EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

PST-Mamba: Parallel Spatio-Temporal Mamba With Adaptive Multi-Band Transformer Fusion for EEG-Based Schizophrenia Diagnosis

Fatemeh Makhloughi, Ateke Goshvarpour

IEEE Access · 2026 · 14 · 114381-114399

DOI: 10.1109/access.2026.3717222 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Fatemeh Makhloughi und 1 weiteren Autor:innen. Die Arbeit nennt 56 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Diagnosing schizophrenia from electroencephalography (EEG) signals is challenging due to complex spatial and temporal patterns in different frequency bands. This study proposes parallel spatio-temporal Mamba (PST-Mamba), a novel parallel multi-branch deep learning framework for automated schizophrenia classification. The architecture decomposes EEG signals into four frequency bands, processing each through dedicated spatio-temporal pathways. Each pathway comprises spatial convolutional modules for learning data-driven electrode configurations and temporal Mamba modules that use selective state space models to capture long-range temporal dependencies with linear computational complexity. Furthermore, the framework employs an adaptive Transformer mechanism that uses multi-head attention and a learnable classification token to dynamically fuse band-specific features. This attention-based fusion enables adaptive, sample-specific weighting of frequency bands and provides an interpretable visualization of diagnostic contributions. Two datasets were analyzed: Dataset 1 (28 participants: 14 schizophrenia patients, 14 healthy controls) and dataset 2 (64 participants: 32 schizophrenia patients, 32 healthy controls), both with 19-channel resting-state EEG recordings. Evaluation on two independent datasets showed that, on dataset 1, PST-Mamba achieved 90.53% accuracy with leave-one-subject-out (LOSO) cross-validation and 94.83% accuracy with 5-fold cross-validation at the optimal 10-second segment length. On dataset 2, the model attained 91.19% (LOSO) and 94.15% (5-fold) accuracy, outperforming Transformer, Bidirectional Long Short-Term Memory (BiLSTM), and temporal convolution network baselines. Our findings confirm that both temporal duration and frequency-specific features are integral to achieving accurate classification. Our approach shows potential for automated schizophrenia classification and provides interpretable insights through frequency band attention analysis.

Zitieren

APA

Makhloughi, F., & Goshvarpour, A. (2026). PST-Mamba: Parallel Spatio-Temporal Mamba With Adaptive Multi-Band Transformer Fusion for EEG-Based Schizophrenia Diagnosis. IEEE Access, 14, 114381-114399. https://doi.org/10.1109/access.2026.3717222

Vancouver

Makhloughi F, Goshvarpour A. PST-Mamba: Parallel Spatio-Temporal Mamba With Adaptive Multi-Band Transformer Fusion for EEG-Based Schizophrenia Diagnosis. IEEE Access. 2026;14:114381-114399. doi:10.1109/access.2026.3717222

BibTeX
@article{Makhloughi2026,
  author = {Makhloughi, Fatemeh and Goshvarpour, Ateke},
  title = {PST-Mamba: Parallel Spatio-Temporal Mamba With Adaptive Multi-Band Transformer Fusion for EEG-Based Schizophrenia Diagnosis},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {114381--114399},
  doi = {10.1109/access.2026.3717222}
}
Kostenlos starten →