Action and semantic tool knowledge – Effective connectivity in the underlying neural networks

Abstract

Evidence from neuropsychological and imaging studies indicate that action and semantic knowledge about tools draw upon distinct neural substrates, but little is known about the underlying interregional effective connectivity. With fMRI and dynamic causal modeling (DCM) we investigated effective connectivity in the left‐hemisphere (LH) while subjects performed (i) a function knowledge and (ii) a value knowledge task, both addressing semantic tool knowledge, and (iii) a manipulation (action) knowledge task. Overall, the results indicate crosstalk between action nodes and semantic nodes. Interestingly, effective connectivity was weakened between semantic nodes and action nodes during the manipulation task. Furthermore, pronounced modulations of effective connectivity within the fronto‐parietal action system of the LH (comprising lateral occipito‐temporal cortex, intraparietal sulcus, supramarginal gyrus, inferior frontal gyrus) were observed in a bidirectional manner during the processing of action knowledge. In contrast, the function and value knowledge tasks resulted in a significant strengthening of the effective connectivity between visual cortex and fusiform gyrus. Importantly, this modulation was present in both semantic tasks, indicating that processing different aspects of semantic knowledge about tools evokes similar effective connectivity patterns. Data revealed that interregional effective connectivity during the processing of tool knowledge occurred in a bidirectional manner with a weakening of connectivity between areas engaged in action and semantic knowledge about tools during the processing of action knowledge. Moreover, different semantic tool knowledge tasks elicited similar effective connectivity patterns.

1. INTRODUCTION

Differential impairments of tool knowledge due to neurological disease indicate that knowledge about action and semantic aspects of tools draw upon distinct neural networks. Impaired manipulation knowledge (action knowledge), that is, knowing how to handle/manipulate a tool, is a key deficit in patients suffering from apraxia (Buxbaum, Giovannetti, & Libon, 2000), who often exhibit parietal lesions (Martin et al., 2016b; Niessen, Fink, & Weiss, 2014). The role of the left inferior parietal cortex (IPL) and left intraparietal sulcus (IPS) in action representation is corroborated by many imaging studies that contrast tasks requiring manipulation knowledge about tools with function knowledge (i.e., knowing what a tool is used for, semantic knowledge; Boronat et al., 2005; Canessa et al., 2008; Chen, Garcea, Jacobs, & Mahon, 2017b; Kellenbach, Brett, & Patterson, 2003) and by non‐invasive brain stimulation studies (Andres, Pelgrims, & Olivier, 2013; Evans, Edwards, Taylor, & Ietswaart, 2016; Ishibashi, Lambon Ralph, Saito, & Pobric, 2011).

In contrast, patients with semantic dementia (SD) or other causes of temporal lobe damage often show deficits concerning the semantic knowledge about tools, for example, impaired identification of tools and/or impaired function knowledge about tools, while praxis skills of the same tool can remain unaffected (Baumard et al., 2017; Buxbaum, Schwartz, & Carew, 1997; Lauro‐Grotto, Piccini, & Shallice, 1997; Magnie, Ferreira, Giusiano, & Poncet, 1999; Martin et al., 2016a; Negri et al., 2007; Sirigu, Duhamel, & Poncet, 1991). Imaging studies contrasting function knowledge with manipulation knowledge revealed activity in the lateral anterior infero‐temporal lobe (ATL; Canessa et al., 2008; Chen, Garcea, & Mahon, 2016) and the medial fusiform gyrus (FFG; Chen et al., 2017b). Consistent with these imaging findings, non‐invasive brain stimulation over the ATL affected function judgment tasks (Andres et al., 2013; Ishibashi et al., 2011; Ishibashi, Mima, Fukuyama, & Pobric, 2017).

This double dissociation regarding differential tool‐related deficits in patients with apraxia (and lesions of the parietal cortex) versus those of SD patients (with lesions of the temporal cortex) are consistent with the hypothesis of two segregated functional visuo‐motor streams: a dorsal stream and a ventral stream, that were originally presumed as processing pathways for vision‐for‐action (“where”) and vision‐for‐perception (“what”), respectively (Goodale, & Milner, 1992; Ungerleider, & Mishkin, 1982). In this framework, the ventral stream mediates semantic aspects of tools and the dorsal stream mediates online control of tool‐associated actions.

