The scalp-recorded frequency-following response (FFR), an auditory-evoked potential with putative neural generators in the rostral brainstem, provides a robust representation of the neurophysiologic encoding of complex stimuli. The FFR is rapidly becoming a valuable tool for understanding the neural transcription of speech and music, language-related processing disorders, and brain plasticity at initial stages of the auditory pathway. Despite its potential clinical and empirical utility, determining the presence of a response is still dependent on the subjective interpretation by an experimenter/clinician. The purpose of the present work was to develop and validate a fully objective procedure for the automatic detection of FFRs elicited by complex auditory stimuli, including speech. Mutual information (MI) was computed between the spectrographic representation of neural FFRs and their evoking acoustic stimuli to quantify the amount of shared time-frequency information between electrophysiologic responses and stimulus acoustics. To remove human subjectivity associated with typical response evaluation, FFRs were first simulated at known signal-to-noise ratios using a computational model of the auditory periphery. The MI at which model FFRs contained +3 dB Signal-to-noise ratio was taken as the criterion threshold (θMI) for the presence of a response. θMI was then applied as a binary classifier on actual neurophysiologic responses recorded previously in human participants (n = 35). Sham recordings, in which no stimulus was presented to participants, allowed us to determine the receiver operating characteristics of the MI metric and the capabilities of the algorithm to segregate true evoked responses from sham recordings. RESULTS showed high overall accuracy (93%) in the metric's ability to identify true responses from sham recordings. The metric's overall performance was considerably better than trained human observers who, on average, accurately identified only ∼75% of the true neural responses. Complementary results were found in the metric's receiver operating characteristic test performance characteristics with a sensitivity and specificity of 97% and 85%, respectively. Additionally, MI increased monotonically and was asymptotic with increasing trials (i.e., sweeps) contributing to the averaged FFR and, thus, can be used as a stopping criteria for signal averaging. The present results demonstrate that the mutual information between a complex acoustic stimulus and its corresponding brainstem response can provide a completely objective and robust method for automated FFR detection. Application of the MI metric to evoked potential speech audiometry testing may provide clinicians with a more robust tool to quantitatively evaluate the presence and quality of speech-evoked brainstem responses ultimately minimizing subjective interpretation and human error.
Read full abstract