Around one-third of epilepsy patients develop drug-resistant seizures; early detection of seizures could help improve safety, reduce patient anxiety, increase independence, and enable acute treatment. In recent years, the use of artificial intelligence techniques and machine learning algorithms in different diseases, including epilepsy, has increased significantly. The main objective of this study is to determine whether the mjn-SERAS artificial intelligence algorithm developed by MJN Neuroserveis, can detect seizures early using patient-specific data to create a personalized mathematical model based on EEG training, defined as the programmed recognition of oncoming seizures before they are primarily initiated, usually within a period of a few minutes, in patients diagnosed of epilepsy. Retrospective, cross-sectional, observational, multicenter study to determine the sensitivity and specificity of the artificial intelligence algorithm. We searched the database of the Epilepsy Units of three Spanish medical centers and selected 50 patients evaluated between January 2017 and February 2021, diagnosed with refractory focal epilepsy who underwent video-EEG monitoring recordings between 3 and 5 days, a minimum of 3 seizures per patient, lasting more than 5 s and the interval between each seizure was greater than 1 h. Exclusion criteria included age <18 years, intracranial EEG monitoring, and severe psychiatric, neurological, or systemic disorders. The algorithm identified pre-ictal and interictal patterns from EEG data using our learning algorithm and was compared to a senior epileptologist's evaluation as a gold standard. Individual mathematical models of each patient were trained using this feature dataset. A total of 1963 h of 49 video-EEG recordings were reviewed, with an average of 39.26 h per patient. The video-EEG monitoring recorded 309 seizures as subsequently analyzed by the epileptologists. The mjn-SERAS algorithm was trained on 119 seizures and split testing was performed on 188 seizures. The statistical analysis includes the data from each model and reports 10 false negatives (no detection of episodes recorded by video-EEG) and 22 false positives (alert detected without clinical correlation or abnormal EEG signal within 30 min). Specifically, the automated mjn-SERAS AI algorithm achieved a sensitivity of 94.7% (95 %; CI 94.67-94.73), and an F-Score representing specificity of 92.2% (95 %; CI 92.17-92.23) compared to the reference performance represented by a mean (harmonic mean or average) and a positive predictive value of 91%, with a false positive rate of 0.55 per 24 h in the patient-independent model. This patient-specific AI algorithm for early seizure detection shows promising results in terms of sensitivity and false positive rate. Although the algorithm requires high computational requirements on specialized servers cloud for training and computing, its computational load in real-time is low, allowing its implementation on embedded devices for online seizure detection.
Keywords: Artificial intelligence; EEG analysis; Epilepsy; Epileptic seizures; Machine learning; Seizure early detection.
© 2023 The Authors.