Abstract
<jats:p>Determining hearing thresholds from auditory brainstem responses (ABR) in clinical practice requires manual interpretation, particularly the assessment of wave V presence across successive stimulus intensity levels. This work presents a comprehensive comparison of the impact of discrete wavelet selection in the two-dimensional discrete wavelet transform (2D-DWT) and region of interest (ROI) window size on automatic hearing-threshold estimation from graphical ABR representations. A total of 106 wavelet variants from 7 families available in PyWavelets were evaluated alongside 6 ROI window variants. Features were extracted via 2D-DWT decomposition and template matching, and classified using a support vector machine (SVM). Two strategies were compared: a multi-template approach, in which a separate template was constructed for each intensity group, and a single-template approach based on one global template. The ABR-based hearing threshold was defined as the lowest intensity at which wave V was detected; performance was measured as the percentage of cases in which the estimation error did not exceed ±10 dB HL. The multi-template approach proved more effective, achieving a peak accuracy of 89.94% in reproducing the ABR-based (electrophysiological) hearing threshold. Both wavelet selection and ROI window size significantly affected result stability and overall performance. The analysis may serve as a practical reference for wavelet selection in future 2D-DWT-based ABR studies.</jats:p>