Machine Learning for Auditory Brainstem Analysis

The Auditory Brainstem Response (ABR) is an evoked potential measurement which is used clinically in the objective estimation of hearing thresholds, particularly in newborns. It is typically interpreted by the visual inspection of a clinician which may be aided by objective statistical measures. Visual interpretation is subjective and is known to be variable between clinicians, potentially leading to incorrect clinical decision making.

This project aims to use machine learning algorithms to help provide an objective measure of the presence or absence of a response which may be used to aid clinicians in their interpretation of the ABR. Designing an objective measure using machine learning techniques has the potential to aid clinicians in more accurately and reliably interpreting ABRs, which would improve clinical decision making and could be used to improve the sensitivity and specificity of hearing screening devices.

It is proposed to collect ABR data from healthy individuals at a set range of sensation levels to train and test the machine learning algorithm. No stimulus recordings will allow the specificity of the model to be assessed and the module will also be testing on simulated data where we can be sure that a response is present in the data. Nested k-fold cross-validation will likely be used to evaluate the performance of the model during its construction and to select the best performing feature extraction and dimensionality reduction techniques. Different feature extraction and dimensionality reduction techniques such as principal component analysis and independent component analysis may help to optimise model performance. Bootstrap analysis will be performed to provide a p-value for any given output of the machine learning algorithm. The model's performance will be compared to that of traditional statistical confidence measures.