PaperlayerView exact revision on arXiv

Source-linked AI summary

SDoH-Aware Narrative Anchoring Bias in Medical LLMs for Trustworthy Clinical Decision Support

Ahnaf Atef Choudhury, Ramkrishna Saha

arXiv:2608.22802v1cs.CLcs.AI

TL;DR

Medical LLM accuracy does not reveal whether equivalent clinical cases receive stable answers when patient narratives differ. This paper evaluates that risk with counterfactual persona variants, three Qwen2.5 models, and case-grouped robustness metrics, finding that Qwen2.5 7B is strongest but remains narrative-sensitive.

  • Problem

    Medical LLM evaluation often emphasizes clinical question-answering accuracy without establishing whether answers remain stable across medically equivalent patient narratives.

  • Method

    The study uses NarrativeShield SDoH MedQA persona variants with fixed answer keys to compare three Qwen2.5 instruction-tuned models across matched cases and prompting conditions.

  • Results

    Qwen2.5 7B achieves 56.33 percent accuracy and 40.33 percent correct consistency, while narrative sensitivity remains despite its stronger performance.

  • Takeaways & Limitations

    Trustworthy clinical decision-support evaluation should report both average correctness and stability across equivalent patient narratives.

  • Takeaways & Limitations

    The evaluation uses multiple-choice questions and controlled persona variants, and clinical equivalence was not independently verified by physicians.

Abstract

from arXiv · show

Medical large language models are often judged by how many clinical questions they answer correctly. That view is useful, but it misses a practical risk. A model may know the right answer and still change its response when the same case is written in a different patient voice. This paper evaluates that risk as SDoH aware narrative anchoring bias. We use NarrativeShield SDoH MedQA, a counterfactual medical question answering dataset in which each case appears in persona based narratives while the answer key remains fixed. The dataset is reshaped from wide format into case grouped persona rows. We evaluate three open source instruction tuned LLMs from the Qwen2.5 family: 1.5B, 3B, and 7B. The final experiment uses 300 clinical cases and produces 8,100 model responses across three prompting conditions. We report persona level accuracy, counterfactual consistency, correct consistency, and narrative sensitivity error. Qwen2.5 7B achieves the best accuracy at 56.33 percent and the best correct consistency at 40.33 percent. Paired McNemar exact tests show significant accuracy gains for 7B over 3B in all prompt settings. Even so, narrative sensitivity remains, with the lowest error still at 31.67 percent. These results suggest that trustworthy clinical decision support should be evaluated by both average correctness and stability across medically equivalent patient narratives.

I. INTRODUCTION

Clinical LLM evaluation must test whether answers remain stable when identical medical facts are expressed through different patient narratives. This study formalizes that risk as SDoH-aware narrative anchoring bias and evaluates it with counterfactual cases, matched models, and case-grouped consistency metrics.

  • Clinical support systems should preserve answers when clinical facts are identical, while distinguishing clinically meaningful context from distracting narrative cues.
  • The study names answer changes caused by patient voice or social context, despite unchanged clinical facts and gold labels, SDoH-aware narrative anchoring bias.
  • NarrativeShield SDoH MedQA supplies persona-based variants of medical questions while keeping the answer key unchanged.
  • The evaluation compares three open-source instruction-tuned Qwen2.5 models on the same 300 clinical cases using matched sampling and deterministic decoding.
  • The analysis separates stable correct behavior from unstable responses using case-grouped consistency metrics, bootstrap confidence intervals, and paired McNemar exact tests.

A. Medical Question Answering and Clinical LLMs

Medical QA benchmarks measure clinical knowledge, but prior work also shows that performance can vary with context, social cues, and presentation. This motivates evaluating stability across cases, not only persona-level accuracy.

  • MedQA, MedMCQA, and PubMedQA established medical question-answering benchmarks spanning exam and biomedical research questions.
  • Scaling and instruction tuning improve medical benchmark results, while high scores do not guarantee careful reasoning under changed context or presentation.
  • Biomedical and clinical pretraining improves several medical NLP tasks but does not establish whether answers remain stable across socially varied patient narratives.
  • SDoH extraction research treats social factors as potentially useful care information, whereas this study examines whether related narrative cues create answer instability.
  • Prior medical-AI studies show sensitivity to demographic and social cues, motivating a focus on stability across cases rather than accuracy at the persona level.

