Beyond Accuracy: Counterfactual Fragility and Demographic Bias in Clinical Evaluation of LLMs

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.

Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline ``meaning-preserving'' modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted interventions, we propose a framework in which we compare (via statistical testing) differences observed under target interventions to those induced by paraphrasing inputs. We then use this framework to revisit a analysis done on the MedPerturb dataset, which reported evidence of model sensitivity to patient demographics and stylistic cues. We find that these effects largely dissipate when we account for general model sensitivity, with only 5 of 120 tests reaching statistical significance. Applying the same framework to occupational biography classification, we detect clearly significant directional gender bias, showing that the framework identifies real directional effects even when they are small. We evaluate a range of metrics -- aggregate, per-sample distributional, and regression -- and find that per-sample metrics are dramatically more powerful than aggregate metrics and regression powerfully and uniquely characterizes effect direction and magnitude.

Background: Expert-annotated benchmarks for non-English open-response clinical questions are scarce. LLM-as-a-judge systems may scale evaluation but require validation. Objective: To introduce MedQADE, a standardized German open-response clinical benchmark with physician reference annotations, and evaluate LLM-as-a-judge alignment, self- and intra-family bias, and abstention. Methods: The benchmark contains 3,800 question-answer sets with answers from five student LLMs and annotations from 10 physicians. All 10 rated the 200-question core; two primary raters assessed each of 3,600 extension questions, with the tenth resolving disagreements. Nine LLM evaluators assessed all sets. We assessed physician reliability, student-model accuracy, evaluator alignment, bias, and abstention. Results: Physicians showed moderate-to-substantial agreement on answer correctness (unweighted mean pairwise Cohen's kappa = 0.612) but limited agreement on question difficulty (Krippendorff's alpha = 0.208 using squared numeric-score distances). Student-model accuracy was 17.8%-66.0% and generally decreased with physician-rated difficulty. Gemini 3 Flash approached the leave-one-out physician reference (kappa = 0.694 vs 0.709). Four of five models rated their own responses more favorably than out-of-family evaluators; five of six intra-family comparisons were positive. Physician abstention increased with perceived difficulty. Seven of nine LLM evaluators abstained in no more than 0.51% of evaluations; the two strongest evaluators assigned definitive labels to every response. Conclusions: Strong LLM evaluators approached physician agreement, but evaluator bias and low observed abstention warrant physician validation and further assessment of selective deferral before fully automated evaluation. These results do not establish clinical safety.