A Noise-Induced Permanent Threshold Shift Prediction Model Based on Deep Convolutional Multi-Task Learning

Authors: Heingjiang Liu1, Wei Qiu2, Jingsong Li3

1Zhejiang Lab
2Yangtze Delta Region Institute of Tsinghua University
3Zhejiang University

Background: The Noise Induced Permanent Threshold Shift (NIPTS) is the primary measure of the noise-induced hearing loss (NIHL). Among the currently approaches for evaluating NIPTS, such as the ISO 1999 prediction model, rely mainly on overall noise sound pressure level and duration of exposure. However, these approaches fail to consider the intricate time-frequency characteristics of noise. This study proposes an innovative NIPTS prediction model based on detailed features extracted by a convolutional neural network and then using a multi-task learning algorithm to predict the NIPTS under different frequencies simultaneously.

Method: 3,081 noise exposure data were collected from Chinese workers with 8 h work noise recordings and pure tone audiometry (PTA) results. First, a 60 s time window was employed to extract the detailed features in the noise sequence. This yielded a matrix comprising detailed features such as short-time sound pressure level, peak sound pressure level, kurtosis, and so on. Then, a convolutional network was employed on the noise sequence data and mapped it into a vector. Finally, the vector was fed to the multi-task learning network of the Multi-gate Mixture-of-Experts architecture. The learning targets were the NIPTS values of workers in different frequency (3, 4, 6 kHz), which were obtained by comparing PTA results with the control group.

Results: The model achieved an RMSE of 12.42 on the NIPTS346 after averaging all outputs, which is a significant improvement over the models in ISO1999 (RMSE=13.79) and Lempert’s paper (RMSE=13.18). The results showed that the model is capable of fully learning the hazard information hidden in the noise and accurately predicting the NIPTS of noise-exposed workers.

Conclusion: The deep convolutional multi-task learning model exhibits superior prediction performance compared to traditional methods. In addition, multi-task learning can be extended to learn simultaneously for a wider range of hearing loss measures.