Brain activity underlying tool-related and imitative skills after major left hemisphere stroke

Abstract

Apraxia is a debilitating cognitive motor disorder that frequently occurs after left hemisphere stroke and affects tool-associated and imitative skills. However, the severity of the apraxic deficits varies even across patients with similar lesions. This variability raises the question whether regions outside the left hemisphere network typically associated with cognitive motor tasks in healthy subjects are of additional functional relevance. To investigate this hypothesis, we explored regions where functional magnetic resonance imaging activity is associated with better cognitive motor performance in patients with left hemisphere ischaemic stroke. Thirty-six patients with chronic (>6 months) large left hemisphere infarcts (age ± standard deviation, 60 ± 12 years, 29 male) and 29 control subjects (age ± standard deviation, 72 ± 7, 15 male) were first assessed behaviourally outside the scanner with tests for actual tool use, pantomime and imitation of tool-use gestures, as well as for meaningless gesture imitation. Second, functional magnetic resonance imaging activity was registered during the passive observation of videos showing tool-associated actions. Voxel-wise linear regression analyses were used to identify areas where behavioural performance was correlated with functional magnetic resonance imaging activity. Furthermore, lesions were delineated on the magnetic resonance imaging scans for voxel-based lesion–symptom mapping. The analyses revealed two sets of regions where functional magnetic resonance imaging activity was associated with better performance in the clinical tasks. First, activity in left hemisphere areas thought to mediate cognitive motor functions in healthy individuals (i.e. activity within the putative ‘healthy’ network) was correlated with better scores. Within this network, tool-associated tasks were mainly related to activity in supramarginal gyrus and ventral premotor cortex, while meaningless gesture imitation depended more on the anterior intraparietal sulcus and superior parietal lobule. Second, repeating the regression analyses with total left hemisphere lesion volume as additional covariate demonstrated that tool-related skills were further supported by right premotor, right inferior frontal and left anterior temporal areas, while meaningless gesture imitation was also driven by the left dorso-lateral prefrontal cortex. In summary, tool-related and imitative skills in left hemisphere stroke patients depend on the activation of spared left hemisphere regions that support these abilities in healthy individuals. In addition, cognitive motor functions rely on the activation of ipsi- and contralesional areas that are situated outside this ‘healthy’ network. This activity may explain why some patients perform surprisingly well despite large left brain lesions, while others are severely impaired.

The severity of apraxia varies between patients with similar lesions. Using functional MRI in 36 chronic left-hemisphere stroke patients, Martin et al . reveal that tool-related and imitative skills depend not only on activation of spared left-hemisphere regions associated with cognitive motor functions, but also on activity within other mainly contralesional areas.

Open in new tabDownload slide

Introduction

Apraxia, a cognitive motor disorder, is characterized by the inability to perform skilled motor acts. Despite intact basic sensorimotor functions, apraxic patients may be unable to correctly pantomime or imitate tool-use gestures, use actual tools, or to imitate meaningless gestures ( Heilman and Valenstein, 2003 ; Goldenberg, 2013 ). Apraxia results mainly from left hemisphere damage, often due to ischaemic or haemorrhagic stroke ( Leiguarda and Marsden, 2000 ; Heilman and Valenstein, 2003; Goldenberg, 2009 ). However, only roughly one-third of patients with substantial left hemisphere damage develop apraxia ( Donkervoort et al. , 2000 ; Hoeren et al. , 2014 ; Martin et al. , 2015 ).

Lesion studies have attempted to account for this variability by identifying left hemisphere regions where damage is particularly likely to be associated with apraxic deficits ( Tessari et al. , 2007 ; Goldenberg and Spatt, 2009 ; Manuel et al. , 2013 ; Mengotti et al. , 2013 ; Buxbaum et al. , 2014 ; Hoeren et al. , 2014 ; Goldenberg and Randerath, 2015 ; Martin et al. , 2015 ). Impairments affecting tool-associated skills were linked to the posterior middle and superior temporal gyrus, the inferior parietal lobule (IPL) and the ventral premotor cortex (PMC) ( Goldenberg et al. , 2007 ; Tessari et al. , 2007 ; Buxbaum et al. , 2014 ; Hoeren et al. , 2014 ). By comparison, imitation of meaningless gestures may be more susceptible to superior parietal lobe (SPL) lesions ( Buxbaum, 2001 ; Hoeren et al. , 2014 ; but see Goldenberg, 2009 , 2013 ; Niessen et al. , 2014 for more comprehensive reviews including other areas that may be involved in tool-associated and imitative tasks). Overall, the areas identified in the lesion studies largely overlap with the foci of activity registered during functional imaging experiments in healthy subjects performing cognitive motor tasks ( Hermsdörfer et al. , 2001 ; Heiser et al. , 2003 ; Chaminade et al. , 2005 ; Johnson-Frey et al. , 2005 ; Mühlau et al. , 2005 ; Fridman et al. , 2006 ; Bohlhalter et al. , 2009 ; Vingerhoets et al. , 2012 ; Yoon et al. , 2012 ; Brandi et al. , 2014 ; Vry et al. , 2015 ). Therefore, the results from lesion studies suggest that cognitive motor functions in stroke patients decline as a function of damage to the ‘healthy’ network (i.e. the regions that normally support the tasks in which apraxic patients fail).

Integrating the results of the lesion studies outlined above with insights from functional MRI investigations in healthy subjects, recent frameworks proposed that the ‘healthy’ left hemisphere network consists of three parallel streams between parietal and frontal cortices. Grossly, these streams may be differentiated into dorsal (either dorso-dorsal or ventro-dorsal), when the connecting long association tracts run above the sylvian fissure [superior longitudinal fascicle (SLF) II and III], and in ventral, when the tracts are below the sylvian fissure (extreme capsule/inferior fronto-occipital fascicle, uncinate fasciculus) ( Catani et al. , 2002 ; Makris and Pandya, 2009 ; Makris et al. , 2009 ; Vry et al. , 2012 , 2015 ; Hamzei et al. , 2015 ). Involving the SPL and the dorsal PMC, the dorso-dorsal stream is thought to be particularly relevant for online movement control, and, possibly, meaningless gesture imitation ( Buxbaum, 2001 ; Brandi et al. , 2014 ; Hoeren et al. , 2014 ). Conversely, the ventro-dorsal stream, which traverses through the IPL towards the ventral PMC, may support learned skilled actions ( Heilman et al. , 1982 ; Buxbaum et al. , 2007 ; Niessen et al. , 2014 ). Lastly, the ‘ventral stream’, as defined in this context, includes inferior, middle and superior temporal cortices as well as the anterior inferior frontal gyrus (IFG) ( Weiller et al. , 2011 , 2015 a ; Rijntjes et al. , 2012 ; Hamzei et al. , 2015 ; Vry et al. , 2015 ). The ventral stream has been linked to conceptual aspects of tool-use such as tool–recipient associations ( Ochipa et al. , 1992 ; Hodges et al. , 1999 ; Martin et al. , 2015 ).

