EEG-based seizure detection in patients with intellectual disability: Which EEG and clinical factors are important?. (March 2019)

• Record Type: Journal Article
• Authors: Wang, Lei; Long, Xi; Aarts, Ronald M.; van Dijk, Johannes P.; Arends, Johan B.A.M.
• Is Part Of: Biomedical signal processing and control. Volume 49(2019)
• Page Start: 404
• Page End: 418
• Publication Date: 2019-03
• Subjects: AL seizure alarm length -- AUCPR area under curve (AUC) of P–R curve -- DT detection threshold -- EMG seizure discharge with EMG activity -- FDs false detections -- FDt/h time of FD per hour of recording -- ID intellectual disability -- LOOCV leave-one-out cross validation -- LDA linear discriminant analysis -- NCSS non-convulsive seizure states -- PPV positive predictive value -- P–R precision and recall -- PS prediction score -- RUSBoost random undersampling AdaBoosting -- RF random forests -- RBF radial basis function -- SP fast spike seizures -- SPWA spike-wave seizures -- SVM support vector machines -- WA wave seizures; EEG -- Seizure detection -- Intellectual disability -- Imbalanced data -- Post-processing -- Multilevel analysis -- LDA -- SVM; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2018.12.003
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We constructed the first, long-term EEG dataset of ID patients with the hierarchical annotation (in .XML files) including both EEG and non-EEG information. This work evaluated the real-life data (i.e. highly imbalanced data) by using proper performance criteria, and we showed the relationship between the epoch detection performance (i.e., classification performance) and event detection performance. This work optimized the seizure detection on the imbalanced data by employing a post-processing process (i.e., patient-specific detection thresholds), and we also evaluated the performance using predefined detection thresholds (DTs) (e.g., DT = 0.5). We employed multi-domain EEG features that showed a better discriminative power in our dataset[13], and compared the linear and nonlinear classifiers (LDA vs. SVM) on this heterogeneous dataset by using LOOCV. Important EEG and non-EEG factors were recognized by using a multilevel analysis, which evaluates mixed effects of hierarchical factors. Abstract: Epilepsy is a commonly secondary disability in people with an intellectual disability (ID), affecting 22% of the ID population while 1% of general population. Surprisingly, EEG-based automated seizure detection in the ID population has not yet been sufficiently studied. The reasons are twofold. Firstly, long-term EEG recordings are few due to behavioral problems. Secondly, the annotation of EEG recordings has been proved difficult due to the complex EEG signal abnormalities caused by brain development disorders. As a result, the performance of automated seizure detection for ID people is largely unknown. In this work, we performed automated seizure detection on a retrospective dataset containing 615 h ambulatory scalp EEG from 29 participants with ID, including 91 seizures. To design a generic seizure detector for the ID people, we need to deal with three major problems: highly imbalanced data, heterogeneous dataset and difficult annotation. (1) For the imbalanced data, we used proper performance criteria (e.g., precision and recall curve) and employed a post-processing process (i.e., patient-specific detection thresholds). (2) For the heterogeneous dataset, we employed multi-domain EEG features that showed a better discriminative power in our dataset, and compared the linear and nonlinear (LDA vs. SVM with Gaussian kernel) classifiers and validated using a leave-one-out cross validation (LOOCV). (3) A stepwise EEG annotation procedure was used to improve the accuracy of annotation due to the presence of numerous seizure imitators and unclear contrast between ictal and interictal EEG activities. Results showed that LDA outperformed SVM with a clear margin of sensitivity, and achieved overall sensitivities 63.1–81.3%, a median FD/h of 1.0 and median latency of 11.5 s. Finally, we conclude that EEG signals of the ID population form a heterogeneous entity with respect to important factors: EEG discharge patterns, EEG backgrounds and EEG seizure visibility. The performance of the seizure detection varies significantly with these factors. The results presented here can serve as prior knowledge for designing a generic seizure detector for the ID patients and the non-convulsive seizure states (NCSS).