Deep Neural Networks for Seizure Detection: A Study on Training Strategies and Architectural Designs

Abstract

Epilepsy is a neurological disorder affecting over 50 million people globally. There is significant need for efficient automatic seizure detection algorithms for timely intervention and accurate diagnosis. This study examines the potential of deep neural networks for seizure detection from electroencephalography (EEG) signals,integrating architectural designs and training techniques from other domains. Our results suggest a limit to the benefits of increased architectural complexity when training on the TUH EEG seizure corpus without augmentations. Training strategies such as mixup, segment translation and ensembling led to substantial improvements in model performance over baseline, particularly when considered in conjunction with suitable model architectures. These methods enhanced both the generalization performance and the calibration of the models. These findings underscore the importance of a balanced approach in designing seizure detection models by considering network architecture and training methods simultaneously.integrating architectural designs and training techniques from other domains. Our results suggest a limit to the benefits of increased architectural complexity when training on the TUH EEG seizure corpus without augmentations. Training strategies such as mixup, segment translation and ensembling training strategies led to substantial improvements in model performance over baseline, particularly when considered in conjunction with suitable model architectures. These methods enhanced both the generalization performance and the calibration of the models. These findings underscore the importance of a balanced approach in designing seizure detection models by considering network architecture and training methods simultaneously.

Flogaveiki er taugasjúkdómur sem hrjáir um 50 milljónir manna á heimsvísu. Þörf er á sjálfvirkum aðferðum við greiningu á flogaköstum til að bæta greiningu og meðhöndlun á sjúkdómnum. Í þessu verkefni voru djúp tauganet notuð til að greina flogaköst út frá heilaritum (EEG). Áhersla var lögð á að kanna áhrif þjálfunaraðferða og netarkitektúrs sem reynst hafa vel í skyldum verkefnum. Niðurstöður benda til þess að netarkitektúr einn og sér hafi minna að segja um nákvæmni líkana en þjálfunaraðferðir. Þjálfunaraðferðir á borð við mixup, tímahliðrun og safnaðferðir gáfu umtalsverða bætingu á frammistöðu, en bætingin var einnig háð netarkitektúr. Bætingin fólst bæði í aukinni greiningarnákvæmni og kvörðun á líkönum. Niðurstöður verkefnisins sýna mikilvægi þess að skoða netarkitektúr og þjálfunaraðferðir samtímis þegar verið er að hanna greiningarlíkön fyrir flogaveiki.