Density Spectral Array of Single-channel EEG Can Differentiate Idiopathic Generalized Epilepsy and Psychogenic Nonepileptic Seizures with Deep Learning
Abstract number :
2.14
Submission category :
3. Neurophysiology / 3C. Other Clinical EEG
Year :
2024
Submission ID :
322
Source :
www.aesnet.org
Presentation date :
12/8/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Kazutoshi Konomatsu, MD – Tohoku University Graduate School of Medicine
Yuki Kashiwada, MA – Tohoku university graduate school of medicine
Jin Kazutaka, MD, PhD – Tohoku university graduate school of medicine
Takafumi Kubota, MD – Tohoku University Graduate School of Medicine
Kazushi Ukishiro, MD, PhD – Tohoku University Graduate School of Medicine
Yosuke Kakisaka, MD, PhD – Tohoku University Graduate School of Medicine
Masashi Aoki, MD, PhD – Tohoku university graduate school of medicine
Nobukazu Nakasato, MD, PhD – Tohoku University Graduate School of Medicine
Rationale: To determine whether the density spectral array (DSA) of ictal EEG can differentiate between epileptic seizures caused by idiopathic generalized epilepsy (IGE) and nonepileptic events associated with psychogenic nonepileptic seizures (PNES) using the deep learning technique.
Methods: We retrospectively reviewed consecutive patients with IGE and PNES admitted to the epilepsy monitoring unit at Tohoku University Hospital from 2014 to 2022. Seizures/events recorded by long-term video EEG monitoring were analyzed, including generalized tonic-clonic, myoclonic, and typical absence seizures of IGE, and akinetic or motor events of PNES. A maximum of three seizures/events from the first seizure/event were incorporated in each patient. The time of clinical onset was defined as zero, and the EEG recordings were clipped from -3 to +3 minutes. Frequency analysis of Cz as well as mean values of C3 and C4, Fp1 and Fp2, and O1 and O2 was performed to generate the DSA with linked ear reference. ResNet34, a convolutional neural network (CNN) model, was trained and tested on datasets from different inclusion years. Statistical performance was assessed with the area under the curve (AUC), accuracy, sensitivity, and specificity.
Results: The present study included 53 patients with IGE and 51 with documented PNES. The CNN architecture was created from 42 patients with IGE (94 seizures) and 40 patients with PNES (82 seizures) from 2014 to 2020 as train data; and the model was evaluated using 11 patients with IGE (24 seizures) and 11 patients with PNES (18 events) from 2021 to 2022 as test data (Table 1). The EEG recordings from -1 to +1 minute showed the optimum discrimination, after evaluating several different time windows (Figure 1). CNN frequency analyses of Cz, C3-C4, Fp1-Fp2, and O1-O2 achieved AUCs of 79.9, 76.6, 69.0, and 66.7, accuracies of 73.8, 76.2, 59.5, and 61.9, sensitivities of 66.7, 66.7, 66.7, and 50.0, and specificities of 83.3, 88.9, 50.0, and 77.8, respectively (Table 2).
Conclusions: DSA of single-channel EEG (Cz electrode) successfully differentiated between epileptic seizures of IGE and nonepileptic events of PNES using a deep learning neural network. The present result provides the basis for a simple automated technique to differentiate IGE and PNES using emergency or ambulatory EEG.
Funding: Research Grants From Japan Foundation of Institute for Neuropsychiatry.
Neurophysiology