Comparison of Deep Learning Models for Objective Auditory Brainstem Response Detection: A Multicenter Validation...

Area of Science:

• Audiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Clinical auditory brainstem response (ABR) interpretation relies on subjective visual inspection, leading to variability.
• Deep learning (DL) shows potential for objective ABR detection but struggles with real-world clinical data due to small, non-diverse datasets.

Purpose of the Study:

• To evaluate the generalizability of nine DL models for ABR detection using large, multicenter clinical datasets.
• To compare the performance of convolutional neural networks (CNNs), transformer-based models, and hybrid architectures for ABR detection.

Main Methods:

• Utilized a primary dataset of 128,123 labeled ABRs from 13,813 participants for training and testing DL models.
• Assessed nine DL models, including CNNs (AlexNet, VGG, ResNet) and transformers (Transformer, PatchTST, Differential Transformer, Differential PatchTST), on held-out and external datasets.
• Performance metrics included accuracy and area under the receiver operating characteristic curve (AUC), with additional analysis of dataset size, diversity, and auxiliary input features.

Main Results:

• The ResPatchTST model achieved the highest performance on the primary test set (accuracy: 91.90%, AUC: 0.976).
• Transformer-based models, especially Patch Time Series Transformer (PatchTST), demonstrated superior generalization across diverse external clinical datasets.
• Larger, more diverse datasets and the inclusion of acquisition parameters/demographic features improved model robustness and cross-center generalization.

Conclusions:

• Transformer-based DL models show significant potential for accurate and generalizable ABR detection in clinical practice.
• The development of clinically reliable ABR interpretation systems necessitates the use of large, heterogeneous datasets.
• DL offers a path towards more objective and consistent ABR analysis, reducing inter-practitioner variability.