Ictal Periods Detection in Photoplethysmographic and Electrodermal Signals

• Published in: International Conference on Electrical Engineering, Computing Science, and Automatic Control (Online) pp. 1 - 6
• Main Authors: Ramirez-Peralta, Maria Fernanda, Romo-Fuentes, Maria Fernanda, Tovar-Corona, Blanca, Silva-Ramirez, Martin Arturo, Garay-Jimenez, Laura Ivoone
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.11.2021
• Subjects: Analytical models; Computational modeling; electrodermal activity; epilepsy; Feature extraction; heart rate variability; ictal period detection; Nonhomogeneous media; Support vector machines; Training; Wearable computers; wearables
• ISSN: 2642-3766
• DOI: 10.1109/CCE53527.2021.9633091

The occurrence of an epileptic crisis can generate changes in the autonomous nervous system, given the relation between the zones in which an epileptic crisis generates and propagates, and the zones of the brain that control the involuntary responses of the body. Thus, an investigation aiming at identifying ictal periods in autonomous nervous system-controlled signals that can be, also, continuously recorded through a wearable device is proposed. Two signals are considered, electrodermal activity, obtained from the measurement of the galvanic skin response, and heart rate variability, derived from the analysis of the inter beat interval computed from the photoplethysmographic signal. A database of 11 subjects, composed by these two signals recorded simultaneously with electroencephalography is employed. Time and frequency-domain features were extracted from the electrodermal and heart rate variability signals by taking 4-minute segments previous to the beginning of a seizure as interictal, and 4-minute segments after as ictal, for training, and through a 4-minute windows with a 30-second slide segmentation for testing, while the electroencephalographic signal was taken as reference to obtain the tags for the ictal and interictal periods. The features were classified using a perceptron multilayer neural network trained with scaled conjugate gradient backpropagation algorithm, obtaining the following results, accuracy: 25.15%, recall: 97.49%, specificity: 0.44% precision: 24.86%, and F2 score: 60.12%.