However, the associations between apraxic deficits and the left hemisphere lesion locations described above are significant only at the group level as the severity of apraxia may vary considerably even between patients with similar lesions. For example, in clinical experience, some patients with left parietotemporal lesions (i.e. extensive damage to the ‘healthy’ network) are severely apraxic while others have only mild impairments. This variability suggests that residual cognitive motor functions may be supported by additional regions outside the ‘healthy’ network at least in patients with favourable outcomes. In the absence of specific data on apraxia, indirect support for this conjecture comes from investigations on post-stroke language deficits where activity, for example, in contralesional homologues ( Weiller et al. , 1995 ; Musso et al. , 1999 ; Leff et al. , 2002 ; Blank et al. , 2003 ; Saur et al. , 2006 ), or in areas related to domain-general cognitive control mechanisms ( Brownsett et al. , 2014 ) was demonstrated to positively affect performance.

To investigate this hypothesis, we explored where functional MRI activity is correlated with off-line performance in different cognitive motor tasks. Thirty-six patients with chronic left hemisphere ischaemic stroke ( Fig. 1 ) were first examined outside the scanner using established tests for imitation of meaningless hand and finger postures, pantomime and imitation of tool-use gestures, as well as for actual tool use. Secondly, functional MRI activity was registered during a paradigm involving the passive observation of audio-visually presented tool-associated actions ( Fig. 2 ).

Our approach was based on using the stimulus-related functional MRI activity as a measure of the different regions’ functional state with respect to action cognition. We designed our functional MRI experiment based on action observation because action observation and action execution activate highly congruent networks and are thought to rely on similar cognitive processes ( Lewis, 2006 ; Caspers et al. , 2010 ; Ferri et al. , 2015 a , b ); however, functional MRI activity elicited by action observation may be less susceptible to confounding influences. For example, experiments involving active motor performance in the scanner may require participants to execute precise movements in response to cues in a well-timed manner ( Cavina-Pratesi et al. , 2010 ; Gallivan et al. , 2013 , 2016 ; Brandi et al. , 2014 ; Hamzei et al. , 2015 ) that may be challenging for patients with large lesions ( Fig. 1 ), even when the left (ipsilesional) hand can be used. Patients may thus show a reduced in-scanner performance, both in terms of movement features (e.g. abnormal trajectory or speed) and in terms of general compliance with the task (e.g. longer reaction times, mixing up conditions). Moreover, patients with neuropsychological deficits may rely more on domain-general functions for executive control and attention to cope with task demands ( Geranmayeh et al. , 2014 ). Consequently, in functional MRI experiments involving active movement execution, activity changes may simply reflect that more impaired patients have greater difficulties following the task. By contrast, the functional MRI activity elicited by the (passive) observation of tool-action videos ( Fig. 2 ) was not dependent on the subjects’ active participation and, therefore, could be interpreted without further need to adjust for in-scanner performance, compliance with the functional MRI paradigm or cognitive strategy.

The experimental approach of relating measures of active performance obtained outside the scanner to functional MRI activity elicited by analogous perceptive tasks was previously used to elucidate adaptive activity changes in stroke patients with aphasia or hand paresis ( Saur et al. , 2006 ; Garrison et al. , 2013 ). In the former study, a composite score of language abilities, which included several subscores for active speech production, was paralleled with functional MRI activity elicited by listening to speech ( Saur et al. , 2006 ); in the latter investigation, the degree of motor impairment was aligned with activation during the observation of simple object-directed actions ( Garrison et al. , 2013 ). However, our study is the first to use this approach for the investigation of apraxic deficits.

Materials and methods

Subjects

Patients were recruited by searching a database of stroke patients treated at the Department of Neurology at the University Medical Centre Freiburg. Inclusion criteria were unilateral left hemisphere embolic stroke of a volume of ≥15 ml at least 6 months prior to the examination. Exclusion criteria were: (i) age ≥80 years; (ii) concurrent brain damage (e.g. traumatic injury); (iii) coexisting neuropsychiatric conditions (e.g. dementia); (iv) hypoperfusion of viable brain tissue (e.g. carotid occlusion with insufficient collateralization; intracranial duplex ultrasound examinations were available for all patients); (v) inability to sustain attention for the duration of the examination; (vi) severe visual impairment with inability to recognize the content of the functional MRI paradigm; and (vii) contraindications for MRI examination. To increase the power to detect correlations between performance and activity in regions outside the ‘healthy’ left hemisphere networks, particular efforts were made to include patients with very large infarcts [i.e. total or near-total left middle cerebral artery (MCA) infarcts; n = 10; Fig. 1 ]. Initially, 38 patients were included. Of these, two were excluded from further analyses because of poor performance in the scanner (see below) and a temporary scanner dysfunction. Thus, we report data from 36 patients ( Fig. 1 ).

Seven patients had an occlusion or high-grade stenosis of the left internal carotid artery. Of these, four had complete MCA infarctions with perfusion of the left anterior cerebral artery via the right A1-segment and anterior communicating artery as documented by intracranial duplex sonography. In the remaining three patients, the left MCA was perfused via anterior and/or posterior communicating arteries with symmetrical Doppler frequencies in right and left MCAs, indicating highly effective collateralization without impaired vasomotor reserve. Drawing on previous investigations on chronic stroke patients, which did not find alterations of neurovascular coupling in the chronic phase ( van Oers et al. , 2010 ), vascular reactivity was not explicitly assessed. Two patients had partial or complete hemianopia, but had no difficulties recognizing the actions depicted in the videos or the white circle ( Fig. 2 ).

In addition, 30 aged healthy volunteers who provided normative behavioural data for several apraxia tests as reported previously ( Hoeren et al. , 2014 ) underwent the same behavioural testing and participated in the same functional MRI experiment as the patients. One subject was excluded due to excessively low performance in imitation of meaningless gestures ( Hoeren et al. , 2014 ) resulting in 29 controls for further analyses.

Full written consent was obtained from all patients and control subjects. In cases of severe aphasia, detailed information was given to the patient’s relatives or the legal guardian. The study was approved by the local ethics authorities (University of Freiburg, Germany).

Clinical and behavioural testing

Behavioural testing was conducted on the day of the MRI examination by either M.H. or one of three specially trained occupational therapists. For scoring, the performances of patients and controls were videotaped and evaluated separately by two independent raters (M.H. and V.M.L.); V.M.L. was blind to location and extent of the stroke. Items scored differently by the two raters were re-evaluated jointly to establish consensus. Subjects who declined videotaping (one patient), or could not be recorded for technical reasons (one control subject) were scored directly by the examiner. All examiners were familiarized with the scoring system prior to the study initiation.

Pantomime of tool use

Patients were examined using a modified version of the test developed by Bartolo et al. (2008) (see Hoeren et al. , 2014 ). Patients were asked to mime the use of 14 common tools depicted as line drawings ( Supplementary material ). Each item was scored as correct (1 point) or incorrect (0 points).

