Understanding Clinician Edits to Ambient AI Draft Notes: A Feasibility Analysis Using Large Language Models

This study demonstrates that few-shot prompted large language models can effectively categorize specific types of clinician edits to ambient AI draft notes, achieving strong performance for medication and symptom modifications while identifying complex, context-dependent edits as better suited for human review triage.

The Gist

Imagine a doctor's office where a super-smart, tireless robot assistant sits in on every patient visit. This robot listens to the conversation and instantly types up a draft of the medical notes. It's fast and saves the doctor time, but it's not perfect. Sometimes the robot misses a detail, gets a dosage wrong, or phrases a symptom in a way that doesn't sound quite right.

So, the doctor has to read the robot's draft, fix the mistakes, and hit "sign."

The Problem:
The researchers wanted to know: What exactly is the doctor fixing? Are they mostly changing the medicine list? Are they tweaking the diagnosis? Are they adding social details like "patient lives alone"?

In the past, to find this out, a team of human experts had to sit down and read thousands of these drafts, manually tagging every single change. It was like hiring a team of editors to check every sentence of a novel by hand—slow, expensive, and exhausting.

The Experiment:
The team asked: "Can we teach a different, smaller AI (a Large Language Model) to act as a 'detective' and automatically sort these changes for us?"

They didn't train the AI with thousands of examples (which takes a lot of data). Instead, they used a technique called "few-shot prompting." Think of this like giving the AI a cheat sheet with just a few examples of what to look for, saying, "Here is a medicine change. Here is a symptom change. Now, look at this new text and tell me what kind of change it is."

The Results: The "Easy" vs. "Hard" Cases

The study found that the AI detective was a master at some jobs but struggled with others.

1. The "Easy" Cases (Medications & Symptoms):

   • Analogy: Imagine looking for a red apple in a fruit bowl. It's bright, distinct, and easy to spot.
   • Result: When the doctor changed a drug name or a symptom description (e.g., changing "chest pain" to "chest tightness"), the AI was very good at spotting it. It got the job right about 78% of the time.
   • Why? These edits usually have "anchors"—specific words like drug names or symptom terms that stand out clearly.

2. The "Hard" Cases (Diagnoses, Tests, & Social History):

   • Analogy: Now imagine looking for a chameleon that blends perfectly into the leaves. Or trying to tell the difference between a "planned trip" and a "completed trip" just by reading a vague travel log.
   • Result: When the edits were about complex diagnoses, ordering tests, or social details (like housing or family support), the AI got confused. It kept shouting "I found a change!" when there wasn't one (false alarms), or it missed subtle shifts in meaning.
   • Why? These edits often rely on context and implication. A doctor might change a sentence from "possible virus" to "confirmed flu" without using the word "diagnosis." The AI, lacking the deep clinical intuition of a human, struggled to understand the weight of that change.

The "Safety Gate" Strategy
To make the AI more reliable, the researchers added a "verification gate." They told the AI: "Don't just guess. If you say 'Yes, this is a medicine change,' you must point to the exact word in the text that proves it."

This was like asking a security guard: "Don't just say 'that person is a VIP.' Show me their badge." This rule helped stop the AI from making wild guesses, though it couldn't fix the confusion on the "chameleon" edits.

The Bottom Line: A New Workflow

The paper concludes that we shouldn't try to use this AI to do everything automatically. Instead, we should use it as a smart filter:

• For the "Red Apples" (Medications/Symptoms): Let the AI do the work. It's fast and accurate enough to automatically track trends (e.g., "Hey, doctors are changing the dosage of Drug X in 20% of notes").
• For the "Chameleons" (Diagnoses/Social): Use the AI as a triage tool. Let it flag the likely changes and say, "Human, please take a quick look at these specific notes." It doesn't replace the human; it just tells the human where to look, saving them from reading every single page.

In short: The AI is a great assistant for spotting obvious, concrete changes, but for the subtle, complex medical reasoning, it still needs a human partner to double-check its work.

Technical Summary

1. Problem Statement

Ambient AI systems generate draft clinical notes from visit audio, which clinicians must review and edit before finalizing in Electronic Health Records (EHR). While these systems improve efficiency, there is a lack of scalable methods to characterize how and where clinicians modify these drafts.

• The Gap: Manual annotation of edits is resource-intensive and does not scale. Traditional Natural Language Processing (NLP) classifiers struggle with the heterogeneity of edits and the need for deep clinical context (e.g., distinguishing between a diagnosis change and a symptom description).
• The Challenge: Existing supervised classifiers require large, task-specific training datasets and constant maintenance as clinical language evolves.
• The Goal: To evaluate the feasibility of using few-shot prompting with Large Language Models (LLMs) to automatically categorize sentence-level edits into specific clinical domains (Medication, Symptom, Diagnosis, Orders/Tests, Social History) under realistic deployment constraints (HIPAA compliance, limited GPU resources, no external APIs).

