Predicting seizure recurrence after status epilepticus: a multicenter exploratory machine learning approach.

Purpose: Tools to predict seizure recurrence after status epilepticus (SE) are lacking. In this multicenter cohort study, we explored the ability of different machine learning (ML) models to predict seizure recurrence following a first episode of SE in patients without a prior history of seizures (de novo SE, dnSE).

Methods: Consecutive SE patients aged ≥16 years without a previous history of seizures admitted to Vall d'Hebron University Hospital (Barcelona, Spain) from 2011 to 2021 and Modena Academic Hospital (Baggiovara, Italy) from 2013 to 2022 were reviewed. Different machine learning techniques (k-Nearest Neighbors, Naïve Bayes, Artificial Neural Network, Support Vector Machines, Decision Tree, Random Forest) as well as the classic logistic regression model were built using clinical and neurophysiological variables and applied to develop predictive models of seizure recurrence. Seventy percent of the total sample was randomly selected to train the models; the remaining 30 % was used for validation. The area under the receiver operating characteristic curves (AUROC) with a 95 % confidence interval (95 % CI) was calculated to assess their predictive capability.

Results: A total of 386 patients were included, of which 136 (35.2 %) had seizure recurrence within 2 years after the SE episode. Factors significantly associated with two-year seizure recurrence included progressive symptomatic SE etiology (p < 0.001), non-convulsive SE with coma (p = 0.021), and out-of-hospital SE (p = 0.033). Acute symptomatic SE etiology resulted a protective factor (p < 0.001). Among ML techniques, all were slightly superior to the logistic regression model in predicting two-year seizure recurrence, except for Artificial Neural Network. The Random Forest algorithm (AUROC 0.687, 95 %CI = 0.580 - 0.793) demonstrated the best predictive capability in the validation dataset, significantly outperforming the logistic regression model (AUROC 0.594, 95 %CI = 0.478 - 0.710).

Conclusions: In our study, the Random Forest algorithm showed the best predictive capability for two-year seizure recurrence after a dnSE in adult patients. Further studies and additional data are needed to improve its predictive performance.

Plain Language Summary: This multicenter international retrospective study found that over one-third (35.2 %) of patients who develop status epilepticus without a prior history of epilepsy will have a new seizure within two years after the event. By providing clinical and demographic variables, machine learning algorithms were built and their ability to predict seizure recurrence compared with that of classical statistical model. This is the first study to show that artificial intelligence - specifically, Random Forest algorithm - can outperform traditional statistical methods in predicting seizure recurrence in these patients. However, their performance remains unsatisfactory, requiring further research and additional data to improve predictive accuracy.