A recent extension of the two‐stream model posits a further subdivision of the dorsal stream into a dorso‐dorsal stream and a ventro‐dorsal stream (Binkofski & Buxbaum, 2013). Here, the dorso‐dorsal stream is thought to process structural object properties for prehensile actions (“grasp system”) and supports online‐motor control, while the ventro‐dorsal stream represents the “use system” for skilled actions with (familiar) objects and supports long‐term tool action representations (Binkofski et al., 2013; Buxbaum & Kalenine, 2010; Hoeren et al., 2014; Rizzolatti & Matelli, 2003). In this context, optic ataxia is a typical disorder caused by lesions to the dorso‐dorsal stream resulting in deficient reaching, whereas limb apraxia is commonly associated with ventro‐dorsal stream lesions, in which online motor control remains intact (Binkofski et al., 2013).

Anatomically, the origin of the visuo‐motor streams is the primary visual area (V1). From here, the ventral stream projects along the occipital and temporal cortices, including the FFG and the ATL (Mahon, et al., 2007). The dorsal stream projects from V1 towards the parietal lobe, in which the superior parietal lobe (SPL) serves as a key player in the dorso‐dorsal stream, while the inferior parietal lobe (IPL) and parts of the anterior intraparietal sulcus (aIPS) are important nodes within the ventro‐dorsal stream (Binkofski et al., 2013; Grefkes, & Fink, 2005; Sakreida, et al., 2016).

Nevertheless, the concept of segregated action and semantic representations falls short in explaining various complex findings. For instance, the well‐known patient DF, suffering from visual agnosia due to ventral lesions, fails to grasp objects in a functionally appropriate manner despite an intact parietal and frontal lobe (Carey, Harvey, & Milner, 1996; Milner, 1997). Likewise, in SD the degraded conceptual semantic information about tools is sometimes associated with impaired tool use/action knowledge (Hodges et al., 2000; Hodges, Spatt, & Patterson, 1999). A 4‐year longitudinal study showed that such tool‐use deficits in SD developed with the decline of conceptual knowledge about tools and indicate the relevance of semantic knowledge for praxis skills (Coccia et al., 2004). In the same vein, apraxic patients may also show deficits in function knowledge about tools as reflected, e.g., in content errors during actual tool use (De Renzi & Lucchelli, 1988; Martin et al., 2016a; Ochipa, Rothi, & Heilman, 1989). These findings suggest that besides the integrity of the visuo‐motor streams, also an adequate information exchange between the streams may be requisite for the proper processing of tool knowledge and tool use.

However, to date only few studies addressed the issue of (functional or effective) connectivity within and between the nodes of the visuo‐motor streams. An interesting set of imaging studies found that in the context of tool action representations, the IPL gets input via ventral areas (Almeida, Fintzi, & Mahon, 2013; Garcea, Kristensen, Almeida, & Mahon, 2016; Kristensen, Garcea, Mahon, & Almeida, 2016; Mahon, Kumar, & Almeida, 2013). These findings suggest that an object's identification and function is first decoded by ventral regions, and then this information is communicated to the IPL, where it is combined with information processed by dorsal regions (e.g., position, orientation) in order to perform adequate tool manipulation (Kristensen et al., 2016).

Here, we aimed at extending those previous findings by investigating the effective connectivity during the processing of action and semantic knowledge about tools with fMRI and dynamic causal modeling (DCM).

In the imaging experiment, we used a manipulation (action) knowledge task (i.e., which of two hand postures is appropriate for using the target tool?) and two semantic knowledge tasks: a function knowledge task (i.e., which of two recipient objects is typically used together with the target tool?), and a value estimation task (i.e., what is the approximate monetary value of the target tool?).

Due to the fact that (i) semantic knowledge about tools comprises various aspects in addition to function knowledge and (ii) that many studies found action knowledge to be contingent on function knowledge, the value estimation task was chosen as an additional semantic task, with clearly no relevance for manipulation. Separate exploration of the effective connectivity patterns during these two semantic tasks (processing of function and value knowledge) was of interest to reveal whether processing of semantic knowledge is generalizable across the different semantic aspects of tool knowledge.

The resulting fMRI data were first evaluated by a conventional general linear model (GLM). Based on these GLM results, we then focused on the context‐dependent effective connectivity (Friston, Harrison, & Penny, 2003).

