Sixty-two patients (age: M = 53.7, SD = 10.7, range 22- 83 years; 34 males; 57 right-handed, see Table 1) met the inclusion criteria: All patients were older than 18 years, presented with first-ever ischemic (83%) or haemorrhagic (17%) stroke and behavioural deficits as assessed by a neurological examination. Patients who had a history of neurological or psychiatric presentations (e.g. transient ischemic attack), multifocal or bilateral strokes, or had MRI contraindications (e.g. claustrophobia, ferromagnetic objects) were excluded from the analysis (n = 131 patients, see the enrollment flowchart in the supplementary materials from Corbetta et al. 2015). We further limited our analysis to the patients whose motor functions were systematically assessed at 2 weeks, 3 months, and 1 year after their stroke for optimal longitudinal comparisons. This study was approved by the Washington University in Saint Louis Institutional Review Board and all participants gave their signed informed consent.
Table 1 Demographics of patients
Abbreviations: SD = standard deviation
Motor abilities assessment
Motor functions were examined for the upper and lower extremities. For upper extremities, active range of motion against gravity was measured by goniometer at shoulder flexion and wrist extension (Dreeben 2008). During the examination of shoulder flexion, patients are asked to raise their arm against gravity as high as possible. The movement amplitude is recorded as the angle between the goniometer centred on the shoulder and the lateral torso. The wrist extension examination requires patients to sit with their arm on the table in a resting position with their palms down, and they are asked to bend back their wrist against gravity. Wrist extension is measured as the angle between the goniometer centred on the wrist and the forearm.
Grip strength was measured using a dynamometer (Demeurisse et al. 1980). Each patient's examined arm was placed with the elbow flexed at 90°. Their fingers flexed for a maximal contraction over the dynamometer handle, while the forearm and wrist were kept in a neutral position. The patients were asked to take a breath while exerting the maximum grip effort during three consecutive trials. The strength score is recorded in kilograms, and the total score is calculated as the average over three trials (Fess and Moran 1981).
The patients’ dexterity and bimanual coordination was measured with the 9-Hole Peg Test (pegs/second) assessing gross movements of arms, hands, and fingers, and fine motor extremity. The test includes a one-piece board with a concave folded dish containing nine pegs next to a 9-holes matrix for the pegs. The task instructions require patients to place and remove the nine pegs one at a time and in random order as quickly as possible (Mathiowetz et al. 1985; Oxford Grice et al. 2003). The final score is calculated as the time in seconds elapsed from the touch of the first peg to when the last peg is placed back into the dish.
The Action Research Arm Test (ARAT, van der Lee et al. 2001) assesses the ability to perform purposeful movements with the upper limb extremities. Patients had to grasp, grip, pinch objects of different weights and shapes and perform gross motor movements. The ARAT’s four subtests have 19 items in total. Each item is rated on a four-point scale (0-3) where higher scores indicate better performance. If the patient scores <3 on the first item, the examiner advances to the second item two (the most accessible item). If the score for the second item is 0, the rest of the items will automatically be scored as 0 for and the test is stopped. If the patients score <3 on the first item but >0 on the second item, the remaining items are administered (Lyle 1981).
For lower extremities, a combined walking index (Kempen et al. 2011; Perry et al. 1995), left/right total motricity index, ankle dorsiflexion goniometry for left/right active range of motion against gravity was recorded (Dreeben 2008).
MRI acquisition and preprocessing
1. MRI scan acquisition
Neuroimaging was performed on a Siemens 3T Tim-Trio scanner at the School of Medicine of the Washington University in St. Louis. All structural scans were collected 2 weeks after the stroke and included (1) a sagittal MP-RAGE T1-weighted image (repetition time = 1950 msec, echo time = 2.26 msec, flip angle = 9 degrees, voxel size = 1.0 x 1.0 x 1.0 mm, slice thickness = 1.00 mm); (2) a transverse turbo spin-echo T2-weighted image (repetition time = 2500 msec, echo time = 435 msec, voxel-size = 1.0 x 1.0 x 1.0 mm, slice thickness = 1.00 mm); and (3) a sagittal FLAIR (fluid attenuated inversion recovery) (repetition time = 7500 msec, echo time = 326 msec, voxel-size = 1.5 x 1.5 x 1.5 mm, slice thickness = 1.50 mm).
2. Lesion segmentation (native space)
Lesions were manually segmented on the T1-weighted MRI images using the Analyze biomedical imaging software system (Robb and Hanson, 1991). Two board-certified neurologists (Drs Corbetta and Carter) reviewed all segmentations blinded to the individual behavioural data.
