The application of statistical learning models to the prediction of noise-induced hearing loss.
Qiu-W; Ye-J; Hamernik-RP
J Acoust Soc Am 2006 Nov; 120(5)(Part 2):3128
Three powerful nonlinear statistical algorithms [a support vector machine (SVM), radial basis function network (RBFN), and regression tree] were used to build prediction models for noise-induced hearing loss (NIHL). The models were developed from an animal (chinchilla) database consisting of 322 animals exposed to 30 Gaussian and non-Gaussian noise conditions. The inputs for the models were either energy or energy plus kurtosis. The models predict inner hair cell (IHC) loss, outer hair cell (OHC) loss, and postexposure threshold shift (PTS) at 0.5, 1, 2, 4, and 8 kHz. The models incorporating both energy and kurtosis improved the prediction performance significantly. The average performance improvement for the prediction of IHC loss was as much as 55%, for OHC loss it was 66% and for PTS, 61%. The prediction accuracy of SVM and RBFN with energy plus kurtosis for all three outputs (predictions) was more than 90% while for the regression tree model it was more than 85%. Energy is not a sufficient metric to predict hearing trauma from complex (non- Gaussian) noise exposure. A kurtosis metric may be necessary for the prediction of NIHL.
Noise-exposure; Noise-induced-hearing-loss; Hearing-loss; Hearing-impairment; Hearing-disorders; Models; Animal-studies; Laboratory-animals; Models
Wei Qiu, PhD, Auditory Research Laboratory, State University of New York at Plattsburgh, 101 Broad Street, Plattsburgh, New York, 12901
Disease and Injury: Hearing Loss
Journal of the Acoustical Society of America