Deep Convolutional Neural Networks Compared to Bandpass Spectral Power for Seizure Detection in Subcutaneous EEG
Abstract number :
3.2
Submission category :
2. Translational Research / 2B. Devices, Technologies, Stem Cells
Year :
2024
Submission ID :
82
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Jordan Clark, BS – Department of Neurology, Mayo Clinic, Rochester MN USA
Pedro Viana, MD, PhD – Institute of Psychiatry, Psychology & Neuroscience, King’s College London, UK
Vlad Sladky, BS – Department of Neurology, Mayo Clinic, Rochester MN USA
Vaclav Kremen, PhD, MS, EMBA – Department of Neurology, Mayo Clinic, Rochester MN USA
Jonas Duun-Henriksen, PhD – UNEEG medical A/S, Borupvang 2, Denmark 3450 Allerod
Troels Kjaer, MD, PhD – UNEEG medical A/S, Borupvang 2, Denmark 3450 Allerod
Andreas Schulze-Bonhage, MD – University Hospital Freiburg
Gregory Worrell, MD, PhD – Mayo Clinic
Mark Richardson, MD, PhD – Massachusetts General Hospital
Benjamin Brinkmann, PhD – Department of Neurology, Mayo Clinic, Rochester MN USA
Rationale: Many people with epilepsy have difficulty in maintaining accurate records of their seizure events, even with the help of a dedicated caregiver. Subcutaneous EEG recording devices represent a possible means of objectively recording seizure activity but generate enormous amounts of data to review. The purpose of this study was to create a seizure detection algorithm for ultra long-term subcutaneous (subQ) EEG recordings in patients with temporal and frontal lobe epilepsy. Both convolutional neural network and conventional spectral bandpass power approaches were used to compare performance. Given the novelty of the subQ data training data is limited in availability, and several approaches to augmenting this data were investigated.
Methods: Twenty-one subjects were used in this study with 16 being used for training data and 5 held out for testing. There were 418 seizures used for training and 72 seizures used for testing with a total 9.03 years of data. Five-minute normalized spectrograms were created containing subQ seizures, artifacts, and interictal background data, as well as spectrograms containing scalp EEG recorded seizures and interictal periods. These spectrograms were used to train and test neural network architectures trained with intracranial EEG data, subQ EEG data, and subQ EEG data augmented with scalp EEG data to identify data segments containing seizures. The model was then compared to a weighted spectral power in band detector for reference.
Results: The intracranial model when tested on the held-out subjects produced an AUROC of 0.955 and AUPRC of 0.288 with a positive sample fraction of 0.00020. The spectral power in band model produced an AUROC of 0.913 and AUPRC of 0.334 with a positive sample fraction of 0.00036 when tested on the held-out subjects. The 2-class CNN-biLSTM model trained with additional artifacts and scalp EEG data produced an AUROC of 0.959 and AUPRC of 0.485 with a positive sample fraction of 0.00021 when tested on the held-out subjects.
Conclusions: Reliable seizure detectors are essential in ultra-long-term EEG monitoring applications with subQ EEG. Despite limited availability of training data, a CNN-biLSTM spectrogram classifier performed slightly better than a conventional spectral bandpass filter model on a held out set of patients.
Funding: This project was supported by NIH NS123066 and the Epilepsy Foundation of America’s My Seizure Gauge grant.
Translational Research