Abstracts

This abstract has been invited to present during the Broadening Representation Inclusion and Diversity by Growing Equity (BRIDGE) poster session

Rationale: Psychogenic nonepileptic (functional) seizures (FS) are paroxysmal events which clinically resemble epileptic seizures (ES) but do not involve ictal discharge. Recently, machine learning (ML) has emerged as a means of extracting useful biomarkers for individual diagnoses and may address current limitations of a rule-out diagnosis in FS. Prior ML applications to FS detection may have obscured subtle morphological changes. Specifically, FS is often preceded by traumatic brain injury (TBI) which may serve as an important model to assess the neural basis of FS. In this study, a random forest (RF) classifier was applied to morphological brain imaging metrics to test the hypothesis that ML can derive biomarkers that reliably discern patients with TBI and FS (FS-TBI) from those without FS (TBI-only).

Methods: Participants with FS-TBI (n = 62) or TBI-only (n = 59) underwent magnetic resonance imaging in a 3T Siemens Prisma MRI device. T1-weighted images were analyzed by voxel- and surface-based morphometry, which estimated the mean gray matter volume (GMV), cortical thickness, sulcal depth, fractal dimension, and gyrification of 36 bilateral ROIs. A RF classifier was trained on the full set of morphological features, which were ranked according to the Gini impurity index, where a greater mean decrease in Gini impurity equates to greater predictive power. An importance score threshold of 0.058 was chosen to define a subset of only the most predictive features on which to train a second RF classifier.

Results: RF model performance is quantified by the “out-of-bag” (OOB) error, or the mean prediction error on each training sample (created via random subsampling with replacement) using only trees fit from the OOB samples generated from unselected data. The RF classifier trained on all features displayed a mean OOB accuracy of 54.8% (OOB error = 45.2%). After feature selection, a total of 20 features were used to train an identical RF classifier, which attained a mean OOB accuracy of 79.8% (OOB error = 20.2%). These results demonstrate moderate performance improvement over a previous RF classifier applied on MRI metrics for the detection of FS against healthy controls, which returned a post-feature-selection OOB accuracy of 74.5%. Feature importance analysis revealed that the most predictive features for FS classification were GMVs in the left inferior frontal gyrus, left inferior temporal gyrus, and bilateral anterior orbital gyri, and the thickness of the right isthmus of the cingulate gyrus.

Conclusions: A RF classifier trained on morphological MRI metrics classified FS-TBI and TBI-only patients with 79.8% accuracy following feature selection. While controlling for TBI status marginally improved model performance, accuracy appears to be limited to 75%-80% for ROI-based approaches. As such, FS classification may benefit from a more sophisticated ML pipeline which trains a RF classifier on voxel-wise features extracted from MRI data by a deep learning network.

Funding: This work was supported by the U.S. Department of Defense (W81XH-17-0619) to WCL and JPS.