Actual tool use

Testing was conducted as described by Goldenberg and Spatt (2009) . Patients were seated in front of a rack on which a nail, a screw, a bolt, a thread and a padlock (recipients) were fixed. One after another, patients were handed one of five tools that corresponded to one of the recipients (e.g. hammer or screwdriver), and were prompted to demonstrate the use of the tool on the suitable recipient. If patients were unable to select the matching recipient, it was indicated by the examiner. Separate subscores were obtained for recipient selection (ToolSelect; 2 points for prompt and flawless recipient selection, 1 point for correct selection after a period of hesitation or trial and error) and action execution (ToolExec; 2 points for the flawless application of the tool, 1 point for success after trial and error or hesitation). One minor modification to the original test consisted of two visible wing screws; these were necessary to disassemble the test for transportation as required for other studies ( Martin et al. , 2015 ).

Imitation of tool-use gestures

To reduce duration of the testing session, we administered a shortened version of the test described by Bartolo et al. (2008) , which comprised 10 easily recognizable tool-associated movements involving different grips, movement dynamics and semantic categories (ImiTool; see Supplementary material ). Items were scored correct or incorrect (1 or 0 points).

Imitation of meaningless gestures

We used a previously described, slightly adapted version of a test by Goldenberg and Strauss (2002) (see Hoeren et al. , 2014 ). Patients imitated 10 positions of the hand relative to the head with invariant finger position (ImiHand), and 10 finger postures with invariant position of the hand (ImiFinger). The examiner presented the gestures with the right hand and asked the patient to imitate ‘like a mirror’, i.e. using the ipsilesional (left) hand. The examiner upheld the posture until the patient had reached the final position. In case of an error, the demonstration was repeated. Two points were given for correct imitation; one for correct imitation at second chance. The sum of ImiHand and ImiFinger was used as an overall performance measure (ImiMeaningless) ( Goldenberg and Karnath, 2006 ).

Inter-rater reliability of the cognitive motor tests

By the definitions of Landis and Koch (1977) , inter-rater reliability calculated with Cohen’s kappa was moderate for ImiTool and ImiFinger (0.499 and 0.561, respectively), substantial for PantoTool and ImiHand (0.671 and 0.616, respectively), and almost perfect for ToolExec and ToolSelect (0.847 and 0.844, respectively); for all tests, P < 0.001. For comparison with other studies using similar behavioural tests ( Goldenberg et al. , 2007 ), Pearson correlations were also calculated (range between 0.936 for ImiFinger to 0.995 for ToolExec; for all scores P < 0.001; see Supplementary material for single values).

Additional tests

Patients completed the Corsi block tapping test for short-term and working memory ( Kessels et al. , 2000 , 2008 ) and the Token Test of the Aachen Aphasia Battery ( Huber et al. , 1984 ). Overall neurological impairment was assessed using National Institutes of Health Stroke Scale (NIHSS), the Barthel-Index and the modified Rankin Scale (mRS).

Analysis

Statistical analyses were calculated using SPSS 21 (IBM Germany). To extract the common variance of the tool associated tasks, the scores of PantoTool, ImiTool, ToolSelect and ToolExec of the 36 patients were entered into a Principal Components Analysis (PCA) with direct Oblimin rotation. One factor was extracted based on the Scree plot and an Eigenvalue > 1. Adequacy of the data for the PCA was demonstrated by Bartlett’s test of sphericity ( P < 0.001) and the Kaiser-Meyer-Olkin measure (0.811).

MRI data acquisition

See the online Supplementary material for the specifications of all sequences. MRI scans were obtained on a 3 T Trio scanner (Siemens) using an 8-channel head coil. Functional echo planar images (EPI) were collected using a T ₂ *-weighted EPI sequence (306 volumes per run). In-house algorithms were applied for online correction of motion and distortion artefacts ( Zaitsev et al. , 2004 ) so that functional images were unwarped and spatially aligned already before preprocessing in SPM8. Subjects lay supine in the scanner bed with head and neck fixated using pillows, and responded by pressing a magnetic resonance-compatible button device held in the left hand. Stimuli were projected onto a screen mounted on the rear of the scanner bore and were viewed via a mirror system. The sounds of the videos were played via pneumatic headphones (MR confon) with the volume adjusted to a comfortable level. Stimulus presentation was controlled by the Presentation software (version 16; Neurobehavioural Systems Inc, Berkeley, CA; https://www.neurobs.com ) with the screen resolution set to 800 × 600 pixels. As a prerequisite for spatial normalization, a high-resolution T ₁ anatomical scan was obtained. Acute or subacute asymptomatic ischaemia was excluded with diffusion-weighted images, and fluid-attenuated inversion recovery (FLAIR) sequences were registered for lesion mapping. In addition, already existing structural images were reviewed to optimally delineate ischaemic lesion.

Functional MRI paradigm

Stimuli

Fifty-six videos showing typical tool-associated actions were recorded with a Panasonic V100 camcorder and cut to a length of 4 s. Actions were performed with the right hand, although the left hand sometimes held or steadied a recipient. To give the viewers as best as possible the impression that they were using the tool themselves, actions were depicted in first person perspective ( Fig. 2 ).

Forty-eight video stimuli were chosen for the actual paradigm. The remaining eight videos were used for a training run ( Supplementary material and below).

Experimental design

The functional MRI paradigm consisted of two runs, each containing 16 blocks of three videos separated by 1-s intervals (condition ‘Video’). The intervals between the blocks were jittered from 9.7 to 25.2 s based on in-house efficiency optimization algorithms. The sequence of the videos was adjusted to maximize distances between semantically-related tools (e.g. hammer and screwdriver), and between similar movements (e.g. sawing and cutting bread). Timing and stimuli used were identical during either session, but the order of the videos varied. Total length of each run was 459 s.

To monitor attention to the screen, subjects responded by button press to intermittent, 400 ms long appearances of a white circle (diameter 125 pixels) in the middle of the screen (condition ‘Circle’). Each run contained 32 attention events. Twenty-two occurred during videos and 10 in the between-block intervals. If appearing during a video, the circle was shown in the last two-thirds of the video (i.e. after a minimum of 1.3 s) to avoid interference with the prompt recognition of the depicted action, but no later than 400 ms before the end of the video to prevent that the change from video to background interfered with the perception of the circle. To reduce predictability, the exact time of the appearance within this time window was specified according to a Gaussian distribution [standard deviation (SD) 1.1 s] with its peak on 3.5 s into the video. In-house algorithms were used to make the circle appearances completely independent from the videos so that during subsequent analyses, the variance in blood oxygen level-dependent (BOLD) signal attributable to either Circle or Video conditions could be unambiguously determined. In summary, our paradigm comprised two experimental conditions (Video and Circle) without further baseline conditions (i.e. rest was used as baseline).

Prior to the actual experiment, subjects were familiarized outside the scanner with a training run including the eight tool-use videos not used in the functional MRI paradigm (see above). Patients were instructed to watch the videos and press the button upon appearance of the white circle.

For pilot testing, the experiment was performed with six healthy students of the University of Freiburg (three males, age mean ± SD 24.2 ± 2.7 years, range 19–27). All subjects reported very good recognizability of all actions and tools. In addition, none of the patients and controls participating in the actual study reported difficulties with recognizability during debriefing after the experiment.

