Application of artificial intelligence (AI) and machine learning (ML) in pediatric epilepsy: a narrative review

Objective: The purpose of this narrative review is to introduce artificial intelligence (AI) and machine learning (ML) to pediatricians in the field of epilepsy. Background: There has been significant interest in AI and ML in the field of medicine. The number of AI research in the field of pediatrics is also increasing rapidly. AI research team often asks pediatricians to review and label the data for AI research and provide insights for planning the AI/ML algorithms. Ever-increasing medical data such as medical imaging data and digitalized physiologic monitoring data and advanced computing power enabled AI and ML research to increase rapidly. The chronic nature of epilepsy care is another reason AI/ML research is increasing using digitized big data such as magnetic resonance imaging (MRI) and electroencephalography (EEG). Methods: This review provides examples of AI/ML research in epilepsy, focusing on clinical implications. The purpose of AI/ML research in epilepsy encompasses increasing diagnostic accuracy and precision, detecting and predicting seizures, supporting treatment decisions, improving treatment outcomes, and predicting seizure and non-seizure outcomes. We will review various AI/ML research on automated EEG interpretation, seizure detection and forecasting. Conclusions: Understanding the strength and limitations of AI/ML research will help pediatricians understand and contribute to AI/ML research of their field of expertise. We must find useful clinical implications and suggestions that affect our medical knowledge and change our clinical practice from the research as clinicians participate in AI/ML research. rate is an essential measure of performance. In case of high false positive alarms, the patient will have to anticipate seizure unnecessarily, and lowering this false alarm rate is relevant. Many studies in seizure prediction and the sensitivity and specificity have increased remarkably for few decades (60-63). We developed deep convolutional neural network-based interictal/preictal EEG prediction and applied to 9 pediatric patients with surgically proven focal cortical dysplasia type II. The best accuracy was 5 minutes as preictal period, all intracranial channels for analysis, and 512 Hz sampling rate for EEG acquisition. When we change the preictal period from 2 hours before seizure to 1 minute, accuracy increased to 5 minutes and showed a small decline in 1 minute. These findings tell us that the best functioning was when we set the preictal period to 5 minutes before the seizure started. When we changed the analyzing electrode from whole intracranial to most relevant four electrodes, accuracy declined, but the amount was 2%. Increasing the sampling rate from 128 to 512 Hz, the increase of accuracy was trivial. We found that we could reduce the number of electrodes and sampling rate with a slight decline in performance. The decreasing number of electrodes helps reduce surgical risk, and sampling rate reduction is related to computational efficiency (64).