A hybrid spiking neural network with contrastive pretraining for interpretable seizure forecasting using explainable AI

Neonatal seizures remain one of the most diagnostically demanding problems in intensive care, partly because the EEG signatures are subtle and partly because expert readers miss roughly one in four events under standard monitoring conditions. We address this gap with a framework that integrates three ideas not previously combined in the neonatal seizure literature: self-supervised contrastive pretraining, neuromorphic spiking neural networks, and a multi-method explainability suite grounded in anatomical electrode mapping. Working with the Helsinki University Hospital Neonatal EEG Dataset—approximately 5800 h of continuous recordings from 79 term neonates—we first apply a SimCLR-style pretraining stage on the majority interictal class to learn general EEG representations without requiring seizure labels. These learned embeddings are concatenated with spectral features derived from Welch’s method, producing a 224-dimensional input for an attention-augmented spiking neural network whose core units are leaky integrate-and-fire neurons simulated over 50 timesteps. The combination preserves temporal structure in a way that standard feedforward classifiers cannot. Class imbalance (approximately 6.48: 1) is handled through focal loss rather than resampling, retaining the full training distribution. A strict patient-level stratified split—55 training, 12 validation, 12 test patients—ensures no patient appears in more than one partition, which prevents the data leakage that inflates reported performance in a notable share of published seizure-detection work. On the held-out test set, the model achieves 90.39% accuracy, 90.02% preictal recall, and an AUC-ROC of 0.910. Interpretability is provided by five complementary attribution methods—Integrated Gradients, SHAP, LIME, saliency gradients, and attention profiling—all mapped back to the International 10–20 electrode system. Temporal channels account for approximately 40% of model decisions, central channels 25%, and frontal channels 20%, a spatial pattern consistent with established neonatal neurophysiology. With roughly one million parameters and 15–20% spike sparsity, the architecture is sized for deployment on resource-constrained NICU hardware. Graphical abstract