Banca de DEFESA: JUSCIAANE CHACON VIEIRA

Classification of normal and epileptic seizures from EEG signals using Explainable Artificial Intelligence

machine learning, explainable AI, electroencephalography, epilepsy.

Epilepsy is a neurological condition that affects millions of people worldwide and significantly impacts the quality of life of individuals. Epileptic seizures, transient events, vary in manifestations, including motor, sensory, and consciousness alterations, and present challenges in both diagnosis and management. This work proposes an innovative methodology for detecting epileptic seizures, utilizing machine learning approaches and Explainable Artificial Intelligence (XAI) to optimize the identification process. The proposal is divided into two approaches: a generalist approach, which uses simplified models with feature and channel reduction, and a specific approach, which customizes detection for each patient based on a single EEG channel. In the generalist approach, feature and channel reduction was explored, achieving performance above 0.95 in accuracy, precision, recall, and F1-core, using only six features and five channels. The methodology proved effective, ensuring good generalization across the database with different patients.
In the specific approach, a personalized supervised learning model, eXtreme Gradient Boosting, was developed for each patient, using only one EEG channel. The results for the three patients investigated were remarkable, with accuracies of 1, 0.99, and 0.88, highlighting the feasibility of detecting seizures using a single channel, considering the topographic location of the seizures in each individual. This study highlights the potential to substantially reduce the number of features and channels required for epileptic seizure detection without compromising accuracy, and reinforces the importance of personalized models for each patient. Furthermore, the research contributes to the advancement of wearable devices for continuous epilepsy monitoring, enabling early detection and real-time follow-up.