A Dual-input deep learning architecture for classification and latency estimation in ABR signals.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Auditory brainstem response (ABR) assesses auditory pathway function.
• Manual analysis of ABR wave V is time-consuming and subjective.
• Automated detection methods are needed to improve efficiency.

Purpose of the Study:

• Develop a multi-task deep learning pipeline for simultaneous wave V detection and latency prediction.
• Introduce a paired-signal approach using high-intensity reference signals to enhance model performance.
• Improve the clinical usability and accuracy of ABR analysis.

Main Methods:

• A multi-task deep learning model with a backbone and two branches (classification and regression) was designed.
• A paired-signal approach was implemented, feeding the model with test signals and their 80 dB references.
• Transfer learning was used, initializing the classification branch with features from the trained latency-prediction network.

Main Results:

• The joint multi-task model outperformed single-task approaches.
• Achieved an F1-score of 0.92 for wave V classification.
• Attained an R-squared value of 0.90 for wave V latency regression.

Conclusions:

• Deep learning, particularly convolutional neural networks, shows significant promise for ABR analysis.
• The proposed methods can streamline clinical workflows for diagnosing auditory disorders.
• Automated ABR analysis can enhance diagnostic accuracy and efficiency.