Performance and reliability evaluation of an improved machine learning-based pure-tone audiometry with automated...

Automated air-conduction pure-tone audiograms through Bayesian estimation and machine learning (ML) classification have recently been proposed in the literature. Although such ML-based audiometry approaches represent a significant addition to the field, they remain unsuited for daily clinical settings, in particular for listeners with asymmetric or conductive hearing loss, severe hearing loss, or cochlear dead zones. The goal here is to expand on previously proposed ML approaches and assess the performance of this improved ML audiometry for a large sample of listeners with a wide range of hearing status.

First, we describe the changes made to the ML method through the addition of: (1) safety limits to test listeners with a wide range of hearing status, (2) transient responses to cater for cochlear dead zones or nonmeasurable thresholds, and importantly, (3) automated contralateral masking to test listeners with asymmetric or conductive hearing loss. Next, we compared the performance of this improved ML audiometry with conventional and manual audiometry in a large cohort (n = 109 subjects) of both normal-hearing and hearing-impaired listeners.

Our results showed that for all audiometric frequencies tested, no significant difference was found between hearing thresholds obtained using manual audiometry on a clinical audiometer as compared to both the manual and automated improved ML methods. Furthermore, the test-retest difference was not significant with the automated improved ML method for each audiometric frequency tested. Finally, when examining cross-clinic reliability measures, significant differences were found for most audiometric frequencies tested.

Together, our results validate the use of this improved ML-based method in adult clinical tests for air-conduction audiometry.