Abstract

Conventional mass-univariate analyses have been previously used to test for group differences in neural signals. However, machine learning algorithms represent a multivariate decoding approach that may help to identify neuroimaging patterns associated with functional impairment in “individual” patients. We investigated whether fMRI allows classification of individual motor impairment after stroke using support vector machines (SVMs). Forty acute stroke patients and 20 control subjects underwent resting-state fMRI. Half of the patients showed significant impairment in hand motor function. Resting-state connectivity was computed by means of whole-brain correlations of seed time-courses in ipsilesional primary motor cortex (M1). Lesion location was identified using diffusion-weighted images. These features were used for linear SVM classification of unseen patients with respect to motor impairment. SVM results were compared with conventional mass-univariate analyses. Resting-state connectivity classified patients with hand motor deficits compared with controls and nonimpaired patients with 82.6–87.6% accuracy. Classification was driven by reduced interhemispheric M1 connectivity and enhanced connectivity between ipsilesional M1 and premotor areas. In contrast, lesion location provided only 50% sensitivity to classify impaired patients. Hence, resting-state fMRI reflects behavioral deficits more accurately than structural MRI. In conclusion, multivariate fMRI analyses offer the potential to serve as markers for endophenotypes of functional impairment.

Introduction

Stroke is a leading cause of permanent motor disability with an incidence of >15 million people worldwide (World Health Organization 2004; Go et al. 2013). Neuroimaging techniques including structural and functional MRI (fMRI) have shown that motor deficits after stroke are associated with disturbances in anatomical pathways and brain activation patterns (Chollet et al. 1991; Ward et al. 2003, 2006; Stinear et al. 2007; Grefkes et al. 2008; Rehme, Fink et al. 2011; Rehme et al. 2012). Recent studies examined stroke-induced changes in functional connectivity between brain areas based on resting-state fMRI (Friston 1994; Rehme and Grefkes 2013; Grefkes and Fink 2014). Resting-state functional connectivity is usually determined by correlating regional low-frequency blood oxygenation level-dependent time-series assessed in the absence of an active task (Friston 1994). Resting-state studies after stroke provided evidence for a close relationship between disturbed functional connectivity of cortical motor areas distant from the lesion site and motor impairment. In particular, interhemispheric resting-state connectivity between primary motor area (M1) of both hemispheres has been shown to be reduced in stroke patients with motor impairments (Carter et al. 2010, 2012; Wang et al. 2010; Park et al. 2011; Golestani et al. 2013).

To date, neuroimaging studies in stroke research have frequently used conventional mass-univariate analyses within the framework of the General Linear Model (GLM) focusing on the statistical relationship between behavioral performance or group membership and neural signals (Friston et al. 1995). These mass-univariate analyses represent a special form of encoding models explaining how the signal measured at a given voxel is generated by underlying hidden factors (e.g., sensory, cognitive, or motor processes) based on experimental manipulations or clinical groups (Kay and Gallant 2009; Naselaris et al. 2011). Statistical inference in encoding models is thus drawn from forward models using experimental information to predict brain activity (Haufe et al. 2014). However, classical encoding analyses do not provide neuroimaging markers that can be used for the diagnosis and prognosis of functional impairment in individual patients (Orrù et al. 2012). In contrast, decoding approaches use brain activity to discriminate groups or experimental conditions (Kay et al. 2008; Kriegeskorte 2011; Naselaris et al. 2011). Here, statistical inference is based on a backward model as the data-generating process is reversed to predict experimental categories from measured brain signals (Naselaris et al. 2011; Haufe et al. 2014). Thus, decoding models test how phenotypical information can be decoded from neuroimaging data (Haufe et al. 2014). Importantly, decoding models allow classifying at the level of individual subjects. In contrast to conventional GLMs, multivariate analyses also consider voxelwise signal dependencies, which increase the power to detect signals of interest (Lemm et al. 2011; Haufe et al. 2014). This raises the clinically relevant question whether a multivariate decoding approach considering the combined multivoxel information from neuroimaging data may be used to classify behavioral performance at the level of individual patients (Orrù et al. 2012).

Support vector machines (SVMs) represent a multivariate decoding technique, which has recently been introduced into the field of clinical neuroimaging (Cortes and Vapnik 1995; Pereira et al. 2009; Lemm et al. 2011; Orrù et al. 2012). SVMs are supervised machine learning algorithms that are most commonly used for classification but can also be applied, for example, in regression analyses (Pereira et al. 2009; Chang and Lin 2011; Lemm et al. 2011; Naselaris et al. 2011). Technically, SVMs consist of a training step where data are nonlinearly mapped into a high-dimensional feature space to optimize a separation boundary between predefined classes. After training, the trained classifier is used to predict the class level of previously unseen data (Cortes and Vapnik 1995; Lemm et al. 2011; Orrù et al. 2012).

Previous studies showed that SVM analyses of resting-state fMRI data can distinguish neuropsychiatric patients from healthy controls for diseases including mild cognitive impairment, multiple sclerosis, major depression, or schizophrenia (Shen et al. 2010; Richiardi et al. 2012; Wee et al. 2012; Zeng et al. 2012). Furthermore, there is evidence that SVM techniques for task-based fMRI data may be used to predict the functional outcome of stroke patients with aphasia (Saur et al. 2010). The use of task-based fMRI in a clinical setting is, however, difficult because it requires both a sophisticated technical set-up and a sufficiently high compliance of the patient. In contrast, resting-state fMRI can be acquired within a few minutes with minimum cooperation of the patient. Therefore, classification of clinical populations based on resting-state data offers potential for providing noninvasive diagnostic markers to predict individual outcome and treatment effects in a standard clinical environment.

The aim of this study was to identify endophenotypes of motor disability after stroke by decoding motor outcome from individual resting-state fMRI connectivity patterns using SVM. Furthermore, we computed mass-univariate decoding analyses to test whether multivariate classification is superior to univariate classification (Kriegeskorte 2011; Lemm et al. 2011; Naselaris et al. 2011). To this end, we implemented a resting-state fMRI scan into a standard MRI stroke protocol and examined 3 groups: acute stroke patients with hand motor deficits, acute stroke patients without hand motor deficits, and nonstroke neurological control patients. We computed the resting-state functional connectivity of ipsilesional M1 and trained linear two-class SVMs to 1) classify new patients and 2) identify resting-state neuroimaging markers of motor deficits after stroke. Individual resting-state markers of motor impairment were compared with the results of a conventional univariate group analysis of resting-state connectivity. Finally, we also classified motor impairment based on the size and location of structural lesions to compare the sensitivity of structural and functional MRI scans.

Materials and Methods

Sample

The study was approved by the local ethics committee (File No. 11–191). From November 2011 to March 2013, we screened patients from the stroke unit, Department of Neurology, University of Cologne, for study eligibility. Inclusion criteria for stroke patients were as follows: 1) first-ever ischemic stroke as verified by diffusion-weighted imaging (DWI) (Fig. 1; for the location of individual lesions, see Supplementary Fig. 1), 2) cortical and/or subcortical lesions sparing M1, 3) symptom onset <8 days, and 4) age: 30–99 years. Exclusion criteria were as follows: 1) hemorrhagic stroke, 2) contraindications to MRI, 3) previous stroke as detected on FLAIR images, 4) bi-hemispheric stroke lesions, 5) carotid artery stenosis >50% according to the NASCET criteria, 6) intracranial stenosis as tested by transcranial Doppler, 7) cognitive impairment or dementia, 8) impaired consciousness, and 9) other neurological or psychiatric disease. Forty acute right-handed stroke patients either with unilateral motor deficits of the upper limb (n = 20) as well as patients without motor deficits (n = 20) were finally enrolled into the study (cf. Table 1 for clinical data). Additionally, we recruited 20 right-handed nonstroke patients from the stroke unit who showed no lesions in FLAIR or DWI images (cf. Table 2). Patients gave informed written consent according to the Declaration of Helsinki.

