In medRxiv : the preprint server for health sciences
Mesial temporal lobe epilepsy (mTLE) is associated with variable dysfunction beyond the temporal lobe. We used functional anomaly mapping (FAM), a multivariate machine learning approach to resting state fMRI analysis to measure subcortical and cortical functional aberrations in patients with mTLE. We also examined the value of individual FAM in lateralizing the hemisphere of seizure onset in mTLE patients. Methods: Patients and controls were selected from an existing imaging and clinical database. After standard preprocessing of resting state fMRI, time-series were extracted from 400 cortical and 32 subcortical regions of interest (ROIs) defined by atlases derived from functional brain organization. Group-level aberrations were measured by contrasting right (RTLE) and left (LTLE) patient groups to controls in a support vector regression models, and tested for statistical reliability using permutation analysis. Individualized functional anomaly maps (FAMs) were generated by contrasting individual patients to the control group. Half of patients were used for training a classification model, and the other half for estimating the accuracy to lateralize mTLE based on individual FAMs. Results: Thirty-two right and 14 left mTLE patients (33 with evidence of hippocampal sclerosis on MRI) and 94 controls were included. At group levels, cortical regions affiliated with limbic and somatomotor networks were prominent in distinguishing RTLE and LTLE from controls. At individual levels, most TLE patients had high anomaly in bilateral mesial temporal and medial parietooccipital default mode regions. A linear support vector machine trained on 50% of patients could accurately lateralize mTLE in remaining patients (median AUC =1.0 [range 0.97-1.0], median accuracy = 96.87% [85.71-100Significance: Functional anomaly mapping confirms widespread aberrations in function, and accurately lateralizes mTLE from resting state fMRI. Future studies will evaluate FAM as a non-invasive localization method in larger datasets, and explore possible correlations with clinical characteristics and disease course.
Gholipour Taha, DeMarco Andrew, You Xiaozhen, Englot Dario J, Turkeltaub Peter E, Koubeissi Mohamad Z, Gaillard William D, Morgan Victoria L
2023-Feb-08
