A Comprehensive Systematic Review and Meta-Analysis: The Future of Artificial Intelligence in Diagnostic Otology — Opportunities and Barriers to Clinical Implementation

Abstract

Introduction: Diagnostic otology relies on operator-dependent interpretation of otoscopy, temporal bone computed tomography (CT), audiometry and vestibular testing, and diagnostic accuracy outside specialist settings remains suboptimal. Artificial intelligence (AI) has been proposed as a scalable decision-support solution, but its pooled accuracy, robustness beyond development data, comparative value against clinicians and barriers to clinical implementation have not been synthesised across all diagnostic otology modalities. This review aimed to quantify the diagnostic performance of AI across otoscopic, radiological, audiological and vestibular applications and to map opportunities and barriers to implementation.

Methods: The study strictly adhered to the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) 2020 guidelines. Randomised controlled trials, cohort, case–control and cross-sectional diagnostic studies (prospective or retrospective) evaluating an AI index test against a clinical reference standard were eligible; systematic reviews were excluded from synthesis. Risk of bias was appraised with QUADAS-2 supplemented by QUADAS-AI considerations. Accuracy, sensitivity and specificity were pooled on the logit scale using random-effects models (restricted maximum likelihood with modified Hartung–Knapp–Sidik–Jonkman confidence intervals) with 95% prediction intervals (PI); risk differences (RD) quantified AI–clinician comparisons. Subgroup analyses, meta-regression, influence diagnostics, small-study tests and GRADE certainty assessment were performed.

Results: Of 139 unique records, 64 primary studies (2019–2026; otoscopy 37, temporal bone CT 14, audiology 6, vestibular 7) were included. No randomised trial of an AI diagnostic tool in otology was identified. Pooled internal-validation accuracy was 92.5% (95% CI 90.0–94.5) (k = 36; PI 64.7–98.8%; I² = 97.8%, τ² = 0.858): otoscopy 93.2%, temporal bone CT 90.2%, audiometry/tympanometry 97.6% and vestibular testing 84.0% (between-modality P = 0.003). Pooled sensitivity and specificity were 92.6% (95% CI 89.4–94.8) and 89.5% (95% CI 74.3–96.1), respectively (diagnostic odds ratio 105; summary AUROC 0.957). Accuracy fell to 81.2% (95% CI 59.1–92.8) in external, prospective or point-of-care cohorts (P = 0.026 versus internal validation), and studies using surgical, histopathological or objective reference standards reported lower accuracy than those using expert annotation (89.0% vs 93.6%; P = 0.041). Task-specific AI outperformed clinicians on identical cases (RD +18.2 percentage points, 95% CI 7.3 to 29.0), and AI assistance increased clinician accuracy (RD +9.4 points, 95% CI 1.4 to 17.4), whereas general-purpose multimodal large language models underperformed specialists (RD -25.0 points). Of 47 meta-analysed studies, 34 (72.3%) had at least one high risk-of-bias domain; certainty of evidence was low to very low.

Discussion: Task-specific AI achieves high diagnostic accuracy across diagnostic otology and consistently exceeds non-specialist performance, supporting its role as a triage and decision-support adjunct, particularly in primary care, tele-otology and resource-limited settings. However, extreme heterogeneity, wide prediction intervals, performance attenuation under domain shift, spectrum and reference-standard bias, scarce prospective evidence and the absence of outcome trials constitute the principal barriers to implementation.

Conclusion: AI is a promising but not yet implementation-ready component of diagnostic otology. Multicentre prospective validation with surgical or objective reference standards, standardised image acquisition, transparent reporting (STARD-AI/TRIPOD+AI-aligned) and pragmatic randomised trials measuring patient-important outcomes are required before routine deployment.