Abstract
Traditional sensorineural hearing loss identification use the framework of feature extraction and classification. Nevertheless, this framework needs manual feature engineering. In this study, we proposed an improved convolutional neural network model to identify hearing loss. Our nine-layer deep convolutional neural network contains six conv layers and three fully-connected layers. We used batch normalization to reduce the impact caused by Internal Covariate shift and dropout techniques to prevent over-fitting to increase the performance in terms of accuracy. Data augmentation was used to enlarge the size of training set. The average results of 10 runs on test set show our method secured sensitivities of left-sided hearing loss, right-sided hearing loss, and healthy controls are 96.33 +/- 2.46%, 96.67 +/- 2.22%, and 96.67 +/- 2.72%, respectively. The overall accuracy of all three classes was 96.56 +/- 0.63%. Deep learning can effectively build the identification model. The performance of our proposed nine-layer convolutional neural network model yields better performance than five state-of-the-art approaches.