The accuracy of outcome prediction models for childhood-onset epilepsy

Two large prospective cohort studies of childhood epilepsy (Nova Scotia and the Netherlands) each developed a statistical model to predict long-term outcome. We sought to evaluate the accuracy of a prognostic model based on the two studies combined.

Analyses with classification tree models and stepwise logistic regression produced predictive models for the combined dataset and the two separate cohorts. The resulting models were then externally validated on the opposite cohort. Remission was defined as no longer receiving daily medication for any length of time at the end of follow-up.

The combined cohorts yielded 1,055 evaluable patients. At the end of follow-up (>or=5 years in >96%), 622 (59%) were in remission. By using the combined data, the classification tree model and the logistic regression model predicted the outcome correctly in approximately 70%. The classification tree model split the data on epilepsy type and age at first seizure. Predictors in the logistic regression model were: seizure number before treatment, age at first seizure, absence seizures, epilepsy types of symptomatic generalized and symptomatic partial, preexisting neurologic signs, intelligence, and the combination of febrile seizures and cryptogenic partial epilepsy. When the prediction models from each cohort were cross-validated on the opposite cohort, the outcome was predicted slightly less accurately than did the model from the combined data.

Based on currently available clinical and EEG variables, predicting the outcome of childhood epilepsy may be difficult and appears to be incorrect in about one of every three patients.