III. DATASET AND STUDY DESIGN

The study evaluates medically matched persona narratives from NarrativeShield SDoH MedQA through a case-grouped counterfactual pipeline. It retains all variants for 300 sampled cases and compares three prompting conditions.

  • NarrativeShield SDoH MedQA contains multiple-choice medical cases with fixed answer keys and persona-based narrative variants that change presentation style.
  • Wide-format records are reshaped into long-format persona rows so variants sharing a question_id can be grouped directly after inference.
  • The evaluation assumes matched persona variants have the same clinical answer because the dataset supplies a fixed answer key, without independently verifying clinical equivalence.
  • The final sample contains 300 unique cases, 900 persona-level rows, and 8,100 responses across three models and three prompt settings.
  • Zero-shot, few-shot, and bias-aware prompts provide the core comparison, with few-shot prompts adding two solved examples.
  • The prompting comparison is intentionally lightweight and tests whether brief instructions reduce narrative anchoring rather than serving as a full fairness intervention.

B. Model Selection

The study uses three Qwen2.5 instruction-tuned models from one family to isolate model-scale effects under controlled evaluation conditions. Its metrics distinguish answer agreement, uniformly correct cases, and narrative sensitivity.

  • B. Model Selection: Qwen2.5 1.5B, 3B, and 7B Instruct models enable a controlled within-family comparison using the same data, decoding, and prompts.
  • B. Model Selection: The model selection supports reproducibility with typical academic computing resources and avoids dependence on paid clinical APIs.
  • C. Evaluation Metrics: Counterfactual consistency measures whether all persona variants of a case receive the same predicted answer.
  • C. Evaluation Metrics: Correct consistency is stricter because every persona variant must receive the correct answer.
  • C. Evaluation Metrics: Narrative sensitivity error counts cases where at least one persona is correct while another equivalent persona is wrong.

V. EXPERIMENTAL SETUP

The experiment uses paired, deterministic evaluation with bootstrap intervals and exact McNemar tests to compare models under matched conditions.

  • All models see the same case, persona, and prompt combinations, enabling paired comparisons at that level.
  • Deterministic decoding reduces sampling noise, and each response is parsed into one answer letter.Invalid outputs are counted as incorrect; they were rare across all models.
  • Bootstrap 95 percent confidence intervals use 2,000 resamples.
  • Paired McNemar exact tests compare accuracy because every model receives the same case, persona, and prompt combination.

VI. RESULTS

Accuracy improves with model scale, and the 7B model also improves case-level stability, but moderate accuracy limits clinical-readiness claims.

  • 56.33 percent is Qwen2.5 7B's best accuracy, compared with 38.44 percent for 1.5B and 48.78 percent for 3B under zero-shot prompting.The same scale pattern appears under few-shot and bias-aware prompting.
  • 40.33 percent is the highest correct consistency, achieved by Qwen2.5 7B under zero-shot and bias-aware prompting.
  • 57.67 percent is Qwen2.5 7B's counterfactual consistency under bias-aware prompting.Scaling to 7B improves both accuracy and case-level stability, although robustness gains from 1.5B to 3B are smaller and uneven.
  • Moderate absolute accuracy means these models should not be interpreted as clinically ready systems.The experiment supports controlled comparison between accuracy and robustness, not clinical validation.
  • Accuracy rises steadily from 1.5B to 7B across prompts, while counterfactual consistency also improves at 7B but varies by prompt.Average correctness and case-level stability should therefore be reported separately.

B. Narrative Sensitivity

Narrative sensitivity measures whether equivalent persona variants receive inconsistent answers, and it remains common even for the strongest model.

