Accurate detection of speech auditory brainstem responses using a spectral feature-based ANN method

The speech auditory brainstem response (sABR) is a promising tool that can be used for objectively assessing auditory function. The main problem in obtaining the sABR is the high background noise, especially noise associated with general brain activity. In practice, a very long recording is needed to detect the sABR. We therefore propose a new detection method of the sABR based on spectral feature extraction that will reduce the detection time without reducing the accuracy. This method involves a constructed feature-frequency vector fed to an artificial neural network. The performance of the proposed method is compared to four other methods reported in the literature: optimal linear filtering, online estimator, Mutual Information, and artificial neural network based on discrete wavelet transforms and approximate entropy. All the methods were evaluated with several datasets of recorded and simulated sABRs ranging from extremely noisy to relatively clean. The proposed method performed very well in terms of sensitivity, specificity, and overall accuracy in detecting the sABR, compared with the other methods The reduction in the required recording time promises to facilitate the application of this measurement technique in clinical settings.