💻 computer science

Implicit Framing in Obstetric Counseling Notes: A Grounded LLM Pipeline on a VBAC-Eligible Cohort

This paper utilizes a grounded LLM-based pipeline to analyze obstetric clinical notes, revealing that physicians use significantly more risk-focused language when documenting counseling for repeat cesarean sections compared to vaginal birth after cesarean.

Original authors: Baris Karacan, Barbara Di Eugenio, Patrick Thornton, Joanna Tess, Subhash Kumar Kolar

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Baris Karacan, Barbara Di Eugenio, Patrick Thornton, Joanna Tess, Subhash Kumar Kolar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The "Menu" Problem: How Doctors Talk About Your Choices

Imagine you walk into a restaurant. You’re hungry, and you have two main options on the menu: Option A (a healthy, home-cooked meal) and Option B (a quick, processed fast-food burger).

Now, imagine how the waiter describes them to you:

For Option A: "This meal is delicious, gives you long-lasting energy, and helps you feel great all day!"
For Option B: "This burger is fast, but it’s high in salt, might make you feel sluggish later, and could lead to health issues if you eat it too often."

Even though the waiter didn't tell you what to pick, they just nudged you toward Option A without saying a single word. They used "framing." They framed one choice with "wins" (benefits) and the other with "losses" (risks).

This paper is about how doctors "frame" the conversation when a mother is deciding how to give birth.

The Medical Context: The VBAC vs. RCS Choice

When a woman has had a C-section in the past, she often faces a big decision for her next baby:

VBAC: Trying for a vaginal birth (the "natural" route).
RCS: Having another C-section (the "repeat" route).

Both are valid medical choices. However, the researchers wanted to see if the way doctors write about these choices in their medical notes subtly pushes patients toward one or the other.

The Problem: The "Messy Notebook"

Doctors don't just fill out checkboxes; they write long, rambling notes (called "narratives"). These notes are like a messy, handwritten diary. They are full of abbreviations, typos, and complicated medical jargon.

It’s very hard for a computer to read a "messy diary" and figure out, "Did the doctor tell the patient about the risks, or did they encourage the patient's preference?"

The Solution: Using an AI "Translator"

The researchers used a powerful AI (a Large Language Model, similar to ChatGPT) to act as a super-smart assistant. They used it in two steps:

The Fact-Checker: First, they used the AI to scan the messy notes to find missing details—like whether a patient was actually medically allowed to try a VBAC. This ensured they were comparing "apples to apples" (comparing patients who actually had a real choice).
The Tone-Detector: Then, they asked the AI to read the "counseling" parts of the notes and categorize the vibe. Was the doctor being "Risk-Focused" (talking about what could go wrong), "Benefit-Focused" (talking about the upsides), or "Shared Decision-Making" (treating the patient like a partner)?

The Big Discovery: The "Fear Factor"

The researchers found a significant pattern.

When doctors were documenting notes for women who ended up having a repeat C-section (RCS), the language was overwhelmingly Risk-Focused. It was like the waiter focusing heavily on the "sluggishness" and "health issues" of the burger. The notes were filled with warnings about complications and dangers.

However, for women who chose a VBAC, the language was a bit more balanced. The doctors used more "Shared Decision-Making" language and talked more about the "Benefits."

Why does this matter?

Even if a doctor thinks they are being neutral, the way they write things down—and the way they speak—can create a "gravity" that pulls a patient toward one decision.

If the conversation is mostly about "what might go wrong," a patient might feel too scared to try a vaginal birth, even if it is a safe and healthy option for them. This study shows that by using AI, we can catch these subtle patterns and help doctors communicate in a way that truly empowers patients to make the best choice for their bodies.

Technical Summary: Implicit Framing in Obstetric Counseling Notes

Problem Statement

In obstetrics, deciding between a Vaginal Birth After Cesarean (VBAC) and a Repeat Cesarean Section (RCS) is a "preference-sensitive" decision involving complex trade-offs. Clinical "framing"—the linguistic manner in which risks and benefits are presented—can significantly bias patient decision-making. However, analyzing these subtle linguistic patterns at scale is difficult because:

Data Heterogeneity: Counseling information is buried in unstructured, "noisy" free-text Electronic Health Record (EHR) narratives.
Confounding Variables: Comparing counseling in VBAC vs. RCS notes is often biased because many patients who undergo RCS are medically ineligible for VBAC (e.g., due to uterine incision types), meaning their notes wouldn't contain VBAC counseling by necessity rather than choice.

Methodology

The researchers developed a multi-stage Large Language Model (LLM) pipeline to ensure a "cohort-controlled" analysis, meaning they only compared patients for whom both delivery modes were clinically viable.

1. Cohort Construction (Grounded Extraction):

Initial Filtering: Used structured data to identify patients with prior cesareans.
LLM-based Eligibility Recovery: To resolve missing data (specifically uterine incision types), they used Llama 3.1 (8B) and Llama 3.3 (70B) to extract verbatim evidence from H&P narratives.
Hallucination Auditing: They implemented a rigorous "verbatim-matching" procedure to distinguish between true hallucinations, benign paraphrases (e.g., "hx" $\rightarrow$ "history"), and typos.

2. Framing Quantification (Zero-Shot Classification):

Framework: They defined seven linguistic framing categories: Risk-Focused, Benefit-Focused, Directive, Shared Decision-Making, Balanced Information, Statistical Evidence, and Reassuring.
Classification: Used Llama 3.3 70B in a zero-shot setup to categorize short counseling segments from the H&P "Plan" sections. The model was required to provide a category and a grounded rationale.
Validation: An expert clinician reviewed a subset of segments, achieving a moderate Cohen’s $\kappa$ of 0.56, confirming the LLM's ability to identify dominant framing cues.

Key Contributions

A Grounded LLM Pipeline: A method for using LLMs to recover missing clinical eligibility data from free text while strictly controlling for hallucinations through verbatim verification.
Cohort-Controlled Design: A framework that isolates "framing" as a variable by ensuring both delivery options were medically appropriate for all analyzed patients.
Scalable Framing Measurement: A zero-shot NLP approach to quantify complex, qualitative communication styles (like "Shared Decision-Making") in large-scale clinical corpora.

Results

Extraction Reliability: Llama 3.3 70B (with a concise prompt) proved highly reliable, producing zero hallucination variants in several fields. Smaller models (Llama 3.1 8B) showed "instruction-overload" failures, where complex prompts actually increased hallucination rates.
Dominant Framing: In both cohorts, Risk-Focused language was the most prevalent.
Statistical Divergence: A Pearson $\chi^2$ $χ^{2}$ test revealed a significant difference in framing distributions between the two groups ( $p \approx 1.48 \times 10^{-11}$ $p \approx 1.48 \times 1 0^{- 11}$ ):
- RCS Notes: Significantly over-represented by Risk-Focused language (loss-framing).
- VBAC Notes: Showed significantly higher proportions of Shared Decision-Making, Benefit-Focused, Statistical Evidence, and Balanced Information compared to the RCS group.

Significance

The study demonstrates that even when both delivery options are medically viable, the documentation for patients who choose RCS is significantly more "risk-heavy" than for those who choose VBAC. This suggests a potential implicit bias in clinical documentation where the presentation of risks may steer patients toward repeat cesareans. Technically, the paper provides a blueprint for using LLMs in healthcare to move beyond simple entity extraction (e.g., finding "medications") toward analyzing the nuanced, communicative intent of clinical discourse.