Multi-heads Self-attention AutoEncoder Using Delta with Beta Bands of EEG Enables Highly Accurate Prediction of Seizure Outcome in West Syndrome of Unknown Etiology
Abstract number :
3.436
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2023
Submission ID :
1421
Source :
www.aesnet.org
Presentation date :
12/4/2023 12:00:00 AM
Published date :
Authors :
Presenting Author: Ryosuke Suzui, MD – Department of Pediatrics, Nagoya University Graduate School of Medicine, Nagoya, Japan
Jun Natsume, Professor – Developmental Disability Medicine – Nagoya University Graduate School of Medicine, Nagoya, Japan; Misae Yamada, MD – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; tsu Yoshimura, MD – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Takamasa Mitsumatsu, MD – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Hajime Narita, MD – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Sumire Kumai, MD – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Fumi Sawamura, MD – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Yuji Ito, Dr – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Hiroyuki Yamamoto, Assistant professor – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Tomohiko Nakata, assistant professor – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Hiroyuki Kidokoro, associate professor/lecturer – Pediatrics – Nagoya University Graduate School of Medicine, Nagoya, Japan; Itsuki Saito, Master of Engineering – Human and Process Systems – Nagoya University Graduate School of Engineering, Nagoya, Japan; Koichi Fujiwara, Associate Professor – Human and Process Systems – Nagoya University Graduate School of Engineering, Nagoya, Japan
Rationale: West syndrome is a developmental epileptic encephalopathy in infants. Predicting long-term seizure outcomes based on EEG findings at the onset of West syndrome is challenging, particularly in patients with West syndrome of unknown etiology. Recent studies on EEG analysis using machine learning have identified EEG features undetectable by visual inspection. This study aims to predict long-term seizure outcomes in West syndrome of unknown etiology by analyzing EEG obtained at the onset with machine learning.
Methods: This study included eighteen patients diagnosed with West syndrome of unknown etiology who were admitted to our hospital. The median age at onsets was five months, ranging from two to eight months. The median follow-up period was eleven years, with a range between five to fourteen years. Thirteen patients whose seizures ceased after initial treatments were categorized as the “good outcome” group, while the remaining five patients who experienced persistent or recurrent seizures were categorized as the “poor outcome” group. Anomaly detection is a machine learning task and aims to detect abnormal samples from data using only normal data for training, which is effective when abnormal sample number is small. For operating anomaly detection, the data from the good outcome group was treated as "normal data." Patients in the good outcome group was then randomly divided into training, validation, and test datasets in a 5:3:5 ratio. Burst portions of hypsarrhythmia were identified using an algorithm designed for the present study. Phase information of each electrode during each burst was used as an input feature for the machine learning model. The delta and beta bands, constructing basis of hypsarrhythmia, in the EEG during sleep, were extracted through bandpass filtering. The instantaneous phase of each band was computed using the Hilbert transformation. We employed a Self-attention AutoEncoder model, designed for anomaly detection by using "attention." It was trained using the training dataset, and its hyperparameters were determined using the validation dataset. The average of mean squared error in each patient was used as an index of anomaly detection. This trial was repeated ten times, randomly exchanging datasets. The average of weights of the attention layer in ten trials was calculated and visualized.
Results: No patient was detected genetic cases by additional genetic test or treated with anything other than ACTH or antiseizure medications. The dose and duration of ACTH did not differ between two groups. In identifying the poor outcome group, the model achieved a sensitivity, specificity, and accuracy of 1.00±0.00, 0.90±0.09, and 0.95±0.05, respectively. The area under the ROC curve was 0.96±0.05 (Fig1). In the visualized weights, the attention from the beta phase to the delta phase in the same or neighboring channels tend to be higher than attention directed by other channels (Fig2)
Conclusions: The Multi-heads Self-attention AutoEncoder, focusing on the delta and beta bands of the hypsarrhythmia at the onset, enabled highly accurate prediction of seizure outcomes. Relationships among delta waves and beta waves in neighboring channels differ in seizure outcome.
Funding: None
Neurophysiology