Latency Localization of Auditory Brainstem Response by Different Deep Learning Methods

Auditory brainstem response (ABR) is an evoked potential generated by the auditory brainstem nuclei in response to auditory stimuli. Hearing loss or impairment in the auditory pathway can be objectively reflected by extracting the latencies of the peaks of characteristic waves (especially wave V) in the clinic. The commonly used clinical method of peak localization is the manual labeling approach by subjective visualization. However, this method is time-consuming and highly dependent on the experience of the clinician. In this study, three different types of neural networks were proposed to realize the automatic intelligent recognition of the wave V position. Firstly, during the data preprocessing, the ABR data of humans was clipped and normalized. Then three types of neural networks, namely DNN, CNN, and RNN, were proposed for the latency localization of wave V. The performances of these three networks were further evaluated under different sound pressure levels and different error scales. The results showed that DNN was the best network for localizing the wave V position. The highest overall accuracy rate was 91.70% and the error scale was 0.1ms. This study showed that the proposed automatic method with neural networks can greatly reduce the workload of medical personnel in finding the ABR latencies and can assist doctors in further diagnosis.