Improving objective analysis of the auditory brainstem response

Abstract

Auditory brainstem response (ABR) testing is a form of electrophysiological assessment used clinically to evaluate the auditory system. One of the main uses of ABR testing is in the evaluation of hearing thresholds in patients for whom behavioural hearing assessments are unreliable, e.g. newborns. Accurate interpretation of the ABR is important, as this will inform clinical decision making and potentially be used to prescribe hearing aid amplification. The overall aim of this research project was to explore methods for improving objective analysis of the ABR. The first study in this thesis evaluated machine learning approaches for ABR detection. Using simulation, based on data recorded from participants, a range of machine learning algorithms were evaluated using nested k-fold cross-validation. The best algorithm, a stacked ensemble, was evaluated on previously unseen test set data. Using the bootstrap method to set the critical value for determining whether a response is present or absent, the stacked ensemble was able to achieve a high and stable level of specificity across ensemble sizes. Additionally, the detection rate of the stacked ensemble was statistically significantly better across all ensemble sizes, compared to the statistical detection methods evaluated. These results suggest that the proposed stacked ensemble algorithm may have the potential to assist clinicians in interpreting ABR waveforms, as well as in improving the performance of automated detection algorithms in ABR screening devices. Due to the low signal-to-noise ratio of the ABR, detection of a response using visual inspection and statistical detection methods can be extremely challenging. Weighted averaging has been proposed as a method of maximising the signal-to-noise ratio in the averaged waveform. A second study aimed to further the understanding of weighted averaging, optimise the parameters of this technique, and quantify its effects on ABR detection using the Fmp statistical detection method. In this second study, the noise level estimation method was optimised, as was the parameter for the number of epochs in each block. As well as being used for hearing threshold estimation, the ABR test may be used diagnostically in the functional assessment of the auditory brainstem pathway, e.g. for the detection of pathologies affecting the structures of this pathway. In a bid to reduce subjectivity in waveform interpretation, in study three of this thesis, several machine learning algorithms were compared in their ability to correctly predict ABR wave latencies. A convolutional recurrent neural network performed best, with 95.9% of predictions being within 0.1 milliseconds of the target label. Overall, this thesis provides three main approaches for improving objective analysis of the ABR. Further work is recommended to help translate this research into clinical practice.

More information

Identifiers

ORCID for Richard Michael McKearney: orcid.org/0000-0001-7030-5617

ORCID for David Simpson: orcid.org/0000-0001-9072-5088

Catalogue record

Export record