Objectives: In occupations that involve hearing critical tasks, individuals need to undergo periodic hearing screenings to ensure that they have not developed hearing losses that could impair their ability to safely and effectively perform their jobs. Most periodic hearing screenings are limited to pure-tone audiograms, but in many cases, the ability to understand speech in noisy environments may be more important to functional job performance than the ability to detect quiet sounds. The ability to use audiometric threshold data to identify individuals with poor speech-in-noise performance is of particular interest to the U.S. military, which has an ongoing responsibility to ensure that its service members (SMs) have the hearing abilities they require to accomplish their mission. This work investigates the development of optimal strategies for identifying individuals with poor speech-in-noise performance from the audiogram.
Design: Data from 5487 individuals were used to evaluate a range of classifiers, based exclusively on the pure-tone audiogram, for identifying individuals who have deficits in understanding speech in noise. The classifiers evaluated were based on generalized linear models (GLMs), the speech intelligibility index (SII), binary threshold criteria, and current standards used by the U.S. military. The classifiers were evaluated in a detection theoretic framework where the sensitivity and specificity of the classifiers were quantified. In addition to the performance of these classifiers for identifying individuals with deficits understanding speech in noise, data from 500,733 U.S. Army SMs were used to understand how the classifiers would affect the number of SMs being referred for additional testing.
Results: A classifier based on binary threshold criteria that was identified through an iterative search procedure outperformed a classifier based on the SII and ones based on GLMs with large numbers of fitted parameters. This suggests that the saturating nature of the SII is important, but that the weights of frequency channels are not optimal for identifying individuals with deficits understanding speech in noise. It is possible that a highly complicated model with many free parameters could outperform the classifiers considered here, but there was only a modest difference between the performance of a classifier based on a GLM with 26 fitted parameters and one based on a simple all-frequency pure-tone average. This suggests that the details of the audiogram are a relatively insensitive predictor of performance in speech-in-noise tasks.
Conclusions: The best classifier identified in this study, which was a binary threshold classifier derived from an iterative search process, does appear to reliably outperform the current thresholds criteria used by the U.S. military to identify individuals with abnormally poor speech-in-noise performance, both in terms of fewer false alarms and a greater hit rate. Substantial improvements in the ability to detect SMs with impaired speech-in-noise performance can likely only be obtained by adding some form of speech-in-noise testing to the hearing monitoring program. While the improvements were modest, the overall benefit of adopting the proposed classifier is likely substantial given the number of SMs enrolled in U.S. military hearing conservation and readiness programs.