JMIR Publications

ABSTRACT

Background:

LLMs are increasingly used to generate clinical teaching cases in all health professions education, but the patient demographics they assign when no particular demographic group is specified remain poorly characterized. Unexamined defaults may encode demographic bias into medical curricula.

Objective:

To quantify default patient demographics across 55 clinical conditions and 17 medical specialties; compare generated distributions with demographic and epidemiological benchmarks; evaluate three prompting strategies; and assess whether model class and medical fine-tuning alter demographic defaults.

Methods:

We conducted a computational audit of nine LLMs: three frontier closed-source models (GPT-5.4-mini, Claude Sonnet 4.6, and Gemini 2.5 Flash), five general-purpose open-source models (Llama 3.1 8B, Mistral 7B, Gemma 2 9B, Phi-3.5 Mini, and Qwen2.5 7B), and one medically fine-tuned open-source model (Llama3-Med42 8B). Each model generated 20 clinical teaching cases per medical condition under four prompts: unprompted baseline, demographically neutral, diversity-instructed, and epidemiological calibration. The resulting 39,600 vignettes were analyzed using a standards-grounded natural language processing pipeline validated against an independent LLM judge. Demographic distributions were compared with US Census 2020 proportions and condition-specific US epidemiological benchmarks (where available).

Results:

All nine models produced usable vignettes. Extraction agreement was substantial for race (κ=0.80) and gender (κ=0.85). Gender defaults were model-specific: as indicated by representation index (RepI), seven models generated predominantly female patients (RepI_"male" =0.51-0.99), whereas two generated predominantly male patients (RepI_"male" =1.32-1.66); Claude Sonnet 4.6 showed male-female parity. Race was specified in only 5.6% of baseline vignettes, with predominating White patients. Frontier models mentioned Black patients more often than open-source models (1.30% vs 0.52%; mixed-effects OR =0.21). Neutral prompting did not diversify race. Diversity prompting overrepresented Hispanic patients in open-source models and Black and Hispanic patients in frontier models. Epidemiological calibration produced the best racial alignment but overrepresented Hispanic patients (pooled RepI=2.9) and induced female overcorrection in open-source models (RepI_"male" =0.20), more weakly in frontier models (0.67). Accurate condition-specific targets calibrated better than flat Census targets (L_1=0.76 vs 1.17; two-tailed Welch t_394.1=-9.94, P<.001). Where interpretable, stereotype amplification overassociated Black patients with type 2 diabetes and White patients with neurodegenerative conditions. In the single analyzable comparison, medical fine-tuning changed rather than improved racial calibration.

Conclusions:

Models most often left race unspecified; when steered, mitigation traded one demographic distortion for another. Medical fine-tuning was not a consistent remedy in the single analyzable comparison. None of the three prompting strategies reliably corrected demographic defaults, inviting deeper research into a balanced, epidemiologically valid generation of clinical teaching cases. Demographic audits should therefore assess both model behavior and the extraction tools used to measure it before LLM-generated cases are incorporated into medical curricula.