Evaluating Large Language Models in Clinical Audiology (AUDIOLOGYBENCH): Benchmark Development and Validation Study

Background: Large language models (LLMs) are increasingly being explored for clinical decision support, but their performance in audiology has not been systematically benchmarked using clinically grounded case materials and rubric-based safety evaluations. Objective: This study aimed to develop and evaluate AUDIOLOGYBENCH, a 3-tier benchmark for characterizing frontier LLM capability in clinical audiology along (1) curated domain knowledge, (2) literature-derived evidence, and (3) clinical reasoning under multimodal case input, with an explicit human audit of the automated adjudicator on the primary end point. Methods: The benchmark comprises 3139 objective items from educational resources, 3175 research article-derived items from peer-reviewed articles published between 2015 and 2025, and 67 multimodal clinical case studies graded against a standardized A-F rubric with 6 prespecified critical-error types that cap scores at D or F. Eight models were evaluated on the educational objective items: 4 frontier multimodal models (Gemini 2.5 Pro, Grok 4, OpenAI O3, and Claude Sonnet 4 Thinking) were evaluated on the research article-derived items, and on 804 case study evaluations. Adjudication used Gemini 2.5 Pro (objective and research-derived items) and Claude Opus 4.5 (case studies). The case study adjudicator was independently audited against PhD-level audiologist consensus on blinded subsamples, supplemented by a post-stratified human-calibrated sensitivity analysis. Results: ≥0.679). Interreviewer reliability between audiologists was high (quadratic-weighted κ of 0.78 and 0.85 across the 80-item and 50-item audits, respectively). When 2 audiologists regraded all 80 model Q1 responses with the diagnostic images available, the adjudicator's per-item Q1 labels diverged from human judgment (κ=0.05; overflagging; sensitivity: 19/26, 73%; positive predictive value: 19/53, 36%), yet its reweighted Q1 critical-error rate (36.2%) was broadly consistent with the image-grounded human estimates (28%-34%), suggesting no systematic inflation of the headline rate. The principal Q3>{Q1, Q2} ranking was preserved under post-stratified human calibration. Web-style multiple-choice items showed ceiling effects (>95% accuracy); short-answer prompts remained challenging (best 30%). Conclusions: Current frontier LLMs show strong recommendation generation but substantial limitations in audiometric numerical interpretation that are shared across models and that an automated adjudicator partially miscalibrated at the per-item level. AUDIOLOGYBENCH characterizes capability boundaries rather than certifying clinical readiness. Deployment of LLM-assisted audiology workflows requires structured human verification of all numerical findings and awareness of fabrication and severity misclassification failure modes documented here.