3. Spatial normalisation (MNI152)
To align T1-weighted MRI scans of patients to a standard stereotaxic space (Montreal Neurological Institute space, MNI152 Grabner et al. 2006), it is necessary to first address the issue of space deformation caused by brain lesions during spatial normalisation (Brett et al. 2001; Ripolles et al. 2012; Volle et al. 2013). An enantiomorphic approach was implemented in the current data analysis: the native-space lesions were replaced with healthy tissue of the same region in the contralateral hemisphere (Nachev et al. 2008). Subsequently, affine and diffeomorphic deformations were applied to co-register scans and lesions to the MNI152 space using the Advanced Normalization Tool (ANTS, Avants et al. 2011; Klein et al. 2009). These analyses are available as part of the ‘Normalisation’ part of BCBtoolkit (Foulon et al. 2018; http://toolkit.bcblab.com).
4. Generation of disconnection maps.
Methodological details are available from Thiebaut de Schotten et al. (2020). In brief, each lesion serves as the input for the BCBtoolkit’s Disconnectome tool that computes maps of white matter pathway disconnection probabilities and its impact on loss of function (Foulon et al. 2018; http://toolkit.bcblab.com). Probabilities of white matter pathways were derived from a normative population of 163 healthy controls (44.8% males) using a diffusion-weighted imaging dataset acquired on a 7T scanner as part of the Human Connectome Project (Vu et al. 2015). The pattern of brain areas that were disconnected in each stroke was subsequently characterised by measuring the average level of disconnection in subcortical areas and areas derived from a multimodal atlas of the brain surface (Glasser et al. 2016). Subcortical areas included manually defined by MTS and included the thalamus, the putamen, the pallidum, the hippocampus, the caudate nucleus, and the amygdala. Hence, for each of the 62 patients in this study, the disconnection probability of 372 grey matter structures (186 structures in each hemisphere: 180 cortical and 6 subcortical) was obtained.
Dimensionality reduction with Principal Component Analysis
1. Behavioural components
To summarise behavioural measurements data while keeping as much variability as possible and minimising noise, behavioural components were extracted as described in Corbetta et al. (2015). In brief, an oblique rotation principal component analysis was applied to the motor scores obtained with different neuropsychological tests described above. Two components (left and right side of the body) resulted from the analysis explaining 77% of the observed variance in the data. The analysis was computed in Matlab (MathWorks Inc.).
2. Disconnection map components
The Disconnectome maps components consist of 46 components, where 30 components have already been shown to capture more than 90% of the variance in the distribution of disconnection maps in stroke (Thiebaut de Schotten et al. 2020). These components were derived from an independent normative dataset of 1333 disconnection maps of ischemic stroke patients (M = 63.89, SD = 15.91, range 18-97 years; 56.1% males) fully described in (Xu et al. 2018) and (Thiebaut de Schotten et al. 2020).
We used the Disconnectome 46 components to decode the profile of disconnection in our sample of 62 patients. Linear regression with 46 independent variables (i.e. the disconnectome components) was performed in RStudio (v.1.3.1093; RStudio Team, 2020) to predict the disconnection patterns for each patient. In doing so, beta coefficients for each predictor corresponded to the component's scores of each patient. Adjusted r-squared values represented the percentage of each patient’s disconnection map variance explained by the 46 components.
The linear regression analysis demonstrated that the Disconnectome (46 principal components) was able to capture the variance in the disconnection patterns of each patient from the longitudinal dataset. The mean adjusted r-squared across the group is 0.94, with a standard deviation of 0.076.
Hierarchical linear regression analyses
The analyses for the anatomical prediction of the motor outcome were carried out in RStudio (v.1.3.1093; RStudio Team, 2020).
Six hierarchical linear regression analyses were performed. Each hierarchical linear regression used the patients’ estimated scores of each component (i.e. beta coefficients from the linear regression) to predict two principal components of motor scores (the left and right side of the body, n = 2) two weeks, three months, and one year after the stroke (n = 3 time points). The least contributively predictor was removed iteratively by the hierarchical linear regression. The goodness of fit of consecutive linear models was compared statistically using an F-test. The process was repeated until no significant difference could be identified between two consecutive models.
This procedure is available as supplementary code with the manuscript (see https://github.com/lidulyan/Hierarchical-Linear-Regression-R-).
Validation of the predictive model
To avoid inflation of the significance of our results (i.e. overfitting), the hierarchical linear regression analyses were performed on 78% of the original dataset (training set).
To achieve the generalisability of our model and avoid overfitting, which is one of the limitations of currently available predictive models in the stroke literature, the optimised model with the reduced number of predictors was tested on the unseen by the model data (22% from the original dataset). This process was repeated 5000 times with different data splits by varying the seed randomly to control the potential error induced by the split. A median from the R-squared distribution, representative of the most representative model, was used to describe model fitness.
Disconnectome components with highly significant contributions to motor impairment (p<0.001) were displayed using Surfice (https://www.nitrc.org/projects/surfice/).
The white matter was identified and labelled manually by expert anatomists (MTS/SJF) according to the Atlas of Human Brain Connections (Rojkova et al. 2016; Thiebaut de Schotten et al. 2015).