2. Methodology

Data and Task Formulation

• Dataset: Derived from 200 clinical encounters at University of California Health (late 2023–mid 2025). The dataset consists of 713 sentence-level edit units (paired "Before" AI draft and "After" clinician-finalized text).
• Task: Five label-specific binary classification tasks:
  1. E-Med: Medication-related edits.
  2. E-Sym: Symptom-related edits.
  3. E-Dx: Diagnosis-related edits.
  4. E-Test: Test/Order/Procedure-related edits.
  5. E-Soc: Social history edits.
• Splits: Data was partitioned into Training (n=313, for example selection), Development (n=200, for prompt iteration), and a Held-out Test set (n=200, for final evaluation).

Model and Infrastructure

• Model: meta-llama/Llama-3.2-3B-Instruct (3B parameters).
• Environment: HIPAA-compliant AWS environment on a single NVIDIA T4G GPU (16 GB VRAM).
• Constraints: No fine-tuning; inference limited to 30 seconds per request; deterministic decoding.

Prompt Engineering Strategy

The authors employed an iterative refinement strategy to improve performance:

1. Few-Shot Learning: Used 4–6 positive and 18–20 negative examples per category.
2. Adversarial Negatives: Added specific "negative" examples that mimic the target category but should be classified as negative (e.g., mentioning a drug name without changing the regimen) to reduce False Positives (FPs).
3. Verification Gate: A mandatory self-check requiring the model to output present=true only if it can extract verbatim "anchor" evidence (e.g., a specific drug name or diagnosis code) that directly corresponds to the edit.
4. Structured Output: Enforced JSON output containing a binary decision and the evidence span.

3. Key Contributions

• Feasibility Framework: Demonstrated that few-shot prompting can categorize clinical edits without fine-tuning, provided strict prompt engineering controls are applied.
• Prompt Refinement Protocol: Established a workflow involving adversarial negatives and evidence verification gates to manage the precision-recall trade-off, specifically improving precision in high-risk categories.
• Granular Error Analysis: Identified that performance is highly dependent on the "explicitness" of the edit. Categories with clear textual anchors (medications) perform well, while context-dependent categories (diagnosis, social history) suffer from boundary ambiguity.
• Operational Constraints: Validated the approach under strict resource constraints (3B model, single GPU, HIPAA environment), proving it is viable for privacy-preserving, on-premise deployment.

4. Results

Performance Metrics (Held-out Test Set, n=200)

The models showed a clear divide between "anchor-rich" and "context-dependent" categories:

Edit Category	Precision	Recall	F1 Score	Performance Note
Medication (E-Med)	0.774	0.800	0.787	Strongest. High precision due to explicit drug names/doses.
Symptom (E-Sym)	0.657	0.959	0.780	High recall, but lower precision due to overlap with diagnosis/plan text.
Diagnosis (E-Dx)	0.560	0.836	0.671	Moderate; struggles with framing vs. management language.
Test/Order (E-Test)	0.523	0.831	0.642	Precision-limited; confuses planned vs. performed tests.
Social History (E-Soc)	0.483	0.933	0.636	Weakest. High recall but very low precision; often misinterprets general narrative as social history.

Error Analysis

• False Positives (FPs): Primarily caused by boundary ambiguity. For example, symptom models triggered on treatment plans; diagnosis models triggered on medication changes.
• False Negatives (FNs): Occurred in delete-only edits (where the "After" text is empty) and long, complex edit units where the target anchor was buried in a list of multiple actions.
• Runtime: Inference approached the 30-second limit as prompts grew longer with more examples and verification steps.

5. Significance and Implications

• Hybrid Workflow Recommendation: The study concludes that prompt-only LLMs are not a "drop-in" replacement for human review across all categories.
  • High-Confidence Automation: Suitable for Medication and Symptom edits where explicit anchors exist. These can be used for scalable quality monitoring.
  • Triage/Assistive Role: For Diagnosis, Orders, and Social History, the models should be used for triage (high-recall candidate retrieval) to flag edits for targeted human review, rather than for automated scoring, due to high false-positive rates.
• Scalability vs. Accuracy Trade-off: Adding detailed verification gates improves precision but increases latency. Future work suggests using larger models or parameter-efficient fine-tuning (PEFT) for context-heavy categories.
• Clinical Impact: This approach provides a pathway for health systems to continuously monitor AI draft quality, identify systematic errors in AI generation, and prioritize workflow improvements without requiring massive labeled datasets or compromising patient data privacy.

Conclusion: The paper establishes that while LLMs can effectively categorize edits with explicit textual cues, their reliability diminishes for complex, context-dependent clinical reasoning. A practical deployment strategy involves using these models for automated monitoring of clear edits and as a triage tool for ambiguous ones.