Clinical Validation of a Neural Network Trained for Interictal Epileptiform Discharge Detection
Abstract number :
1.212
Submission category :
4. Clinical Epilepsy / 4B. Clinical Diagnosis
Year :
2021
Submission ID :
1826172
Source :
www.aesnet.org
Presentation date :
12/4/2021 12:00:00 PM
Published date :
Nov 22, 2021, 06:52 AM
Authors :
Catarina da Silva Lourenco, MSc - University of Twente; Marleen Tjepkema-Cloostermans - University of Twente; Michel van Putten - University of Twente
Rationale: Deep learning methods have shown potential in automating interictal epileptiform discharge (IED) detection in electroencephalograms (EEGs). To implement such a method in a clinical setting, it needs to reach performance levels comparable to visual analysis. We aim to perform this comparison in order to validate a neural network developed for IED detection and subsequent identification of patients with epilepsy.
Methods: We processed 56 ambulatory EEGs (14 from patients with focal epilepsy, 7 from patients with generalized epilepsy and 35 normal EEGs, average duration = 20h) with a previously developed deep learning algorithm, trained to detect IEDs [1]. The EEG epochs identified as IEDs by the neural network were flagged and reviewed by two experts. We timed this process and used the flagged epochs to draw a conclusion about the entire recording (i.e. whether it included IEDs or not). This conclusion was then compared to the original report associated with the recording.
Results: Based on the epochs flagged by the algorithm, it was possible to correctly identify 12 EEGs which included epileptiform discharges and 34 which did not have IEDs. This corresponded to a sensitivity of 57% and specificity of 97%. The average review time per ambulatory EEG was 1 minute and 40 seconds. This value could be further reduced by implementing more effective strategies for a posteriori artefact removal, as this would largely decrease the number of incorrectly flagged epochs. Aside from artefacts, k complexes were often flagged by the algorithm. Conversely, the neural network missed IEDs in 9 EEG recordings. This should be corrected through active learning to further improve performance. Following artefact removal, lowering the probability threshold used for detection can also aid in increasing sensitivity.
Conclusions: The use of a neural network for IED detection greatly reduces the review time of ambulatory EEGs. However, sensitivity should be improved for this algorithm to be useful in a clinical setting.
Funding: Please list any funding that was received in support of this abstract.: Epilepsiefonds, grant number WAR16-08.
Clinical Epilepsy