Detecção e localização de crises epilépticas em sinais de EEG utilizando aprendizado de máquina e inteligência artificial explicável

Abstract

Epilepsy is a neurological condition that affects millions of people worldwide and significantly impacts individuals’ quality of life. Epileptic seizures, which are transient events, vary in manifestations, including motor, sensory, and consciousness alterations, and represent a challenge both in diagnosis and management. This work proposes an innovative methodology for the detection and localization of epileptic seizures, using machine learning and explainable artificial intelligence approaches to optimize the identification process. The proposal is divided into two approaches: a generalist one, which uses simplified models with explainable feature and channel reduction, and a specific one, which personalizes detection for each patient based on a single electroencephalogram (EEG) channel. In the generalist approach, feature and channel reduction was explored, achieving performance above 0.95 in accuracy, precision, recall, and f1-score metrics using only six features and five channels. The use of the SHAP method allowed for the interpretation of each feature’s contribution by channel, reinforcing the explainability of the models and aligning the results with the visual knowledge of EEG specialists. The methodology proved effective, ensuring good generalization to the dataset with different patients. In the specific approach, a bipolar montage with adjacent electrodes was introduced, creating 58 channel combinations, and a centrality measure with topographic visualization was applied to identify the most relevant channels for each patient. Additionally, a 3-second temporal filter was developed to reduce the model’s false positives. A personalized supervised learning model, extreme gradient boosting, was trained for each patient using only one EEG channel. The results for the three investigated patients were remarkable, with accuracy scores of 1; 0.99; and 0.88, highlighting the feasibility of seizure detection using a single channel, considering the topographic location of 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, facilitating the detection and localization of patients with epilepsy.