Functional Ultrasound Imaging of Epileptic Dynamics: A Model-Based Investigation

Functional ultrasound imaging (fUSI) enables access to deep brain activity through neurovascular coupling, complementing the surface-biased perspective of EEG [1, 2]. In this study, we evaluated whether synthetic fUSI signals, derived from stereo EEG (sEEG), could aid seizure phase classification, particularly for detecting preictal states relevant to early intervention. We trained one-dimensional convolutional neural networks (1D-CNNs) on sEEG recordings from 34 drug-resistant epilepsy patients from the HUP iEEG Epilepsy Dataset [2]. Signals were resampled to 500 Hz, high-pass filtered at 1 Hz, and cleaned using spectral interpolation and bipolar rereferencing. For each seizure, 2-minute preictal and 30-second postictal segments were extracted, along with matched interictal epochs. Data were split using 3-second input windows. Synthetic fUSI signals were obtained by convolving sEEG data with an experimentally derived hemodynamic response function (HRF), modeling neurovascular dynamics [3]. To support training, we employed a Wilson-Cowan neural mass model with dynamic excitatory-inhibitory coupling to simulate transitions across interictal, preictal, and ictal phases [4]. Synthetic electrophysiological signals were convolved with the HRF to produce fUSI-like data. The CNN architecture consisted of three convolutional layers with ReLU activation, batch normalization, dropout, and global average pooling. Regularization and early stopping were used to prevent overfitting. Pretraining on the synthetic dataset significantly enhanced performance on real sEEG recordings. On the test set, the model achieved 0.85 accuracy for interictal vs. ictal classification and 0.75 for interictal vs. preictal, demonstrating strong discriminative capability, including during early seizure phases. Comparable results were obtained using synthetic fUSI signals derived from the sEEG data, with accuracies of 0.78 and 0.70, respectively. Our work suggests that fUSI data may provide phase-specific information relevant to seizure monitoring, indicating its potential utility in clinical neuroimaging applications. Beyond classification, such models may support interpretability by revealing phase-specific dynamics in latent space. Future work will explore scaling this approach to larger cohorts and integrating scalp EEG with synthetic or real fUSI to develop robust, non-invasive, deep-aware seizure prediction systems suited for closed-loop neuromodulation.