Randomized Trial Protocol: Epic Generative AI Chart Summarization Tool to Reduce Ambulatory Provider Cognitive Task Load

This protocol outlines a 90-day, two-arm, parallel-group randomized controlled trial designed to evaluate whether access to a native Epic generative AI chart summarization tool reduces cognitive task load and improves professional fulfillment among ambulatory providers compared to usual care.

The Gist

Imagine you are a doctor. Every morning, before you see your first patient, you have to do something called "pre-charting." This is like a detective reviewing a massive, messy pile of case files from the last few years to remember everything about the person sitting in your waiting room. You have to scroll through hundreds of pages of notes, lab results, and old emails just to get up to speed.

This process is exhausting. It's like trying to find a specific needle in a haystack while someone is shouting questions at you. Many doctors feel burned out because they spend so much time staring at screens that they have less time to actually talk to patients or go home to their families.

Enter the "Magic Summary Bot."

Epic, the company that makes the computer system most hospitals use, has built a new tool powered by Artificial Intelligence (AI). Think of this tool as a super-fast, super-smart intern. Instead of you reading 30 different notes from the past, you click a button, and this AI reads them all for you in seconds. It then hands you a neat, one-page "cheat sheet" summarizing the most important parts of the patient's history.

The Big Question

The doctors at UCLA Health are asking: Does this magic intern actually help, or does it just add more noise?

While many doctors are already using this tool, no one has done a strict scientific experiment to prove if it really reduces stress and saves time. So, they are running a Randomized Controlled Trial (the gold standard of science experiments).

How the Experiment Works

Imagine a giant lottery. The researchers have invited 284 doctors and nurse practitioners to join.

• Group A (The AI Team): These doctors get the "Magic Summary Bot" for 90 days. They can use it whenever they want to help them prepare for their patients.
• Group B (The Control Team): These doctors continue doing things the old-fashioned way, scrolling through notes manually, without the AI help.

They will switch groups or compare them after three months to see who is happier, less tired, and more efficient.

What Are They Measuring?

The researchers aren't just guessing; they are measuring specific things:

1. Mental Load: They use a survey to ask, "How hard did your brain have to work to get ready for your patients?" (This is called "Cognitive Task Load").
2. Time: They look at the computer logs to see exactly how many minutes doctors spend staring at charts before a visit.
3. Burnout: They ask, "Do you feel more exhausted or more fulfilled in your job?"
4. Patient Experience: They check patient surveys to see if patients feel like their doctor "knew their story" better.

The Catch (Limitations)

The authors are very honest about the flaws in their plan:

• It's a "Try It" Tool: Doctors in the AI group can choose not to use it if they don't like it. This is good for real life, but it makes the science a bit messy because some people might not use the tool enough to show a difference.
• One Hospital: The study only happens at UCLA. What works there might not work at a small rural clinic.
• No Blinding: The doctors know if they have the AI tool or not. This is like a taste test where one group knows they are eating the "new flavor" and the other knows they are eating the "old flavor." Their expectations might change how they feel.

Why This Matters

If this study proves that the AI tool actually saves doctors time and reduces their stress, it could change how hospitals everywhere use technology. It could mean doctors spend less time fighting with computers and more time looking their patients in the eye.

However, if the AI makes mistakes (like "hallucinating" facts) or takes too long to load, it could actually make the doctors' jobs harder. This study is the safety check to make sure the "Magic Summary Bot" is truly a helper and not a headache.

In a Nutshell
This paper is a plan for a 3-month experiment to see if a new AI tool that summarizes patient notes is the "superpower" doctors need to stop burnout, or if it's just another gadget that doesn't quite work as advertised.

Technical Summary

Based on the provided protocol paper, here is a detailed technical summary of the study.

1. Problem Statement

Electronic Health Record (EHR) documentation and chart review are significant drivers of physician burnout, cognitive overload, and "pajama time" (after-hours work). In ambulatory settings, the "pre-charting" process—where providers review historical notes, labs, and results to contextualize a patient's visit before an appointment—is particularly time-consuming and mentally demanding.

While Epic Systems has deployed generative AI (genAI) tools to summarize historical clinical notes to alleviate this burden, there is a lack of rigorous empirical evidence regarding their efficacy, safety, and impact on cognitive task load. Existing data is limited to small, quasi-experimental pre/post studies. Furthermore, concerns regarding AI hallucinations, inaccuracies, and workflow integration risks necessitate a controlled evaluation before widespread adoption.