2. MATERIALS AND METHODS

2.1. Participants

Twenty healthy participants gave written informed consent to participate in the study. Due to technical problems two subjects had to be excluded from the final analyses. Therefore, data from 18 subjects were analyzed (10 female; mean age 25.3 years, range 18–35 years). All participants had normal or corrected to normal vision and were right‐handed (Oldfield, 1971). The study was conducted in accordance with the Declaration of Helsinki and was approved by the local ethics committee.

2.2. Stimuli and task

The study featured three experimental conditions. For each of these (manipulation knowledge condition [M], function knowledge condition [F], and monetary value knowledge condition [V]), 40 stimuli per task were created based on an identical set of uni‐manually manipulable tools used in all three conditions (see Supporting Information, Table S4). During the functional magnetic resonance imaging (fMRI) experiment, visual stimuli were presented on a 30″ shielded TFT (thin film transistor) monitor mounted 245 cm behind the scanner. Stimuli were viewed via a mirror installed on top of the head coil. The size of the stimuli displayed on the TFT screen corresponded to a visual angle of 13.8° × 8.2°. For stimulus presentation and response monitoring, the Presentation Software package (Neurobehavioral Systems Inc., Berkeley, CA) was used.

In the “manipulation” condition (M), stimuli consisted of a target tool and photographs of two different hand postures. The target tool, surrounded by a black frame, was presented centrally in the lower part of the stimulus display. The pictures of the two hand postures were presented simultaneously in the upper right and left corners of the stimulus display, one being suitable for manipulating the tool and the other one not. In the “function” condition (F), stimuli were composed of a target tool and pictures of two objects, with one of these two objects functionally related to the target tool and the other not. Finally, stimuli of the “monetary value” condition (V) contained pictures of the same target tools in combination with two pictures of coins (1 or 2 Euro coin) or banknotes (5, 10, 20, or 50 Euro note). One of the depicted amounts of money approximated the true value of the target tool whereas the other was clearly too high or low (see Figure 1 for experimental stimuli in the three tasks and Supporting Information, Table S4 for displayed items, coins and banknotes).

Figure 1.

Experimental stimuli. This graphic shows examples of the experimental stimuli used drawing upon action tool knowledge and semantic tool knowledge. In the manipulation condition (M), the appropriate hand posture for manipulating the target tool (here: a hammer) had to be selected (here: the hand posture on the right side of the stimulus). In the function condition (F), subjects selected the appropriate recipient object for the target tool (here: the nail on the right side of the hammer). In the value knowledge condition (V), subjects estimated the monetary value of the target tool and selected the appropriate banknote or coin (here: the 10 Euro note on the left side of the stimulus). For a list of the 40 target tools with the recipient objects (F) and respective coins/banknotes (V), see Supporting Information, Table S4 [Color figure can be viewed at http://wileyonlinelibrary.com]

Participants were asked to examine the stimuli and to judge whether the hand posture (M), the object (F), or the monetary value (V) presented on the left or right side fitted the target tool better. Participants indicated their choice by pressing one of two buttons with the index or the middle finger of the left hand. The left hand was used to minimize any confounding effects of motor execution on left‐hemispheric activations. The chosen picture corresponded spatially to the left‐/right‐sided response key (i.e., left‐sided picture/middle finger, right‐sided picture/index finger).

2.3. Procedure and design

Before entering the scanner, participants were familiarized with the tasks. Practice trials did not re‐appear in the actual fMRI experiment to avoid learning effects. Participants were asked to respond both as accurately and as fast as possible.

The study employed a blocked within‐subject design alternating experimental blocks (duration 28 s) with baseline (duration 22 s) to maximize design efficiency (Mechelli, Price, Henson, & Friston, 2003). During the low‐level baseline period, subjects were shown a white screen with a black frame (surrounding the target tool in the experimental conditions) only. After 20 s the color of the frame changed to red indicating the start of the next task‐block in 2 s. Five blocks, each containing eight trials, were presented per experimental condition (M, F, and V), yielding a total of 15 experimental blocks. Stimulus duration was fixed (3,500 ms) with no inter‐stimulus interval. The order of the stimuli was randomized, while the order of blocks was pseudo‐randomized. During the fMRI‐experiment reaction times and accuracy were recorded.

