Deep Motion Analysis for Epileptic Seizure Classification.

Area of Science:

• Neurology
• Computer Science
• Biomedical Engineering

Background:

• Epileptologists use visual semiology like facial expressions and pose to identify seizures.
• Automatic seizure detection is challenging due to variations in appearance.
• Facial and pose dynamics are underutilized in current automated systems.

Purpose of the Study:

• To develop a multi-modal deep learning approach for classifying mesial temporal lobe epilepsy (MTLE) and extra-temporal lobe epilepsy (ETLE).
• To quantitatively fuse facial expressions and pose dynamics for improved epilepsy classification.
• To create a virtual assistant for objective clinical analysis and decision-making.

Main Methods:

• A deep learning model combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) was used.
• Spatiotemporal features were extracted from facial and pose semiology in video recordings.
• A dataset of 18 patients (12 MTLE, 6 ETLE) was utilized for experiments.

Main Results:

• Facial semiology and body movements were effectively recognized and tracked.
• The fusion model achieved an average test accuracy of 92.10% using multi-fold cross-validation.
• A leave-one-subject-out cross-validation achieved 58.49% accuracy, demonstrating robust generalization.

Conclusions:

• The proposed approach effectively models semiology features to discriminate between temporal and extra-temporal epilepsy.
• This technology can serve as a virtual assistant, enhancing patient safety and clinical decision support.
• The method offers objective clinical analysis, saving time and improving diagnostic accuracy.