A Dynamic Time Warping-Aware Series-Temporal Transformer for Automated Thresholding of Auditory Brainstem Responses.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Auditory brainstem response (ABR) is crucial for hearing assessment but relies on subjective visual inspection for threshold determination.
• Current statistical methods for ABR analysis often overlook valuable inter-waveform correlations, limiting accuracy and efficiency.
• Human bias in ABR threshold estimation can impact diagnostic reliability.

Purpose of the Study:

• To develop a novel deep learning framework for automated and objective auditory brainstem response (ABR) threshold estimation.
• To leverage inter-waveform dependencies and temporal dynamics in ABR signals for improved threshold prediction.
• To overcome the limitations of manual ABR analysis and existing statistical approaches.

Main Methods:

• Proposed the Dynamic Time Warping (DTW)-Aware Series-Temporal Transformer (DTWA-STformer) framework.
• Utilized a DTW similarity-aware Series Transformer with stimulus level-informed positional encodings to capture waveform dependencies.
• Employed a hierarchical multi-scale Temporal Transformer for feature extraction and a multi-class classifier for threshold prediction.

Main Results:

• DTWA-STformer achieved high accuracies on large-scale human and mouse datasets (e.g., 92.08% exact-match on Dataset I).
• The model demonstrated superior performance compared to state-of-the-art methods in ABR threshold estimation.
• Achieved excellent performance across different ABR types, including click and tone-pip stimuli in mice.

Conclusions:

• DTWA-STformer provides accurate and objective ABR threshold estimation from pre-recorded waveforms.
• The framework can serve as a valuable post hoc tool for verifying clinician-estimated hearing thresholds.
• Future integration with active learning could further enhance efficiency in threshold determination.