2.4. fMRI data acquisition and preprocessing

A 3‐Tesla MRI System (Trio, Siemens, Erlangen, Germany) was used to obtain T2*‐weighted gradient echo‐planar images (EPI) with BOLD contrast (matrix size: 64 × 64; voxel‐size: 3.1 × 3.1 × 3.0 mm³; field of view: 200 mm; repetition time: 2,200 ms; echo time: 30 ms; flip angle: 90°). Thirty‐six transversal slices of 3 mm thickness were acquired sequentially with a 0.3 mm interslice gap (whole‐brain coverage). A total of 362 functional volumes were collected for each subject in a single functional run. FMRI data were analyzed using the Statistical Parametric Mapping software package (SPM8, Wellcome Department of Imaging Neuroscience, London; http://www.fil.ion.ucl.ac.uk/spm). The first six EPI volumes were omitted to allow for T1 equilibration effects. In order to correct for inter‐scan movements, EPI images were first spatially realigned. Then, the mean EPI image for each participant was computed and spatially normalized to the Montreal Neurological Institute (MNI) template using the “unified segmentation” function in SPM8 (Ashburner, & Friston, 2005). Finally, the data were smoothed using a Gaussian kernel of 8 mm full width half maximum (FWHM) to suppress noise. The effective connectivity analysis was conducted using the latest DCM code as implemented in SPM12 (version: Oct 20, 2016).

2.5. Data analyses

2.5.1. Data processing

Analyses of the behavioral data were performed using SPSS (IBM SPSS Statistics, Version 21). For the statistical analysis of the BOLD data, three regressors of interest containing the onset and duration of the three experimental conditions were defined. The BOLD response was modeled using a canonical hemodynamic response function and its first derivative. Moreover, head movement parameters were included as additional regressors of no interest in the design matrix. Baseline periods were not explicitly modeled.

The three simple contrasts (i.e., the three experimental conditions each compared with the implicit baseline) were specified at the first‐level and then transferred to a second‐level ANOVA model. At the second level, all six differential contrasts were calculated (M > F, M > V, F > M, V > M, F > V, V > F). As expected, the semantic tool knowledge tasks (F and V) revealed similar behavioral effects and engaged similar brain regions in fMRI, affirmed by a conjunction analysis of the contrasts F > M and V > M (for Results see Supporting Information, Tables S1 and S2). Therefore, and in order to identify robust activations of brain regions engaged in action knowledge (here: M) and semantic knowledge (here: F and V) about tools, the following additional differential contrasts were specified: 2 × M > (F + V) and vice versa (F + V) > 2 × M (cf. Table 1; see Section 3.2.1).

Table 1.

Brain regions showing significant relative increases of BOLD response associated with each comparison of interest as used for the time series extraction

				MNI coordinates
	Hemisphere	Cluster size (voxels)	Max. T‐value	x	y	z
Manipulation > (Function + Value), 2 × M > (F + V)
Middle temporal gyrus (LOTC, hOc4la)	R	1908	17.56	52	−70	−2
Middle temporal gyrus (LOTC, hOc4la)	L	1671	14.47	−58	−62	4
Intraparietal sulcus (hIP3)	L	1634	9.60	−36	−42	54
Supramarginal gyrus (PFt)	L	1634	8.22	−56	−28	40
Supramarginal gyrus (PFt)	L	1634	8.22	−56	−28	40
Inferior frontal gyrus (BA 44)	L	149	8.41	−50	8	20
(Function + Value) > Manipulation, (F + V) > 2 × M
Fusiform gyrus (FG3)	R	570	9.67	34	−32	−26
			9.07	30	−50	−14
Fusiform gyrus (FG3)	L	170	8.68	−26	−48	−20
Medial prefrontal cortex (Fp2)	M	400	8.04	0	56	8
	L	400	7.30	−8	62	0
		400	6.49	−2	66	4
	L	400	6.49	−2	66	4
Angular gyrus (PGa)	R	242	6.17	58	−64	32
Angular gyrus (PGa)	L	155	5.26	−40	−72	48
			5.23	−50	−68	36

For each activation cluster, the coordinates in MNI space are given referring to the maximally activated voxel within an area of activation as indicated by the highest T‐value. Note that in some cases sub‐maxima have been used as reference coordinates for the time series extraction for DCM. All activations are significant at p < .05 (family wise error [FWE] corrected at the voxel level) using an extent threshold of 100 voxels. The precise functional/cytoarchitectonic location of the coordinates, assessed with the anatomic toolbox (Eickhoff et al., 2005) is given in parenthesis.

LOTC, lateral occipito‐temporal cortex; hOc, human occipital cortex; PFt, inferior parietal area PFt; hIP3, human intra‐parietal area 3; BA, Brodmann area; FG3, fusiform gyrus area 3; Fp2, fronto‐polar area 2; PGa, inferior parietal area PG anterior.