Table 1

Clinical data of stroke patients

Patient number Age (years) Gender (M, male; F, female) Time after stroke onset (days) Affected hemisphere MI (affected hand score) Mean MI (affected upper limb) Mean MI (affected lower limb) Trunk control test Grip force Index (% affected/unaffected) NIHSS (total) Modified Rankin Scale Lesion volume (mm³) 
Patients with hand motor impairment (n = 20) 
 Patient 1 91 21 38 1160 
 Patient 2 86 15 26 706 
 Patient 3 65 12 4894 
 Patient 4 78 12 11 14 388 
 Patient 5 71 11 13 14 891 
 Patient 6 89 11 13 21 1860 
 Patient 7 64 57 122 
 Patient 8 72 22 27 25 13 12 57 621 
 Patient 9 74 22 17 14 15 19 36 123 
 Patient 10 60 22 20 13 21 6183 
 Patient 11 50 22 27 28 25 23 2892 
 Patient 12 74 26 29 33 25 31 7064 
 Patient 13 69 19 29 33 25 35 1920 
 Patient 14 63 26 19 19 19 38 2363 
 Patient 15 89 26 22 22 42 1762 
 Patient 16 72 11 44 12 77 709 
 Patient 17 57 22 29 33 25 50 4361 
 Patient 18 84 26 29 33 25 55 12 346 
 Patient 19 70 26 24 30 25 56 2433 
 Patient 20 64 26 25 33 25 60 3098 
 MEAN 72.1  3.5  15.9 16.5 18.6 13.0 24.8 35 145 
 SD 11.4  2.1  10.6 11.4 12.4 10.1 21.7 84 527 
Patients without hand motor impairment (n = 20) 
 Patient 21 39 33 33 33 25 90 1222 
 Patient 22 67 33 31 33 25 90 7125 
 Patient 23 85 33 33 33 25 91 537 
 Patient 24 81 33 33 33 25 94 992 
 Patient 25 69 33 33 33 25 94 8583 
 Patient 26 74 33 33 33 25 95 1151 
 Patient 27 74 33 33 33 25 99 1070 
 Patient 28 60 33 33 33 25 106 4678 
 Patient 29 76 33 33 33 25 106 1306 
 Patient 30 82 33 33 33 25 107 1627 
 Patient 31 69 33 33 33 25 109 1677 
 Patient 32 72 33 33 33 25 112 4776 
 Patient 33 84 33 33 33 25 116 2761 
 Patient 34 43 33 33 33 25 119 4860 
 Patient 35 68 33 33 33 25 121 105 560 
 Patient 36 81 33 33 33 25 122 71 226 
 Patient 37 79 33 33 33 25 122 749 
 Patient 38 82 33 30 33 25 132 3895 
 Patient 39 75 33 31 30 22 136 1168 
 Patient 40 69 33 33 33 25 138 6642 
 MEAN 71.5  2.8  33 32.6* 32.9 24.8 111.0 11 580 
 SD 12.3  2.1  0.9 0.6 0.7 15.1 26 958 
Patient number Age (years) Gender (M, male; F, female) Time after stroke onset (days) Affected hemisphere MI (affected hand score) Mean MI (affected upper limb) Mean MI (affected lower limb) Trunk control test Grip force Index (% affected/unaffected) NIHSS (total) Modified Rankin Scale Lesion volume (mm³) 
Patients with hand motor impairment (n = 20) 
 Patient 1 91 21 38 1160 
 Patient 2 86 15 26 706 
 Patient 3 65 12 4894 
 Patient 4 78 12 11 14 388 
 Patient 5 71 11 13 14 891 
 Patient 6 89 11 13 21 1860 
 Patient 7 64 57 122 
 Patient 8 72 22 27 25 13 12 57 621 
 Patient 9 74 22 17 14 15 19 36 123 
 Patient 10 60 22 20 13 21 6183 
 Patient 11 50 22 27 28 25 23 2892 
 Patient 12 74 26 29 33 25 31 7064 
 Patient 13 69 19 29 33 25 35 1920 
 Patient 14 63 26 19 19 19 38 2363 
 Patient 15 89 26 22 22 42 1762 
 Patient 16 72 11 44 12 77 709 
 Patient 17 57 22 29 33 25 50 4361 
 Patient 18 84 26 29 33 25 55 12 346 
 Patient 19 70 26 24 30 25 56 2433 
 Patient 20 64 26 25 33 25 60 3098 
 MEAN 72.1  3.5  15.9 16.5 18.6 13.0 24.8 35 145 
 SD 11.4  2.1  10.6 11.4 12.4 10.1 21.7 84 527 
Patients without hand motor impairment (n = 20) 
 Patient 21 39 33 33 33 25 90 1222 
 Patient 22 67 33 31 33 25 90 7125 
 Patient 23 85 33 33 33 25 91 537 
 Patient 24 81 33 33 33 25 94 992 
 Patient 25 69 33 33 33 25 94 8583 
 Patient 26 74 33 33 33 25 95 1151 
 Patient 27 74 33 33 33 25 99 1070 
 Patient 28 60 33 33 33 25 106 4678 
 Patient 29 76 33 33 33 25 106 1306 
 Patient 30 82 33 33 33 25 107 1627 
 Patient 31 69 33 33 33 25 109 1677 
 Patient 32 72 33 33 33 25 112 4776 
 Patient 33 84 33 33 33 25 116 2761 
 Patient 34 43 33 33 33 25 119 4860 
 Patient 35 68 33 33 33 25 121 105 560 
 Patient 36 81 33 33 33 25 122 71 226 
 Patient 37 79 33 33 33 25 122 749 
 Patient 38 82 33 30 33 25 132 3895 
 Patient 39 75 33 31 30 22 136 1168 
 Patient 40 69 33 33 33 25 138 6642 
 MEAN 71.5  2.8  33 32.6* 32.9 24.8 111.0 11 580 
 SD 12.3  2.1  0.9 0.6 0.7 15.1 26 958 

Note: *Note that the MI score for the upper limb was slightly reduced in 3 patients without hand motor impairment as they had difficulties to move the proximal muscles of elbow and shoulder against external resistance despite full grip force and hand motor performance.

MI, Motricity Index; NIHSS, National Institutes of Health Stroke Scale.

Table 2

Clinical data of nonstroke neurological patients (control group)

Patient number age (years) gender (M, male; F, female) Mean MI (affected upper limb) Mean MI (affected lower limb) Trunk Control Test Grip force Index (% affected/unaffected) Modified Rankin Scale Clinical diagnosis 
Control 1 79 33 33 25 90 TIA 
Control 2 72 33 33 25 93 Meningitis 
Control 3 75 33 33 25 94 TIA 
Control 4 46 33 33 25 95 AION 
Control 5 27 33 33 25 96 TIA 
Control 6 50 33 33 25 99 TIA 
Control 7 56 33 33 25 100 Peripheral facial nerve palsy 
Control 8 74 33 33 25 104 AION 
Control 9 54 33 33 25 104 Migraine 
Control 10 70 33 33 25 104 TIA 
Control 11 72 33 33 25 105 TIA 
Control 12 72 33 33 25 105 Amaurosis fugax 
Control 13 81 33 33 25 105 TIA 
Control 14 57 33 33 25 105 TIA 
Control 15 53 33 33 25 105 TIA 
Control 16 38 33 33 25 106 TIA 
Control 17 81 33 33 25 116 TIA 
Control 18 67 33 33 25 124 TIA 
Control 19 71 33 33 25 127 TIA 
Control 20 92 33 33 25 148 TIA 
MEAN 64.4  33.0 33.0 25.0 106.3 0.0  
SD 16.2  0.0 0.0 0.0 13.6 0.0  
Patient number age (years) gender (M, male; F, female) Mean MI (affected upper limb) Mean MI (affected lower limb) Trunk Control Test Grip force Index (% affected/unaffected) Modified Rankin Scale Clinical diagnosis 
Control 1 79 33 33 25 90 TIA 
Control 2 72 33 33 25 93 Meningitis 
Control 3 75 33 33 25 94 TIA 
Control 4 46 33 33 25 95 AION 
Control 5 27 33 33 25 96 TIA 
Control 6 50 33 33 25 99 TIA 
Control 7 56 33 33 25 100 Peripheral facial nerve palsy 
Control 8 74 33 33 25 104 AION 
Control 9 54 33 33 25 104 Migraine 
Control 10 70 33 33 25 104 TIA 
Control 11 72 33 33 25 105 TIA 
Control 12 72 33 33 25 105 Amaurosis fugax 
Control 13 81 33 33 25 105 TIA 
Control 14 57 33 33 25 105 TIA 
Control 15 53 33 33 25 105 TIA 
Control 16 38 33 33 25 106 TIA 
Control 17 81 33 33 25 116 TIA 
Control 18 67 33 33 25 124 TIA 
Control 19 71 33 33 25 127 TIA 
Control 20 92 33 33 25 148 TIA 
MEAN 64.4  33.0 33.0 25.0 106.3 0.0  
SD 16.2  0.0 0.0 0.0 13.6 0.0  

AION, anterior ischemic optic neuropathy; TIA, transitory ischemic attack.

Figure 1.

