Artikel
Developping an epileptic seizure warning system
Entwicklung eines Warnsystems für epileptische Anfälle
Suche in Medline nach
Autoren
Veröffentlicht: | 25. Mai 2022 |
---|
Gliederung
Text
Objective: Background: One third of epilepsy patients do not become seizure free from medical therapy and only few benefit from surgery or brain stimulation. They suffer from the unpredictability of seizures, falls, social stigmatisation etc, and would profit by a seizure warning device, offering opportunities of protection and intervention.
Aim: to develop an implantable mobile warning system, we analysed algorithms for binary classification of intracranial EEG data (iEEG) to identify preictal states.
Methods: iEEG from 14 patients with longstanding pharmacorefractory epilepsy (n=13 depth electrodes, n=1 subdural electrodes, seizures recorded 3 – 38, duration of recording 139-257 h per patient) were analysed. Seizure were annotated by an experienced epileptologist (GL). A preictal episode was defined as the time-period from 65 to 5 min before a seizure and an interictal episode was defined beginning 240 min after a seizure and before the next seizure, respectivelly.
Patient specific models were estimated based on the patient’s previous seizures using discrete spectral density bands, inter-channel correlation in the time and frequency domain, and eigenvalues of the correlation matrices; followed by analysis of variance (ANCOVA f-value) and finally classification by a support vector machine (SVM).
Main outcome parameter was the identification of a true preictal episode (=prediction) of the last seizure of the patient. The previous seizures were used to train the system
Results: The last seizure could be predicted correctly in 10 out of 14 pts (i.e. ROC AUC (SVM) > 0.5, Table 1 [Tab. 1]) with a mean of 0.71 over all pts. Further tests showed that a prediction based only on the principal component (PC) was possible for 8 of 14 pts. (ROC AUC (PC), mean 0.683).
Conclusion: Our data shows that seizures can be predicted in the majority of patients but that the algorithm is sensitive to other dynamics such (PC) such as the sleep-wake-cycle. These confounding factors contribute to the unstationarity of the ground truth. This has negative effects on the training in rather short-term data sets (Müller et al. [1], Eberlein et al. [2]).
References
- 1.
- Müller J, Yang H, Eberlein M, Leonhardt G, Uckermann O, Kuhlmann L, Tetzlaff R. Coherent false seizure prediction in epilepsy, coincidence or providence? Clin Neurophysiol. 2022 Jan;133:157-164. DOI: 10.1016/j.clinph.2021.09.022
- 2.
- Eberlein M, et al. Evaluation of machine learning methods for seizure prediction in epilepsy. Curr Dir in Biomed Eng. 2019;5(1):109-112.