All differential contrasts were thresholded at p < .05, family wise error (FWE)‐corrected for multiple comparison at the voxel level, to adjust for falsely positive voxels (Nichols, & Hayasaka, 2003). Note that voxel‐wise inference is considered more robust than cluster‐size inference, since voxel‐wise inference is not affected by the recently discussed problems caused by too liberally defined cluster thresholds (Eklund, Nichols, & Knutsson, 2016; Woo, Krishnan, & Wager, 2014). In addition, an extent threshold of 100 voxels was applied for the six simple contrasts and 2 × M > (F + V) and (F + V) > 2 × M.

2.5.2. Dynamic causal modeling (DCM)

We used dynamic causal modelling (Friston et al., 2003) to study the underlying effective connectivity in the neural network engaged in processing different aspects of tool knowledge. DCM is a hypothesis‐driven approach to make interferences about neural connectivity in terms of direction and effect size (using the parameter “rate constant” with the unit Hertz (1/s), while the influence that one region exerts upon another can be positive or negative). Note that obtained connectivity parameters may not necessarily reflect monosynaptic anatomical connections but rather the net effect, for example, transmitted via direct connections, a single relay area or more extensive loops (Stephan et al., 2009b).

For DCM analyses, the “model space” must be defined a priori, that is, different plausible network models need to be generated beforehand. Subsequently, the competing models are fed with the underlying data of all subjects in form of subject‐specific volumes of interest (VOI). Finally, based on Bayesian model selection (BMS) the model that best explains the underlying fMRI data is identified as “winning model.”

The (dynamic) changes in the neural model over time (ż) assumed by DCM are represented in the bilinear state equation ż =  $(\mathrm{A}+\sum _{\mathrm{j}}{\mathrm{u}}_{\mathrm{j}} {\mathrm{B}}^{\mathrm{j}})$ z + Cu (Friston et al., 2003). The A‐matrix represents the task‐independent intrinsic (fixed) connectivity, the B‐matrix the task‐dependent modulations of A, and the C‐matrix the direct input to the system.

Regions of interest (ROI)

The selection of regions of interest was based on the results of the GLM analysis. Since the number of regions of interest (ROI) in the computation of a DCM analysis is limited (Daunizeau, David, & Stephan, 2011; Stephan et al., 2010), we included only the eight regions in the left hemisphere (LH) that had shown significant activation in the differential contrasts 2 × M > (F + V) or (F + V) > 2 × M, which is clinically plausible, since tool use deficits are most often observed after LH lesions (Martin et al., 2016a; Martin et al., 2016b).

Thus, the VOI for the eight left hemispheric regions were obtained for each participant at the individual level. Given the results of the GLM analysis, the VOI for the lateral occipito‐temporal cortex (LOTC), intraparietal sulcus (IPS), supramarginal gyrus (SMG), and inferior frontal gyrus (IFG) of the LH were based on the contrast 2 × M > (F + V), while the reverse contrast (F + V) > 2 × M was applied to obtain the VOI for the left fusiform gyrus (FFG), the left angular gyrus (AG), and the medial prefrontal cortex (mPFC; cf. Table 1). The coordinates for the early visual region (visual area 1 and 2, V1/2), as common input region, were selected based upon a conjunction analysis of all three conditions (M, F, and V) versus baseline (Friston, Penny, & Glaser, 2005; Nichols et al., 2005).

Mostly the individual local maximum, in some cases a sub‐maximum to fit the anatomical constraints of the identified ROI was chosen for time‐series extraction (see Supporting Information, Table S3 for the individual MNI coordinates and the respective anatomical localization assessed with the Anatomy Toolbox (Version 2.2c), implemented in SPM12 (Eickhoff, et al., 2005). For a visual depiction of the individual maxima see Figure 3. Computation of the individual Euclidean distances between the MNI coordinates ruled out any potential VOI overlap. The shortest distance between two VOI was 17 mm (between IPS and AG).

Figure 3.