Summary map of DWI lesions. (A) For the entire sample of stroke patients, the maximum overlap (n = 9/40) was at the posterior limb of the internal capsule of the CST. (B) Considering each stroke group separately, the maximum overlap in patients with hand motor deficits (n = 7/20) was also found at the posterior limb of the internal capsule. Thus, the overlap of lesion locations causing hand motor deficits was rather low. (C) Stroke patients without hand motor deficits also showed a maximum, albeit low lesion overlap at the internal capsule region (n = 5/20). This indicates that the lesion location within the internal capsule is not always associated with impairment of upper limb function. Lesion maps were normalized to an MNI reference brain. Lesion overlap was created using MRIcron (http://www.mccauslandcenter.sc.edu/mricro/mricron/index.html). For individual lesion maps, see Supplementary Figure 1.

Figure 1.

Summary map of DWI lesions. (A) For the entire sample of stroke patients, the maximum overlap (n = 9/40) was at the posterior limb of the internal capsule of the CST. (B) Considering each stroke group separately, the maximum overlap in patients with hand motor deficits (n = 7/20) was also found at the posterior limb of the internal capsule. Thus, the overlap of lesion locations causing hand motor deficits was rather low. (C) Stroke patients without hand motor deficits also showed a maximum, albeit low lesion overlap at the internal capsule region (n = 5/20). This indicates that the lesion location within the internal capsule is not always associated with impairment of upper limb function. Lesion maps were normalized to an MNI reference brain. Lesion overlap was created using MRIcron (http://www.mccauslandcenter.sc.edu/mricro/mricron/index.html). For individual lesion maps, see Supplementary Figure 1.

Magnetic Resonance Imaging

All sequences were part of a clinical MRI protocol for stroke diagnostics acquired on a 1.5T scanner (Philips, Guildford, UK). Axial whole-brain resting-state fMRI scans were obtained using a gradient echo planar imaging (EPI) sequence: TR = 2100 ms, TE = 50 ms, FOV = 250 mm, 24 axial slices, voxel size: 3.9 × 3.9 × 3.9 mm³, 183 volumes, ∼6 min. Prior to scanning, patients were instructed to rest motionless and awake. The MRI stroke protocol further included DWI, T2-weighted, T2*-weighted, FLAIR, and time-of-flight angiography sequences (see Supplementary material for details).

Behavioral Tests

We assessed motor function in a brief bedside test at the day of the MRI. Global neurological impairment was rated using the National Institutes of Health Stroke Scale (NIHSS). Motor performance was measured using the Motricity Index (MI) (Demeurisse et al. 1980). The MI is a brief rating scale based on movements of the proximal, middle, and distal joints of arms and legs, which are classified according to whether they can be performed against gravity or even against resistance. Normal motor performance is rated by a maximum score of 33 per item. Additionally, the maximum grip force was assessed for each hand in 3 consecutive trials using a vigorimeter. We computed a grip force index representing the percent grip force of the stroke-affected relative to the unaffected hand (i.e., mean grip force [stroke-affected hand]/mean grip force [unaffected hand] × 100).

Clinical Definition of Motor Impairment Groups

To decode neuroimaging markers of hand motor impairment after stroke based on M1 resting-state connectivity patterns, we defined 2 groups of stroke patients either with hand motor impairment or without hand motor impairment based on the grip force index and the MI score of the stroke-affected hand. Patients without hand motor impairment were defined to have 1) preserved grip strength as indicated by a grip force index of ≥ 90% and 2) the maximum MI hand score of 33 (Table 1). These criteria guaranteed that none of the patients in this group featured significant hand weakness.

In contrast, patients with hand motor deficits were defined to show 1) considerable hand weakness as indicated by a grip force index of ≤2/3 (66%) and 2) an MI score of the stroke-affected hand of ≤26. The grip force criterion ensured that reduced grip strength was indeed caused by stroke as it has been previously shown that differences in handedness may result in reductions of the grip force index of up to 30% (Crosby et al. 1994). The MI criterion implied that hand movements could not be performed against resistance or even gravity. Finally, 40 patients fulfilled the criteria of being assigned to either group. Nonstroke controls showed normal motor performance (grip force ≥ 90%, MI = 33) (Table 2). We compared clinical parameters, age, and gender between groups using ANOVA or chi-squared tests in SPSS21 (IBM, New York).

Resting-State fMRI Analysis

Resting-state fMRI data were analyzed using Statistical Parametric Mapping (SPM8; http://www.fil.ion.ucl.ac.uk/spm/) and Matlab R2012a (The Mathworks, Natick, MA, USA) according to standard preprocessing protocols (Rehme et al. 2013). The first 3 images of the time-series were discarded. The remaining 180 resting-state EPI volumes were spatially realigned to the mean image of the time-series and co-registered with the structural T2-weighted image. Then, all images were spatially normalized to the standard template of the Montreal Neurological Institute (MNI, Canada) using the unified segmentation approach with masked lesions (Ashburner and Friston 2005). Finally, data were smoothed with an isotropic Gaussian kernel of 8 mm. After preprocessing, the effect of known confounds was removed from the time-series. Confound regressors included the mean-centered global, gray matter, white matter, and cerebrospinal fluid signal intensities and their squared values, the 6 head motion parameters from image realignment, their squared values as well as their first-order derivatives (Satterthwaite et al. 2013). Furthermore, the time-series was adjusted for the first 5 components of a principal component analysis (Behzadi et al. 2007; Chai et al. 2012). Data were then band-pass filtered between 0.01 and 0.08 Hz. The majority of the stroke patients had left-hemispheric stroke lesions (n = 31/40 patients). Patients with right-hemispheric lesions were equally distributed in the 2 stroke groups (n = 5 with hand motor deficits, n = 4 without hand motor deficits). Thus, for group comparisons, data of these 9 patients were flipped along the midsagittal plane. Consequently, the left hemisphere corresponded to the ipsilesional hemisphere in all patients. We tested for group differences in head motion parameters from image realignment by comparing the framewise displacement and the root mean squared error (RMSE) (Power et al. 2012). There was no difference between groups neither in the framewise displacement (F2,57 = 0.126, P = 0.882) nor in the RMSE (F2,57 = 0.158, P = 0.854).

Resting-state connectivity maps of ipsilesional M1 were obtained by computing voxelwise whole-brain correlations of time-courses with the ipsilesional M1 seed region, which were transformed into Fisher Z-scores (Biswal et al. 1995). The M1 seed coordinate was defined from a meta-analysis on neural activity for hand motor tasks and was located at the rostral wall of the central sulcus at the “hand knob” area (Rehme et al. 2012). Note that none of the subjects suffered from lesions in M1 as verified by structural scans (DWI, FLAIR, T2).

Multivariate SVM Analysis

SVM Features

We used 1) voxels of individual resting-state connectivity maps representing the functional connectivity between M1 seed region and whole-brain target voxels and 2) voxels of binary DWI lesion maps as input variables ( = features) for classification.

To focus on the cortical resting-state motor network and ensure an identical number of voxelwise features per patient, individual resting-state maps were masked by cytoarchitectonic probability maps of frontoparietal sensorimotor areas (Eickhoff et al. 2005) (Brodmann areas 6, 4a/b, 3a/b, 1, 2) using FSLv5.0 (http://fsl.fmrib.ox.ac.uk). Thus, after masking, the resting-state functional connectivity map consisted of 67 957 voxelwise feature variables.

Voxelwise connectivity as a linear parameter was scaled to range between 0 and 1. To increase predictive power and reduce noise, we computed Fisher's criterion score for feature selection (Müller et al. 2004; Feis et al. 2013). This separation index reflects the squared difference between group means in relation to intragroup variance for each feature. In the leave-one-subject-out cross-validation approach described later, feature selection was carried out for each training sample separately. The classification of one left-out subject was based on the features selected in the respective training sample. This procedure guaranteed that the classification of data in the testing step was inherently independent from the selection criteria applied in the training step to prevent “double dipping” (Kriegeskorte et al. 2009). For each training sample, we entered 5% of the most discriminative voxels of the resting-state connectivity maps (i.e., 3398 features per patient) into the SVM algorithm.

To test whether lesion location can be used to classify motor impairment after stroke, one SVM was set up based on the whole-brain DWI lesion maps, which were masked by a summary map of all lesions (178 437 voxelwise features per patient). In this way, noninformative voxels where none of the patients showed a lesion were removed from the analysis. Furthermore, to consider only corticospinal tract (CST) lesions for classification, we masked the lesion maps with a probabilistic map of the CST (Oishi et al. 2010), which yielded 27 437 features per patient.

SVM Classification

We computed 3 different linear two-class SVMs using the Library for SVMs (Chang and Lin 2011) (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) to classify patients according to 1 of 3 groups: stroke patients without hand motor impairment, stroke patients with hand motor deficits, and nonstroke controls.

