Deep Learning for Patient-Independent Epileptic Seizure Prediction Using Scalp EEG Signals

Epilepsy produces unexpected, recurrent, and irregular nerve cell activities in the brain, known as seizures. These seizures can cause unconsciousness, memory loss, strokes, and brain tumors. The World Health Organization (WHO 2022) suggests that more than 50 million people worldwide have epilepsy.

Researchers have been working on classifying different epileptic stages by monitoring the electrical activity in the brain. Early prediction of the possible occurrence of seizures can help reduce the episodes of epilepsy and its adverse effects.

Electroencephalography (EEG) is a low-cost, viable, and non-invasive solution for monitoring electrical activities in the brain. Researchers in the biomedical field have conducted substantial research on EEG-based seizure prediction.

Medically, a seizure event passes through four brain stages: interictal, pre-ictal, ictal, and postictal. "Interictal stage" refers to the normal brain state of a person. The stage before a seizure event is called the "pre-ictal stage." The brain will be in an "ictal state" during the seizure, and it will shift to the "postictal stage" following the seizure event.

The proposed epileptic seizure prediction model is a patient-independent, deep-learning model. The model uses algorithms to distinguish between the pre-ictal and interictal states of the patient's brain. Detection of a pre-ictal EEG signal indicates the occurrence of a seizure within a defined pre-ictal time. However, the duration of the pre-ictal stage is patient-dependent. It is often assumed to be a constant value when designing models.

The strength of a seizure prediction system depends on its predictive capability and prediction horizon. The predictive capability is the measure of the classification accuracy of the model, where the model tries to classify the given EEG segment as interictal or pre-ictal. Patient-independent seizure prediction models offer accurate performance across multiple subjects within a dataset. Patient-specific models implement subject-specific solutions.

The proposed model uses two patient-independent Convolution Neural Network (CNN) architectures with different learning strategies: multi-task learning and contrastive learning. CNN architectures can automatically learn salient features from the time-frequency representation of biosignal data. The models can accurately recognize seizures with a one-hour prediction window.

The proposed model achieved state-of-the-art performance for seizure prediction on the CHB-MIT-EEG dataset, demonstrating 88.81% and 91.54% accuracy, respectively, with an up to one-hour early seizure detection capability. The models show superior performance for patient-independent seizure prediction. The architecture also works as a patient-specific classifier with some adaptation.

The learning models were trained on the Mel Frequency Cepstral Coefficients (MFCCs) of the samples. The MFCCs provide a viable feature map for seizure prediction that contains predictive biomarkers. Furthermore, we use model interpretation to understand EEG channel-related variations when the brain stage shifts from the interictal to the pre-ictal state.

The proposed model could achieve a favorably generalized classifier for completely unseen subjects, which is quite a complex task. However, more research is required to develop patient-independent models with an effective real-world solution to the seizure prediction problem.