Seizure recurrence risk after discontinuation of anti-seizure medications: prognostic factors and development of a new predictive scale

Background: The discontinuation of anti-seizure medications (ASMs) in seizure- free patients is an important achievement given their side effects and the social stigma associated with epilepsy. It remains however a complex decision considering all the medical, psychological, and social consequences of a possible seizure relapse. In the first part of our study, we assessed: the frequency of seizure relapses in our population after ASM withdrawal, any associated risk factor, and the accuracy of a model for predicting the risk of relapse after ASM discontinuation already available in the literature (Lamberink et al,2017). In the second part of our study, since in our cohort the Lamberink prediction model has shown inadequate performance, we propose a method for implementing a scale with a clear-cut threshold to accurately predict the risk of seizure recurrence at 2 years after ASM withdrawal. Material and Methods: We retrospectively enrolled patients, followed by the Epilepsy Center of the Careggi University Hospital in Florence, who have discontinued ASMs because they have been seizure-free, usually for at least two years. In the first part of the study we evaluated the onset of seizure relapses, the presence of new effective pharmacological control of epilepsy after relapse (defined as a seizure-free period of at least one year at the maximum follow-up), and the possible predictive role of several clinical-demographic variables. Finally, the predictive accuracy of the Lamberink prediction model at an individual level was assessed at various possible probability cut-offs. In the second part of the study, we used the aforementioned collected data to develop a new scale. The collected clinical and demographic features were incorporated into a customized R algorithm. This algorithm assessed each covariate and its corresponding weight to determine the combination of weighted scores with the highest sensitivity and specificity. Moreover, the algorithm generated a clear-cut threshold that differentiates between low- and high-risk patients. Results: The enrolled patients (n=133) had been followed for a median of 3 years (range 0.8-33 years) after ASM discontinuation and 60 (45%) of them relapsed. On univariate analysis, risk factors for seizure relapse were: the presence of febrile seizures in childhood (HR=3.927;95%IC:1.403-10.988;p=0.009), a seizure-free period on therapy shorter than 2 years (HR=2.313;95% IC:1.193-4.486;p=0.013), and the presence of persistent motor deficits (HR=4.568;95% IC:1.412 14.772;p=0.011). Multivariate analysis revealed that only a seizure-free period on therapy shorter than 2 years was significantly associated with the onset of relapse (HR=2.365;95%IC1.178-4,7444;p=0.015). Among the 60 patients who relapsed, 59 agreed to start ASMs again. Of these, 51 patients were followed for at least one year after relapse (median 5 years; range 1-33). During this period, 42/51 patients obtained a new effective pharmacological control of epilepsy (82.4%); 5/51 did not (10%); 4/51 (7.8%) developed drug-resistant epilepsy. Using univariate analysis, no clinic-demographic features were significantly associated with new pharmacological control of epilepsy. The prediction model of Lamberink et al showed an unsatisfactory accuracy in our population even using different probability thresholds. In the second part of our study, our customized R algorithm progressively selected the combination of the 10 covariates and weights that composed the scale with the maximized Youden index: “duration of epilepsy”, “duration of the seizure-free period on therapy”, “duration of ASM tapering”, “development delay”, “age at ASM withdrawal over 50 years”, “gender”, “age at ASM withdrawal over 40 years”, “structural etiology of epilepsy”, “failure of previous ASM discontinuations”, “age at seizure onset” (Youden index = 0.50; sensitivity = 0.86; specificity = 0.64). The scale provides a cut-off of 5. Patients whose scores exceed this value are deemed to be at high risk of experiencing seizure recurrence within 2 years of ASM withdrawal. Discussion: The percentage of patients who relapsed in our cohort and the risk factors for seizure recurrence that we identified are consistent with those already published. The percentage of our patients who obtained a new effective pharmacological control of epilepsy (82.4%) is also consistent with the literature: 91% of patients in the study of Lamberink et al and between 64 and 91% of patients in the meta-analysis by Schmidt et al. In the study of Lamberink et al, 9% of the cohort became drug-resistant while this value ranged between 7 and 23% in Schimdt's meta-analysis. These data are in agreement with our 7.8% of patients with drug-resistant epilepsy. The fact that in our work no clinic-demographic features were significantly associated with a greater chance of a new pharmacological control can be explained by the small sample size. Regarding the prediction model of Lamberink et al, we verified in our cohort that no probability threshold value yields a sufficient model accuracy: there is either an excessive risk of recurrence if high specificity is chosen but consequently low sensitivity, or an excessive conservative attitude if reverse values are selected. For these results and since prediction models that produce approximate probability values could lead to confusion among both patients and clinicians, in the second part of the study we develop a new scale with a clear cut-off. Some of the variables included in our scale are in agreement with the Lamberink prediction model and with previous literature. Others instead are not considered by Lamberink et al and are also not coherent with previous literature (e.g., “failure of previous ASM discontinuations”). However, it has to be considered that here we are not proposing a definitive scale, but rather a method to implement a scale with a clear cut-off for the prediction of seizure recurrence risk after ASM withdrawal. Conclusions: In the first part of our study we have confirmed the predictive strength only of some of the factors known to be associated with the risk of relapse. In addition, the poor accuracy of the prediction model of Lamberink et al. suggested the development of another scale with a clearer cut-off and more easily applicable in clinical practice. For this reason in the second part of the study, we proposed a data- driven method that allows for the implementation of a scale with a clear-cut threshold. This approach offers the advantage of being easily applicable in clinical practice and understandable to patients. As a data-driven method, future multicenter and international studies are necessary to develop powerful and generalizable scales.