Each SVM was trained and tested using a nested leave-one-subject-out cross-validation procedure (Fig. 2) (Müller et al. 2004; Dosenbach et al. 2010; Feis et al. 2013). In the outer loop, one subject was considered as the test sample and thus set aside whereas all remaining subjects formed the training sample (N-1). Feature selection was repeated for each training sample. Then, the classification model was optimized based on different soft margin constants of the linear separation boundary (i.e., C-parameters ranging from small [C = 0.0001] to large [C = 30]). For this optimization, we used another leave-one-subject-out cross-validation procedure by successively omitting and classifying one subject of the training data. In the outer loop, the model with the highest accuracy and the highest generalizability (i.e., smallest C-parameter) of the inner loop was used to classify the left-out, unknown subject (test sample). The outer loop was repeated for each subject left out and being classified based on the model optimized for the rest of the sample. Thus, training and testing were entirely independent steps (Kriegeskorte et al. 2009). We then computed the posterior balanced accuracy of classifications across all outer loops and reported 95% confidence intervals (CIs) (Brodersen et al. 2010). The significance of this classification was tested using chi-squared tests for equal distributions of correct and incorrect classifications. Finally, we reconstructed the SVM feature weights for the feature selection of each training data set and averaged these weights across all 40 training sets to visualize the voxels that mostly contributed to classification.

Figure 2.

SVM design. (A) Preparing feature variables for the SVM analysis. The left panel shows the seed region at the hand knob formation in primary motor cortex (M1) of the ipsilesional (affected) hemisphere, which was used to compute voxelwise correlation maps with every voxel within frontoparietal brain areas (cf. Methods). The right panel shows individual examples of resting-state functional connectivity maps. Each map consisted of 67 957 feature variables. Based on Fisher's criterion (cf. Methods), 5% of the most discriminative voxelwise features ( = connectivity values) for each training data set were entered into the SVM analyses. (B)Linear two-class SVM with a nested leave-one-subject-out cross-validation procedure. The outside loop was for model evaluation. Here, in a leave-one-subject-out cross-validation procedure, one subject was left out as a test sample and the rest of the sample constituted the training sample (N-1). The inner loop was for model optimization with a second leave-one-subject-out cross-validation procedure based on the training data. The model was optimized using different C-parameters of a linear decision function. The model with the highest accuracy and generalizability of the inner loop was used for classifying the left out, unknown subject of the outer loop. The posterior balanced accuracy was computed across all outer model evaluation loops. (C) The SVM weights of the linear separation function were reconstructed for all patients (cf. Methods). These weights show voxelwise connectivity values that contribute to SVM classification.

Figure 2.

SVM design. (A) Preparing feature variables for the SVM analysis. The left panel shows the seed region at the hand knob formation in primary motor cortex (M1) of the ipsilesional (affected) hemisphere, which was used to compute voxelwise correlation maps with every voxel within frontoparietal brain areas (cf. Methods). The right panel shows individual examples of resting-state functional connectivity maps. Each map consisted of 67 957 feature variables. Based on Fisher's criterion (cf. Methods), 5% of the most discriminative voxelwise features ( = connectivity values) for each training data set were entered into the SVM analyses. (B)Linear two-class SVM with a nested leave-one-subject-out cross-validation procedure. The outside loop was for model evaluation. Here, in a leave-one-subject-out cross-validation procedure, one subject was left out as a test sample and the rest of the sample constituted the training sample (N-1). The inner loop was for model optimization with a second leave-one-subject-out cross-validation procedure based on the training data. The model was optimized using different C-parameters of a linear decision function. The model with the highest accuracy and generalizability of the inner loop was used for classifying the left out, unknown subject of the outer loop. The posterior balanced accuracy was computed across all outer model evaluation loops. (C) The SVM weights of the linear separation function were reconstructed for all patients (cf. Methods). These weights show voxelwise connectivity values that contribute to SVM classification.

Confound Analysis

We carried out SVM control analyses with the same leave-one-subject-out cross-validation approach to test whether group classification may already be explained by a sampling bias (e.g., idiosyncrasies of the study groups). Age and the RMSE of the head motion parameters were included as potential confounds for all 3 groups into the analysis. For stroke patients, we also considered lesion volume and time after stroke onset as potential confounds. All parameters were linearly scaled to range between 0 and 1. We first conducted univariate SVM analyses for each confound, separately, followed by an SVM with all confounds but without neuroimaging data. We then combined the features of interest (i.e., resting-state connectivity and DWI lesion maps) with the different confound variables.

Mass-Univariate GLM Group Analyses

We were interested whether resting-state markers for individual motor impairment revealed by multivariate SVM feature weights were comparable with the results of a conventional group analysis computed within the framework of the GLM. To this end, we used the same connectivity maps as for the SVM analysis consisting of positive correlations between seed voxel time-courses in ipsilesional M1 and whole-brain target voxels, which were transformed into Fisher Z-scores and then masked by frontoparietal sensorimotor areas (Eickhoff et al. 2005). To warrant consistency between multivariate and univariate procedures, we computed separate independent samples t-tests with SPM to obtain T-contrast images for comparisons between controls and stroke patients (P < 0.05).

Mass-Univariate Decoding Analysis

To directly compare multivariate SVM classification with univariate decoding, we tested the classification accuracy of individual voxels for the discrimination of the 3 groups. We used the same approach as in the SVM training consisting of a leave-one-subject-out cross-validation procedure with a training sample and one independent test sample. For each voxel of the average SVM weight image, the mean of the average connectivity in each group of the training data set was used as a separation boundary to classify the respective left-out subject. Finally, the classification accuracy for each voxel was averaged across all cross-validation loops.

Expert Rating of DWI Lesions

Additionally, we performed a clinical rating with 9 certified consultant neurologists (Department of Neurology, University of Cologne, Germany; >5 years of clinical experience) using printouts of the axial DWI slices showing the entire extent of the lesion. The raters were asked to indicate whether the respective patient has a hand motor deficit or not. We then computed the posterior accuracy of these clinical ratings (Brodersen et al. 2010).

Comparison of Classification Accuracies

We compared classification accuracies between 1) resting-state connectivity data, 2) DWI lesion maps, and 3) clinical ratings of each of the 9 raters using McNemar's chi-squared test of the null hypothesis that classification is equal across these modalities (P < 0.05, one-tailed) (McNemar 1947).

Results

Sample

There was no significant difference in age and gender between groups (P ≥ 0.139) (Tables 1 and 2; see Supplementary Table 1 for statistical comparisons). Likewise, the two stroke groups did not differ in time after stroke onset or lesion volume (P ≥ 0.242). The stroke group with hand motor deficits showed a mean grip force index of 24.8% (SD: 21.7%) and a mean MI hand score of 15.9 (10.6). In contrast, stroke patients without hand motor deficits had an MI hand score of 33 and a mean grip force of 111% (15.1%). Hence, patients with motor deficits exhibited significantly worse motor performance than stroke patients with normal hand motor performance and nonstroke controls (P < 0.001). There was no difference in motor performance between stroke patients without hand motor deficit and controls (P > 0.5).

Resting-State Functional Connectivity

Univariate Group Analyses (GLM)

SPM-T-contrast images for the comparison of stroke patients with hand motor deficits and nonstroke controls showed that nonstroke controls featured greater interhemispheric resting-state connectivity between ipsilesional (left) M1 and contralesional sensorimotor areas including contralesional M1, primary sensory cortex (S1), superior parietal lobe (SPL) as well as contralesional dorsal and ventral premotor cortex (PMC) (P < 0.05; Fig. 3A, blue color scale). This means that patients with hand motor deficits showed reduced connectivity between these areas. In contrast, stroke patients with hand motor deficits as compared with nonstroke controls featured greater resting-state connectivity between ipsilesional M1 and supplementary motor area (SMA), as well as ventral and dorsal PMC, particularly in the ipsilesional hemisphere (Fig. 3A, red color scale). The SPM-T-contrasts for the comparison between stroke patients with and without hand motor impairment revealed similar results (Fig. 3B). Again, stroke patients with hand motor deficits were characterized by reduced interhemispheric resting-state connectivity between ipsilesional M1 and contralesional sensorimotor areas (Fig. 3B, blue color scale) but showed enhanced connectivity between ipsilesional M1 and ipsilesional as well as contralesional premotor areas including SMA and ventral PMC (Fig. 3B, red color scale).

Figure 3.

GLM results. (A) Group differences (independent samples t-test) between stroke patients with hand motor impairment and nonstroke control subjects. Blue: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “controls > patients.” Red: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “patients > controls.” (B) Group differences between stroke patients with and without hand motor impairment. Blue: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “patients without impairment > patients with hand motor impairment.” Red: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “patients with hand motor impairment > patients without hand motor impairment.” The green dot indicates the seed voxel coordinate in ipsilesional M1 (-38,-24,58). The color bar indicates the respective t-value (P < 0.05 uncorrected).