Functional magnetic data analysis

SPM8 ( http://www.fil.ion.ucl.ac.uk/spm/software/spm8 ) was used for imaging preprocessing and analyses. Motion-corrected functional images were despiked using the ArtRepair toolbox (v4; http://cibsr.stanford.edu/tools/human-brain-project/artrepair-software.html ) and were then coregistered to subjects’ anatomical scans. Parameters for normalization were calculated using the Automated Lesion Identification method ( Seghier et al. , 2007 ; Sanjuán et al. , 2013 ) that includes a normalization procedure based on a modified implementation of the unified segmentation-normalization approach used in SPM8 ( Ashburner and Friston, 2005 ).

First-level estimates of haemodynamic activation changes were computed based on the General Linear Model (GLM). Individual regressors were built for the two trial types, i.e. Video and Circle. The respective regressors were built by convolving with a canonical haemodynamic response function either boxcar functions indicating onsets and offsets of the videos, or stick functions indicating the onsets of the circle presentations. In addition, head-motion parameters and their first derivatives, as well as first to fourth order polynomial regressors of slow drift, were entered as nuisance regressors. Before estimation, a standard 128 s high-pass filter was applied to the data and the model. The functional MRI activity elicited by watching tool-associated actions and by responding to the appearance of the white circle was defined by the contrasts Video > rest and Circle > rest, respectively. Contrasts between the two conditions (e.g. Video > Circle) were not calculated. We used a post-statistics normalization approach in which the first-level analyses were conducted in individual space and resulting beta images for Video (i.e. Video > rest) and Circle (i.e. Circle > rest) were transformed into stereotactic MNI space using the normalization parameters from the anatomical scans ( Poldrack et al. , 2011 ). Images were then resampled to a spatial resolution of 3 × 3 × 3 mm ³ and smoothed with an isotropic Gaussian kernel with a full-width at half-maximum of 9 mm. For further analyses, the first level results of the two runs were averaged (see Supplementary Fig. 4 for a schematic overview over the sequence of steps). For second level analyses, GLM Flex2 (Release 1 June 1 2014) was used ( http://mrtools.mgh.harvard.edu/index.php/GLM_Flex ). As all ‘missing value’ voxels were set to 0 during normalization (see below), all voxels had complete datasets (i.e. no missing data). First, to make sure that the two conditions (Video and Circle) elicited plausible patterns of activity, one-sample t -tests were performed for each condition in patients and controls separately, and differences between patients and controls were determined for each condition using two-sample t -tests ( Fig. 3 A and Supplementary Fig. 1 ). Second, regions associated with overall behavioural performance were delineated using simple regression models correlating the different test scores of the 36 patients with the amplitude of the BOLD signal change elicited by watching tool-associated actions ( Fig. 3 B–D). Third, regression analyses were repeated using total lesion volume as an additional covariate ( Fig. 6 ). Thus, only data from patients, but not controls, were entered into the regression analyses. Plots illustrating correlations at different peak voxels were created using the viewer FIVE distributed with GLM Flex and modified for illustration using Matlab (release 2012a) ( Figs 4 and 7 ).

In addition, the regression analyses between functional MRI activity and different behavioural scores were repeated taking the lesion volume within predefined regions of interest as covariate. Thus, functional MRI activity and behavioural scores were not corrected for the total left hemisphere lesion volume, but for the localized lesion volume within particular region of interests [posterior middle temporal gyrus, anterior temporal lobe (ATL), SPL, anterior IPL, posterior IPL, PMC, and anterior IFG]. As before, all analyses were carried out across the entire brain. These analyses were performed to check for different compensatory patterns depending on the lesion location (e.g. the patterns of activity in patients who perform relatively well despite inferior parietal lesions may be different from the activation associated with relatively preserved behavioural performance in patients with left premotor lesions). However, as the lesion volume within these regions of interest was tightly correlated with overall lesion volume [Pearson’s r ranging from 0.696, P < 0.001 (SPL) to 0.904, P < 0.001 (anterior IFG)], the results were highly similar to the results yielded by the analyses with total lesion volume as covariate (data not shown).

Results are displayed as surface renderings on an in-house average template of 50 normalized T ₁ scans from a sample of healthy subjects who had participated in other studies in our lab (age, mean ± SD 47 ± 20.75, range 22–84 years; 25 male) ( Beume et al. , 2015 ) or, for slices, using an average image of the normalized T ₁ scans of the 36 patients was used. Throughout the paper, we use a threshold of P < 0.001 uncorrected; this threshold allows for good comparability with past functional MRI studies in stroke patients, which also reported values below P < 0.001 uncorrected ( Saur et al. , 2006 , 2010 ; Snaphaan et al. , 2009 ; Umarova et al. , 2011 ; Stagg et al. , 2012 ), or below the more liberal threshold of P < 0.005 uncorrected ( Robson et al. , 2014 ).

Assignment of functional imaging results to anatomical structures was based on the Automated Anatomical Labelling (AAL) atlas ( Tzourio-Mazoyer et al. , 2002 ).

Evaluation of potential confounds

One potential pitfall of our post-statistics normalization approach ( Poldrack et al. , 2011 ; Supplementary Fig. 4 ) was that voxels set to ‘missing value’ (Not a Number, NaN) in the first-level beta- and con-images were converted to zero during the subsequent normalization (i.e. in our case, the Automated Lesion Identification method, Seghier et al. , 2007 ). ‘Missing values’ may be assigned by SPM during first-level analyses to voxels situated in areas affected by lesions or artefacts, and, when first-level statistics are calculated on previously normalized and smoothed EPI images, lead to the exclusion of these voxels from SPM second level analyses. We tolerated the NaN-to-zero conversions during normalization because in the case of lesions, zeros correctly reflect the absence of task-related activity. However, to exclude potential bias particularly in regions regularly affected by susceptibility artefacts, such as the inferior temporal lobe and fronto-basal regions ( Du et al. , 2007 ; Viallon et al. , 2015 ), the spatial distribution of voxels classified as NaN during first-level analyses was visualized ( Supplementary Fig. 5 ), and all second level analyses based on the Video condition ( Figs 3 and 6 ) were repeated including only voxels where in no patient a NaN-to-zero conversion had taken place ( Supplementary Figs 4, 6 and 7 ). The analyses revealed that NaN values were assigned only very rarely due to lesions (i.e. max. one patient in the peak voxel regions within the posterior parietal lobule and ventral PMC), but frequently within anterior temporal and fronto-basal regions, presumably due to susceptibility artefacts (20 and 11 patients, respectively, at the peak voxels found for ToolSelect corrected for lesion volume) ( Fig. 6 and Supplementary Fig. 5 ). Accordingly, the additional regression analyses yielded very similar results for all regions except the ATL and pars orbitalis of IFG (compare Figs 3 and 6 with Supplementary Figs 6 and 7 ). In these regions, however, an additional analysis across the 16 and 26 patients with intact functional MRI signal showed that the correlations between functional MRI activity and ToolSelect scores corrected for lesion volume were robust [ Supplementary Fig. 8 ; r-values across 36 patients including patients with NaN-to-zero conversions for ATL and IFG orbitalis were 0.74 and 0.72, respectively ( P < 0.001); r-values across the 16 and 22 patients without NaN values were 0.81 ( P < 0.01) and 0.64 ( P < 0.05), respectively]. In summary, these supplementary analyses corroborated that that our approach to handling ‘missing values’ due to lesions did not influence the main findings of the study; however, due to susceptibility artefacts, the results with respect to the ATL/IFG orbitalis should be interpreted with greater caution.

