Abstract: The primary goal of this research project was to demonstrate the feasibility of developing an optimal prediction model for noise-induced hearing loss (NIHL) using a radial basis function neural network (RBFNN). The model was developed to predict noise-induced hearing loss (NIHL) from an archive of animal noise exposure data, which contains 900 chinchillas exposed to various noise environments. There were basically two different kinds of noises that were used in these modeling studies: High level (peak>140dB) impulse noise transients, which typically occur in military environments, and low and medium level (peak<140dB) long term complex noise exposures, which usually occur in industrial environments. Both noise metrics and biological parameters were used as input variables to the prediction models. Since the prediction models consist of individual biological information, it should be possible to predict noise-induced hearing loss in an individual. Two frequency specific prediction models were considered: One was a specific frequency model (SF model), the other was a wide band frequency model (WF model). In the SF model a prediction model was built for each specified frequency. In the WF prediction model contiguous frequency band information on either upper or lower side(s) of a specified frequency band were considered as additional input(s) for the models. Both SF and WF models were built and tested. Two partition methods were used to stratify the data sets for training, validating and testing the prediction models. Partition method 1 considers all the data sets from the different noise types. The database is stratified according to exposure noises before sampling to make sure that the samples from different types of noise are allocated to both training and test data sets more or less equally. Partition method 2 neglects the fact that all available training data are collected under different exposure conditions. The samples from different groups are divided into disjoint sample sets of equal size randomly, without taking into account the different exposure groups. The prediction models using partition 1 and 2 were built and tested. For long term complex noise exposures 10 noise metrics and 5 biological parameters were used as the inputs of the prediction model in the initial stage of the research project. Four prediction results at specific center frequencies were produced, i.e. noise-induced permanent threshold shift (PTS), inner hair cell (IHC) loss, outer hair cell (OHC) loss, and permanent change in cubic distortion product otoacoustic emission (DeltaDPOAE). It was found that energy alone was not a sufficient metric to predict complex noise induced hearing hazard. Kurtosis is a complementary metric. With the kurtosis metric performance of the prediction models is improved as much as 66%. It was also found that some biological parameters, such as asymptotic threshold shift and pre-exposure DPOAE did not contribute much to increasing the prediction accuracy (in many cases these variables made the prediction results worse). In the final prediction model for complex noise exposures 6 noise metrics and one biological parameter were used as input variables to predict the NIHL in terms of PTS, IHC loss and OHC loss. It was found that the prediction models using the WF method would yield the best average prediction accuracy. The percent prediction accuracy for IHC loss using the RBF model was 92%, OHC loss: 93% and PTS: 96%. The performance of the RBF models using partition method 1 and method 2 were not significantly different. For impulse noise exposures parameters, such as peak, duration, number and rate, are important to the prediction of NIHL. The RBF model with the SF method was selected as the best prediction model. The percent prediction accuracy for IHC loss was 85%, OHC loss: 87% and PTS: 93%.
Roger P. Hamernik, PhD, Auditory Research Laboratory, State University of New York at Plattsburgh, 101 Broad Street, Plattsburgh, New York, 12901