Predicting the Outcome of Surgery in Patients with Medically Refractory Temporal Lobe Epilepsy ?Artificial Neural Networks Model

BACKGROUND AND AIMS To use an artificial neural networks (ANN) model based entirely on presurgical clinical and investigation variables for predicting postoperative surgical outcome for patients who underwent surgery for medically refractory temporal lobe epilepsy (TLE), and at the same time to compare with binary logistic regression model (BLR) using the Engel outcome. METHODS The subjects included were 115 patients with temporal lobe epilepsy who underwent surgery and had at least 1 year post surgery follow up. Initially 17 presurgical variables were coded on binary scale and depending on p value (less than0.05) with logistic regression forward selection 3 predictors were selected namely imaging (MRI), partial seizures with secondary generalization and seizure frequency for developing the models. Outcome was assessed using ANN and BLR according to Engel outcome classifications on binary scale. RESULTS The 115 datasets of the patients were used for classification by BLR and ANN methods for predicting the Engel outcome. BLR model sensitivity 80 %, specificity 85 % and that of ANN Sensitivity 80 % specificity 85 %, however the ROC area under curve for BLR is 0.703 and ANN is 0.732. The ANN model the ROC area under curve is higher compared to BLR model. CONCLUSIONS Using artificial neural networks, prediction models were developed to predict the outcome of surgery in patients with refractory temporal lobe epilepsy by using simple pre-operative clinical and investigation parameters. The ANN classifier performed better than BLR classifier.