Figure 3.

GLM results. (A) Group differences (independent samples t-test) between stroke patients with hand motor impairment and nonstroke control subjects. Blue: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “controls > patients.” Red: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “patients > controls.” (B) Group differences between stroke patients with and without hand motor impairment. Blue: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “patients without impairment > patients with hand motor impairment.” Red: Areas of voxelwise resting-state connectivity for the SPM-T-contrast “patients with hand motor impairment > patients without hand motor impairment.” The green dot indicates the seed voxel coordinate in ipsilesional M1 (-38,-24,58). The color bar indicates the respective t-value (P < 0.05 uncorrected).

Multivariate SVM Classification: Stroke Patients with Motor Impairment versus Nonstroke Controls

The SVM classifying stroke patients with hand motor deficits and controls showed that resting-state connectivity of ipsilesional (left) M1 yielded a posterior classification accuracy of 82.6% (P < 0.001, CI = 66.9–90%) (Fig. 4A). We observed a similar accuracy when patients with right-hemispheric lesions who had been flipped for group comparisons were omitted (i.e., 79.6%, P < 0.001, CI = 62.6–88.3%). The sensitivity to correctly identify stroke patients with hand motor impairment was 80%. The specificity to identify nonstroke controls was 85%. The mean weight image across all training samples—which represents the regional contribution of voxelwise connectivity to classification—corresponded to the direction and spatial distribution of univariate group differences in the conventional GLM analysis (Fig. 3A). Accordingly, controls were characterized by stronger interhemispheric connectivity between ipsilesional M1 and contralesional sensorimotor areas including M1, S1, and SPL (Fig. 4A). In contrast, enhanced connectivity between ipsilesional M1 and ipsilesional ventral and dorsal PMC, SMA, and contralesional medial S1 contributed to the classification of stroke patients with hand motor impairment. In contrast, the SVM for classifying stroke patients without hand motor impairment and nonstroke controls showed a posterior classification accuracy of 58.5%, which was almost at chance level (P = 0.155, CI = 42.7–71.9%). Thus, based on M1 resting-state connectivity, stroke patients without motor impairment could not be distinguished from controls.

Figure 4.

Multivariate SVM classification. (A) Results in stroke patients with hand motor impairment and nonstroke controls. Resting-state functional connectivity of ipsilesional M1 provided 82.6% mean accuracy for the classification of unknown patients. The SVM weight image shows the regional contribution of voxelwise resting-state connectivity to the classification of stroke patients with hand motor deficits and nonstroke controls. (B) SVM results in stroke patients. Resting-state functional connectivity of ipsilesional M1 provided 87.6% mean accuracy for the classification of stroke patients into hand motor impairment groups. Blue: Areas of voxelwise resting-state connectivity, which contribute to the classification of nonstroke controls (Fig. 4A) or stroke patients without hand motor deficit (Fig. 4B) (i.e., stroke patients with hand motor impairment were characterized by reduced connectivity with these areas). Red: Areas of voxelwise resting-state connectivity, which contribute to the classification of stroke patients with hand motor impairment. The scaling of the weights reflects the relative contribution of voxelwise resting-state connectivity with ipsilesional M1 to group classification. The green dot indicates the seed voxel coordinate in ipsilesional M1 (-38,-24,58).

Figure 4.

Multivariate SVM classification. (A) Results in stroke patients with hand motor impairment and nonstroke controls. Resting-state functional connectivity of ipsilesional M1 provided 82.6% mean accuracy for the classification of unknown patients. The SVM weight image shows the regional contribution of voxelwise resting-state connectivity to the classification of stroke patients with hand motor deficits and nonstroke controls. (B) SVM results in stroke patients. Resting-state functional connectivity of ipsilesional M1 provided 87.6% mean accuracy for the classification of stroke patients into hand motor impairment groups. Blue: Areas of voxelwise resting-state connectivity, which contribute to the classification of nonstroke controls (Fig. 4A) or stroke patients without hand motor deficit (Fig. 4B) (i.e., stroke patients with hand motor impairment were characterized by reduced connectivity with these areas). Red: Areas of voxelwise resting-state connectivity, which contribute to the classification of stroke patients with hand motor impairment. The scaling of the weights reflects the relative contribution of voxelwise resting-state connectivity with ipsilesional M1 to group classification. The green dot indicates the seed voxel coordinate in ipsilesional M1 (-38,-24,58).

Multivariate SVM Classification of Motor Impairment in Stroke Patients

In the next step, we tested whether SVM can be used to differentiate between the 2 stroke groups. Here, resting-state connectivity of ipsilesional M1 correctly classified unseen stroke patients as either with or without hand motor deficit with a posterior accuracy of 87.6% (P < 0.001, CI = 72.4–93.2%) (sensitivity: 90%, specificity: 85%). A similar weight pattern as described earlier for classifying nonstroke controls versus patients with hand motor deficits separated the two groups (Fig. 4B). These weights were again similar to the respective univariate group comparison (Fig. 3B). Accordingly, patients with hand motor deficits were characterized by reduced interhemispheric connectivity of ipsilesional M1 with contralesional sensorimotor areas. In contrast, higher resting-state connectivity between ipsilesional M1 and contralesional ventral PMC and ipsilesional premotor areas (i.e., dorsal and ventral PMC, SMA) discriminated stroke patients with impaired hand function compared with stroke patients without hand motor impairment.

Mass-Univariate Group Classification

The results of the mass-univariate decoding analyses, using the mean of the average voxel connectivity in each group in each training sample as a boundary to classify the respective left-out subject in a cross-validation approach, showed that more than half of the voxels provided classification accuracies of >50% (Supplementary Fig. 2). However, for the discrimination of stroke patients with hand motor deficits and controls, only 0.2% of the voxels were equal to or better than the classification accuracy of the multivariate SVM classifier of 82.6%. Furthermore, the classification accuracy of 87.6% for separating stroke patients with or without hand motor deficits was only reached by one single voxel in the univariate analysis (i.e., 0.01%). This small proportion was clearly below a 5%-threshold of all voxels. Thus, multivariate decoding seems to outperform univariate decoding in classifying individual patients.

DWI Lesions

The SVM trained with the whole-brain DWI lesion maps considering white matter, cortex, and basal ganglia classified stroke patients with or without hand motor impairment with a posterior accuracy of 67.9% (P = 0.021, CI = 50.6–79%). When considering DWI lesions along the CST using probabilistic maps (Oishi et al. 2010), the posterior classification accuracy was 73.8% (P = 0.003, CI = 56.4–83.1%). However, the sensitivity to identify patients with motor deficits was low (i.e., 45–50%), whereas the specificity to classify patients without motor deficits was ≥90% (see Supplementary Fig. 3 for the SVM weights). Together, classifications based on DWI lesions tended to be lower than the classification accuracy of resting-state connectivity (chi2(1) = 3.27, P = 0.07).

We further tested how stroke-experienced consultant neurologists (n = 9; Department of Neurology, University of Cologne, Germany) classified motor function based on DWI. The posterior accuracy of the raters was 71% (P < 0.001, CI = 66–75.3%) (Supplementary Table 2). The SVM for M1 resting-state connectivity classified patients significantly more accurate than 6 out of 9 raters (P < 0.03). In contrast, one rater classified patients significantly better than the SVM for DWI lesions, whereas the classifications of the rest of the raters were equal to the SVM results for DWI lesions.

Confound Analysis

We carried out separate and joint SVM control analyses for potential confounds including age, RMSE of head motion parameters, lesion volume, and time after stroke onset. None of these confounds provided a significant group classification (maximum accuracy: 60.1%, P = 0.106). Furthermore, combined SVM analyses of confound variables and resting-state connectivity or DWI lesion maps of the CST did not exceed the classification accuracy of each of these feature maps alone (see Supplementary Table 3 for details).

Discussion

This study provides evidence that individual resting-state fMRI scans can be used to decode individual hand motor impairment based on multivariate SVM pattern classification. As expected, conventional univariate analyses of resting-state data revealed a similar pattern of differences between groups as identified by multivariate SVM weights for individual classification. However, only the multivariate SVM analysis approach allowed a significant classification of individual patients with respect to motor impairment. Importantly, motor performance was more accurately classified by M1 resting-state connectivity than by lesion location or by expert rating.