As amplitude and shape of the haemodynamic response function may be abnormal in patients with cerebrovascular disease ( Altamura et al. , 2009 ; Bonakdarpour et al. , 2015 ), the temporal evolution of the BOLD responses at left and right hemisphere peak voxels was evaluated using a finite impulse response (FIR) model spanning 27 s ( Supplementary Fig. 9 A–E). The shapes of the BOLD responses were overall similar between patients and controls, with the first local maximum reached within 4.5–7.5 s in either group at most peak voxels.

Lesion delineation and voxel-based lesion–symptom mapping

If available, lesions were delineated on diffusion-weighted imaging scans from the acute period ( n = 24) and spatially normalized using the procedure implemented in the VBM8 toolbox (r435, http://dbm.neuro.uni-jena.de/vbm/download/ ). This approach was chosen whenever possible because it may lead to highly accurate results ( Hoeren et al. , 2014 ; Martin et al. , 2015 ) (see Supplementary material for details).

If suitable diffusion-weighted imaging scans from the acute period were unavailable ( n = 12), the lesions were delineated on the fluid-attenuated inversion recovery (FLAIR) images and were normalized using clinicaltbx ( http://www.nitrc.org/plugins/mwiki/index.php/clinicaltbx :MainPage). This toolbox was specially developed for the processing of lesions and, like the automated lesion identification method ( Seghier et al. , 2007 ; Sanjuán et al. , 2013 ), uses the unified segmentation normalization routine implemented in SPM8 ( Ashburner and Friston, 2005 ) with cost function masking ( Brett et al. , 2001 ; Andersen et al. , 2010 ).

After normalization, all lesions were carefully reviewed to ensure that the lesion maps accurately reflected the extent of the lesions in MNI space. Manual adjustments were made if necessary.

Voxel-based lesion–symptom mapping (VLSM) was conducted as described previously ( Hoeren et al. , 2014 ; Martin et al. , 2015 ) using the non-parametric statistics implemented in MRIcron (version 12/12/2012) ( Rorden et al. , 2007 ). Voxelwise Brunner-Munzel Tests were calculated to identify voxels where lesions were significantly associated with reduced performance in the different overall scores (ToolPCA and ImiMeaningless) and subscores (PantoTool, ImiTool, ToolSelect, ToolExec, ImiHand, ImiFinger) ( Fig. 8 ). In addition, logistic regression analyses with lesion volume as covariate were performed to control for the association between lesion volume and scores. Only voxels affected in at least four patients were included in the analyses.

Results

Lesion anatomy

See Fig. 1 for an overlap of the lesions and the lesions of the individual subjects. Most lesions were the territory of the MCA, and the highest lesion overlap was found in white matter below the insula cortex (28/36). Conversely, regions outside the MCA territory, i.e. inferotemporal, inferior frontal, dorsolateral frontal and superior parietal regions were lesioned only infrequently. Overall, the lesion distribution was similar to previous lesion studies ( Goldenberg, 2009 ).

Demographic and behavioural results

See Supplementary Table 2 for a summary of demographic and behavioural data, and Supplementary Tables 3 and 4 for correlations between different apraxia test scores, additional tests as well as clinical and demographic data. Compared to the control group, the patient group was male-predominant and younger (patients: 29/36 male, age ± SD 59.7 ± 12.4 years; controls: 15/29 male, age 72 ± 7.2 years). However, significant test score differences between males and females were observed neither in patients (Mann-Whitney U-tests, range from P = 0.12 for PantoTool to P = 0.82 for ToolExec), nor in controls (range from P = 0.06 for ImiFinger to P = 0.82 for ToolSelect). Moreover, no significant correlations between test scores and age were found in patients (range from r = −0.10 to r = 0.11, P > 0.51; see Supplementary Table 4 ) or controls (range from r = −0.17 to r = 0.15; P > 0.38). Consequently, a significant confounding influence of age and gender on the test score differences observed between patients and controls seems unlikely. Using PCA, one factor was extracted from patients’ performances in the tasks involving tool-associated actions (PantoTool, ImiTool, ToolSelect, ToolExec), which explained 84.2% of the total variance (Tool-PCA). The factor loadings were 0.924, 0.935, 0.898 and 0.914, respectively. By means of the PCA, we aimed to separate the variance more generally attributable to the processing of tool-associated knowledge across the different subtasks from the residual variance individually related to specific subtests (PantoTool, ImiTool, ToolSelect and ToolExec). This individual variance may be caused by subtest-specific features such as the type of stimuli (i.e. line drawings versus gestures versus actual tools and objects for PantoTool, ImiTool, and ToolSelect/ToolExec, respectively), or the mode of movement execution (gestured tool use versus reaching towards an object with a tool versus actual tool use for PantoTool/ImiTool, ToolSelect and ToolExec, respectively), but also by the differential involvement of cognitive processes including executive functions or the retrieval of semantic knowledge (see Discussion). Similarly, we combined the subscores for imitation of meaningless hand and finger gestures into one score (ImiMeaningless) to achieve some generalizability over different gesture types.

Performance during functional MRI

Responses were counted as hits when the reaction time was within the limits of the mean reaction time ±3 SD of the current run; additional responses within or outside these limits were classified as false alarms. For further analyses, data were considered only from subjects in whom neither the number of false alarms nor the number of missed Circle events exceeded 20% of the total number of trials across the two runs (i.e. <13/64). Thus, one patient was excluded. There were no significant differences between controls and patients for the numbers of correct answers, false alarms, or response latencies ( Supplementary Table 2 ).

Functional MRI results

Activity elicited by the functional MRI task

Watching tool-use videos resulted in widespread bilateral activation with accompanying deactivation of areas largely corresponding to the default mode network ( Yeo et al. , 2011 ; Fox et al. , 2015 ) ( Fig. 3 A). Given the ischaemic defects, patients showed markedly less activity in the left hemisphere compared to controls; however, right hemisphere activity was also lower in patients. The activation patterns elicited by responding to the appearance of the white circle were similar for patients and controls ( Supplementary Fig. 1 ); only a small region in the left insula cortex was more activated in controls > patients.

Functional MRI activity correlated with uncorrected behavioural scores

The analysis using the common variance extracted from the tool associated tasks by PCA (Tool-PCA) was positively correlated with activity in left posterior middle temporal gyrus, supramarginal gyrus (SMG), ventral PMC, anterior intraparietal sulcus and SPL ( Fig. 3 B and Supplementary Table 5 ). Similar results were found for single tool-associated tasks except for ToolExec for which the behavioural variance was very low as many subjects performed close to ceiling. Compared to tool-associated tasks, ImiComplete as well as ImiHand alone were correlated with more extensive activity around anterior intraparietal sulcus and SPL whereas fewer voxels were found in SMG. No significant results were found for ImiFinger. A possible reason for this negative finding is that ImiFinger involves additional executive and attentional processes that make it less specifically dependent on the visuo-motor mapping mechanism commonly thought to underlie imitation ( Goldenberg and Karnath, 2006 ). Alternatively, we cannot exclude that the functional MRI task did not sufficiently activate the regions involved in ImiFinger, although this explanation seems less likely given the additional absence of significant results for ImiFinger in the current VLSM analyses (see below), and also given the considerable overlap between lesion locations associated with ImiHand and ImiFinger deficits in a recent VLSM study ( Hoeren et al. , 2014 ). Negative correlations were not found. In summary, correlations between functional MRI activity and cognitive motor performance emerged in the left hemisphere regions that are thought to support these functions in healthy subjects (i.e. the ‘healthy’ network).

