همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Mamba Meets Sleep: Do State Space Models Outperform CNNs for EEG Classification?

نویسندگان :

Mostafa Mehrabi¹ Hamed Malek²

1- Shahid Beheshti University 2- Shahid Beheshti University

کلمات کلیدی :

Sleep stage classification،EEG signals،state space models،Mamba،deep learning،convolutional neural networks،subject-independent validation

چکیده :

Automatic sleep stage classification from EEG signals is critical for diagnosing sleep disorders, yet most studies evaluate architectures in isolation. We present a systematic evaluation of 15 models across five families: traditional machine learning, residual CNNs, hybrid ResNet-BiLSTM, efficient CNNs, and state space models. Using Sleep-EDF with leave-one-subject-out cross-validation on 20 subjects, we assessed subject-independent generalization. Mamba Base achieved the highest accuracy (93.32% ± 3.01%) and recall (87.11%), while ResNet8 provided comparable accuracy (93.20%) with lowest variance (2.47%). Surprisingly, adding recurrent layers or increasing depth degraded performance, suggesting simpler architectures better capture temporal dependencies in 30-second EEG epochs. State space models combine transformer expressiveness with linear complexity, making them suitable for clinical deployment. Traditional ML methods achieved 90-91% accuracy with minimal computational cost, offering interpretable alternatives. Our comprehensive comparison reveals clear trade-offs between accuracy, stability, efficiency, and deployment constraints, providing practical guidance for architecture selection based on clinical requirements and resource availability. Implementation code is available at: https://github.com/mostafamehrabii/sleep-stage-classification .