Applications of Machine Learning Methods in Retrospective Studies on Hearing

摘要

Hearing healthcare professionals rely on the audiograms produced through pure tone audiometry, among other tests, to diagnose and treat hearing loss. Researchers also rely on audiograms to study the prevalence of hearing loss in various populations. Notably, due to the available test time, intraoctave frequencies are not often recorded, even though they can contribute to certain diagnoses. Previous work has proposed the imputation of these thresholds using a simple average of neighboring thresholds. In this work, we present an alternative approach for addressing missing intra-octave thresholds that relies on a pmbk -nearest neighbors algorithm and show that accuracy can be slightly improved using a data-driven approach to imputation. We also present a Gaussian mixture model-based approach to flagging atypical or potentially unreliable audiograms to produce high quality datasets. Our method allows the imputation of intra-octave thresholds with an accuracy no worse than simple averaging. For the more challenging 6000 Hz threshold, our method appears to be particularly effective. Overall, our method allows for improved presentation of complete audiogram datasets.