Abstracts

Rationale: In the last decade, a lot of efforts were made to build a system for home monitoring of epileptic patients. Since the gold standard, video/EEG (electroencephalography), is uncomfortable for long-term monitoring and labor-intensive for the care givers, our main focus lies on the detection of motor convulsions related to the epileptic seizures using user-friendly motion sensors. Herein, four tri-axial accelerometers attached to the wrists and ankles are used.Methods: Our study includes 10 refractory epilepsy children with (tonic-)clonic seizures. Patients were monitored during several nights using both video/EEG and accelerometry (ACM). Annotations were determined by an EEG specialist applying the ILEA (International League against Epilepsy) recommendations. The recordings are divided into 2-second epochs with 75% overlap, but only epochs with motion activity (standard deviation within the epoch higher than 10 mg) are kept for further analysis. Multiple features from the time, frequency, continuous wavelet transform, packet wavelet transform and cross-recurrence plot domain are extracted from each epoch. Using feature selection methods (filter method based on mutual information in the first stage, and wrapper LS-SVM (Least-Square Support Vector Machines) method with forward search in the second stage), this feature set is reduced to two most relevant features. In addition to the first five clonic seizures of patients 30, 40 and 51 (see Table 1), five more clonic seizures from four patients are included in the feature selection and LS-SVM tuning and training (which are not included in Table 1 since their data were not used for testing). Tuning of the LS-SVM model with a radial basis function kernel is performed with 5-fold crossvalidation and maximization of the F1-score parameter. The obtained model is tested both on the rest of the data of three patients with clonic seizures, and on data from patients with tonic-clonic seizures. The classification performances are estimated describing sensitivity (at least one epoch within the seizure is detected), mean latency (delay between the first detected epoch and clinical onset of the seizure) and specificity on the level of epoch detection. Results: Table 1 summarizes the classification results. Even though the model is built for the detection of clonic seizures, it is able to detect the tonic-clonic seizures of other patients. However, 24 clonic seizures are missed. This is partially explained by the morphology of these seizures. While the seizures B and C on Figure 1 are labeled as clonic seizures, compared to the seizure A, these seizures do not involve the four limbs so much; in addition, they are quite weak and short. Excluding seizures with lower motor manifestations of monitored limbs would enhance the system performance.Conclusions: Accelerometry is capable of detecting motor seizures with a repetitive rhythm. Using only an ACM system, nocturnal monitoring of epileptic children is both comfortable and user-friendly. It enables long-term monitoring, better treatment management and easier follow-ups.