See Me, Believe Me: Causality, Intersectionality, and Interventions Improving the Appearance of Patients

arXiv:2410.01227 · cs.LG, cs.AI · Submitted 2026-08-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "See Me, Believe Me: Causality, Intersectionality, and Interventions Improving the Appearance of Patients".

Jane: The paper was written by Kenya S. Andrews, Mesrob I. Ohannessian and Elena Zheleva from Brown University and University of Illinois Chicago.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and Authors: Tom: Welcome back to the show, everyone! Today we’re digging into a paper that’s got a title that just grabs you: “See Me, Believe Me: Causality, Intersectionality, and Interventions Improving the Appearance of Patients.” Jane, I have to say, that title alone made me sit up.

Jane: It really does, Tom. And it’s from Kenya Andrews at Brown, along with Mesrob Ohannessian and Elena Zheleva at UIC. They’re tackling something huge here—how the words doctors write about patients can actually undermine those patients’ experiences.

Tom: Right, and they call that “testimonial injustice.” Basically, when a patient says something and the person listening just doesn’t take them seriously because of who they are.

Jane: Exactly. And what’s wild is that this isn’t just about being rude. It’s about how clinical notes get written. Words like “complains” or “claims” instead of “reports” or “describes.” Those little choices can make a patient sound less credible.

Tom: So the paper is saying that your race, your gender, your age—these things can actually predict how you’re going to be written about in your own medical record?

Jane: That’s the core of it. And they’re using something called causal discovery to actually map out how those features lead to unjust language. It’s not just correlation; they’re trying to show cause.

Tom: Cause and effect in medical notes. That’s a big step. I mean, we’ve known for a while that bias exists in healthcare, but to actually model the causal pathways?

Jane: Yeah, and that’s what makes this paper exciting. They’re not just pointing at the problem. They’re building a map of how it happens, so we can figure out where to step in.

Tom: And stepping in is exactly what they try to do. They actually edit the notes and see if it changes how doctors and even AI systems perceive the patient.

Jane: Right. So it’s not just a diagnosis of the problem. It’s a treatment plan. And that’s what we’re going to get into.

Tom: I’m hooked already. Let’s keep going.

Paper Summary: Jane: So, Tom, let’s break down what this paper actually does. They start with a dataset called MIMIC-III, which is a huge collection of real ICU medical records.

Tom: ICU records—so these are really sick patients. High stakes.

Jane: Very high stakes. And they look at four categories of unjust language: evidential terms like “complains,” judgmental terms like “claims,” negative terms like “combative,” and stigmatizing terms like “addict” or “non-compliant.”

Tom: And they’re counting how often these show up in notes for different groups of patients?

Jane: Exactly. They look at race, gender, and age—specifically whether someone is in a marginalized group. So, for race, that’s Black or Latino patients. For gender, that’s women. And for age, that’s children under fifteen or seniors over sixty-five.

Tom: So they’re looking at the intersections too. A young Black woman, an older Latino man—those combinations.

Jane: Precisely. And they use a causal discovery algorithm called FCI to build a structural causal model. That’s a fancy way of saying they map out which demographic features actually cause which types of unjust language.

Tom: And what did they find?

Jane: Some really interesting stuff. Race directly causes judgmental and stigmatizing terms. Age directly causes evidential and judgmental terms. And gender is linked to negative terms. But here’s the kicker—these don’t work alone.

Tom: Right, because intersectionality is the whole point.

Jane: Yes. For example, age influences stigmatizing terms, but only through judgmental terms first. So it’s not just that being older gets you stigmatized—it’s that being older gets you judged, and that judgment opens the door to stigma.

Tom: So the pathways matter. It’s not just “old people get bad words.” It’s a chain reaction.

Jane: Exactly. And that’s what causal discovery gives them that a simple correlation study couldn’t. They can see the order of operations, so to speak.

Tom: That’s powerful. So once they have this map, what do they do with it?

Jane: That’s the intervention part, and that’s what we’re going to talk about next.

Improvements Suggested: Tom: Okay, so they’ve built this causal map. Now they want to actually fix things. What’s their approach?

