Performance, safety, and repeated-query stability of large language models versus early-career urologists in urological oncology: a source-masked vignette benchmark

Technical Summary: Performance, Safety, and Repeated-Query Stability of LLMs vs. Early-Career Urologists

Problem Statement
While Large Language Models (LLMs) are increasingly evaluated for medical decision support, aggregate accuracy metrics often obscure consequential errors and variability across repeated queries. In urological oncology, where management relies on the nuanced integration of disease stage, risk groups, and patient-specific modifiers, high performance on standardized knowledge questions does not guarantee reliable, context-sensitive clinical decision-making. Existing literature indicates that LLMs may fail to meet non-inferiority criteria against multidisciplinary tumor boards and that safety-critical errors can persist even when overall performance appears acceptable. Furthermore, the stability of LLM outputs—whether repeating a query yields the same classification or recommendation—remains under-explored as a distinct dimension of clinical reliability.

Methodology
This study employed a comparative, cross-sectional, vignette-based benchmark to evaluate three consumer-facing LLMs against two fixed early-career urologist comparators.

• Benchmark Design: The study utilized 100 synthetic clinical vignettes covering five urological cancers (prostate, bladder, renal cell carcinoma, testicular, and upper tract urothelial carcinoma) and five decision categories (diagnosis/staging, risk classification, initial treatment, sequencing, and follow-up/safety).
• Participants:
  • LLMs: GPT-5.2 Instant, Claude Sonnet 4.5, and Gemini 3 Pro Preview.
  • Human Comparators: Two early-career urologists (one with ~1 year and one with ~3 months of post-specialization experience).
  • Evaluators: Two source-masked urologists with uro-oncology expertise (one professor, one associate professor) independently scored all 500 responses.
• Conditions: All sources answered the 100 vignettes using an identical standardized English prompt without external resources, web browsing, or iterative correction.
• Evaluation Framework: Responses were scored against a prespecified 2025 European Association of Urology (EAU)-anchored framework across four domains: guideline concordance, clinical safety, decision completeness, and risk/stage accuracy.
  • Primary Endpoint: A binary "composite success" requiring a score ≥2 in all four domains and the absence of any prespecified disqualifying clinical error.
  • Safety Endpoint: A separate assessment of "safety-critical errors" defined as errors that could plausibly cause serious harm, major undertreatment, or failure to recognize urgency.
  • Stability Analysis: A subset of 20 vignettes was queried three times per platform to assess exact three-run stability for primary classification, safety status, and full four-domain score profiles.
• Statistical Analysis: Comparisons utilized Cochran's Q test for overall differences and McNemar tests for pairwise contrasts, with Holm adjustment for multiplicity. Effect sizes were estimated via nonparametric bootstrap resampling.

Key Results

• Primary Composite Success: Early-career urologists significantly outperformed LLMs. Urologist A achieved 85.0% success, and Urologist B achieved 89.0%. In contrast, LLM success rates were 70.0% (GPT-5.2), 64.0% (Claude), and 60.0% (Gemini). While Urologist A vs. GPT-5.2 did not remain significant after Holm adjustment (adjusted P = 0.095), all other human-vs-LLM contrasts were significant.
• Safety-Critical Errors: A stark disparity was observed in safety. LLMs produced safety-critical errors in 25.0% (GPT-5.2), 25.0% (Claude), and 26.0% (Gemini) of responses. Urologists produced such errors in only 2.0% and 1.0% of cases, respectively. All human-vs-LLM safety contrasts remained significant after adjustment.
• Error Phenotypes: The most frequent LLM error was the omission of a management-changing factor (CE-6), followed by diagnostic/staging omissions and undertreatment. Harmful or contraindicated recommendations (CE-9) occurred in 1, 3, and 3 LLM responses (1%, 3%, and 3%, respectively) and in neither urologist comparator.
• Repeated-Query Stability: Stability was limited. Across 60 vignette-platform combinations, exact three-run stability for primary classification was 45.0%, and for safety status, it was 48.3%. Only 28.3% of combinations maintained stability across the full four-domain score profile. Notably, stability included both consistently successful and consistently unsuccessful (or unsafe) outputs.

Key Contributions

1. Decoupling Performance from Reliability: The study demonstrates that aggregate performance metrics do not fully capture response reliability. High average scores can coexist with a high burden of safety-critical errors and low repeatability.
2. Safety-Gated Evaluation: By defining safety-critical errors separately from general guideline concordance, the study highlights that LLMs frequently generate responses that are "broadly acceptable" in structure but clinically dangerous due to specific omissions or contraindications.
3. Stability as a Clinical Metric: The research quantifies the instability of consumer-facing LLMs, showing that repeating a query often changes the clinical classification or safety status, a factor not captured by correlation-based repeatability measures.
4. Source-Masked Benchmarking: The use of a balanced, source-masked design with expert uro-oncology evaluators provides a rigorous comparison between current consumer AI and early-career clinicians in a complex subspecialty.

Significance and Claims
The paper concludes that while consumer-facing LLMs often generate acceptable responses to synthetic urological oncology vignettes, they currently exhibit a clinically relevant rate of safety-critical errors and limited repeatability. The findings suggest that these models are not yet ready for autonomous clinical use. Instead, the results support a model of clinician-supervised support, where LLM outputs are treated as draft suggestions that must be verified against the complete clinical context and current guidelines.

The authors emphasize that performance, consequential error, and repeatability should be evaluated as separate dimensions of reliability. They advocate for the incorporation of safety-gated endpoints and repeated-query stability metrics alongside conventional accuracy measures in future high-stakes clinical evaluations of AI. The study explicitly avoids claiming population-level superiority for the specific human comparators tested, noting that two individuals cannot represent the broader urologist population, but the direction of the results aligns with other studies showing LLMs struggle with complex, context-sensitive oncology decisions.