A Bayesian averaged response-driven multinomial model for lateralization of temporal lobe epilepsy

Purpose: To develop a Bayesian averaged multinomial model for lateralization of epileptogenicity in temporal lobe epilepsy (TLE) patients based upon features extracted from preoperative T1-weighted and FLAIR imaging. Methods: A retrospective cohort of seventy-six TLE patients with surgical outcome of Engel class I was quantitatively analyzed to extract hippocampi volumetrics and FLAIR intensity. Using multinomial logistic regression, single response-driven models were estimated. Based on Bayesian model averaging (BMA), a model was developed and its performance was compared with the single response models. Results: The Bayesian averaged model achieved a lateralization rate of 84.2% for TLE patients that was higher than any single response model. Out of the thirty-four patients who underwent phase II intracranial monitoring, the epileptogenic side was correctly lateralized in nineteen cases. Conclusion: The proposed response-driven model can improve the decision-making for surgical resection and may reduce the need for implantation of intracranial monitoring electrodes.