Abstract
Epilepsy is one of the most common neurological disorders, affecting up to 1% of the world population. In cases where epileptics are not responsive to medication, resective surgery has been found to be an effective treatment for seizure freedom or increase in control. For this, epilepsy must be correctly diagnosed. That evaluation is typically performed by VideoEEG monitoring by physicians, a subjective task that can easily be aided by movement quantification and pattern recognition techniques. Regarding their onset location, seizures can be classified as having extratemporal or temporal origin. This contribution contains an analysis of infrared data from 143 seizures from 31 different patients, recorded in the Epilepsy Monitoring Unit of University of Munich, for seizure classification. Regarding classification, using the seizures' duration and the existence of movements of interest, an 86% +/- 17% AUC is obtained using 10-fold cross validation and a Support Vector Machine model. Using the video recordings, the region of interest (bed) was first detected with 88% correct detections and 22% overdetections using a threshold-based method, and all beds were rotated to a vertical position for consistency. Lastly, seizures were classified with a Convolutional Neural Network and a Multilayer Perceptron, obtaining a 65% AUC and showing that the model is better at classifying extratemporal seizures, mostly due to class imbalance. The developed approach shows potential for clinical decision support using a non-intrusive and low cost solution.
Item Type: | Journal article |
---|---|
Faculties: | Medicine |
Subjects: | 600 Technology > 610 Medicine and health |
Language: | English |
Item ID: | 81469 |
Date Deposited: | 15. Dec 2021, 14:58 |
Last Modified: | 15. Dec 2021, 14:58 |