Predicting seizures in pregnant women with epilepsy: Development and external validation of a prognostic model

Abstract

Background: Seizures are the main cause of maternal death in women with epilepsy; but there are no tools for predicting seizures in pregnancy. We set out to develop and validate a prognostic model, using information collected during the antenatal booking visit, to predict seizure risk at any time in pregnancy and until six weeks postpartum in women with epilepsy on anti-epileptic drugs.

Methods and Findings: We used datasets of a prospective cohort study (EMPiRE) of 527 pregnant women with epilepsy on medication from 50 hospitals in the UK (Nov 2011 - Aug 2014). Model development dataset included 399 women whose anti-epileptic drug doses were adjusted based on clinical features only; validation dataset included 128 women whose drug dose adjustments were informed by serum drug levels. The outcome was epileptic (non-eclamptic) seizure captured using diary records. We fitted the model using LASSO (Least Absolute Shrinkage and Selection Operator) regression, and reported the performance using C-statistic (scale 0-1, values > 0.5 show discrimination) and calibration slope (scale 0-1, values near 1 show accuracy) with 95% confidence interval (CI). We determined the net benefit (a weighted sum of true positive and false positive classifications) of using the model for various probability thresholds to aid clinicians make individualised decisions such as referral to tertiary care, frequency and intensity of monitoring, and changes in anti-epileptic medication. Seizures occurred in 183 women (46%, 183/399) in the model development dataset, and in 57 women (45%, 57/128) in the validation dataset. The model included age at first seizure, type of seizure, history of mental health disorder or learning difficulties, occurrence of any seizure three months pre-pregnancy, previous admission to the hospital for seizures, and dose of lamotrigine and levetiracetam. The C-statistic was 0.79 (95% CI 0.75, 0.84). On external validation, the model showed good performance (C-statistic 0.76, 95% CI 0.66, 0.85; calibration slope 0.93, 95% CI 0.44, 1.41) but with imprecise estimates. The EMPiRE model showed the highest net proportional benefit for predicted probability thresholds between 12% and 99%. Limitations of this study include the varied gestational ages of women at recruitment, retrospective patient recall of seizure history, potential variations in seizure classification, small number of events in the validation dataset, and the clinical use restricted to decision-making thresholds above 12%. The model findings may not be generalisable to low and middle-income countries, or when information on all predictors is not available.

Conclusions: The EMPiRE model showed good performance in predicting the risk of seizures in pregnant women with epilepsy who are prescribed antiepileptic drugs. Integration of the tool within antenatal booking visit, deployed as a simple nomogram, can help to optimise care in women with epilepsy.