Generalizable Sample-Efficient Siamese Autoencoder for Tinnitus Diagnosis in Listeners With Subjective Tinnitus.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalogram (EEG)-based neurofeedback is a key area for tinnitus therapy research.
• Current machine learning and deep learning models for EEG analysis are often data-intensive or lack generalizability.
• Expert cognitive prediction remains a common, yet potentially limited, approach in existing studies.

Purpose of the Study:

• To develop a robust and data-efficient model for distinguishing tinnitus from healthy states using EEG neurofeedback.
• To improve the generalizability and reduce data requirements of machine learning models in this domain.
• To address the limitations of expert-driven analysis and data-hungry deep learning methods.

Main Methods:

• Introduction of 'trend descriptor' for noise reduction in EEG signals.
• Utilizing a Siamese encoder-decoder network for accurate alignment and transferable feature learning.
• Supervised boosting to enhance the network's performance across subjects and EEG channels.

Main Results:

• The proposed model achieved high accuracy (91.67%-94.44%) in predicting tinnitus versus control subjects.
• Demonstrated superior performance compared to state-of-the-art algorithms in a subject-independent setting.
• Experiments analyzed EEG neurofeedback responses to 90dB and 100dB sound stimuli.

Conclusions:

• The developed model offers a significant advancement in data-efficient and robust EEG-based tinnitus detection.
• The method shows strong potential for reliable, subject-independent tinnitus diagnosis and neurofeedback therapy.
• Ablation studies confirmed the model's stability and consistent performance across various conditions.