Analysis of correlation between window duration for kurtosis computation and accuracy of noise-induced hearing loss prediction.

Abstract

Kurtosis is considered an important metric for evaluating noise-induced hearing loss (NIHL). However, how to select window duration to calculate kurtosis remains unsolved. In this study, two algorithms were designed to investigate the correlation between window duration for kurtosis computation and the accuracy of NIHL prediction using a Chinese industrial database. Pure-tone hearing threshold levels (HTLs) and full-shift noise were recorded from each subject. In the statistical comparison, subjects were divided into high- and low-kurtosis groups based on kurtosis values computed over different window durations. Mann-Whitney U test was used to compare the difference in group HTLs to find the optimal window duration to best distinguish these two groups. In the support vector machine NIHL prediction model, kurtosis obtained from different window durations was used as a feature of the model for NIHL evaluation. The area under the curve was used to evaluate the performances of models. Fourteen window durations were tested for each algorithm. Results showed that 60 s was an optimal window duration that allows for both efficient computation and high accuracy for NIHL evaluation at test frequencies of 3, 4 and 6 kHz, and the geometric mean of kurtosis sequence was the best metric in NIHL evaluation.