Validation of the predictive model for seizure recurrence after withdrawal of antiepileptic drugs.

Purpose: The purpose of this study was to validate the practicability of Lamberink's prediction model in risk assessment of antiepileptic drug (AED) withdrawal in a real, seizure-free population and to find a practical cutoff value to guide clinical withdrawal.

Methods: A group of seizure-free patients from West China Hospital was recruited. Each patient had been seizure-free for at least two years. The seizure recurrence risk among the patients was calculated by an online AED withdrawal risk calculator. The predictive ability of Lamberink's model was assessed by analyzing discrimination and calibration with receiver operating characteristic (ROC) curves and calibration plots, respectively.

Results: A total of 184 seizure-free patients received risk evaluation, all of whom were followed up for at least two years or had an earlier report of seizure relapse. Of these patients, 128 patients were followed up for at least five years or had an earlier report of relapse within five years. Sixty-two of 184 (33.7%) patients relapsed within two years, while 81 of 184 (44.0%) patients relapsed within five years after the start of AEDs' withdrawal. Cox regression analyses showed that seizure duration before remission and the age of seizure onset were independent predictors of relapse at two years. For predictors of recurrence at five years, seizure duration before remission, age at onset, and withdrawal were significant. For discrimination, ROC curve analysis showed that the area under the curve (AUC) for the seizure recurrence within two and five years was 0.605 (95% confidence interval [CI]: 0.518-0.692, p = 0.02) and 0.656 (95% CI: 0.563-0.749, p = 0.003), respectively. For calibration, it was poor in two-year prediction; the observed number was considerably lower than the predicted number. However, the calibration plot showed good calibration with the five-year prediction except for the second, fourth, and eighth deciles. With a cutoff two-year recurrence risk of 47%, the model had a sensitivity of 0.758 and a specificity of 0.410; the largest Youden index was 1.168. With a cutoff five-year recurrence risk of 77%, the model had a sensitivity of 0.358 and a specificity of 0.979; the largest Youden index was 1.337.

Conclusions: Lamberink's prediction model has a general discrimination ability. The model overestimated the actual recurrence events when predicting the two-year recurrence risk, but it showed relatively good calibration with five-year prediction. The cutoff value found in this study may be used to guide patients and clinicians towards a decision regarding the withdrawal of AEDs. The model appears to be a useful tool for predicting seizure recurrence for the five-year recurrence risk.