Development and internal validation of a prognostic model for predicting tonic-clonic seizures during pregnancy in...

We aim to develop a model to predict the probability of tonic-clonic seizures in women with epilepsy (WWE) at any point during pregnancy until six weeks postpartum.

We conducted a screening of patients diagnosed with epilepsy and who were pregnant, at a tertiary hospital in China, during the period of 1 January 2010 to 31 December 2020. We then followed up with these patients for at least one year postpartum. A total of 271 eligible patients were included in the cohort. The outcome was the occurrence of a tonic-clonic seizure during pregnancy or within six weeks postpartum. Predictors were screened through univariate analysis, and models were fitted through multivariate logistic regression analysis. Further, we compared the WMU model with the AntiEpileptic drug Monitoring in PREgnancy (EMPiRE) model in terms of discrimination (the area under receiver operating characteristic curve [AUC]), accuracy (GiViTI calibration belt), decision curve analysis (DCA), net reclassification index (NRI), and integrated discrimination improvement (IDI). Finally, we plotted a nomogram of the WMU model.

Of the 271 pregnant WWE, 62 patients (22.9%) had the outcome. The WMU model included three predictors: age at the time of pregnancy, admission to hospital for seizures in previous pregnancy, and seizures in the 12 months before pregnancy. Compared to the EMPiRE model, the AUC value of the WMU model was higher (0.76 vs. 0.639, P < 0.05). GiViTI calibration belt showed that the predicted risks of the WMU model were mostly consistent with the observed risks. In terms of DCA, the WMU model revealed the highest net proportional benefit for predicted probability thresholds between 10% and 90%. Additionally, our model exhibited better reclassification performance than the EMPiRE model (NRI: 0.331, P < 0.01 and IDI: 0.129, P < 0.01).

We attempted to develop a prognostic model for predicting the risk of tonic-clonic seizures in pregnant WWE. The WMU model showed good performance, but without external validation, it is unclear whether WMU model could be generalized.