Jane: They call it “intentional interventions.” They edit the physicians’ notes to remove or replace the unjust language. And they do it in two different ways.

Tom: Two ways—what are they?

Jane: The first is rule-based. They use the causal model to figure out which patients are most likely to experience which type of injustice, and then they apply targeted edits. So if the model says a patient is likely to get stigmatizing language because they’re marginalized by race, they focus on removing stigmatizing terms from that patient’s notes.

Tom: So it’s like a precision strike. You only edit what needs editing.

Jane: Right. They don’t want to just scrub all language everywhere, because that could erase nuance. They want to be selective. And they use things like replacing “complains” with “expressed concerns about,” or removing unnecessary adverbs like “blatantly.”

Tom: And the second approach?

Jane: That’s the context-based approach. They use an LLM—GPT-4o-mini—to edit the notes with instructions to reduce negative and stigmatizing language and inject empathy where needed.

Tom: So you’ve got a scalpel and a sledgehammer. Or maybe a scalpel and a very smart assistant.

Jane: Ha, that’s a good way to put it. The rule-based is precise, but it only catches what’s in their lexicon. The LLM can understand context, but it might make changes that are less predictable.

Tom: And then they test both approaches. They ask human experts—doctors, nurse practitioners—and another LLM, Gemini, to read the edited notes and answer questions about the patient.

Jane: Exactly. They want to see if the edits change how the patient is perceived. Do they seem more credible? Does their condition seem more urgent? Does the blame shift away from the patient?

Tom: And what did they find? I’m on the edge of my seat here.

Jane: Well, the results were pretty encouraging. Both types of edits made the LLM express more urgency and provide more detailed reasoning about the patient’s condition. And for the human experts, the rule-based edits specifically shifted blame away from the patient and toward systemic factors.

Tom: That’s huge. Just changing a few words can change whether a doctor thinks the patient is responsible for their own illness.

Jane: Exactly. And that’s the kind of impact that could change patient outcomes. But we should get into the specifics of how they measured all this.

First Page Discussion: Tom: So Jane, let’s go back to the very beginning of the paper. The abstract sets up this whole idea of testimonial injustice in medical records.

Jane: Right. And they make a really important point early on. They say that patients are not just vulnerable because they’re sick—they’re vulnerable because they’re dependent on the person writing about them.

Tom: That’s a profound observation. You’re in the hospital, you’re scared, and the person who’s supposed to help you is writing down words that might make other people not believe you.

Jane: And they cite studies showing that clinicians are more likely to dismiss the concerns of Black and female patients. But the key contribution here is that they’re looking at the intersections—younger Black females, senior White males, Latina females—because the experience isn’t the same for everyone.

Tom: So it’s not just “women get dismissed.” It’s “young Black women get dismissed in this specific way, and senior White men get dismissed in that specific way.”

Jane: Exactly. And they also bring up a really timely concern: LLMs are being used more and more in clinical settings. If these models are trained on biased notes, they’re going to absorb those biases.

Tom: So it’s not just about human doctors reading the notes. It’s about AI systems making decisions based on them.

Jane: Right. And they cite research showing that even when race isn’t explicitly mentioned, LLMs can infer it from the language used in the notes. So the bias gets baked into the model.

Tom: That’s terrifying, but also motivating. If we can fix the language, we can fix the model.

Jane: That’s the hope. And they frame their research questions really clearly: Can we identify how demographic features influence language in medical settings? And can we quantify that influence to make intentional interventions?

Tom: And that’s exactly what they do. They build the causal model, they quantify the influence, and they test the interventions.

Jane: Right. And one thing I love about this paper is that they’re not just talking about fairness in the abstract. They’re actually measuring it. They’re showing that these edits change how patients are perceived, which is a concrete, measurable outcome.

Tom: So it’s not just “be nicer.” It’s “here’s the data showing that being nicer changes the outcome.”

Jane: Exactly. And that’s what makes this paper so important. It’s not just a critique; it’s a solution.

