Optimising seizure prediction with reduced computational resources using depthwise CNN

Existing deep learning models for epileptic seizure prediction are accurate but parameter-heavy, which limits their deployment on wearable and other resource-constrained edge devices. We present DSCNNNet, a 3D depthwise separable convolutional network operating on Mel-frequency cepstral coefficient (MFCC) features extracted from scalp EEG. On the CHB-MIT dataset DSCNNNet reaches 89. 58% sensitivity with 11, 714 parameters and 45. 75 KB of weight memory – roughly an order of magnitude fewer parameters than comparable CNN baselines at similar sensitivity. Replacing standard 3D convolution with its depthwise separable form reduces the per-layer multiply – accumulate cost by approximately 10× without a loss of predictive performance, supporting real-time operation on low-power edge platforms.