Validation of hemispherectomy outcome prediction scale in treatment of medically intractable epilepsy

Abstract

The Hemispherectomy Outcome Prediction Scale (HOPS) was developed to aid both clinicians and patients in determining the chance of success after hemispheric surgery for medically refractory epilepsy. The original study generating HOPS had a multi-institutional, large cohort format yielding near perfect patient stratification. Evidence suggests that methodologies utilized to create such predictive models, including cross-validation as well as stratification utilizing the same data employed for model generation, may be at risk of an undesirable modeling phenomenon known as overfitting. We posed the question of whether overfitting may be influencing HOPS results and aimed for preliminary evidence of external validation with parameters from patients at our institution not included in the original HOPS study. We found HOPS to stratify our limited post-operative cohort adequately. However, the likelihood of complete seizure freedom among the patients predicted by HOPS to be at greatest chance of success was ∼75 %, about 20 points lower than in the original HOPS cohort. This reduction in absolute chance of success predicted by HOPS may represent some degree of overfitting. It will be informative to aim for external validation of HOPS utilizing patient cohorts entirely separate from those used for model generation. External validation of HOPS and similar models could optimize realistic prediction of success after intervention.