Prediction of Long-term Prognosis in Infantile Epileptic Spasms Syndrome of Unknown Etiology based on Hypsarrhythmia by Self-Attention Autoencoder.

Abstract

Infantile epileptic spasm syndrome (IESS) is developmental epileptic encephalopathy occurring in childhood. It is often difficult to predict long-term seizure outcomes at the time of onset, particularly in patients with unknown etiology. Predicting long-term seizure outcomes from electroencephalograms (EEG) at onset may be helpful for decision-making on therapeutic strategy by clinicians and families. This study included twenty-two patients with IESS of unknown etiology with a median follow-up period of 11 years (range 5-14 years). We categorized patients diagnosed with IESS of unknown etiology into good and poor outcome groups according to the presence or absence of seizures at the time of the last follow-up. Fifteen patients were categorized into the good outcome group, and the rest into the poor outcome group. A machine learning (ML) model was developed using scale EEG data during sleep to identify patients in the poor outcome group. A self-attention autoencoder (SA-AE), which is an anomaly detection method for time-series data, was adopted because the numbers of patients in the poor and good outcome groups were unbalanced. The ML model achieved a sensitivity of $1.00\pm 0.05$, specificity of $0.88\pm 0.11$, and accuracy of $0.95\pm 0.06$. In addition, the AUROC was $0.91\pm 0.09$. Thus, the SA-AE model under development enabled highly accurate prediction of long-term prognosis in patients with IESS of unknown etiology. We confirmed that trained attention among the EEG channels could be explained from the viewpoint of pathophysiology of IESS. This study contributes to the determination of appropriate treatment for patients with IESS of unknown etiology.