Abstracts

Rationale: A third of patients do not become seizure free after temporal lobe surgery for hippocampal sclerosis. Here, we used random forest (RF) classifiers for predicting epilepsy surgery outcome in individuals, based on preoperative PET using FDG and [11C]flumazenil (FMZ), and compared its performance against previous method based on assessment of periventricular white matter (WM) signal increases (Hammers et al., 2005; Yankam Njiwa et al., 2012).Methods: Sixteen patients with MRI-defined HS who had preoperative FDG and [11C]FMZ PET and at least 23 (median 68) months follow-up were retrospectively identified. They were compared with controls (30 FDG, 41 FMZ). Feature data from voxel-wise data acquired with each tracer were independently used as inputs to a RF classifier (Breiman, 2001) for comparison. Similarities estimated by the RF classifiers were used to classify different groups based on all available features. Performance was compared with the previous method based on detection of periventricular increases using Statistical Parametric Mapping (SPM8) and periventricular masks. For detecting patient who would not become seizure free after temporal lobe surgery, this previous method had obtained up to 63% sensitivity / 88% specificity for [11C]FMZ and 50% sensitivity / 63% specificity for FDG. Results: Eight of the 16 patients were not Engel IA during follow-up. Classification performance of the RF classifiers was assessed between several relevant pairs of diagnostic groups (NSF_TL/C, NSF_TR/C, SF_TL/C, SF_TR/C; see Table 1 for abbreviations). Both tracers were able to distinguish patients from controls above chance level, except for FDG and seizure free patients. FMZ performed better than FDG in all pair comparisons except for the NSF_TR/C comparison (see Table1). The RF classifiers correctly identified 6/8 individual NSF patients using FDG, but only 5/8 NSF for FMZ. Two SF patients were misclassified for both FMZ and FDG. Therefore, the performed test led to 63% sensitivity/75% specificity for FMZ and 75% sensitivity/75% specificity for FDG. The most important voxels used in RF classifiers were in the periventricular regions (see Fig. 1). Further, Well-predicted patients with the method described in (Yankam Njiwa et al., 2012) were added to the well-classified patients with RF classifiers. Combining the RF classifier prediction with the prediction of the previously used method based on periventricular masks (Yankam Njiwa et al., 2012) helped increase the prediction performance, with a sensitivity of up to 75% for FMZ and a specificity of up to 100% for both FDG and FMZ. Conclusions: This study confirms the association between periventricular WM increases of [11C]FMZ and [18F]FDG binding with NSF after temporal resection for HS. Periventricular concentration of [11C]FMZ- but less so [18F]FDG-may usefully predict individual postoperative outcome. Combining the RF classification with periventricular masking further improved performance.