XTinnitusNet: Multi-view robust model ensemble for EEG-based tinnitus diagnosis with noisy labels

Abstract

Tinnitus is a common neurological disease that seriously affects the quality of life of patients. Current deep learning-based tinnitus diagnosis methods face two key challenges: difficulty in extracting high-dimensional tinnitus-related features from non-stationary and low signal to noise ratio(SNR) EEG signals, and vulnerability to noisy labels in training data. To address these, we propose XTinnitusNet, a multi-view graph robust model ensemble. It integrates a Graph Attention Neural Network (GANN) for analyzing brain functional connectivity features and a Multi-scale Convolutional Neural Network (MSCNN) for extracting multi-scale time-series features. This dual-component architecture enhances the model’s ability to capture complex EEG patterns. Additionally, the co-teaching plus mechanism is incorporated into the training process, enabling cross-updating between the MSCNN and GANN components using disagreement data with minimal loss. We evaluated the EEG data of 24 tinnitus patients and 24 healthy subjects using a five-fold cross-validation strategy, with metrics including area under the curve (AUC) and expected calibration error (ECE). Results show XTinnitusNet outperforms baseline models in diagnostic accuracy and robustness across different noise levels. Specifically, it achieves an AUC of 88.33% at a noise rate of 0.1 and maintains a high AUC of 84.02% when the noise rate increases to 0.3. Feature interpretability analysis also shows that XTinnitusNet effectively extracts meaningful functional connectivity features from EEG signals, particularly in the left temporal and frontal cortices. This work provides a robust and interpretable framework for automated tinnitus diagnosis using EEG signals, enhancing diagnostic accuracy and reliability even with noisy labels.