The peak voxel correlations of functional MRI activity with Tool-PCA and ImiMeaningless scores are shown in Fig. 4 (see Supplementary Fig. 2 for subscores). Patients with structural damage at the different peak voxels were marked to differentiate whether low functional MRI activity resulted from direct lesion or from a functional downregulation of structurally intact regions due to remote lesions (e.g. by depriving the remote region of input; diaschisis) ( Weiller et al. , 2015 b ). The majority of low functional MRI values coincided with structural lesions. The region with the largest number of patients with low activity despite undamaged tissue (i.e. possible diaschisis) was the SPL ( Fig. 4 ).

Two strands of analyses were pursued to assess the specificity of the results. First, correlations between video-related functional MRI activity and lesion volume were explored to evaluate the potential confound of lesion size ( Fig. 3 E). As structural damage prevents stimulus-correlated activity in affected regions ( Fig. 4 ) and also impairs behavioural performance ( Fig. 5 ), the correlations between test scores and functional MRI activity ( Fig. 3 B–D) may have reflected the dependence of the apraxia tests on the structural integrity of the left hemisphere in general (indicated by an intact pattern of fronto-parietal functional MRI activation) rather than a specific association with distinct regions. When lesion volume was regressed on functional MRI activity, significant voxels appeared within regions that were unlikely to be involved in apraxia tests (Heschl-gyrus, primary visual cortex; Fig. 3 E). Conversely, the ventral PMC emerged only in the analyses with tool-related scores, but not in the analyses with lesion volume (compare Fig. 3 B–D with Fig. 3 E). These differences between the analyses with apraxia scores versus lesion volume suggest that the results for the apraxia scores were not mediated by lesion volume alone (compare Fig. 3 E with Fig. 3 B–D). Moreover, the dissociation between SMG and anterior intraparietal sulcus/SPL for Tool-PCA and ImiMeaningless, respectively (see above and Fig. 3 B and C), cannot be explained by the confound of lesion volume, as lesion volume would be expected to level differences between scores.

Second, we correlated apraxia scores as well as lesion volume with activity elicited by the Circle task ( Supplementary Fig. 1 ). This was done to explore possible additional correlations between apraxia scores and areas that were not significantly modulated by action observation (e.g. the more posterior IPL/angular gyrus). However, only lesion volume but not apraxia scores were significantly correlated with activity in the more posterior IPL (border between SMG and angular gyrus (AG), Supplementary Fig. 1 C). The finding that in contrast to lesion volume, apraxia scores were not correlated with the Circle-related activity at the SMG/AG junction, but were significantly associated with the activation of adjacently situated SMG, anterior intraparietal sulcus and SPL (see above and Fig. 3 B–D) points to a specific importance of these latter regions for cognitive motor functions.

Partial correlations between functional MRI activity and behavioural scores adjusted for lesion volume

Lesion volume was added as a covariate because the highly significant correlation between behavioural scores and lesion size ( Fig. 5 ) may obscure a relationship between behavioural scores and activity uncorrelated (or even positively correlated, see Fig. 3 E) with lesion volume ( Conger, 1974 ; Elkana et al. , 2013 ). As illustrated in Fig. 5 , the residual behavioural scores that remain after correcting for lesion volume express how much patients’ performance was better (or worse) than expected from the extent of their left hemisphere brain damage.

Residualized PCA-Tool scores were positively correlated with activity within right pars triangulars and opercularis of the IFG, right dorsal PMC, as well as within left ATL ( Fig. 6 A). For ImiMeaningless, a relationship with left dorso-lateral prefrontal cortex (PFC) and left anterior cingulate cortex emerged ( Fig. 6 B). Analyses of single subscores showed variable patterns ( Fig. 6 C). No regions were significantly correlated with residualized ImiFinger scores. See Fig. 7 and Supplementary Fig. 4 for scatter plots illustrating the partial correlations at the peak voxels of detected clusters. In summary, significant partial correlations between functional MRI activity and apraxia scores were found in infrequently damaged ipsi- and contralesional regions outside the ‘healthy’ left hemisphere network. The positive relationship between behavioural performance and activity in these regions was obscured in the initial regression analyses with uncorrected scores because lesion volume was highly negatively correlated with cognitive motor scores ( Fig. 5 ), but was at the same time uncorrelated with activity in right frontal regions, and positively associated with left dorsolateral frontal regions ( Fig. 3 E). Statistically speaking, lesion volume hence acted as a suppressor variable on the correlation between performance and activity in these regions ( Conger, 1974 ; see Elkana et al. , 2013 for the description of a similar effect in a language study with stroke patients). Significant negative correlations were not observed.

To assess whether these regions were upregulated as part of a compensatory mechanism, the activity levels in patients were compared to control subjects ( Fig. 7 ). No upregulation was observed as the BOLD signal change registered during the Video condition showed a tendency of higher activation in controls compared to patients.

Voxel-based lesion–symptom mapping

While lower tool-related skills (ToolPCA) were strongly associated with anterior IPL/SMG lesions, meaningless posture imitation (ImiMeaningless) was more susceptible to damage around the posterior superior temporal sulcus and the superior temporal gyrus ( Fig. 8 ). With respect to subscores, PantoTool and ImiTool deficits were also mainly related to anterior IPL damage, while the pattern of significant voxels found for ToolSelect and ToolExec impairments was more widespread and likely reflected that actual tool use disturbances occur mainly due to extensive lesions ( Goldenberg, 2013 ). The results for the ImiHand subscore largely corresponded to the regions identified using the combined meaningless gesture imitation score (ImiMeaningless). As in the functional MRI analyses, no regions were significantly associated with ImiFinger deficits (see above). Moreover, no significant results were found in the regression analyses with lesion volume as covariate.

Discussion

We used functional MRI during the passive observation of videos showing tool-use actions ( Fig. 2 ) to explore brain activity correlated with the off-line performances in tool-related tasks and meaningless gesture imitation in a cohort of 36 patients with chronic large left hemisphere infarcts ( Fig. 1 ). The regions where functional MRI activity was associated with better performance in the off-line tasks could be grouped into two categories. The first category comprised left hemisphere areas that are active in functional MRI when healthy subjects perform tasks corresponding to the clinical tests presently used to evaluate the patients (i.e. activity within the putative healthy network, Fig. 3 ) ( Vry et al. , 2015 ). Second, when lesion volume was added as an additional covariate to the regression analyses, positive correlations between behavioural performance and functional MRI activity emerged in infrequently damaged ipsi- and contralesional areas that are not typically associated with cognitive motor skills in healthy subjects ( Fig. 6 ). These positive partial correlations indicated that patients with relatively intact tool-related and imitative abilities despite major left hemisphere stroke benefit from activity within right IFG and PMC, as well as within left ATL and left dorso-lateral PFC.

