AI Driven Hybrid Brain Computer Interface for Epilepsy

다국어 초록 (Multilingual Abstract)

Epilepsy is a neurological disorder prevalent worldwide affecting individuals not just physically but also mentally and socially causing irreversible damages. It is caused when brain neurons are unable to regulate electrical signals resulting in seizures. The diagnosis of this life threatening disorder is thus critical. The current EEG test is one of the most common and effective epilepsy diagnosis medium, however it is prone to human error and time consuming. Therefore, we propose an Artificial Intelligence model that can diagnose these seizures with accuracy. Our model implements a combination of Convolutional Neural Network and Long Short-Term Memory architecture. Trained on 80% of total EEG recordings of 129 files with one or more seizures and 198 seizure events, this model used data of 23 pediatric subjects between the age of 3 -22 with 5 males, 17 females and 1 undisclosed. Data recordings with 10-20 systems of EEG electrode positions were obtained from Children’s Hospital Boston. The EEG signals obtained from the data were first segmented into fixed length windows appropriate for model input. Then they were normalized for a consistent signal amplitude. To remove background noises and artifacts, these processed recordings were filtered and then the recording segments were labeled based on the presence or absence of epileptic seizures. The data was then split into 80% training and 20% testing sets. For spatial feature extraction and capturing temporal dependencies in EEG signals CNN and LSTM models were implemented. The model was then cross validated. A confusion matrix was generated to visualize true and predicted classifications and an accuracy of 90% was achieved for currently available datasets. We are planning to include a wide range of EEG recordings with diverse age ranges and conditions to improve reliability. We are trying enhance accuracy by adding extra preprocessing steps. To conclude, this paper presents a methodological approach to analyzing brain activity via EEG dataset with the goal of epileptic detection without claiming medical accuracy or effectiveness.