Abstract
Epilepsy is one of the most common neurological disorders in children, characterized by recurrent and unpredictable seizures. Accurate prediction of the likelihood of seizure recurrence can significantly contribute to optimizing treatment and improving patients' quality of life. This study aimed to develop a deep learning-based prediction model for forecasting the probability of seizure recurrence in children with epilepsy.Methods: In this retrospective cohort study, simulated data from 300 pediatric epilepsy patients were generated and analyzed. Clinical features included age of onset, type of epilepsy syndrome, EEG and MRI findings, treatment adherence, and other risk factors. A deep neural network model with a multilayer architecture was developed to predict the probability of recurrence. The model's performance was evaluated using AUC-ROC, AUC-PR, accuracy, and precision metrics. Results: The developed model could predict the probability of seizure recurrence with significant accuracy (AUC-ROC: 0.91, AUC-PR: 0.87). The most important predictors of recurrence included the type of epilepsy syndrome (especially LGS and West syndromes), abnormal EEG and MRI findings, poor treatment adherence, and the presence of developmental delay. The overall recurrence rate in the study population was 34.2%. Conclusion: The deep learning model demonstrated high predictive accuracy (AUC-ROC: 0.95) in identifying children at risk of seizure recurrence. Clinically, the model can serve as a decision-support tool to stratify high-risk patients—particularly those with Lennox-Gastaut or West syndromes, abnormal neuroimaging, or poor treatment adherence—enabling physicians to intensify monitoring, personalize anti-epileptic therapy, and allocate healthcare resources more effectively. Prospective validation in real-world clinical settings is recommended before widespread implementation.