To our knowledge, this is the first study using functional MRI activity to investigate the pathophysiology of apraxic deficits in stroke patients. Moreover, also for the first time, we delineate brain areas that may support residual cognitive motor functions in stroke patients suffering from lesions to the left hemisphere regions that mediate these functions in healthy subjects.

Left hemisphere networks for tool-associated skills and meaningless gesture imitation

The left hemisphere regions of the ‘healthy’ network emerged when voxel-wise regression analyses were used to relate the uncorrected behavioural scores to functional MRI activity. Tool-associated tasks were positively correlated with activity within posterior middle temporal gyrus/middle occipital gyrus, SMG and ventral PMC ( Fig. 3 B). Conversely, ImiMeaningless was more strongly associated with a cluster covering anterior intraparietal sulcus and SPL ( Fig. 3 C). These findings confirm that cognitive motor functions in left hemisphere stroke patients depend on the regions that subserve these tasks in healthy subjects (see Mühlau et al. , 2005 ; Caspers et al. , 2010 ; Orban and Caruana, 2014 for overviews).

The relative dissociation between SMG and anterior intraparietal sulcus/SPL for tool-related tasks and ImiMeaningless, respectively, is in accord with the proposed subdivision of the dorsal stream into separate dorso-dorsal and ventro-dorsal pathways dedicated to online movement control and learned skilled actions, respectively (see ‘Introduction’ section) ( Rizzolatti and Matelli, 2003 ; Pisella et al. , 2006 ; Binkofski and Buxbaum, 2013 ; Vingerhoets, 2014 ). The significance of the ventro-dorsal stream for tool-associated abilities was established by converging evidence (see Binkofski and Buxbaum, 2013 ; Vingerhoets, 2014 for reviews). However, the differential involvement of dorso-dorsal and ventro-dorsal areas in meaningless posture imitation has remained uncertain as neuroimaging studies in healthy individuals and lesion studies in stroke patients have variably indicated a specific role of either the superior or the inferior parietal lobule ( Mühlau et al. , 2005 ; Rumiati et al. , 2005 ; Goldenberg, 2009 ; Buxbaum et al. , 2014 ; Hoeren et al. , 2014 ; Goldenberg and Randerath, 2015 ). Our present results confirm the importance of the dorso-dorsal stream for imitation of meaningless gestures ( Hoeren et al. , 2014 ). In addition, our findings suggest that functional effects of remote lesions on the intact-appearing SPL should be considered as an explanation for the discrepancy between the different lesion studies ( Weiller et al. , 2015 b ). While low activity in the SMG was largely caused by structural damage, we observed low SPL activity despite healthy-appearing tissue in several subjects ( Fig. 4 ). Thus, although direct SPL ischaemic lesions are relatively rare ( Figs 1 and 4 ) ( Buxbaum et al. , 2014 ), lesions to the surrounding tissue may disrupt the SPL’s function, possibly by depriving the SPL of input ( Price et al. , 2001 ; Crinion et al. , 2006 ; Weiller et al. , 2015 b ). As this downregulation cannot be registered with lesion delineation techniques, the effect of SPL dysfunction may be falsely attributed to remote or adjacent regions in investigations using methods such as VLSM. This view was supported by the additional VLSM analyses ( Fig. 8 ) that confirmed the relatively greater importance of the SMG for tool-associated skills (ToolPCA), but, likely due to the low lesion density, did not corroborate the comparatively stronger association between anterior intraparietal sulcus/SPL integrity and ImiMeaningless.

On the other hand, the VLSM analyses revealed an association between ImiMeaningless deficits and damage around left posterior superior temporal sulcus/STG ( Fig. 8 ) that was not detected when regressing behavioural scores on functional MRI activity. One possible explanation may be that in contrast to the highly familiar tool-associated actions shown during the functional MRI experiment, meaningless gesture imitation posed particularly high demands on decoding (novel) biological motion ( Puce and Perrett, 2003 ; Saygin, 2007 ). Consequently, the functional MRI task may not have led to a sufficient activation of posterior superior temporal sulcus/superior temporal gyrus (compare Fig. 3 A).

Additional activity outside the ‘healthy’ network

The positive association between cognitive motor performance and activity within ipsi- and contralesional areas outside the ‘healthy’ network became evident after adjusting behavioural scores and functional MRI activity for lesion volume ( Figs 6 and 7 ). The residual behavioural scores after correcting for lesion volume express how much better (or worse) patients performed than predicted from the size of their left hemisphere lesion ( Fig. 5 ). The positive correlations between adjusted scores and activity hence indicate that patients with relatively intact cognitive motor skills despite left hemisphere damage showed higher levels of activity in the ‘additional’ ipsi- and contralesional areas ( Fig. 6 ) compared to patients with similar lesions but lower behavioural scores. The activation in the ‘additional’ regions, which were not or only rarely affected by stroke (compare Fig. 1 ), may therefore explain why some patients with large lesions performed surprisingly well while others were severely impaired. In summary, our results are in accord with the beneficial rather than detrimental effect of activity within peri- and contralesional areas that was reported for language recovery after stroke ( Weiller et al. , 1995 ; Musso et al. , 1999 ; Leff et al. , 2002 ; Blank et al. , 2003 ; see Naeser et al. , 2005 for a contrary proposal).

Given our cross-sectional study design, we could not determine if the activity within the ‘additional’ regions ( Fig. 6 ) reflected pre-existing network features or resulted from the recruitment of these areas during the course of post-stroke recovery. However, even the highest levels of activation in the relevant frontal and left temporal regions in patients did not exceed the range registered in healthy controls ( Fig. 7 ). As, therefore, we observed no signs for an upregulation, the activity more likely represented the preserved function of structurally unaffected areas more distantly related to cognitive motor skills rather than a shift of functions from lesioned areas ( Weiller et al. , 2015 b ). Longitudinal study designs, as have been used for language ( Saur et al. , 2006 ), are needed to further clarify this issue.

