Feature Extraction in Auditory Brainstem Responses using Wavelet Decomposition on a Moving Window of Waveform Data

The auditory brainstem response is a waveform present in a subject´s EEG in response to a heard stimulus. The waveforms are hidden deep within the EEG and a significant body of work has been devoted to the enhancement and automated classification of these responses. This paper investigates the use of features extracted from the wavelet domain to assist in the classification of the ABR waveform. Initially, strong responses were classified without error by combining power features from the time and wavelet domain and applying a negative weighting to test cases where the presence of an artefact was suspected. The remaining ABR waveforms were passed to a second stage of classification. Cross-correlation features were extracted from repeat recordings using wavelet decomposition performed on a moving window of data within the post stimulus waveform. By separating different frequency levels within the decomposition a more representative post stimulus section of the waveform was analysed. When compared with expert opinion, the lower level responses with repeat recordings were classified to an accuracy of 76.4%