  • 31.67 percent is Qwen2.5 7B's lowest narrative sensitivity error under bias-aware prompting.Lower values indicate fewer cases where one equivalent persona variant is correct and another is wrong.
  • Narrative sensitivity occurs when at least one persona variant is correct and at least one is wrong.It is a case-level failure mode in which the model has enough information to answer one version correctly but does not preserve that answer across versions.
  • The case-level error decomposition distinguishes stable correct, stable wrong, narrative sensitive, and unstable wrong outcomes.This separates missing medical knowledge from instability caused by the narrative.
  • About 40 percent is the 7B stable correct rate, while the 3B model remains below 30 percent.Narrative-dependent cases remain common despite the stronger model finding the right answer more often.

D. Persona Level Accuracy

Persona-level accuracy is descriptive, while the main analysis emphasizes case-level consistency and tests whether larger models improve accuracy reliably.

  • Accuracy varies across alpha, beta, and gamma narratives, supporting grouped case analysis beyond aggregate accuracy.Persona-level differences are reported as descriptive evidence rather than the primary evaluation target.
  • The main analysis focuses on case-level consistency, correct consistency, and narrative sensitivity error.
  • Qwen2.5 7B is significantly more accurate than 3B in all prompt settings according to paired McNemar exact tests.
  • Bias-aware prompting does not significantly improve accuracy over zero-shot prompting.Telling the model to ignore irrelevant narrative cues does not reliably improve accuracy.
  • Higher accuracy and counterfactual consistency intervals for 7B show that scaling improves performance without fully solving narrative sensitivity.

VII. DISCUSSION

Scaling improves both clinical answer accuracy and case-level robustness, but narrative sensitivity persists across equivalent patient voices. The results therefore support reporting average correctness alongside stability metrics and treating robustness as a core evaluation target.

  • VII. DISCUSSION: Qwen2.5 7B improves accuracy and correct consistency over smaller models, but most cases remain unstable across all three persona variants.Correct consistency reaches 40.33 percent under zero-shot and bias-aware prompting.
  • VII. DISCUSSION: Bias-aware prompting does not significantly outperform zero-shot prompting and does not remove narrative sensitivity error.A short instruction to ignore irrelevant narrative cues is therefore insufficient for dependable behavior in this evaluation.
  • VII. DISCUSSION: Accuracy and case-level stability should be reported separately because higher average correctness can coexist with answer shifts across equivalent narratives.The same overall score can conceal different levels of stability across patient narratives.
  • VII. DISCUSSION: Narrative robustness is clinically relevant because response differences may affect triage advice, diagnostic suggestions, or follow-up guidance.The concern is not removing clinically meaningful social context, but preventing excessive reactions to unchanged clinical facts and answers.

VIII. LIMITATIONS AND REPRODUCIBILITY

The study improves reproducibility through public data, fixed sampling, available models, and traceable row-level outputs, but its conclusions are constrained by controlled multiple-choice evaluation, persona construction, physician-unverified equivalence, and a narrow model set.

  • The evaluation uses multiple-choice questions, which improves repeatability but does not capture full diagnostic reasoning or treatment planning.
  • Controlled persona variants isolate patient voice but may not represent complex clinical conversations with missing information, mixed languages, or clarification rounds.
  • Clinical equivalence across persona variants comes from dataset structure and fixed answer keys rather than independent physician verification.
  • The model set is limited to three Qwen2.5 instruction-tuned models, so findings should not generalize to all general-purpose or medical-domain LLMs.
  • The reproducible pipeline uses public data, seed 42, available models, and saved answers, labels, prompt conditions, and case identifiers.

IX. CONCLUSION

The study finds that narrative anchoring persists in compact to mid-sized medical LLMs: Qwen2.5 7B is stronger than 3B, yet answer stability remains imperfect. It therefore supports evaluating clinical systems by both correctness and consistency across equivalent patient narratives.

  • Qwen2.5 7B was more accurate and robust than Qwen2.5 3B across the 300-case, 8,100-response evaluation.
  • Bias-aware prompts did not improve model accuracy compared with zero-shot prompts.
  • Even the stronger model remained affected by changes in how patient stories were told.
  • Reliable clinical decision support should report whether answers stay consistent when similar cases are described in different ways, alongside average accuracy.
Loading 2608.22802v1…