Markers of Motor Impairment after Stroke

SVM classifications are always driven by all features of a given weight vector (Naselaris et al. 2011; Haufe et al. 2014). Here, high weights may reflect both “true” neurophysiological signal and noise. That is, to maximize separation, an optimal classifier may require nonzero weights on features containing the signal of interest as well as on features containing additional noise. In this way, the classifier accounts for the entire structure of the data and provides a more accurate classification. Accordingly, the decoding approach usually does not allow drawing conclusions such that one connection reflects a marker of a clinical parameter. In contrast, the parameters of a forward model, for example a GLM analysis, are physiologically interpretable (Haufe et al. 2014). Therefore, when combined with a forward model approach, SVM weights may be interpreted in terms of their neurophysiological meaning. Figures 3 and 4 show that positive and negative SVM weights spatially correspond to the direction of group differences between patients and controls in the GLM analysis. This correspondence between the backward (SVM) and forward (GLM) model therefore suggests that resting-state connectivity between ipsilesional M1 and these areas significantly contributes to the characterization of impaired hand motor function after stroke.

Accordingly, the SVM classified stroke patients with hand motor deficits based on reduced interhemispheric resting-state connectivity between ipsilesional and contralesional motor areas. This finding was reproduced by contrasting patients with hand motor deficits with independent samples of nonstroke controls and stroke patients with normal hand motor performance (Fig. 4). Our results thus corroborate findings of disturbed interhemispheric connectivity between sensorimotor areas in groups of stroke patients with motor deficits as reported in previous fMRI group studies (Carter et al. 2010, 2012; Wang et al. 2010; Park et al. 2011; Golestani et al. 2013).

Changes in interhemispheric resting-state connectivity may play a key role in early cortical reorganization (Carmichael 2006). Intracranial recordings in humans and animals showed that resting-state connectivity reflects spontaneous oscillations of synchronous neuronal activity (He et al. 2008; Nir et al. 2008; van Meer et al. 2012). There seems to be a close relationship between alterations in synchronous activity remote from the lesion and induction of axonal sprouting in the first few days after stroke (Carmichael and Chesselet 2002). Thus, reduced interhemispheric resting-state connectivity at the acute stage may represent an early imaging marker of neural repair processes after stroke. This conclusion is supported by human fMRI studies probing the motor execution system after stroke by means of motor activation tasks and connectivity analyses (Rehme et al. 2012; Grefkes and Fink 2014). These studies frequently reported enhanced neural activity in contralesional M1 during movements of the stroke-affected hand. In the first few days after stroke, contralesional M1 has been shown to exert a positive influence onto ipsilesional M1 activity, which may support early motor recovery (Rehme, Eickhoff, Wang et al. 2011). Over the ensuing weeks and months, this positive influence disappears or may even turn into maladaptive inhibition in some patients (Grefkes et al. 2008; Rehme, Eickhoff, Wang et al. 2011). Although task-based and resting-state fMRI reflect distinct underlying neural processes (Rehme et al. 2013), findings from fMRI motor activation studies emphasize the role of contralesional M1 as a brain region involved into early motor reorganization that corresponds to the present resting-state findings.

A new finding of the present study is that patients with hand motor deficits were also characterized by enhanced resting-state connectivity between ipsilesional M1 and secondary motor areas in frontoparietal cortex including PMC, SMA, and S1, particularly in the ipsilesional hemisphere. The SVM results show that this increase in connectivity is essential for the individual classification of motor impairment. Findings from previous task-based fMRI studies suggest an important role of premotor areas in post-stroke recovery (Johansen-Berg et al. 2002; Rehme, Fink et al. 2011). These studies demonstrated that an increase of activity in premotor areas during movements of the stroke-affected hand predicts greater motor recovery after stroke. Furthermore, effective connectivity data revealed that positive influences between premotor areas, particularly SMA, and ipsilesional M1 correlate with motor recovery and recovered motor function (Grefkes et al. 2008; Wang et al. 2011; Rehme, Eickhoff, Wang et al. 2011). Likewise, Park and colleagues showed that higher resting-state connectivity between ipsilesional M1 and contralesional SMA at the acute stage predicts motor performance at the chronic stage after stroke (Park et al. 2011). Thus, as premotor neurons are engaged in motor preparation and higher-order movement control (Dum and Strick 2002), increased connectivity between M1 and premotor areas may reflect a neural disposition facilitating early cortical reorganization. Indeed, in chronic stroke patients, transcranial magnetic stimulation studies have shown that disruption of both ipsilesional and contralesional dorsal PMC activity can deteriorate recovered motor function as assessed with reaction time or sequential finger movement tasks (Fridman et al. 2004; Lotze et al. 2006). In the present study, we used a rather simple motor assessment that could be easily performed at the bedside, even with patients suffering from severe hemiplegia. It remains to be elucidated whether similar weight distributions would be observed when using more complex motor tasks to quantify the hand motor deficit.

The Diagnostic Use of SVM and fMRI

Ipsilesional M1 resting-state connectivity provided a high sensitivity of ≥80% to identify unknown patients with hand motor deficits when comparing them with stroke patients with normal hand motor performance or nonstroke controls. In contrast, the multivariate SVM classifier did not discriminate stroke patients with normal motor performance and controls. Our data, therefore, suggest that resting-state motor networks are not altered after stroke per se but rather allow specific inference about motor impairment. The direct comparison with mass-univariate decoding analyses showed that most of the individual voxels had some power to discriminate previously unseen patients. However, the accuracy of multivariate SVM classifiers was only exceeded by 0.2% of the voxels. Our findings, therefore, suggest that the combined information from multiple voxels is more powerful than single voxels, which may rather reflect random results. Thus, this finding strengthens the importance of carrying out multivariate rather than univariate decoding analyses. Multivariate machine learning techniques have been successfully applied to resting-state fMRI networks in other clinical populations (Shen et al. 2010; Richiardi et al. 2012; Wee et al. 2012; Zeng et al. 2012). These studies reported accuracies of 70–96%, which are consistent with the ones we observed in our sample. Thus, although resting-state fMRI is measured in a task-free setting, it allows classifying clinical populations at the single subject level.

With respect to neuroimaging markers of functional impairment after stroke, Saur and colleagues showed that SVM can predict good versus poor recovery from aphasia with 86% accuracy based on early fMRI activation during a language task in combination with age and the initial language test score (Saur et al. 2010). In contrast to this study, we used resting-state fMRI, which is particularly useful in clinical routine assessment of acute, severely disabled patients because it yields robust signals after only 5–6 min and requires no task-related behavior (Friston 1994; Carter et al. 2010; Van Dijk et al. 2010; Wang et al. 2010).

Furthermore, our findings are well in line with previous conventional univariate resting-state group studies in stroke patients, which demonstrated that disturbances in connectivity are specific to the functional system that has been damaged. For example, stroke patients with motor deficits, aphasia, or visual neglect show disturbed resting-state connectivity in the corresponding motor, language, or attention resting-state networks (He et al. 2007; Warren et al. 2009; Carter et al. 2010). Importantly, our findings extend the results of previous univariate encoding analyses performed at the group level by showing that multivariate decoding approaches for resting-state fMRI data enable classification at the level of an individual patient. Given that a single resting-state fMRI scan allows simultaneous assessment of different resting-state networks, SVMs for resting-state data might be suited to classify patients with aphasia or attention deficits (He et al. 2007; Warren et al. 2009; Carter et al. 2010) but also to ultimately classify the global neurological deficit and potential for recovery.

Acute Ischemic Lesion

The SVM accuracy for the classification of hand motor deficits based on structural lesions was 73.8%. However, the SVM provided a poor sensitivity of 50% to identify patients with hand motor deficits. Likewise, we observed a similar degree of classification accuracy in the ratings of experienced consultant neurologists, which suggests that expert knowledge of functional brain anatomy is not sufficient to ensure correct classifications.

Hence, stroke lesions provide a less reliable classification of motor deficits at the level of individual patients compared with resting-state connectivity. At first, this finding seems to be at odds with previous diffusion tensor imaging (DTI) studies showing that greater motor impairment was associated with greater disruptions of CST fibers, particularly of those extending from ipsilesional M1 (Stinear et al. 2007; Lindenberg et al. 2012; Wang et al. 2012). However, in contrast to these correlation analyses, individual predictions made by SVM classification are based on a multivariate approach. Here, classifier rules are learned based on a valid training step. Thus, the low sensitivity (50%) revealed by the SVM analysis most likely results from the heterogeneity of lesions that may cause motor deficits in our sample (Fig. 1; Supplementary Fig. 1). This heterogeneity particularly constrains the SVM performance, which depends on the consistency of lesion locations within one group. In contrast to our findings, previous clinical machine learning studies using structural MRI data yielded higher accuracies of 83–96%, particularly for classifying patients with dementia compared with healthy subjects (Klöppel et al. 2008; Gerardin et al. 2009). These differences in classification accuracies may result from different structural data (i.e., DTI, high-resolution T1-weighted images). However, in stroke, even small lesions may cause severe deficits, a phenomenon that contrasts with findings in neurodegenerative disease, which usually affects large parts of the brain (Orrù et al. 2012). Thus, functional neuroimaging markers seem to be superior when classifying functional deficits resulting from confined and heterogeneous lesions, as often encountered after stroke.

