Bootstrapped disconnection-symptom mapping and decoding in stroke patients with motor deficits

1613 

Abstract Submission 

Ardalan Aarabi¹, Maedeh Khalilian¹, Martine Roussel¹, Olivier Godefroy¹

¹University of Picardy Jules Verne, Amiens, Picardie

Ardalan Aarabi  
University of Picardy Jules Verne
Amiens, Picardie

Maedeh Khalilian  
University of Picardy Jules Verne
Amiens, Picardie

Martine Roussel  
University of Picardy Jules Verne
Amiens, Picardie

Olivier Godefroy  
University of Picardy Jules Verne
Amiens, Picardie

Focal structural damage to white matter tracts can lead to functional deficits in stroke patients. Traditional voxel-based lesion-symptom mapping (VLSM) is susceptible to issues such as lesion frequency and collinearity between neighboring voxels (Godefroy et al., 1998). The statistical power of VLSM is also hindered by patient heterogeneity in terms of lesion locations, volumes, and neurological symptoms, especially when lesions are situated in critical areas with a significant impact on structural connectivity between brain regions associated with functional deficits (Boes et al., 2015). In this study, we present a bootstrapped multivariate structural disconnection-symptom mapping (SVR-DSM) and decoding method based on support vector regression (DeMarco et al., 2018) to identify brain structures associated with motor deficits in stroke patients.

Structural imaging data were acquired from 340 stroke patients (mean age 63.9 ± 10.5 years) with left/right motor deficits as part of the GRECogVASC study cohort [4] conducted at Amiens University Hospital. Clinical and neuropsychological examinations, along with MR imaging, were performed six months post-stroke. Among these patients, 64 displayed motor impairment in upper and lower limb items in the NIHSS (32 with left and 29 with right motor deficits, and 3 with both). Lesions for each patient were manually segmented on 3D T1 MRI data using MRIcron and normalized into the MNI152 template with SPM12. For disconnection-symptom mapping, a probabilistic structural disconnection map was created for each patient by utilizing streamlines passing through the patient's lesion. This was estimated through fiber tracking from diffusion-weighted imaging data obtained from over 400 healthy controls (mean age 62.87 ± 13.47 years) from the Cambridge Centre for Ageing and Neuroscience repository (CamCAN, Stage 2) [5]. In these probabilistic maps, each voxel represented a disconnection probability (ranging from 0 to 1) based on the number of healthy subjects exhibiting a disconnection in that specific voxel [6]. For the left and right motor deficit, 100 bootstrapped bags of disconnection maps were then generated from all 340 patients to balance patients with (32 with left, 29 with right and 3 with both) motor deficits compared to 276 without motor deficits. Multivariate SVR regression analysis was then used to obtain a β-map for each bag using logistic regression and permutation testing (p<0.005, 1000 permutations) for left/right motor deficits. Statistical comparison of β-maps (SPM's second level, p<0.005, FWER corrected) using permutation testing produced a t-statistic map of voxels with significant associations with left/right motor deficits across all bags. Finally, a decoding model was constructed to predict right-left motor scores on a voxel-by-voxel basis using a searchlight strategy [7] and the leave-one-out cross-validation strategy across the bags.

The statistical maps obtained using SVR-DSM were more effective in identifying brain structures, including the corticospinal tract, shown to be associated with left/right motor deficits (Arnoux et al., 2018). In comparison with VLSM, for the right motor deficit, no lesioned voxels survived with p<0.005 in the left hemisphere after correction for multiple comparisons. SVR-DSM also demonstrated high sensitivities (0.87 and 0.79) in identifying patients with left and right motor deficits, respectively. The decoding model based on bootstrapped SVR maps achieved high accuracies (up to 100%) for disconnected voxels in identifying patients with motor deficits.

Our results showed that for patient datasets with imbalanced data the bootstrapped bagging could significantly improve the performance of structural disconnection-deficit mapping compared to lesion-symptom mapping methods for identifying brain structures associated with motor deficits in stroke patients with relatively low false positive rates when dealing with unbalanced data

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Diffusion MRI Modeling and Analysis ¹

Cerebrovascular Disease

Computational Neuroscience

Data analysis

Modeling

Motor

Multivariate

STRUCTURAL MRI

Tractography

White Matter

Other - stroke