A Deep Learning Model for Automatic Thresholding of Auditory Brainstem Responses: Multicenter and Multispecies...

This study aims to develop a deep learning model that directly estimates auditory brainstem response (ABR) thresholds from stacked waveform series across multiple stimulus intensities, and to validate its performance on multicenter, large-scale, diverse human and mouse datasets.

Independent of traditional binary waveform-by-waveform detection, we propose a transformer-based model that integrates adjacent-level cross-attention, a series encoder, and a supervised contrastive learning branch to capture both intra-waveform temporal structures and inter-series dependencies, aligning with expert thresholding practices. The model was trained and validated on two heterogeneous human datasets (Human Dataset I: 8350 subjects; Human Dataset II: 136 subjects) and two publicly available mouse datasets (Mouse Dataset I: 8259 mice; Mouse Dataset II: 351 mice). Model performance was evaluated using exact-match accuracy and tolerance accuracy within ±5 and ±10 dB. Ablation studies were conducted to assess the contribution of each architectural component.

On Human Dataset I, the model achieved 93.07% exact-match accuracy and 99.32% accuracy within ±10 dB under hold-out validation. It maintained strong generalization on the external Human Dataset II (94.24% exact; 99.48% within ±10 dB) and demonstrated robust performance across age groups and hearing conditions. On Mouse Dataset I, the model achieved accuracies of 76.23%, 98.79%, and 99.82% for click-evoked ABRs, and 65.30%, 94.16%, and 99.06% for tone-evoked ABRs at exact-match, ±5 dB, and ±10 dB tolerance levels, respectively. On Mouse Dataset II, accuracies reached 57.53%, 91.51%, and 97.55%. These results surpass those of previously published algorithms on the same datasets. Performance degradation in all ablated variants confirmed the contribution of each model component.

Unlike prior methods focusing on waveform-level detection of the presence or absence of an ABR, this study presents the first deep learning model for automatic thresholding of ABRs from multi-level waveform stacks. It demonstrates strong generalizability across centers, species, and stimulus types, highlighting its potential for clinically applicable, efficient, and automated threshold estimation. Future work should focus on broader external validation and the integration of the model into real-time ABR acquisition workflows, enabling concurrent threshold estimation during ongoing measurements for improved clinical utility.