2. Methodology

Study Design:

• Type: Two-arm, parallel-group, pragmatic Randomized Controlled Trial (RCT).
• Duration: 90 days (February 23, 2026 – May 24, 2026).
• Setting: Ambulatory outpatient clinics within the UCLA Health system (single academic health system using Epic EHR).
• Participants: 284 ambulatory physicians and Advanced Practice Practitioners (APPs) across multiple specialties, working at least half a day per week. Residents, fellows, and psychologists are excluded.
• Randomization: 1:1 allocation to Intervention vs. Usual Care. Stratified by active AI scribe license status. Covariate-constrained randomization ensures balance on baseline cognitive task load and chart review time.

Intervention:

• Tool: Native Epic Generative AI Chart Summarization Tool.
• Functionality: Automatically generates summaries of the most recent clinical notes (progress notes, consults, procedures, H&P, discharge summaries, ED notes) for scheduled patients.
  • Constraints: Limited to 24,000 characters or 30 notes.
  • Features: Includes a "focus on" prompt feature and manual note selection.
  • Timing: Batch-generated ~36 hours prior to appointments; on-demand for same-day visits.
  • Scope: Summaries are derived only from clinical notes (excluding labs/imaging) and serve as informational aids, not clinical decision support.
• Control: Usual care (standard pre-visit chart review without AI summarization).
• Blinding: Unblinded (participants know their assignment).

Outcomes:

• Primary Outcome: Change in Physician Task Load (PTL), a 4-item instrument adapted from the NASA-TLX (Mental Demand, Temporal Demand, Physical Demand, Effort) specifically for pre-charting tasks. Scores range 0–400.
• Exploratory Outcomes:
  • Objective Metrics: EHR-derived time metrics (Caboodle/Signal) for pre-charting activities; time outside scheduled hours; Relative Value Units (RVUs).
  • Psychometric Measures: Professional Fulfillment Index (PFI) for burnout; System Usability Scale (SUS); Self-reported pre-charting effectiveness.
  • Patient Experience: CG-CAHPS metric regarding the provider's knowledge of medical history.
  • Safety: Clinician-reported safety events, hallucination frequency, and accuracy ratings via EHR-native feedback.
• Subgroup Analyses: Stratified by AI literacy (MAILS score), ambient scribe access, specialty, panel complexity (RAF score), age, and sex.

Statistical Analysis:

• Model: Linear mixed models with provider random effects to account for repeated measurements.
• Adjustments: Fixed effects for study arm, response period, and interaction terms; adjusted for baseline characteristics (sex, age, specialty, clinic volume, baseline task load, scribe usage).
• Power: 284 participants provide 80% power to detect a small-to-medium effect size (0.33 SD) at a 0.05 significance level.
• Handling Missing Data: Intention-to-treat analysis assuming Missing At Random (MAR); sensitivity analyses via multiple imputation.

3. Key Contributions

• First RCT of EHR-Native GenAI: This is the first randomized controlled trial to rigorously evaluate an EHR-embedded generative AI chart summarization tool in a real-world clinical workflow.
• Validated Cognitive Metrics: Moves beyond simple time-tracking by utilizing a validated, adapted cognitive task load instrument (PTL) to measure the mental burden of pre-charting.
• Comprehensive Safety & Usability Framework: Integrates objective EHR time metrics with subjective clinician feedback (burnout, usability) and safety monitoring (hallucination reporting) to provide a holistic view of the tool's impact.
• Effect Modification Analysis: Explicitly plans to analyze how AI literacy and existing tool usage (ambient scribes) modify the intervention's effectiveness.

4. Results

• Status: No results are reported in this document.
• Reason: This is a study protocol submitted on February 22, 2026, prior to the completion of the 90-day intervention period. The paper outlines the study design, hypotheses, and analysis plan but does not contain outcome data.

5. Significance

• Clinical Workflow Optimization: If successful, the study could validate a scalable method to reduce physician cognitive load and burnout, potentially improving retention and care quality.
• Evidence-Based Adoption: Provides high-quality data to guide healthcare systems (like UCLA Health) and EHR vendors in making data-driven business decisions regarding the procurement and deployment of genAI tools.
• Safety Benchmarking: Establishes a framework for monitoring AI safety (hallucinations, omissions) in real-time clinical settings, addressing a critical gap in current AI implementation strategies.
• Limitations & Generalizability: The authors acknowledge limitations including the single-center design, short duration (3 months), voluntary participation (selection bias), and lack of blinding. However, the rigorous RCT design sets a new standard for evaluating AI in healthcare compared to previous observational studies.