EEG-based Prediction of Neuronal Excitability using Deep Learning

Epilepsy is one of the most common neurological diseases, affecting approximately 50 million people worldwide. Despite the availability of antiseizure medicines, one third of patients develop drug-resistant epilepsy. For them, closed-loop neuromodulation represents a promising alternative, as stimulation is delivered only when needed. However, most existing systems still rely on seizure detection, despite several studies demonstrating that predicting changes in network excitability is feasible, justifying a shift towards more proactive approaches. This work utilises a dataset of 16 rats, each providing 576 samples consisting of a 10-second intracranial electroencephalography (iEEG) recording paired with the slope extracted from a field excitatory postsynaptic potential (fEPSP) waveform. Three machine learning models, namely Ridge regression, Random Forest, and XGBoost, and four deep learning architectures, namely a convolutional neural network (CNN), a dual-branch CNN, a dual-branch CNN combined with a temporal convolutional network (TCN), and a dual-branch CNN-TCN with multi-head self-attention (MHSA), were trained and evaluated under a leave-one-subject-out (LOSO) protocol. The results demonstrate that deep learning approaches consistently outperformed machine learning baselines, with the dual-branch CNN-TCN achieving the strongest performance. High inter-subject variability was confirmed as the main challenge, although the findings may inform future work towards predictive closed-loop neuromodulation in drug-resistant epilepsy.