WPTNet-WPL: A Wave-Prior Transformer with Wasserstein Prior Learning for Interpretable Auditory Brainstem Response...

Auditory brainstem response (ABR) testing typically relies on repeated recordings under multiple sound pressure level (dB) conditions and manual interpretation based on clinical experience. This approach suffers from significant subjectivity, time-consuming procedures, and poor consistency in results, thereby limiting its feasibility for conducting high-frequency, long-term, objective assessments within auditory rehabilitation settings. To address these limitations, this study proposes a deep neural network framework integrating clinical latency priors with adaptive optimal transport regularization, termed the Wave-Prior Transformer Network with Wasserstein Prior Learning (WPTNet-WPL). The framework performs binary normal/abnormal classification and five-class hearing-threshold classification from an ABR recording averaged over 1024 sweeps at a single stimulus intensity within the studied 10-100 dB nHL range. Its prediction target is the hearing-threshold category rather than the location of an individual ABR wave peak. The model first employs a one-dimensional convolutional network to extract local waveform features from ABR waveforms. The Dynamic Weighted Prior Attention (DWPA) mechanism then uses canonical Wave I, Wave III, and Wave V latency windows as soft physiological anchors. These anchors guide attention toward physiologically plausible ABR latency regions while preserving data-driven temporal features relevant to threshold classification. Furthermore, the proposed Wasserstein Prior Learning (WPL) module dynamically adjusts the Gaussian mixture prior distribution based on optimal transport theory, enabling adaptive alignment between the learned attention distribution and the physiological prior. This achieves synergistic optimization of prior knowledge and data-driven learning. Experimental results demonstrate that WPTNet-WPL achieves 90.29% accuracy in the binary classification task (normal/abnormal) and 82.26% accuracy in the five-class auditory threshold classification task, both significantly outperforming mainstream baseline models. The framework showed robust overall performance on single-intensity averaged ABR recordings acquired across the studied 10-100 dB nHL range. Auditory rehabilitation assessment and longitudinal functional monitoring remain potential future applications requiring prospective longitudinal validation across multiple centers.