Cluster of the individual coordinates used for DCM analysis. Graphical demonstration of the individual MNI coordinates for the DCM analysis (listed in Supporting Information, Table S3) projected onto a 3D surface rendering of a standard single subject brain. Regions of interest (ROI): visual area 1 and 2 (V1/2) lateral occipito‐temporal cortex (LOTC), angular gyrus (AG), intraparietal sulcus (IPS), supramarginal gyrus (SMG), inferior frontal gyrus (IFG), medial prefrontal cortex (mPFC). All ROIs are in the left hemisphere. In this orientation not depicted: fusiform gyrus [Color figure can be viewed at http://wileyonlinelibrary.com]

Time series of the VOI were extracted for supra‐threshold voxels as the first eigenvariate within a sphere of 3 mm radius around the individual (sub‐)maxima at a threshold of p < .001 (uncorrected) and, if necessary, gradually lowered to p < .05 (uncorrected). In one subject the eight VOI could not be reliably identified, so data from this subject could not be included into the connectivity analysis.

Generation of models for DCM/the model space

The different models, grouped into model families (see Penny et al., 2010) were specifically generated to examine two main aspects of the network processing tool knowledge.

Given that action and semantic knowledge about tools are processed by different brain regions/nodes, we first created four different model families (F1–F4) that varied with respect to putative interactions between semantic nodes and action nodes, i.e., structural differences in the A‐matrix (see Supporting Information, Figure S1a): model family F1, no interaction; model family F2, interaction within the occipito‐temporal lobe (FFG–LOTC); model family F3, interaction within the parietal lobe, (AG–SMG); and model family F4, interactions within both, the occipito‐temporal and the parietal lobe (FFG–LOTC, AG–SMG).

Second, different hypotheses about the task‐dependent modulation of effective connectivity by the three experimental tasks (B‐matrices for M, F, V) lead to six models within each model family (see Supporting Information, Figure S1b–e): three different levels of modulations were assumed; (i) modulatory effects within the occipito‐temporal lobe only (V1/2, LOTC, FFG); (ii) modulatory effects throughout the occipito‐temporal lobe and parietal lobe (V1/2, LOTC, FFG, IPS, SMG, AG); and (iii) modulatory effects at all (three) levels, that is, throughout the occipito‐temporal lobe, the parietal lobe and the frontal lobe (V1/2, LOTC, FFG, SMG, AG, IFG, mPFC). All these different hypotheses were configured as (a) unidirectional and (b) bidirectional to examine whether the experimental conditions (M, F, and V) modulated the connectivity between involved brain regions/nodes in only one direction or in a bidirectional manner.

Accordingly, the investigated aspects led to a total of 24 models: four model families (F1–F4, different interactions between action nodes and semantic nodes) with six models each (six = three [different levels of modulation] times two [unidirectional versus bidirectional]). For an overview of all 24 considered models see Supporting Information, Figure S1.

Note that similar differential activity patterns in the GLM (as for the F and V task) do not necessarily imply that the activated regions reveal analogous connectivity patterns. Thus, all three tasks (M, F, and V) were implemented as separate B‐matrices to explore whether the effective connectivity patterns for processing of semantic aspects about tools could be generalized across different semantic conditions.

After the different models were established and fed with the underlying fMRI data, a random‐effects Bayesian model selection (BMS) was applied to identify the “winning” model family of the four families, and the “winning” single model, that is, the one with highest evidence in explaining the given data out of all tested model families/single models. This superiority can be expressed by its exceedance probability (in %) in relation to the tested alternatives (Stephan et al., 2009a). Hence, the model parameters (A‐, B‐, and C‐matrices) were extracted for each subject, then averaged across subjects and tested for significance by one‐sample t tests (p < .05, corrected for multiple comparisons by false discovery rate (FDR; Benjamini & Hochberg, 1995).

Finally, the total mean variance between prior and posterior parameters (i.e., the mathematical model and the observed data) explained by the winning model was computed with the spm_dcm_fmri_check.m script provided in the SPM helpline by Karl Friston (https://www.jiscmail.ac.uk/cgi-bin/webadmin? A2 = spm;bebd494.1203; 2012).

3. RESULTS

3.1. Behavioral data

One‐way repeated measures ANOVAs revealed that accuracy was similar across tasks (M: 94.7%, F: 95.3%, V: 94.7%; F _(2,34) < 1). However, a significant difference with respect to reaction times (RTs) was observed (F _(2,34) = 77.9, p < .001). Post hoc comparisons revealed that the RTs of the action knowledge task (M) were systematically longer than those of the semantic knowledge tasks (F and V), which did not differ significantly from each other (M [mean ± SD]: 1,758 ± 176 ms, F: 1,329 ± 172 ms, V: 1,398 ± 233 ms; M vs. F: t ₍₁₇₎ = 17.1, p < .001; M vs. V t ₍₁₇₎ = 8.7, p < .001; F vs. V: t ₍₁₇₎ = 1.7, p = .117).

