Large Language Models Versus Clinicians in Predicting Recovery from Idiopathic Sudden Sensorineural Hearing Loss

Background/Objectives: Early prognostication in idiopathic sudden sensorineural hearing loss (SSNHL) remains difficult, and although the use of large language models (LLMs) in otolaryngology is being increasingly evaluated, data directly comparing their prognostic predictions with those of clinicians assessing the same patients remain limited. This study evaluated the zero-shot, role-prompted performance of three contemporary LLMs against a panel of board-certified otolaryngologists in forecasting SSNHL outcomes. Methods: We retrospectively evaluated 100 consecutive patients with idiopathic SSNHL and randomly selected 30 for a blinded observer comparison. In the primary cohort of 100 patients, recovery was classified as favorable recovery (complete/partial, n = 37, 37.0%) or poor outcome (slight/none, n = 63, 63.0%) based on functional audiometric utility. Identical de-identified multimodal case sets (n = 30; 9 favorable, 21 poor), including the baseline audiogram image, were independently graded for 1-month recovery likelihood (low/moderate/high) by three otolaryngologists and three zero-shot LLMs—ChatGPT (GPT-5.6 Sol; OpenAI), Gemini (Gemini 3.1 Pro; Google), and Claude (Claude Opus 4.7; Anthropic), accessed through their publicly available web interfaces in July 2026—all blinded to outcomes; actual recovery was classified using the Siegel criteria. Results: In the primary cohort, multivariable logistic regression identified baseline pure-tone average (aOR = 0.957, 95% CI: 0.936–0.979, p = 0.0001) and admission delay (aOR = 0.797, 95% CI: 0.656–0.966, p = 0.0211) as significant independent predictors of recovery (model C-statistic = 0.834, Hosmer–Lemeshow p = 0.7925). In the sub-cohort (n = 30), Expert 2 achieved the highest overall diagnostic accuracy (86.7%, 95% CI: 70.3–94.7%; sensitivity 95.2%, specificity 66.7%). All three LLMs achieved identical overall accuracy (70.0%) that did not differ significantly from that of individual experts, although the study was not powered to establish equivalence, and all models remained below the best-performing expert (86.7%). Gemini achieved the highest specificity (77.8%, 95% CI: 45.3–93.7%) and the highest positive predictive value (87.5%, 95% CI: 64.0–96.5%) among all raters, and every case it rated as having low recovery potential failed to recover (10 of 10). Conclusions: Zero-shot, role-prompted LLMs predicted 1-month SSNHL outcome with 70.0% accuracy—not significantly different from that of individual experienced otolaryngologists, though the study was not powered for equivalence—and below the best-performing expert. Their reliable “rule-in” but poor “rule-out” pattern warrants prospective evaluation and does not yet support use in treatment decisions.