Enhanced Classification of Tinnitus Patients Using EEG Microstates and Deep Learning Techniques

Objective: This study aims to deepen the understanding and classification of tinnitus through a comprehensive analysis of EEG signals utilizing innovative microstate analysis techniques and cutting-edge machine learning approaches. Methods: EEG data were meticulously collected from 16 healthy individuals and 20 tinnitus patients, segmented into five frequency sub-bands (delta, theta, alpha, beta, gamma) and categorized into four distinct microstate configurations (4-state, 5-state, 6-state, 7-state). Advanced feature extraction techniques were employed to analyze microstate duration, occurrence, coverage, global field power (GFP), and transition probabilities. A diverse array of classifiers was utilized, including traditional models (SVM, DT, RF), deep learning models (DNN, CNN), and pre-trained architectures (VGG16, ResNet50, Xception). Notably, a novel feature-to-image transformation method was integrated to enhance the interpretability of the microstate data. Results: The DNN classifier achieved the highest accuracy of 97.49% in the beta sub-band, employing a 5-state configuration alongside GFP_TC features. Tinnitus patients demonstrated increased activation in the alpha frequency band and exhibited more dispersed microstate distributions, indicating significant alterations in neural dynamics. Additionally, a detailed numerical analysis of transition matrices revealed reduced self-transition probabilities and heightened transitions between states in tinnitus patients, exemplified by a drop from 0.88 to 0.72 in the Delta band for state A. Furthermore, the SVM classifier, when paired with ResNet50 and Xception, reached perfect performance metrics, achieving accuracy, precision, recall, F1-score, and ROC AUC all equal to 1 across all microstates. Conclusion: The findings suggest that microstate analysis, when integrated with advanced machine learning methodologies, serves as a powerful tool for enhancing the diagnosis and comprehension of tinnitus. The high classification accuracies, coupled with detailed microstate assessments, reveal significant neural dynamics differences between tinnitus patients and healthy controls, offering promising implications for clinical diagnosis and treatment strategies .