Investigating the Need for Pediatric-Specific Machine Learning Approaches for Seizure Detection in EEG

Approximately 1 in every 150 children is diagnosed with epilepsy during the first ten years of life. These children ex- perience seizures, which disrupt their lives and directly harm the developing brain. EEG is a key tool for the non-invasive recording of brain activity and the diagnosis of epilepsy. However, the interpretation of EEGs requires time-consuming expert analysis. Automated seizure detection can help to reduce the time required to annotate EEGs. Research on seizure detection methods mainly focuses on adult EEG; automated seizure detection methods in paediatric EEG has been limited. Research has shown that brain events in EEG change with ageing. Therefore, adult-based seizure detection methods maybe not be suitable for children. In this study, we present a random forest-based seizure detection method developed using TUH adult EEG. 4,449 adult EEG recordings were used to train the method, and 490 adult EEG recordings were used to validate the method. An additional 509 TUH adult EEG and 192 TUH pediatric EEG were used for independent testing of the method. The CHB-MIT pediatric EEG Database (N=668) was used as an external independent test set. Ten channels were selected, and twenty-two features were estimated from each channel to develop the method. The random forest-based method achieved 69.3% balanced accuracy on the independent test set of TUH adult EEG and 70.9% on the independent test set of TUH pediatric EEG. However, balanced accuracy on the paediatric CHB-MIT independent test set was only 50.8%. Additionally, specificity was very low on both the TUH pediatric and CHB-MIT independent test sets (49.8% and 10.3% respectively). These result shows that the adult-based seizure detection method is unsuitable for children. There is a need to develop seizure detection methods specifically for paediatric EEG.