3.2. Functional imaging

3.2.1. GLM analysis

As expected, the simple contrasts of M > F, M > V, F > M, F > V, V > M, and V > F revealed highly similar activation patterns for the two semantic tool knowledge tasks (F and V). The similarity of the activity patterns for F and V was further supported by a conjunction analysis of the contrasts F > M and V > M. These results are reported in the Supporting Information as the focus of this study was on the effective connectivity analysis, which was based on the GLM results.

For the direct comparison of action knowledge and semantic knowledge about tools, we contrasted in the GLM analysis the M task with the F and V task (i.e., 2 × M > (F + V)) yielding significant activation clusters in the LOTC bilaterally, the left IPS, the left SMG, and the left IFG (see Table 1 and Figure 2, activations in red). The reverse contrast ((F + V) > 2 × M) revealed significant activation clusters in the FFG and the AG bilaterally as well as in the mPFC (see Table 1 and Figure 2, activations in blue).

Figure 2.

Visual depiction of the results of the GLM analysis. The activation patterns observed for the Manipulation>(Function+Value) contrast (red) and for the (Function + Value) > Manipulation contrast (blue) are projected onto a 3D surface rendering of a standard single subject brain. (a) View of left hemisphere, (b) view from bottom, (c) view from the front, (d) view from behind

3.2.2. DCM results

Family 4 was superior to the other tested hypotheses with an exceedance probability of 73.6% (see Table 2a). As a common feature, all models of family 4 postulated couplings between the regions processing action knowledge (action nodes) and semantic knowledge about tools (semantic nodes) within the occipito‐temporal lobe (LOTC–FFG) and the parietal lobe (SMG–AG). Out of the 24 tested models the individual model F4M2 belonging to family 4 had the highest evidence for explaining the underlying data with an exceedance probability of 54.4% (see Table 2b). Please note, that in Bayesian Model Selection (BMS), models with more degrees of freedom are penalized to balance the benefit of complexity in fitting the data with the loss of accuracy (Penny, Stephan, Mechelli, & Friston, 2004; Stephan et al., 2009b). With respect to the divergence between prior and posterior parameter distributions, we computed the total mean variance explained. On average across subjects our winning model explained 40.0% ± SD 11.5% of the variance (range 19%–59%) indicating a good fit of predicted and observed responses. ‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬

Table 2.

Exceedance probabilities derived from the comparison of model families (a.) and single models (b.)

a. Model families	Exceedance probability
Family 1	0.0201
Family 2	0.0178
Family 3	0.2263
Family 4	0.7358

b. Single models	Exceedance probability	Single models	Exceedance probability
Family 1		Family 2
F1M1	0.0126	F2M1	0.0003
F1M2	0.0006	F2M2	0.0529
F1M3	0.0005	F2M3	0.0003
F1M4	0.0099	F2M4	0.0004
F1M5	0.0006	F2M5	0.0003
F1M6	0.0003	F2M6	0.0004
Family 3		Family 4
F3M1	0.0003	F4M1	0.0006
F3M2	0.0997	F4M2	0.5441
F3M3	0.0003	F4M3	0.0004
F3M4	0.1219	F4M4	0.1528
F3M5	0.0000	F4M5	0.0001
F3M6	0.0003	F4M6	0.0004

‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬‬Effective connectivity during the processing of tool knowledge

Readout of the winning model F4M2 across subjects revealed the following significant (p < .05, FDR‐corrected) mean intrinsic effective connectivity patterns (A‐matrix): from more caudal nodes to more rostral nodes positive intrinsic couplings were revealed (action nodes: V1/2→LOTC→IPS, SMG→IFG; and semantic nodes: V1/2→FFG, AG→mPFC), and in the opposite direction negative intrinsic couplings (action nodes: IFG→SMG; and semantic nodes: mPFC→AG, FFG→V1/2; see Figure 4a and Table 3).

Figure 4.

