Abstract
The goal of this study was to develop an automated auditory brainstem response (ABR) peak identification and labelling system using a combined decision theoretic and rule-based approach. The filters and labelling rules implemented were derived from both the known spectral composition of ABRs and from modelling. The accuracy of the labelling system was compared with that of three human experts. In addition to labelling ABR peaks, the developed system produced an ABR assessment. The assessment was based on threshold response, interpeak latencies and slope changes in the latency-intensity functions, which were derived from the labelled peaks. The accuracy of the system was also compared to the known hearing status of each of the subjects.A total of 24 ears (15 normal, 9 with sensorineural and various other neuropathologies) yielding 1052 recordings were used in the study. Data were acquired at stimulation levels between 0 and 70 dB HL in 10 dB increments. Data from subjects with rare pathologies were also obtained by digitizing recordings from paper tracings of clinical evaluations 80 dB HL. The following results were obtained: (A) Spectral at analysis of ABRs, revealed 3 primary frequency components at approximately 200, 500 and 900 Hz. The amplitude of the spectral components were found to decrease with decreasing stimulation intensity with the higher frequency components disappearing first. Only the low frequency components were present at low and near threshold stimulation levels. The developed ABR model based on the spectral characteristics generated realistic waveforms yielding insight as to the contribution of each spectral component towards the time domain peaks. (B) The automated ABR peak labelling accuracy was 88% for normal subjects and 83% for abnormals. (C) The automated classification procedure was able to determine correctly latency delays in abnormal recordings, however, neurological evaluation was more difficult. The generated latency-intensity curves showed cross correlations in the range of 0.99 when compared to the human expert curves. The mean threshold error was 2.92 dB.