Evaluation of automatic identification algorithms for auditory brainstem response used in universal hearing loss screening

During the last decade several programs for universal detection of newborn hearing loss (UNHL) have been developed, based on the analysis of either the auditory brainstem response (ABR) and/or otoacoustic emissions (OAEs). At the present time, although there are several systems and algorithms capable of automatically determining the presence of response, there does not exist a good standard to analyze and determine the performance of such algorithms. In the present work a methodology is proposed to compare and to determine the performance of algorithms for automatic identification of ABR used in UNHL. An ABR database of several monoaural stimulation intensities in 7 normal hearing subjects was generated and 4000 epochs were stored for each study. The signals obtained with ∼ 0 dBspl stimulus were considered representative of hearing loss. Three algorithms were programmed in Matlab and applied to the database. The performance of these algorithms was evaluated by means of the analysis of relative operating characteristic (ROC) curves, which offer better information than the metrics currently in use. The area of ROC curves showed to be an appropriate index to evaluate these algorithms, and indicated that the intensity of 40 dBspl could be set as a threshold to obtain a suitable classification in all the algorithms.