RESUMO
OBJECTIVE: The aim of this study is to compare machine learning algorithms and established rule-based evaluations in screening audiograms for the purpose of diagnosing vestibular schwannomas. A secondary aim is to assess the performance of rule-based evaluations for predicting vestibular schwannomas using the largest dataset in the literature. STUDY DESIGN: Retrospective case-control study. SETTING: Tertiary referral center. PATIENTS: Seven hundred sixty seven adult patients with confirmed vestibular schwannoma and a pretreatment audiogram on file and 2000 randomly selected adult controls with audiograms. INTERVENTIONS: Audiometric data were analyzed using machine learning algorithms and standard rule-based criteria for defining asymmetric hearing loss. MAIN OUTCOME MEASURES: The primary outcome is the ability to identify patients with vestibular schwannomas based on audiometric data alone, using machine learning algorithms and rule-based formulas. The secondary outcome is the application of conventional rule-based formulas to a larger dataset using advanced computational techniques. RESULTS: The machine learning algorithms had mildly improved specificity in some fields compared with rule-based evaluations and had similar sensitivity to previous rule-based evaluations in diagnosis of vestibular schwannomas. CONCLUSIONS: Machine learning algorithms perform similarly to rule-based evaluations in identifying patients with vestibular schwannomas based on audiometric data alone. Performance of established rule-based formulas was consistent with earlier performance metrics, when analyzed using a large dataset.