The patterns found for different cognitive motor subscores after adjusting for lesion volume were more varied compared to the relatively uniform results of the regression analyses with raw scores (compare Figs 3 B–D and 6). Together, these findings suggest that while cognitive motor tasks rely on a common left hemisphere network (with relative differences between tool-associated tasks and meaningless gesture imitation; Fig. 3 ), the performance in different subtasks is further modulated by distinct task-specific additional mechanisms ( Fig. 6 ). With respect to a more precise functional characterization of these mechanisms, some of the identified regions, such as the right frontal regions noted for Tool-PCA and several tool-associated subscores ( Fig. 6 ) may be specifically related to cognitive motor functions, e.g. by providing representations of learned skilled actions ( Calvo-Merino et al. , 2005 ; Nelissen et al. , 2005 ; Umiltà et al. , 2008 ; Bohlhalter et al. , 2009 ). Alternatively, as proposed in the original description of motor recovery, in which recovered patients activated dorso-lateral PFC and anterior cingulate above and beyond the normal level to perform simple finger opposition tasks with the formerly paretic hand ( Weiller et al. , 1992 ), the ‘additional’ ipsi- and contralesional activity outside the ‘healthy’ network may be the correlate of different domain-general cognitive functions that are required by the subtests to variable degrees. A similar view was recently brought into focus for language recovery ( Brownsett et al. , 2014 ; Geranmayeh et al. , 2014 ). Thus, consistent with the ATL activation noted in healthy subjects when contrasting cognitive motor tasks with and without dependency on semantic knowledge (e.g. meaningful versus meaningless gesture imitation) ( Rumiati et al. , 2005 ), the ability to associate tools with recipients (ToolSelect) was correlated with activity within left ATL and left anterior IFG after adjusting for lesion volume. These areas are involved in linking modality-specific information to form conceptual semantic knowledge ( Lambon 2014 ), and in the controlled retrieval of content from semantic memory ( Badre et al. , 2005 ), respectively. Following a disruption of action knowledge stored in ventro-dorsal stream areas, these ventral stream regions may, therefore, aid decisions on tool-recipient associations by providing other types of semantic knowledge, such as information about the typical context of use ( Hoeren et al. , 2013 ; Martin et al. , 2015 ). In a similar vein, the association of pantomimed tool use (PantoTool) with right IFG triangularis may result from the high selection difficulty ( Goldenberg et al. , 2007 ; Levy and Wagner, 2011 ), as the abstract line drawings shown to patients during PantoTool hardly constrain the range of possible actions. Conversely, the left dorso-lateral PFC may support meaningless gesture imitation and actual tool use due to its role in visuospatial problem-solving and adaptive motor behaviour ( Lau et al. , 2004 ; Kaller et al. , 2011 ; Nitschke et al. , 2012 ); moreover, anterior cingulate and dorso-lateral PFC activity were associated with higher task difficulty across cognitive motor and language tasks in healthy subjects ( Rumiati et al. , 2004 ), possibly due to higher demands on conflict resolution, response suppression and error monitoring ( Carter et al. , 1998 ; Botvinick et al. , 1999 ). Last, ImiTool may pose particular demands on decoding movement dynamics to identify the observed actions, hence relying on areas involved in the processing of biological motion, such as MT/V5+ ( Grill-Spector and Malach, 2004 ).

Limitations

Supplementary analyses ( Supplementary Figs 4–8 ) indicated that the estimation of the statistical significance may have been overly liberal with respect to the peak voxels within left ATL and anterior IFG ( Fig. 6 ) where in several subjects, task-dependent BOLD modulations could not be reliably measured due to susceptibility artefacts ( Du et al. , 2007 ; Viallon et al. , 2015 ) (see ‘Materials and methods’ section). Consequently, the association between activity within ATL and anterior IFG and relatively preserved tool-recipient matching (ToolSelect, Fig. 6 ) warrants further confirmation in studies using distortion-correction approaches specifically developed for these regions ( Halai et al. , 2014 ; Jackson et al. , 2015 ).

Second, we cannot exclude that the group differences between stroke patients and controls ( Fig. 3 A) reflected differences in the vascular reactivity rather than neurocognitive changes ensuing ischaemic brain damage ( Bonakdarpour et al. , 2015 ). However, the dissociations between the regions where functional MRI activity was associated with performance in different behavioural tasks (i.e. the key findings of the study; Figs 3 and 6 ) cannot be explained by differences in the haemodynamic response function as the same functional MRI data were used for all regression analyses. Nevertheless, future studies should correct the haemodynamic response of individual patients for a measure of vascular reactivity such as the resting-state fluctuation amplitude (RSFA), as this may increase sensitivity and specificity of the results ( Kazan et al. , 2015 ; Tsvetanov et al. , 2015 ).

Third, as discussed above, our study did not allow for an exact characterization of the neuropsychological functions subserved by the regions identified in the regression analyses with lesion volume as covariate ( Fig. 6 ). Future studies should include specific measures of visual, lexical-semantic, and attentional/executive functions to account for the behavioural and functional MRI variance attributable to these domain-general cognitive processes.

Fourth, only areas where task-dependent increases (e.g. anterior intraparietal sulcus, SMG, ventral PMC, SPL) or decreases (e.g. ATL, middle frontal gyrus) of activity were observed could be assessed using correlations between behavioural scores and functional MRI activity ( Fig. 3 A and Supplementary Fig. 8 ). As the functional MRI paradigm was more closely related to the tool-associated tasks, insufficient activity may have particularly occurred in regions involved in meaningless posture imitation. Although most regions thought to mediate ImiMeaningless were activated by the functional MRI paradigm ( Mühlau et al. , 2005 ; Goldenberg, 2009 ; Buxbaum et al. , 2014 ; Hoeren et al. , 2014 ) and, therefore, valuable conclusions could be drawn from correlating ImiMeaningless with functional MRI activity ( Figs 3 and 6 ), this limitation may apply to the superior temporal sulcus/superior temporal gyrus region. Here, an association between lesions and ImiMeaningless deficits was found in the VLSM analyses ( Fig. 8 ), but not in the regression analyses with functional MRI activity (see above).

Conclusion

Our results demonstrate that in patients with left hemisphere stroke, performance in tool-related tasks depends on the integrity of ventro-dorsal stream areas including SMG and ventral PMC, whereas meaningless gesture imitation relies more on dorso-dorsal regions, such as anterior intraparietal sulcus and SPL. When lesion volume was added to the regression analyses as an additional covariate, further positive correlations between test scores and activity emerged in ipsi- and contralateral regions that are less commonly associated with cognitive motor abilities in healthy subjects. Tool-associated skills were correlated with activity in right IFG, right PMC and left ATL, while meaningless gesture imitation was related to activity within left dorso-lateral PFC. These infrequently lesioned areas may support residual tool-associated and imitative abilities after damage to the left hemisphere regions that normally mediate cognitive motor functions in healthy individuals.

Acknowledgements

We thank Hansjörg Mast for assistance in data acquisition. We thank Gabriele Lind, Sarah Höfer and Cornelia Pietschmann for conducting the neuropsychological testing; without their careful examinations, this study would not have been possible. We thank Professor Georg Goldenberg for help with the actual tool use test.

Funding

This work was supported by the BrainLinks-BrainTools Cluster of Excellence funded by the German Research Foundation (DFG, grant #EXC1086).

Supplementary material

Supplementary material is available at Brain online.

References

Abbreviations

Abbreviations
 
• ATL =

  anterior temporal lobe

 
• BOLD =

  blood oxygen level-dependant

 
• IFG =

  inferior frontal gyrus

 
• IPL =

  inferior parietal lobule

 
• MCA =

  middle cerebral artery

 
• PFC =

  prefrontal cortex

 
• PMC =

  premotor cortex

 
• SMG =

  supramarginal gyrus

 
• SPL =

  superior parietal lobule

 
• VLSM =

  voxel-based lesion–symptom mapping

Author notes