Write a PREreview

Objective: Tinnitus affects 10-15% of the global population yet lacks objective diagnostic biomarkers. This study evaluated machine learning and deep learning approaches for automated tinnitus classification using electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) data through parallel unimodal investigations.Methods: Two independent datasets were analyzed: 64-channel resting-state EEG from 80 participants (40 tinnitus patients, 40 controls) and resting-state fMRI from 38 participants (19 tinnitus patients, 19 controls). EEG analysis employed microstate feature extraction across four clustering configurations (4-state through 7-state) and five frequency bands, yielding 440 features per participant. Complementary analysis transformed Global Field Power signals into continuous wavelet transform images for convolutional neural network classification. fMRI analysis utilized slice-wise spatial pattern classification through deep learning models and hybrid architectures combining pre-trained networks (VGG16, ResNet50) with traditional classifiers. Subject-level 5-fold cross-validation with nested slice selection protocols ensured unbiased evaluation and prevented data leakage. Comprehensive artifact assessment confirmed classification was driven by genuine neural signals rather than motion or electromyographic contamination.Results: EEG microstate analysis revealed systematic disruptions in tinnitus-related neural dynamics, with the largest effect in gamma-band microstate B occurrence (healthy: 56.56 vs. tinnitus: 43.81 events/epoch, Cohen's d = 2.11, p < 0.001) and reduced alpha-band coverage. Random Forest and Decision Tree algorithms achieved 98.8% accuracy on microstate features. Deep learning analysis of wavelet-transformed signals demonstrated VGG16 superiority, achieving 95.4% accuracy for delta-band features. fMRI spatial pattern analysis identified slice 17 with optimal discrimination (99.0% ± 0.4% accuracy). The hybrid VGG16-Decision Tree model achieved 98.95% ± 2.94% accuracy with 98.84% ± 2.05% ROC-AUC on combined high-performing slices.Conclusion: Independent EEG and fMRI analyses provided complementary neural signatures for tinnitus classification. Spatially localized fMRI alterations in auditory-related regions and temporally dynamic EEG network disruptions in gamma and alpha bands support conceptualizing tinnitus as a multi-network disorder involving both localized connectivity changes and large-scale network instability. Future research employing simultaneous multimodal acquisition would enable direct spatiotemporal integration of hemodynamic and electrophysiological alterations.