Conclusion: Tom: Alright, Jane, let’s wrap this up. We’ve been talking about “See Me, Believe Me: Causality, Intersectionality, and Interventions Improving the Appearance of Patients” by Andrews, Ohannessian, and Zheleva.

Jane: And what a ride it’s been. They’ve shown that testimonial injustice in medical notes isn’t random—it follows predictable causal pathways based on race, gender, and age.

Tom: And they’ve shown that those pathways can be mapped and understood, which is the first step to fixing them.

Jane: Right. And then they actually fixed them. They edited the notes, and they showed that those edits changed how both human experts and AI systems perceived the patients.

Tom: The rule-based edits shifted blame away from patients, and the context-based edits increased urgency and clarity. Both approaches made a difference.

Jane: And that’s the takeaway—small, targeted changes in language can have measurable effects on how patients are perceived and treated.

Tom: It’s a reminder that words matter, especially in high-stakes settings like the ICU.

Jane: Absolutely. And it’s a call to action for the whole field. We need more research on this, and we need to start implementing these insights in real clinical settings.

Tom: Well said. Thanks for joining us on this deep dive. We’ll be back with more papers soon.

Jane: Until then, keep reading, keep questioning, and remember—see me, believe me.

Tom: See you next time!

Kenya S. Andrews, Mesrob I. Ohannessian, Elena Zheleva

Brown University · University of Illinois Chicago

cs.LG, cs.AI

Submitted: 2026-08-08

Updated: 2026-08-11

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 40/100

The gist: This paper investigates "testimonial injustice" in medical records, a phenomenon "where the textual account undermines the validity of [patients'] experiences." The researchers aim to "use causal

Key concepts

Testimonial Injustice
This occurs when a person's testimony is dismissed or not taken seriously by others because of who they are. The paper focuses on how clinical notes can write about patients in ways that undermine their credibility.
Causal Discovery
A method used to map out cause-and-effect relationships, rather than just correlation. The researchers used this algorithm to show how demographic features (like race or age) actually lead to specific types of unjust language in medical records.
Intersectionality
This concept highlights that a person's experience is not defined by single characteristics, but by the combination of multiple identities (e.g., being a young Black woman). The paper shows that injustice affects these combinations uniquely.
Intentional Interventions
The process of actively editing medical notes to remove or replace unjust language. The hosts discuss two methods: rule-based edits and context-based edits using an LLM.

Terminology

Summary

This paper investigates testimonial injustice in medical records, a phenomenon where the textual account undermines the validity of [patients'] experiences. The researchers aim to use causal discovery to study the degree to which certain demographic features tied to marginalization (namely age, gender, and race), together lead to specific types of testimonial injustice terms and to test strategic interventional methods along the intersectional paths of injustice.

Causal Discovery and the Structural Causal Model (SCM)

Using the MIMIC-III dataset, the researchers performed causal discovery via the Fast Causal Inference (FCI) algorithm to build a Structural Causal Model (SCM). They binarized demographic features to identify marginalized groups: is marginalized gender = 1 if a person is female, is marginalized race = 1 if a person is Black or Latino, and is marginalized age = 1 if a person is a child (age ≤ 15) or a senior (age ≥ 65). To quantify injustice, they focused on four categories of unjust terms: evidential, judgmental, negative, and stigmatizing. Injustice was defined as is testinj = 1 if any of the categories of terms is present in 'high' numbers, specifically the 90th percentile.

The resulting SCM shows a direct causal edge from both race and age to judgmental terms, indicating that these factors jointly influence the outcome in an intersectional manner. The model also revealed that experiencing stigmatizing terms is affected both directly by race and indirectly by age via the mediator of judgmental terms, and experiencing evidential terms is affected both directly by age and indirectly by gender via negative terms, as well as by race via stigmatizing terms.

Intervention Methods

The researchers implemented two distinct editing approaches to address these injustices:

  • Rule-based: Causality-Informed Modifications: These modifications were guided by the insights we gleaned from the SCM and logistic regression models. They utilized three methods: word removal (removing unjust adverbs ending in -ly), empathy reframing (rephrasing content to detach patients from their conditions), and preferred rephrasing and word replacement (using synonyms suggested by experts).

