Optimising objective detection methods for the auditory brainstem response

Abstract

The transient Auditory Brainstem Response (ABR) is a change in neural activity along the auditory pathway in response to a brief acoustic stimulus. It is typically recorded non-invasively using electroencephalography (EEG), and has become an important diagnostic tool in the clinic, e.g. for diagnosing various neurological disorders and hearing screening in new borns. Detecting the ABR, however, can be a challenging task, and is still strongly dependent on highly trained individuals who are given the task to visually examine the EEG data. Besides incurring additional training costs, visual inspection has limitations in terms of specificity and sensitivity. Consequently, test time for some ABR examinations can be quite extensive, and information is often incomplete. This can have significant clinical implications, and has a large impact on the parents/carers of the infants.

The limitations associated with visual inspection have led to the development of many different objective measures for assisting the examiner during the visual inspection task, and improving the reliability and efficiency of the test. The overall goal for this thesis is to further improve the reliability and efficiency of ABR examinations by improving the specificity, sensitivity, and test time of objective ABR detection methods. To achieve this, the focus is firstly on the objective ABR detection methods themselves, i.e. on exploring, evaluating, optimising, and comparing new and existing detection methods, with the goal to find or develop methods with a high sensitivity, a low test time, and a controlled specificity. Important elements within this analysis include the assumptions underlying the detection methods, along with the adopted test and pre-processing parameters. Results demonstrate that the main concern for specificity is the independence assumption between epochs, which is violated as a function of the stimulus rate and the filter’s high-pass cut-off frequency. The best performing method in terms of sensitivity and test time was furthermore a new bootstrapped statistic, consisting of a combination of the Hotelling’s T2 test and a correlation coefficient.

A second route in this thesis for improving the performance of objective ABR detection methods is through the development and optimisation of a new sequential testing framework for ABR detection methods. The approach, called the Convolutional Group Sequential Test (or CGST), controls the specificity of sequentially applied statistical tests, and permits data-driven adaptations (using previously analysed data) to test parameters following each stage of the sequential analysis. This allows the statistical analysis to be tailored specifically to the subject and recording in question, which offers new opportunities to speed up testing with high statistical power and controlled specificity. Results demonstrate relatively large reductions in test time when compared to a ’single shot’ test where the detection method is applied to the data just once.

A final route in this thesis for improving the performance of objective detection methods is through a new adaptive ensemble size re-estimation procedure, integrated within the sequential testing framework. Besides further reductions in test time (relative to non-adaptive sequential test procedures), the adaptive approach can help bring ABR examinations to an unambiguous test outcome in terms of ‘ABR present’ or ‘ABR absent or abnormal’.

More information

Identifiers

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

Catalogue record

Export record