Evaluating Self Supervised Channel Independent JEPA for Pediatric Seizure Detection

Epileptic seizures can be detected as abnormal neuronal electrical activity in the brain and affects wider population including physical safety, mental health, and social and economic life. Timely detection of epileptic seizures is of a major importance. Machine Learning plays an increasing role in accurately detecting seizures from analysing the electrical activity of the brain. Traditionally machine learning algorithms rely on annotations by medical professionals, with quality of the inference accuracy being directly related to the amount of data being annotated. This is a time consuming task. We evaluate the Self Supervised Representation Learning (SSL) via Joint Embedding Predictive Architecture (JEPA) framework on the CHB-MIT dataset. JEPA framework enables learning by predicting latent space representation in order to understand the invariant and causal structure instead of solely relying on label specific cues. This exploratory study investigates whether JEPA, the self supervised learning (SSL) framework, can truly provide comparable results in literature while reducing the amount of annotations required. We explore a hybrid JEPA architecture that combines a hierarchical convolutional tokenizer with a transformer encoder trained using the JEPA self supervised objective. Our baseline model achieves an average AUC of 91.8% for a Leave One Patient Out (LOPO) validation. The results show that the proposed method provides a higher AUC.