Conclusion

We here show that the combination of resting-state fMRI and multivariate machine learning tools constitutes a sensitive technique to infer neurological impairment at the level of single patients. All data were acquired during a routine imaging session, which highlights the clinical practicability of our approach with acute, severely disabled patients (Friston 1994; Carter et al. 2010; Van Dijk et al. 2010; Wang et al. 2010). It should be tested in the future whether this approach offers potential for providing noninvasive diagnostic markers to predict the neurological outcome and to optimize rehabilitation after stroke.

Supplementary Material

Supplementary material can be found at: http://www.cercor.oxfordjournals.org/

Funding

C.G. is supported by the German Research Foundation (DFG GR 3285/2-1; GR 3285/4-1). G.R.F. gratefully acknowledges support by the Marga and Walter Boll Stiftung.

Notes

We thank the MR staff from the Department of Neuroradiology, University of Cologne, for their support in assessing resting-state fMRI scans in acute stroke patients during clinical routine. We further thank Martha Kutscha for her help in data management. Conflict of Interest: None declared.

References

Ashburner
J
Friston
KJ
2005
.
Unified segmentation
.
Neuroimage
 .
26
:
839
851
.
Behzadi
Y
Restom
K
Liau
J
Liu
TT
2007
.
A component based noise correction method (CompCor) for BOLD and perfusion based fMRI
.
Neuroimage
 .
37
:
90
101
.
Biswal
B
Yetkin
FZ
Haughton
VM
Hyde
JS
1995
.
Functional connectivity in the motor cortex of resting human brain using echo-planar MRI
.
Magn Reson Med
 .
34
:
537
541
.
Brodersen
KH
Ong
CS
Stephan
KE
Buhmann
JM
2010
.
The balanced accuracy and its posterior distribution
.
Proceedings of the 20th International Conference on Pattern Recognition
 .
3121
3124
.
Carmichael
ST
2006
.
Cellular and molecular mechanisms of neural repair after stroke: making waves
.
Ann Neurol
 .
59
:
735
742
.
Carmichael
ST
Chesselet
MF
2002
.
Synchronous neuronal activity is a signal for axonal sprouting after cortical lesions in the adult
.
J Neurosci
 .
22
:
6062
6070
.
Carter
AR
Astafiev
SV
Lang
CE
Connor
LT
Rengachary
J
Strube
MJ
Pope
DL
Shulman
GL
Corbetta
M
2010
.
Resting interhemispheric functional magnetic resonance imaging connectivity predicts performance after stroke
.
Ann Neurol
 .
67
:
365
375
.
Carter
AR
Patel
KR
Astafiev
SV
Snyder
AZ
Rengachary
J
Strube
MJ
Pope
A
Shimony
JS
Lang
CE
Shulman
GL
et al
2012
.
Upstream dysfunction of somatomotor functional connectivity after corticospinal damage in stroke
.
Neurorehabil Neural Repair
 .
26
:
7
19
.
Chai
XJ
Castanon
AN
Ongur
D
Whitfield-Gabrieli
S
2012
.
Anticorrelations in resting state networks without global signal regression
.
Neuroimage
 .
59
:
1420
1428
.
Chang
CC
Lin
CJ
2011
.
LIBSVM: A library for support vector machines
.
ACM Trans Intell Syst Technol
 .
27
:
1
39
.
Chollet
F
DiPiero
V
Wise
RJ
Brooks
DJ
Dolan
RJ
Frackowiak
RS
1991
.
The functional anatomy of motor recovery after stroke in humans: a study with positron emission tomography
.
Ann Neurol
 .
29
:
63
71
.
Cortes
C
Vapnik
V
1995
.
Support-vector networks
.
Machine Learn
 .
20
:
273
297
.
Crosby
CA
Wehbe
MA
Mawr
B
1994
.
Hand strength: normative values
.
J Hand Surg Am
 .
19
:
665
670
.
Demeurisse
G
Demol
O
Robaye
E
1980
.
Motor evaluation in vascular hemiplegia
.
Eur Neurol
 .
19
:
382
389
.
Dosenbach
NU
Nardos
B
Cohen
AL
Fair
DA
Power
JD
Church
JA
Nelson
SM
Wig
GS
Vogel
AC
Lessov-Schlaggar
CN
et al
2010
.
Prediction of individual brain maturity using fMRI
.
Science
 .
329
:
1358
1361
.
Dum
RP
Strick
PL
2002
.
Motor areas in the frontal lobe of the primate
.
Physiol Behav
 .
77
:
677
682
.
Eickhoff
SB
Stephan
KE
Mohlberg
H
Grefkes
C
Fink
GR
Amunts
K
Zilles
K
2005
.
A new SPM toolbox for combining probabilistic cytoarchitectonic maps and functional imaging data
.
Neuroimage
 .
25
:
1325
1335
.
Feis
DL
Brodersen
KH
von Cramon
DY
Luders
E
Tittgemeyer
M
2013
.
Decoding gender dimorphism of the human brain using multimodal anatomical and diffusion MRI data
.
Neuroimage
 .
70
:
250
257
.
Fridman
EA
Hanakawa
T
Chung
M
Hummel
F
Leiguarda
RC
Cohen
LG
2004
.
Reorganization of the human ipsilesional premotor cortex after stroke
.
Brain
 .
127
:
747
758
.
Friston
KJ
1994
.
Functional and effective connectivity in neuroimaging: a synthesis
.
Hum Brain Mapp
 .
2
:
56
78
.
Friston
KJ
Holmes
AP
Poline
JB
Grasby
PJ
Williams
SC
Frackowiak
RS
Turner
R
1995
.
Analysis of fMRI time-series revisited
.
Neuroimage
 .
2
:
45
53
.
Gerardin
E
Chetelat
G
Chupin
M
Cuingnet
R
Desgranges
B
Kim
HS
Niethammer
M
Dubois
B
Lehericy
S
Garnero
L
et al
2009
.
Multidimensional classification of hippocampal shape features discriminates Alzheimer's disease and mild cognitive impairment from normal aging
.
Neuroimage
 .
47
:
1476
1486
.
Go
AS
Mozaffarian
D
Roger
VL
Benjamin
EJ
Berry
JD
Borden
WB
Bravata
DM
Dai
S
Ford
ES
Fox
CS
et al
2013
.
Heart disease and stroke statistics--2013 update: a report from the American Heart Association
.
Circulation
 .
127
:
e6
e245
.
Golestani
AM
Tymchuk
S
Demchuk
A
Goodyear
BG
2013
.
Longitudinal evaluation of resting-state FMRI after acute stroke with hemiparesis
.
Neurorehabil Neural Repair
 .
27
:
153
163
.
Grefkes
C
Fink
GR
2014
.
Connectivity-based approaches in stroke and recovery of function
.
Lancet Neurol
 .
13
:
206
216
.
Grefkes
C
Nowak
DA
Eickhoff
SB
Dafotakis
M
Kust
J
Karbe
H
Fink
GR
2008
.
Cortical connectivity after subcortical stroke assessed with functional magnetic resonance imaging
.
Ann Neurol
 .
63
:
236
246
.
Haufe
S
Meinecke
F
Gorgen
K
Dahne
S
Haynes
JD
Blankertz
B
Biessmann
F
2014
.
On the interpretation of weight vectors of linear models in multivariate neuroimaging
.
Neuroimage
 .
87
:
96
110
.
He
BJ
Snyder
AZ
Vincent
JL
Epstein
A
Shulman
GL
Corbetta
M
2007
.
Breakdown of functional connectivity in frontoparietal networks underlies behavioral deficits in spatial neglect
.
Neuron
 .
53
:
905
918
.
He
BJ
Snyder
AZ
Zempel
JM
Smyth
MD
Raichle
ME
2008
.
Electrophysiological correlates of the brain's intrinsic large-scale functional architecture
.
Proc Natl Acad Sci USA
 .
105
:
16039
16044
.
Johansen-Berg
H
Dawes
H
Guy
C
Smith
SM
Wade
DT
Matthews
PM
2002
.
Correlation between motor improvements and altered fMRI activity after rehabilitative therapy
.
Brain
 .
125
:
2731
2742
.
Kay
KN
Gallant
JL
2009
.
I can see what you see
.
Nat Neurosci
 .
12
:
245
.
Kay
KN
Naselaris
T
Prenger
RJ
Gallant
JL
2008
.
Identifying natural images from human brain activity
.
Nature
