Disentangled and Side-Aware Unsupervised Domain Adaptation for Cross-Dataset Subjective Tinnitus Diagnosis

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Electroencephalogram (EEG)-based tinnitus classification is crucial for diagnosis and treatment.
• Current models struggle with generalization due to non-stationary EEG signals and dataset variations.
• Developing cross-dataset adaptable models is essential for robust tinnitus diagnosis.

Purpose of the Study:

• To propose a novel model, Disentangled and Side-aware Unsupervised Domain Adaptation (DSUDA), for cross-dataset tinnitus diagnosis.
• To enhance the generalization capability of EEG-based tinnitus classification models.
• To mitigate distribution discrepancies across different EEG datasets.

Main Methods:

• Developed a disentangled auto-encoder to separate class-irrelevant information from EEG signals.
• Implemented a side-aware unsupervised domain adaptation module to handle domain variance.
• Aligned left and right ear EEG signals to address inherent pattern differences.

Main Results:

• DSUDA demonstrated significant improvements over state-of-the-art methods in cross-dataset tinnitus diagnosis.
• The model successfully generalized to new, unseen datasets.
• Achieved superior performance based on comprehensive evaluation criteria.

Conclusions:

• The proposed DSUDA model effectively generalizes to new datasets for tinnitus diagnosis.
• DSUDA enhances the reliability and applicability of EEG-based tinnitus classification.
• This approach offers a promising solution for robust cross-dataset tinnitus diagnosis.