Award: American Academy of Pediatrics Award
Authors: Ryan Dinh Nguyen; Matthew Smyth, MD; Liang Zhu, PhD; Ludovic Pao, BS; Anish Mitra, BS; Rajan Patel, MD; Jeremy Lankford, MD; Gretchen Von Allmen, MD; Michael Watkins, MD, PhD; Michael Funke, MD; Manish Shah, MD (Houston, TX)
Introduction: Obtained with standard structural MR imaging, resting state MRI (rsMRI) in pediatric epilepsy patients could be useful to identify surgical candidates. This study aims to assess the utility of automated extreme gradient decision-tree boosting (XGBoost) binary classification of pediatric epilepsy disease state specifically using rsMRI latency map analysis. Methods: With IRB approval, preoperative rsMRI and anatomical MRI scans were obtained from multiple centers for 63 epilepsy and 259 healthy control (HC) patients. After atlas transformation, voxel-wise cross-covariant analysis was performed to create a latency map of the temporal difference between the rsMRI signal and the global mean signal. Voxel-wise latency z-score maps were created for each epilepsy and HC patient using HC mean and standard deviation latency maps. Features were obtained from rsMRI latency z-score maps using the FMRIB Software Library statistics tool. The distribution of feature values between epilepsy and HC patient groups were compared to determine viability of use in training. Using an XGBoost machine learning algorithm, mean latency z-score and median latency z-score were used as features for training along with corresponding patient binary labeling (0=HC, 1=epilepsy). Area under the receiver-operating characteristic curve (AUC) was calculated to evaluate the algorithm’s classification of epilepsy disease state (present or absent). Extent of model overfitting was evaluated by calculating the difference between train-test AUC and accuracy. A 14 k-fold cross validation grid-search was performed to optimize XGBoost models. Results: An XGBoost model correctly classified epilepsy disease state in 71% of patients with an AUC of 0.73. Conclusion: An XGBoost machine learning algorithm coupled with rsMRI latency analysis can potentially accurately identify pediatric refractory epilepsy in an automated fashion. Study of additional machine learning algorithms with larger datasets is needed to further establish the value of automated classification in identifying pediatric refractory epilepsy and reducing time to surgical referral.