Support vector machines predict postoperative memory outcomes in temporal lobe epilepsies.

OBJECTIVE

We aimed to predict the side of epilepsy as well as the pre- and postoperative verbal and nonverbal memory performance in a cohort of left and right temporal lobe epilepsy (TLE) patients based on hippocampal activations during three different memory fMRI tasks (verbal, nonverbal and combined verbal and nonverbal) using support vector machines (SVM).

METHODS

Thirty-five patients with unilateral TLE (20 left) were investigated before mesial temporal resection using a verbal, a nonverbal, and a face-name (combined verbal and nonverbal) memory fMRI paradigm. SVM was used to test whether voxel-by-voxel activation patterns within the bilateral hippocampus during each of the three paradigms can be used to classify TLE patients according to their side of epilepsy, preoperative and postoperative verbal and nonverbal memory outcome.

RESULTS

Classification accuracy regarding the side of epilepsy was best for the face-name paradigm closely followed by the verbal paradigm. Classification accuracy of the preoperative verbal memory performance was formally statistically significant for all three paradigms, but specificities were low. Regarding the preoperative nonverbal memory performance, activations during the nonverbal and the verbal paradigm provided high prediction accuracies. The results regarding the postoperative memory outcome revealed that activations during the verbal paradigm can be used to predict postoperative verbal memory outcome, whereas activations during the nonverbal paradigm can be used for the prediction of the nonverbal memory outcome. Preoperative activations during the face-name paradigm were able to predict both the verbal and the nonverbal postoperative memory outcome.

SIGNIFICANCE: It is possible to classify TLE patients according to their side of epilepsy as well as their postoperative memory performance using SVM based on hippocampal activations during task-based memory fMRI. The highest classification accuracies were obtained for the face-name paradigm, suggesting this combined verbal and nonverbal paradigm to be most suitable to address these clinical questions. However, further validation in a larger cohort would be necessary.

PLAIN LANGUAGE SUMMARY: This study aims to investigate whether it is possible to predict the side of epilepsy as well as the preoperative and postoperative verbal and nonverbal memory performance in left and right temporal lobe epilepsy patients using support vector machines (a machine learning technique) based on hippocampal activations during three different memory fMRI tasks. Results showed that classification of patients according to their side of epilepsy and their postoperative memory performance is possible, with the highest classification accuracies being obtained using a face-name association task.