Use of machine learning to analyse auditory evoked electrophysiological data

Abstract

The aim of this research is to use machine learning (ML) in the analysis of auditory electrophysiological potentials (AEP), to investigate whether machine learning on AEP can identify features within the AEP waveforms or identify pathology. ML models were trained to estimate latency of the peak of wave I and V of the auditory brainstem response (ABR), to classify ABR waveforms from subjects with tinnitus and without, and to estimate a lifetime noise exposure score. For each application a range of supervised ML algorithms (regression, support vector machine, gaussian process regression, and artificial neural network based) were used to train and cross-validate ML models on wavelet features extracted from the ABRs, and then the models were tested using new ABR wavelet data. The best performing wave I latency estimate model achieved an excellent mean absolute error (0.0445 ms) but only a moderate fit of the data (R-squared 0.5220). The best wave V latency estimate model gave an excellent mean absolute error (0.0572 ms) alongside a remarkable fit of the data (R-squared 0.8925), thus demonstrating good correlation between the estimated and actual wave V latencies. The best peak latency estimate models were applied to reduced average ABR waveforms which mimics the sub-optimal testing condition of an incomplete ABR acquisition. Two of the wave V ML models gave a low error and good correlation (estimated latency to actual) down to reduced average ABR of 1000 but not lower. The wave I latency estimate models did not perform well on the reduced average ABRs, with particularly poor correlation between estimated and actual latency values. The ML models trained to estimate lifetime noise exposure score performed poorly and even the best ML model was deemed not suitable for the task. The best ABR ML tinnitus or no-tinnitus classifier achieved a sensitivity of 80% and a specificity of 81.82%. In conclusion, ML could identify features (estimate wave peak latency) within the auditory electrophysiological waveforms for wave I and V and could identify wave V in the sub-optimal condition of reduced average ABRs. ML was able to identify pathology within the auditory electrophysiological waveforms by classification of tinnitus from no-tinnitus ABRs, but ML was unable to predict a lifetime noise exposure score.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Chris Plack (Supervisor)