Counterfactual Auditing of Bias in Open-Source Large Language Models for Clinical Triage

The experimental design effectively isolates non-clinical influence by holding the clinical vignette fixed and injecting exactly one templated phrase at a time. The repeated-run stability analysis and deterministic decoding ($do\_sample=False$) ensure that observed shifts are attributable to the injected context rather than sampling variation. The positive-control experiment is compelling: Fine-tuned Qwen2.5-7B appropriately uptriaged nearly all cases after injections indicating active seizure or cyanosis requiring bag-mask ventilation (100.00% shift, all toward higher acuity), confirming that reduced counterfactual sensitivity is not mere model inertia. Additionally, the shared biased-outcome correlation analysis shows that Fine-tuned Qwen2.5-7B had a smaller average absolute off-diagonal correlation score than Qwen2.5-7B (0.0366 vs. 0.0849), indicating fewer clustered failure patterns across injected variables.

A primary weakness is the lack of statistical rigor when ranking models by shift rates. The authors report aggregate percentages—e.g., Fine-tuned Qwen2.5-7B at 5.27% any-shift versus Qwen2.5-7B at 16.02%—but provide no confidence intervals or hypothesis tests, so it is unclear whether the gaps exceed sampling variability.

Second, the evaluation relies on nurse-assigned ESI labels that are themselves imperfect and potentially biased; as the authors note, 'nurse-assigned ESI labels are operational reference standards rather than perfect ground truth, and the fine-tuned model may inherit label bias.' Third, the institutional evaluation set was filtered to 352 encounters with substantive documentation and excludes behavioral-health-dominant presentations and most ESI-1 cases, limiting generalizability to sparse, ambiguous, or highest-acuity notes. Finally, the single-variable design cannot capture intersectional effects, which the authors acknowledge but which leaves a significant real-world bias gap unaddressed.

The empirical evidence supports the descriptive ranking of models: Fine-tuned Qwen2.5-7B had the lowest any-shift rate among all evaluated models at 5.27% and the lowest mean absolute shift at 0.0534, while MedGemma-27B-IT shifted in 41.18% of cases with mean absolute shift of 0.84. The authors fairly situate their work against recent triage-bias literature, correctly noting that Lee et al. examined intersectional sex-by-race preferences in general triage models, whereas this study targets open-source pediatric ESI with a broader set of injected variables and explicit undertriage/overtriage directionality. That said, the absence of effect-size bounds or uncertainty quantification weakens the strength of comparative claims, and the external validation on 245 handbook vignettes, while useful, does not substitute for multi-site real-world deployment testing.

The methodological description is reasonably detailed: the authors specify the prompt template with explicit invalid-determinant instructions, deterministic decoding, and a QLoRA adaptation with prompt masking and label-scoring inference over ESI-1 to ESI-5. However, critical training hyperparameters—rank, learning rate, batch size, number of epochs, and exact optimizer settings—are omitted, and no code, model weights, or evaluation vignettes are released. Reproducing the reported shift rates therefore requires substantial guesswork, especially for the fine-tuned model. Providing the injection pipeline scripts and the filtered 352-case evaluation set would materially improve reproducibility.