CMEpiNet: Complex-Valued Multimodal Epilepsy Detection Network Model

Figures

• 

  Figure 1. Overall framework of CMEpiNet. EEG, ECG, and EMG signals are first encoded by the shared complex-valued feature extractor (CMFE), then aligned across modalities by the dual-level semantic alignment fusion module (DLSAF). The aligned features are stacked into a T×d×3 feature block and classified as seizure or non-seizure by the 3D spatial-attention classifier (3DCSA).

• 

  Figure 2. Architecture of the CMFE module. For each modality, the analytic signal is processed by two parallel complex-valued branches: a 1D time-domain phase branch and a 2D STFT-based time-frequency branch. The resulting amplitude and phase features are pooled and concatenated to form the complex-domain feature vector for that modality.

• 

  Figure 3. Architecture of the DLSAF module. The first level enforces cross-modal consistency in a shared alignment space through reliability gating and window-level softmin matching (Lcm). The second level aligns the mean and covariance of modality-specific features in a latent seizure space (Ldist).

• 

  Figure 4. Architecture of the 3DCSA classifier. Stacked 3D convolutional blocks jointly model the temporal, feature, and modality dimensions of the fused feature block. A spatial attention map then reweights informative regions before a fully connected Softmax layer outputs the seizure probability.

• 

  Figure 5. Threshold-dependent sensitivity–FA/h curve of CMEpiNet on the SeizeIT2 test set.

• 

  Figure 6. Comparison of sensitivity and FA/h on the SeizeIT2 test set.

• 

  Figure 7. Ablation study of CMEpiNet on the SeizeIT2 test set in terms of sensitivity and FA/h.

• 

  Figure 8. AUROC and AUSF trends of CMEpiNet and the baseline model under different perturbation conditions.

• 

  Figure 9. Performance degradation under window-level Gaussian noise replacement.

• 

  Figure 10. Performance degradation under auxiliary-modality temporal shifts.