EEG-VTSFN: An enhanced approach for differentiating between unipolar and bipolar mood disorders.

Accurate differentiation between Bipolar Disorder (BD) and Unipolar Depression (UD) is a significant clinical challenge due to overlapping manifestations and a lack of objective biomarkers. While EEG offers high temporal resolution, its psychiatric utility for diagnostic differentiation remains limited. To address this, we propose the Variational Autoencoder (VAE)-based Time-Space Fusion Network (VTSFN). VTSFN transforms complex EEG signals into latent-space representations to mitigate noise and construct stable latent co-activation patterns. By integrating Long Short-Term Memory (LSTM) and attention mechanisms, the model captures temporal dynamics to identify discriminative patterns. Our results demonstrate that VTSFN achieved the highest accuracy and AUC point estimates among the evaluated models, with a test-set accuracy of 79.3% (95% CI, 65.5%-93.1%) and an AUC of 82.7% (95% CI, 63.2%-98.5%), without significant paired AUC differences at p < 0.01. Validation in a diagnostic-switch (tBD) cohort (n = 22) identified 17/22 converters (sensitivity, 77.3%; 95% CI, 59.1%-90.9%), requiring further longitudinal validation. Exploratory feature analysis revealed that, compared with UD patients, BD and tBD patients exhibited weaker latent co-activation in right parieto-temporal and temporo-occipital regions, alongside stronger latent co-activation in left frontal-right parietal and temporal regions. These model-derived channel similarity patterns require independent validation.