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

Clinical language models are evaluated on vignettes written in a clinician's voice, but patients describe the same illness in their own. We pair 1,000 MedQA cases with three retellings that carry the same audited original facts, in a clinical, a socioeconomic and a cultural voice, and evaluate 12 open models (3B to 24B, including two medical fine-tunes). Every model is less accurate when the patient speaks in a marked voice: 26% more errors, a pooled penalty of $-9.6$ accuracy points (95% CI $[-11.0, -8.2]$), with no detectable reduction at larger scale. The penalty is not explained by any retelling property we measured (length, omitted age, MedQA split, added clinical content) and persists when the marked narrative contains the original vignette verbatim; because generation settings differ between voices, we read it as sensitivity to these marked retellings, not proof that register alone is the cause. Chain-of-thought and explicit debiasing instructions together remove only about a fifth of the gap. Having the same model first rewrite the patient's account as a clinical vignette, then answer from the rewrite, halves the gap while raising marked-voice accuracy by 3.9 points (means over the 11 models that can run the pipeline). A matched pipeline using a structured schema instead of prose closes the same gap but is 7.1 points less accurate in all 11 models, and adding reasoning agents or tools costs more still. Rewriting is not free: it recovers more facts than extraction but also invents plausible details, which we quantify. In our experiments, the representation a model reasons over matters more than how much it reasons.

Clinical Large Language Models (LLMs) achieve strong medical-exam accuracy; however, a correct answer does not guarantee that the explanation names the concepts that actually drove the decision. We introduce three lightweight, directly interpretable metrics for this faithfulness gap: the Explanation Stability Index (ESI), which measures reasoning consistency across repeated queries; the Causal Faithfulness Score (CFS), which tests whether cited concepts drive predictions via concept ablation; and the Perturbation Stability Score (PSS), which measures robustness to semantic-preserving paraphrases. By evaluating six LLMs on 150 MedQA-USMLE questions (900 model-question observations), we found that only 23.3% of the cited clinical concepts were causally necessary. Correct answers had lower CFS than incorrect answers (0.212 vs. 0.398), answer consistency negatively predicted CFS (Spearman r = -0.466), and model pairs could agree on answers while sharing only 8.8% of cited reasoning concepts. These results show that accuracy, consistency, and consensus are incomplete safety signals for clinical decision-making support. The evidence is behavioral rather than mechanistic: concept ablation tests counterfactual sensitivity of outputs, not internal circuits.

Small open-source language models are gaining attention for healthcare applications in low-resource settings where cloud infrastructure and GPU hardware may be unavailable. However, the reliability of these models under different phrasings of the same clinical question remains poorly understood. We evaluate five open-source models (Gemma 2 2B, Phi-3 Mini 3.8B, Llama 3.2 3B, Mistral 7B, and Meditron-7B, a domain-pretrained model without instruction tuning) across three clinical question answering datasets (MedQA, MedMCQA, and PubMedQA) using five prompt styles: original, formal, simplified, roleplay, and direct. Model behavior is evaluated using consistency scores, accuracy, and instruction-following failure rates. All experiments were conducted locally on consumer CPU hardware without fine-tuning. Consistency and accuracy were largely independent across models. Gemma 2 achieved the highest consistency (0.845-0.888) but the lowest accuracy (33.0-43.5%), while Llama 3.2 showed moderate consistency (0.774-0.807) alongside the highest accuracy (49.0-65.0%). Roleplay prompts consistently reduced accuracy across all models, with Phi-3 Mini dropping 21.5 percentage points on MedQA. Meditron-7B exhibited near-complete instruction-following failure on PubMedQA (99.0% UNKNOWN rate), indicating that domain pretraining alone is insufficient for structured clinical question answering. These findings show that high consistency does not imply correctness: models can be reliably wrong, a dangerous failure mode in clinical AI. Llama 3.2 demonstrated the strongest balance of accuracy and reliability for low-resource deployment. Safe clinical AI requires joint evaluation of consistency, accuracy, and instruction adherence.