Abstracts

Rationale: Surgery can lead to seizure-freedom for patients with drug-resistant temporal lobe epilepsy (TLE). However, a significant number of patients still remain with disabling seizures despite surgery, and there are currently no good clinical prediction tools to prognosticate TLE surgical outcomes. We therefore aimed to derive a probabilistic statistical model which could predict both short and long-term TLE surgical outcomes. Methods: Clinical, neurophysiological, and imaging variables known to have an impact on surgical outcomes were collected from 167 patients with drug-resistant TLE who had undergone surgery at the University of California, San Francisco (UCSF) over a fifteen year period. Adults were included if (i) unilateral temporal lobe seizure onset was demonstrated during scalp and/or intracranial electroencephalography (EEG) monitoring, (ii) pre-surgical 1.5 or 3.0 Tesla MRIs were either non-lesional, or showed unilateral hippocampal atrophy, (iii) at least one year of post-surgical follow-up was available. To increase the homogeneity of the selected study population, patients with other structural imaging abnormalities in the temporal lobes such as cortical dysplasias, neoplasms or cavernomas were excluded. Post-surgery Engel outcomes at year one (Y1), two (Y2) and five (Y5) were analysed as primary endpoints. We built a multidimensional Bayesian network classifier (MBC) on the UCSF data and evaluated its predictive performance by comparing it with an established nomogram¹ for epilepsy surgeries by making predictions on an independent dataset from the Montreal Neurological Institute (MNI) (n=64). Results: MBC modelling clearly discriminated better than the nomogram in all time scales (Table 1). It also achieved the highest AUC (0.84) for the MNI data at Y5, and performed reasonably well at Y1 (0.73) and Y2 (0.66), offering useful predictive insights into long-term post-surgical seizure freedom rates. An online calculator based on our MBC (http://manikkavacakar.dcmohan.com:3838/tlesop) offers readers the opportunity to make individualized TLE surgical outcome predictions, guided by calibration curves (Figure 1). Conclusions: The superior performance of the MBC compared to established nomograms highlights the predictive value and discriminative ability of the subset of selected feature variables, and the capability of the MBC to encode probabilistic relationships between predictors and class variables (Engel outcomes). The evaluation and further validation of this model on prospectively acquired data in different patient populations will help optimize its classification accuracy and facilitate its eventual contribution in pre-surgical patient counselling.Reference1. Jehi L, Yardi R, Chagin K, et al. Development and validation of nomograms to provide individualised predictions of seizure outcomes after epilepsy surgery: a retrospective analysis. Lancet Neurol. 2015;14(3):283-290. doi:10.1016/S1474-4422(14)70325-4. Funding: Not applicable