Clinical and EEG features integrated by a deep learning method to investigate treatment outcomes in epilepsy with eyelid myoclonia.

OBJECTIVE: To explore the correlation between clinical and electroencephalogram (EEG) factors and therapeutic outcome in epilepsy with eyelid myoclonia (EEM). METHODS: An exploratory neural network model for EEM treatment response prediction (EEMTRP) was introduced to be observed and validated in a replication cohort. Generalized paroxysmal fast activity (GPFA) was studied and analyzed in EEG recorded under different stimulation conditions. RESULTS: The preliminary EEMTRP neural network prediction model integrating the presence of GPFA, photosensitivity, fixation-off sensitivity, EEM status epilepticus, early-onset EEM, and a history of generalized tonic-clonic seizure (GTCS) achieved a prediction accuracy of 75% in the discovery cohort and was validated in the replication cohort with an initial accuracy of 71.4%. Data from two hospitals were tested with an area under curve (AUC) of 0.75. Twenty-three patients developed GPFA over the course of the disease, 22 of whom developed GPFA with eyes closed and six of whom developed GPFA induced by intermittent photic stimulation (IPS). Drug-responsive EEM was seen in two of all patients. INTERPRETATION: This study established a preliminary EEMTRP model that integrates clinical and EEG features to investigate treatment response in EEM. GPFA is a preliminary, multivariable-dependent indicator of refractory EEM. However, due to the lack of statistical significance in univariate comparisons between treatment groups, this association is exploratory and require further validation.