrides to catch drift. Drift can be sneaky. When VUMC gave the abbreviation "VCH" to a newly acquired hospital in Clarksville, the team had to update Brim, because clinicians had long used VCH to mean the children's hospital main campus. Eligibility changes like a new sleep apnea threshold, need the same care.
Dan's advice for teams just starting: manually review 5 to 10 charts first to get a feel for the data, and begin with one clinic or service line rather than everything at once.
What's next
VUMC is piloting the reverse: finding main campus patients who could safely move to the ASC, starting with ENT, its largest surgical division. That would shorten their waits and open main OR time for complex cases. The team is also exploring rapid rescheduling during disruptions like January's ice storm, and flagging implanted devices placed outside VUMC that Epic can't surface today. A third phase is measuring clinician burnout and trust in AI before and after implementation. These are all possible because Brim is deployed as shared infrastructure at VUMC.
Watch the full conversation
You can find the full recording below or on YouTube. Beth's paper on the retrospective validation is slated for publication in December, and we'll share it here when it's out.
Thank you to Beth Bass and the PATCH team at Vanderbilt University Medical Center for sharing their work.
If your team is weighing AI for chart review, whether for operations, registries, or research, book a demo to see what Brim can do with your criteria.