To improve the efficiency and timeliness in frequency-following response (FFR) testing, the purpose of this study was to investigate the capabilities of machine learning in the detection of FFRs. Continuous brain waves were recorded from 25 Chinese adults in response to a pre-recorded Mandarin monosyllable \yi2\ with a rising frequency contour. A total of 8000 artifact-free sweeps were recorded from each participant. Continuous brain waves sub-averaged from the first sweep up to the first 500 sweeps were considered FFR absent, whereas brain waves sub-averaged from the first sweep up to the last 1000 sweeps were considered FFR present. Six response features (Frequency Error, Slope Error, Tracking Accuracy, Spectral Amplitude, Pitch Strength and Root-Mean-Square Amplitude) were extracted from each recording and served as key predictors. Twenty-three supervised machine-learning algorithms, with a 10-fold cross-validation procedure, were implemented via a Classification Learner App in MATLAB. Two algorithms yielded 100% efficiency (i.e., 100% sensitivity and 100% specificity) and 14 others produced efficiency of 99%. Results indicated that a majority of the machine-learning algorithms provided accurate predictions in whether an FFR was present or absent in a recording.

This content is only available via PDF.