  • Context-based: LLM Modifications: This approach utilized OpenAI’s GPT-4o-mini acting as "an expert editor of clinician notes who wants to reduce negative and stigmatizing language in text... and reduce evidential and judgment, one can inject empathy/use alternative empathic rephrasing or remove words when possible."

Evaluation and Results

The impact of these interventions was assessed through LLM analysis (using Google DeepMind’s Gemini 2.0 Experimental Flash) and a survey of US Board-Certified medical experts (Physicians, Nurse Practitioners, and Physician Assistants).

  • LLM Analysis: The study found that "urgency & causal reasoning increased in the rule-based approach and even more significantly in the context-based approach. While the unedited notes were efficient and generally neutral, the edited versions produced more detailed justifications, linking symptoms to potential conditions or diseases in a causal manner."

  • Human Expert Analysis: Experts reported that the blame for the patient’s condition tended to shift from the patient to poor medical care in the rule-based approach and to unfounded causes in the context-based approach. Specifically, rule-based edits resulted in the least blame attributed to patients, with more blame assigned to unfounded or environmental causes. Like the LLM results, human experts also observed an increase in "urgency & causal reasoning" in both edited versions.

  • Sentiment Analysis: Sentiment analysis revealed that both rule-based edits and LLM-based edits reduced the negative sentiment across the notes compared to the unedited versions. The LLM-based edits introduced a more positive tone, while the rule-based edits softened the negative tone slightly but did not drastically change the overall sentiment distribution, a trait the authors note is much more desirable in medical writing.

The paper concludes that small, targeted interventions in documentation can have measurable downstream effects on perception and decisionmaking, and that making minimal intentional changes accordingly can effect improved patient perception.

Improvements for AI systems

1. Causal-Path-Targeted Debiasing Engines

  • What the improved system can do: Instead of using generic toxicity filters that may strip away necessary clinical nuance, this system uses Structural Causal Models (SCMs) to identify specific intersectional paths of injustice (e.g., the causal link between Race + Age to Judgmental Terms). The AI can then apply surgical, counterfactual edits that specifically break these causal links—such as replacing stigmatizing descriptors with neutral, evidence-based observations—without altering the core clinical facts of the patient's condition.

2. Hybrid Neutrality-Reasoning Clinical Documentation Assistants

  • What the improved system can do: This system functions as a real-time co-pilot for clinicians that balances two competing needs: clinical precision and social neutrality. It uses rule-based logic (word removal and empathy reframing) to ensure the note remains medically neutral and prevents patient-blaming language, while simultaneously using context-aware LLM reasoning to enhance the urgency and causal reasoning of the note. This ensures the final documentation is both socially equitable and clinically high-quality.

3. Intersectional Bias Auditing Frameworks for Medical Datasets

  • What the improved system can do: This is an automated diagnostic tool for developers of medical AI. It uses Fast Causal Inference (FCI) to scan massive medical corpora to map bias topologies. It can specifically alert developers if a training dataset contains hidden intersectional injustices—such as how the combination of Gender + Age might be driving Evidential Injustice—allowing for targeted data cleaning before the model is ever trained.

4. Counterfactual Data Augmentation for Robust Clinical Training

  • What the improved system can do: This system acts as a synthetic data generator to create fair training sets. By applying the interventional methods described in the paper (like empathy reframing and preferred rephrasing), the AI can take biased, real-world medical records and generate counterfactual versions of those same records. These unbiased versions can then be used to train downstream diagnostic models, teaching them to make decisions based on symptoms rather than the biased linguistic patterns associated with marginalized demographics.

5. Real-time Blame-Shift Feedback Loops for Clinicians

  • What the improved system can do: Integrated into Electronic Health Record (EHR) interfaces, this system provides immediate feedback to providers. When a clinician types language that attributes a condition to a patient's character or lifestyle (testimonial injustice), the AI suggests rephrasings that shift the focus from patient blame to environmental or medical causes. This helps clinicians maintain professional neutrality and ensures that the patient's medical history is documented in a way that does not trigger downstream diagnostic bias.

Sources

Related papers