Abstract:

Objective   To construct a predictive model of epileptic seizures based on unbalanced data. Methods   The study included 736 epileptic patients treated in Linyi People′s Hospital from September 2008 to January 2011. Epidemiological investigation on risks factors for seizures was made. As the frequency of seizures in epileptic patients was non-balanced data, the data were made a balance based on the Smote Algorithm. Then the random forest was applied to construct a model to make discriminant prediction on the frequency of seizures. Results   Using the random forest to analyze the data, the correctly classified accuracy was 82.53%, incorrectly classified accuracy was 17.47%, area under the receiver operating characteristic(ROC) curve was 94.2%, and out of bag error(OOB) was 13.3%. Conclusion   The random forest is capable of rapidly discriminating the frequency of seizures after processing unbalanced data， which can provide a scientific basis for the forecast of seizures.

Key words: Unbalanced data； Smote Algorithm； Random forest; Seizures; Cross-validation