 .
452
:
352
355
.
Klöppel
S
Stonnington
CM
Chu
C
Draganski
B
Scahill
RI
Rohrer
JD
Fox
NC
Jack
CR
Jr
Ashburner
J
Frackowiak
RS
2008
.
Automatic classification of MR scans in Alzheimer's disease
.
Brain
 .
131
:
681
689
.
Kriegeskorte
N
2011
.
Pattern-information analysis: from stimulus decoding to computational-model testing
.
Neuroimage
 .
56
:
411
421
.
Kriegeskorte
N
Simmons
WK
Bellgowan
PS
Baker
CI
2009
.
Circular analysis in systems neuroscience: the dangers of double dipping
.
Nat Neurosci
 .
12
:
535
540
.
Lemm
S
Blankertz
B
Dickhaus
T
Muller
KR
2011
.
Introduction to machine learning for brain imaging
.
Neuroimage
 .
56
:
387
399
.
Lindenberg
R
Zhu
LL
Ruber
T
Schlaug
G
2012
.
Predicting functional motor potential in chronic stroke patients using diffusion tensor imaging
.
Hum Brain Mapp
 .
33
:
1040
1051
.
Lotze
M
Markert
J
Sauseng
P
Hoppe
J
Plewnia
C
Gerloff
C
2006
.
The role of multiple contralesional motor areas for complex hand movements after internal capsular lesion
.
J Neurosci
 .
26
:
6096
6102
.
McNemar
Q
1947
.
Note on the sampling error of the difference between correlated proportions or percentages
.
Psychometrika
 .
12
:
153
157
.
Müller
K-R
Krauledat
M
Dornhege
G
Curio
G
Blankertz
B
2004
.
Machine learning techniques for brain-computer interfaces
.
Biomed Eng
 .
49
:
11
22
.
Naselaris
T
Kay
KN
Nishimoto
S
Gallant
JL
2011
.
Encoding and decoding in fMRI
.
Neuroimage
 .
56
:
400
410
.
Nir
Y
Mukamel
R
Dinstein
I
Privman
E
Harel
M
Fisch
L
Gelbard-Sagiv
H
Kipervasser
S
Andelman
F
Neufeld
MY
et al
2008
.
Interhemispheric correlations of slow spontaneous neuronal fluctuations revealed in human sensory cortex
.
Nat Neurosci
 .
11
:
1100
1108
.
Oishi
K
Faria
AV
van Zijl
PCM
Mori
S
2010
.
MRI atlas of Human White Matter
 .
Amsterdam
:
Elsevier
.
Orrù
G
Pettersson-Yeo
W
Marquand
AF
Sartori
G
Mechelli
A
2012
.
Using support vector machine to identify imaging biomarkers of neurological and psychiatric disease: a critical review
.
Neurosci Biobehav Rev
 .
36
:
1140
1152
.
Park
CH
Chang
WH
Ohn
SH
Kim
ST
Bang
OY
Pascual-Leone
A
Kim
YH
2011
.
Longitudinal changes of resting-state functional connectivity during motor recovery after stroke
.
Stroke
 .
42
:
1357
1362
.
Pereira
F
Mitchell
T
Botvinick
M
2009
.
Machine learning classifiers and fMRI: a tutorial overview
.
Neuroimage
 .
45
:
S199
S209
.
Power
JD
Barnes
KA
Snyder
AZ
Schlaggar
BL
Petersen
SE
2012
.
Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion
.
Neuroimage
 .
59
:
2142
2154
.
Rehme
AK
Eickhoff
SB
Grefkes
C
2013
.
State-dependent differences between functional and effective connectivity of the human cortical motor system
.
Neuroimage
 .
67
:
237
246
.
Rehme
AK
Eickhoff
SB
Rottschy
C
Fink
GR
Grefkes
C
2012
.
Activation likelihood estimation meta-analysis of motor-related neural activity after stroke
.
Neuroimage
 .
59
:
2771
2782
.
Rehme
AK
Eickhoff
SB
Wang
LE
Fink
GR
Grefkes
C
2011
.
Dynamic causal modeling of cortical activity from the acute to the chronic stage after stroke
.
Neuroimage
 .
55
:
1147
1158
.
Rehme
AK
Fink
GR
von Cramon
DY
Grefkes
C
2011
.
The role of the contralesional motor cortex for motor recovery in the early days after stroke assessed with longitudinal FMRI
.
Cereb Cortex
 .
21
:
756
768
.
Rehme
AK
Grefkes
C
2013
.
Cerebral network disorders after stroke: evidence from imaging-based connectivity analyses of active and resting brain states in humans
.
J Physiol
 .
591
:
17
31
.
Richiardi
J
Gschwind
M
Simioni
S
Annoni
JM
Greco
B
Hagmann
P
Schluep
M
Vuilleumier
P
Van
DV
2012
.
Classifying minimally disabled multiple sclerosis patients from resting state functional connectivity
.
Neuroimage
 .
62
:
2021
2033
.
Satterthwaite
TD
Elliott
MA
Gerraty
RT
Ruparel
K
Loughead
J
Calkins
ME
Eickhoff
SB
Hakonarson
H
Gur
RC
Gur
RE
et al
2013
.
An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data
.
Neuroimage
 .
64
:
240
256
.
Saur
D
Ronneberger
O
Kummerer
D
Mader
I
Weiller
C
Kloppel
S
2010
.
Early functional magnetic resonance imaging activations predict language outcome after stroke
.
Brain
 .
133
:
1252
1264
.
Shen
H
Wang
L
Liu
Y
Hu
D
2010
.
Discriminative analysis of resting-state functional connectivity patterns of schizophrenia using low dimensional embedding of fMRI
.
Neuroimage
 .
49
:
3110
3121
.
Stinear
CM
Barber
PA
Smale
PR
Coxon
JP
Fleming
MK
Byblow
WD
2007
.
Functional potential in chronic stroke patients depends on corticospinal tract integrity
.
Brain
 .
130
:
170
180
.
Van Dijk
KR
Hedden
T
Venkataraman
A
Evans
KC
Lazar
SW
Buckner
RL
2010
.
Intrinsic functional connectivity as a tool for human connectomics: theory, properties, and optimization
.
J Neurophysiol
 .
103
:
297
321
.
van Meer
MP
Otte
WM
van der Marel
K
Nijboer
CH
Kavelaars
A
van der Sprenkel
JW
Viergever
MA
Dijkhuizen
RM
2012
.
Extent of bilateral neuronal network reorganization and functional recovery in relation to stroke severity
.
J Neurosci
 .
32
:
4495
4507
.
Wang
L
Yu
C
Chen
H
Qin
W
He
Y
Fan
F
Zhang
Y
Wang
M
Li
K
Zang
Y
et al
2010
.
Dynamic functional reorganization of the motor execution network after stroke
.
Brain
 .
133
:
1224
1238
.
Wang
LE
Fink
GR
Diekhoff
S
Rehme
AK
Eickhoff
SB
Grefkes
C
2011
.
Noradrenergic enhancement improves motor network connectivity in stroke patients
.
Ann Neurol
 .
69
:
375
388
.
Wang
LE
Tittgemeyer
M
Imperati
D
Diekhoff
S
Ameli
M
Fink
GR
Grefkes
C
2012
.
Degeneration of corpus callosum and recovery of motor function after stroke: a multimodal magnetic resonance imaging study
.
Hum Brain Mapp
 .
33
:
2941
2956
.
Ward
NS
Brown
MM
Thompson
AJ
Frackowiak
RS
2003
.
Neural correlates of motor recovery after stroke: a longitudinal fMRI study
.
Brain
 .
126
:
2476
2496
.
Ward
NS
Newton
JM
Swayne
OB
Lee
L
Thompson
AJ
Greenwood
RJ
Rothwell
JC
Frackowiak
RS
2006
.
Motor system activation after subcortical stroke depends on corticospinal system integrity
.
Brain
 .
129
:
809
819
.
Warren
JE
Crinion
JT
Lambon Ralph
MA
Wise
RJ
2009
.
Anterior temporal lobe connectivity correlates with functional outcome after aphasic stroke
.
Brain
 .
132
:
3428
3442
.
Wee
CY
Yap
PT
Zhang
D
Denny
K
Browndyke
JN
Potter
GG
Welsh-Bohmer
KA
Wang
L
Shen
D
2012
.
Identification of MCI individuals using structural and functional connectivity networks
.
Neuroimage
 .
59
:
2045
2056
.
World Health Organization
.
2004
.
The Atlas of Heart Disease and Stroke
 .
Zeng
LL
Shen
H
Liu
L
Wang
L
Li
B
Fang
P
Zhou
Z
Li
Y
Hu
D
2012
.
Identifying major depression using whole-brain functional connectivity: a multivariate pattern analysis
.
Brain
 .
135
:
1498
1507
.