A Transfer Learning-based Model for Individualized Clustered Seizure Prediction Using Intracranial EEG

Clustered seizures are prevalent among people with epilepsy and can increase mortality risk. While past research has mainly focused on seizure cluster detection, a few recent studies predict seizure clustering by determining whether there will be more seizures in the next 24 hours after the termination of a seizure. Moreover, personalized prediction of clustered seizures in the presence of limited and imbalanced data remains an outstanding problem. We address this problem using a novel transfer learning model to predict seizure clustering within a 24-hour window. To compensate for the limited and imbalanced available data, for each target patient, the model combines trained individual-level predictive models of the target patient and two other patients whose seizure patterns are similar to those of the target patient. Approximate Kullback-Leibler divergence is used to measure the similarity between patients in high-dimensional data. The proposed model is evaluated on a long-term ambulatory intracranial EEG dataset. Compared with individualized predictive models, the proposed model improves F1 scores for patients with limited or highly imbalanced data by up to 51.0%. In addition, the proposed model achieves an average F1 score of 0.702 and an area under the precision-recall curve of 0.809. Our model can be clinically helpful in guiding the treatment of clustered seizures.