Authors :
Presenting Author: Adam Dickey, MD, PhD – Baylor College of Medicine
Alica Goldman, MD, PhD – Baylor University
Marina Vannucci, PhD – Rice University
Sameer Sheth, MD, PhD – Baylor College of Medicine
Daniel Drane, PhD – Emory University School of Medicine
Rationale:
We previously published an 8-point ordinal score which predicts seizure freedom after minimally invasive epilepsy surgery for medial temporal lobe epilepsy1. This score equally weighted 8 binary clinical variables, such as presence or absence of mesial temporal sclerosis (MTS). Equal weighting of all variables performed better in cross validation than traditional logistic regression, as it avoided overfitting. However, equal weighting is clinically implausible, as quantitative MRI and EEG results should be more predictive than qualitative clinical history (i.e. history of febrile seizures). We hypothesize that Bayesian logistic regression could assign varying weights without overfitting, and thus still perform well in cross validation. We present a revised point score assigning between 1 and 3 points to the 8 clinical variables.
Methods:
The prior distributions for Bayesian logistic regression were derived from meta-analysis. For four variables (MTS on MRI, concordant ictal EEG, concordant interictal EEG, and history of febrile seizures) we used the Cochrane systematic meta-analysis data2. For the other 4 variables (no tonic-clonic seizures, no atypical aura, age of onset < 16 yo, concordant PET) we performed a convenience meta-analysis of literature on hand. We analyzed a cohort of 108 patients who underwent stereotactic laser amygdalohippocampotomy (SLAH) at Emory University3. To facilitate interpretation, we converted the coefficients (log odds-ratios) of the final Bayesian regression into points, where each point represents a log odds-ratio of 0.33, or an odds ratio of 1.4.
Results:
We performed 100 runs of two-fold cross validation, training on a random sample of 54 subjects and testing on the other 54 subjects. The mean area under the curve (AUC) of the receiver operator characteristic (ROC) for the Bayesian model (0.73) was equivalent to the previously published, equally weighted model (0.73). Both models outperformed the mean AUC of traditional logistic regression (0.65), or a model using only presence or absence of MTS (0.64). The weights for the full Bayesian regression (with all 108 patients) were: 3 points for 1) not having an atypical aura (auditory, vertigo, visual); 2 points each for 2) presence of MTS, 3) concordant ictal EEG, 4) concordant inter-ictal EEG, and 5) absence of tonic-clonic seizures; and 1 point each for 6) history of febrile seizures, 7) age of onset < 16, and 8) concordant PET. 39 of 49 (80%) of patients were seizure free at 1 year with a with a score of 9 or more.