Automated auditory brainstem response peak estimation using a convolutional neural net

Abstract

Auditory brainstem responses (ABRs) are a core part of objective functional evaluations of hearing sensitivity and subcortical auditory transmission. Manual assessments of ABR waveforms are still a primary means by which thresholds and peak amplitudes and latencies are measured, which is time-consuming and prone to user variability. Automated methods have offered promising alternatives for ABR classification, but they have sometimes been limited in accuracy or robustness. Here, we developed and tested a supervised convolutional neural network (CNN) based ABR peak classifier that works across sound levels and sound frequencies that can be run quickly on a personal computer using single or dual-channel ABR inputs. For ABR peaks I, III, IV, and V, the classifier achieved over 95% accuracy. High accuracy was maintained even after noise-exposure causing temporary or permanent threshold shifts, and over 90% of peaks were within 0.041 ms (1 sample) of the manually identified peak. Only a few hundred samples were needed to train the network, making it widely amenable to smaller data studies or where the number of subjects or sessions may be low. HighlightsO_LISupervised CNN automatic ABR peak classifier built for 1-2 channel inputs C_LIO_LIAccuracy for peaks was at least 95% for all peaks C_LIO_LIAccuracy was robust across sound levels and noise-exposure C_LIO_LIA relatively small number of labeled samples were sufficient for training C_LI CRediT author contributionsJPM: Conceptualization, Methodology, Investigation, Formal analysis, Writing - original draft, Writing - review & editing. MCZ: Methodology, Investigation. ELB: Conceptualization, Methodology, Resources, Supervision, Writing - original draft, review & editing, Funding acquisition.