Did You Hear That? Detecting Auditory Events with EEGNet

The behavioural nature of pure-tone audiometry (PTA) limits those who can participate in the test, and therefore those who can access accurate hearing threshold measurements. Event Related Potentials (ERPs) from brain signals has shown limited utility on adult subjects, and a neural response that can consistently be identified as a result of pure-tone auditory stimulus has yet to be identified. The in doing so challenge is worsened by the nature of PTA, where stimulus amplitude decrease to a patient's lower threshold of hearing. We investigate whether EEGNet, a compact Convolutional Neural Network, could help in this domain. We trained EEGNet on a dataset collected whilst patients underwent a test designed to mimic a pure-tone audiogram, then assessed EEGNet performance in the detection task. For comparison, we also trained Support Vector Machines (SVMs) and Common Spatial Patterns + Linear Discriminant Analysis (CSPLDA) on the same task, with the same training paradigms. The results show that EEGNet is capable of detecting hearing events with 81.5% accuracy on unseen participants, outperforming SVMs by just over 5%. Whilst EEGNet outperformed SVMs and CSPLDA, it did not, however, always show a statistically significant improvement. Further analysis of EEGNet predictions revealed that, with sufficient test repetition, EEGNet has the potential to accurately ascertain hearing thresholds. The implication of these results is for a brain-signal based hearing test for those with physical or mental disabilities that limit their participation in a PTA. While this research is promising, future research will be needed to address the complexity of test setup, the duration of testing, and to further improve accuracy.