~/yuki-kashiwada

$ cat research/eeg-dsa-seizure-differentiation.md

Differentiation between epileptic and functional/dissociative seizures using density spectral array of ictal single-channel EEG with deep learning

ArticleSecond author (co-author)License: CC BY-NC-ND 4.0PMID 41016125

Authors

  1. Kazutoshi Konomatsu,
  2. Yuki Kashiwada,
  3. Takafumi Kubota,
  4. Kazutaka Jin,
  5. Ryu Koda,
  6. Kento Takahashi,
  7. Temma Soga,
  8. Makoto Ishida,
  9. Naoto Kuroda,
  10. Kazushi Ukishiro,
  11. Yosuke Kakisaka,
  12. Masashi Aoki,
  13. Nobukazu Nakasato

Epilepsy & Behavior, 172:110713 · https://doi.org/10.1016/j.yebeh.2025.110713

  • Received:
  • Revised:
  • Accepted:
  • Published online:
  • Issue:

TL;DR

Ictal EEG was turned into time-frequency images (density spectral arrays, DSA) and a ResNet34 model was trained to tell mesial temporal lobe epilepsy (mTLE) seizures from functional/dissociative seizures (FDS). Among reduced-electrode settings, the single Cz channel gave the highest AUC, 0.941.

##Key numbers

AUC (single channel Cz)
0.941
Middle 1/3 interval
Patients
99
48 mTLE / 51 FDS
Seizures analysed
206
Test set
19 / 33
patients / seizures
Model
ResNet34
EEG window
±3 min
around clinical onset

##Background

Diagnosing epilepsy starts with telling epileptic seizures apart from non-epileptic ones. This study asked whether deep learning on the DSA of ictal EEG can make that distinction.

##Data

Seizures recorded during long-term video-EEG monitoring were analysed retrospectively in consecutive patients with mTLE or FDS: 48 with mTLE and 51 with FDS, 99 in total.

The model was built on 40 mTLE patients (91 seizures) and 40 FDS patients (82 seizures) and tested on 8 mTLE patients (15 seizures) and 11 FDS patients (18 seizures), 206 seizures in total.

##Methods

With clinical onset as time zero, EEG from 3 minutes before to 3 minutes after onset was clipped. Frequency analysis with a linked-ear reference produced DSAs for Cz, the mean of C3/C4, the mean of Fp1/Fp2, the mean of O1/O2, and all electrodes.

A ResNet34 convolutional neural network was trained and tested on these DSA images.

##Results

In exploratory analyses, the Cz electrode combined with the "Middle 1/3" time interval gave the highest AUC among reduced-electrode settings (0.941). Pre-specified DeLong tests with Bonferroni correction statistically confirmed this result.

##Significance

The study showed that deep learning on a single-channel (Cz) DSA can differentiate epileptic from non-epileptic seizures. The authors position it as a practical adjunct for early screening and triage, especially in resource-limited settings.

##Limitations

  • -Retrospective design with a small test set (19 patients, 33 seizures).
  • -The epileptic group includes only mesial temporal lobe epilepsy.
  • -The electrode and interval comparisons were exploratory analyses.
  • -This page summarizes the published abstract; see the full text for detailed metrics.

##FAQ

Q.Can deep learning on EEG distinguish epileptic seizures from psychogenic non-epileptic seizures (PNES)?

In this study, a ResNet34 model applied to the DSA of ictal EEG differentiated mesial temporal lobe epilepsy seizures from functional/dissociative seizures (FDS, often previously called PNES), reaching an AUC of 0.941 with the single Cz channel.

Q.What is a density spectral array (DSA)?

A DSA is a time-frequency display of EEG in which the power of each frequency band over time is shown as colour. In this study, DSAs covering 3 minutes before and after seizure onset were fed to the deep learning model as images.

Q.How accurate was the model?

Among reduced-electrode settings, the best was the Cz electrode with the Middle 1/3 interval, at an AUC of 0.941. The test set had 8 mTLE patients (15 seizures) and 11 FDS patients (18 seizures).

Q.Why does a single EEG channel matter?

If one channel is enough, recording and analysis become simpler. The authors highlight its potential as an adjunct for early screening and triage, particularly in resource-limited settings.

Q.What was Yuki Kashiwada's role in this paper?

Yuki Kashiwada is the second of 13 authors and the only author from the Department of AI and Innovative Medicine, Tohoku University Graduate School of Medicine. The first author is Kazutoshi Konomatsu and the corresponding author is Kazutaka Jin.

Q.How do I cite this paper?

Cite it as Konomatsu K, Kashiwada Y, et al. Epilepsy & Behavior. 2025;172:110713. doi:10.1016/j.yebeh.2025.110713. The article is open access under CC BY-NC-ND 4.0.

##How to cite

APA style

Konomatsu, K., Kashiwada, Y., Kubota, T., Jin, K., Koda, R., Takahashi, K., Soga, T., Ishida, M., Kuroda, N., Ukishiro, K., Kakisaka, Y., Aoki, M., & Nakasato, N. (2025). Differentiation between epileptic and functional/dissociative seizures using density spectral array of ictal single-channel EEG with deep learning. Epilepsy & Behavior, 172, 110713. https://doi.org/10.1016/j.yebeh.2025.110713

BibTeX

@article{konomatsu2025eegdsa,
  author       = {Konomatsu, Kazutoshi and Kashiwada, Yuki and Kubota, Takafumi and Jin, Kazutaka and Koda, Ryu and Takahashi, Kento and Soga, Temma and Ishida, Makoto and Kuroda, Naoto and Ukishiro, Kazushi and Kakisaka, Yosuke and Aoki, Masashi and Nakasato, Nobukazu},
  title        = {Differentiation between epileptic and functional/dissociative seizures using density spectral array of ictal single-channel {EEG} with deep learning},
  journal      = {Epilepsy \& Behavior},
  volume       = {172},
  pages        = {110713},
  issn         = {1525-5050},
  publisher    = {Elsevier Inc.},
  year         = {2025},
  doi          = {10.1016/j.yebeh.2025.110713},
  url          = {https://doi.org/10.1016/j.yebeh.2025.110713}
}

Citation files: konomatsu2025eegdsa.bib (BibTeX) · konomatsu2025eegdsa.ris (RIS / EndNote / Zotero)

Summaries are written by the site owner from the published article or dataset record. Refer to the original for authoritative details.