Effective connectivity (DCM) results. The left‐hemispheric regions of interest considered in the DCM analysis are the visual area 1 and 2 (V1/2), lateral occipito‐temporal cortex (LOTC), fusiform gyrus (FFG), angular gyrus (AG), intraparietal sulcus (IPS), supramarginal gyrus (SMG), inferior frontal gyrus (IFG), and medial prefrontal cortex (mPFC). Significant effective connectivity is shown in quality (green = positive connectivity/strengthening and red = negative connectivity/weakening) and quantity (coupling strength in Hz [1/s]), gray = non‐significant connectivity/modulation (p < .05, FDR‐corrected).(a) Intrinsic connectivity (A‐matrix). (b, c, and d) Modulation of the connectivity by the experimental tasks (B‐matrices). The modulations are the dynamic change of the A‐matrix (a) during the tasks. The three experimental conditions consist of the Manipulation knowledge task (M, b), the Value knowledge task (V, c) and the Function knowledge task (F, d) [Color figure can be viewed at http://wileyonlinelibrary.com]

Table 3.

DCM coupling parameters of the winning model (F4M2) in 1/s (Hz) tab

ROIs (left)		Intrinsics		Manipulation (M)		Function (F)		Value (V)
Origin	Target	Mean	p (FDR)	Mean	p (FDR)	Mean	p (FDR)	Mean	p (FDR)
V1/2	FFG	0.175	0.015	–	–	0.763	<0.001	0.497	<0.001
V1/2	LOTC	0.210	0.007	1.085	<.001	–	–	–	–
FFG	V1/2	−0.213	0.007	–	–	−0.221	n.s.	−0.072	n.s.
FFG	LOTC	−0.085	n.s.	−0.450	.009	0.107	n.s.	−0.004	n.s.
FFG	AG	−0.031	n.s.	–	–	0.093	n.s.	0.074	n.s.
LOTC	V1/2	0.044	n.s.	−0.220	n.s.	–	–	–	–
LOTC	FFG	0.040	n.s.	0.046	n.s.	−0.358	n.s.	−0.020	n.s.
LOTC	IPS	0.257	0.004	0.501	<0.001	–	–	–	–
AG	FFG	0.028	n.s.	–	–	−0.198	n.s.	−0.128	n.s.
AG	SMG	0.024	n.s.	0.204	n.s.	−0.064	n.s.	0.067	n.s.
AG	mPFC	0.315	0.001	–	–	0.200	n.s.	−0.005	n.s.
IPS	LOTC	−0.062	n.s.	−0.370	0.018	–	–	–	–
IPS	SMG	0.063	n.s.	0.626	<0.001	–	–	–	–
SMG	AG	0.055	0.045	−0.452	0.001	0.087	n.s.	0.272	n.s.
SMG	IPS	0.043	n.s.	−0.798	0.001	–	–	–	–
SMG	IFG	0.208	0.014	0.326	<0.001	–	–	–	–
IFG	SMG	−0.117	0.045	−0.500	<0.001	–	–	–	–
mPFC	AG	−0.106	0.008	–	–	−0.194	n.s.	−0.053	n.s.

p Value (FDR‐corrected, false‐discovery‐rate); n.s. = not significant; – = connection not part of the tested model.

V1/2, visual area 1 and 2; FFG, fusiform gyrus; LOTC, lateral occipito‐temporal cortex; AG, angular gyrus; SMG, supramarginal gyrus; IPS, intraparietal sulcus; IFG, inferior frontal gyrus; mPFC, medial prefrontal cortex.

Directed modulatory effects (B‐matrix) during the manipulation task further strengthened the above mentioned intrinsic positive couplings (V1/2→LOTC→IPS→SMG→IFG), while—in opposite direction connectivity was weakened (IFG→SMG→IPS→LOTC; see Figure 4b, action nodes = ROI on the left; and Table 3).

The modulatory effects of both semantic tasks (F and V) were not as pronounced, revealing only one significant modulation: from V1/2 to FFG. Interestingly, this (one) modulatory effect was similar in the F and V task (same direction, same effect (positive) and similar effect strength: V: +0.05 Hz, F: +0.76 Hz; see Figure 4c,d, semantic nodes = ROI on the right; and Table 3), indicating that processing different aspects of semantic knowledge exert similar interregional effective connectivity (see Section 4).

Effective connectivity between action nodes and semantic nodes

Overall, the results indicated superiority of models that assumed couplings between semantic nodes and action nodes, namely couplings within the occipito‐temporal lobe (FFG–LOTC) and within the parietal lobe (AG–SMG). Intrinsic couplings between action nodes and semantic nodes revealed a significant (p < .05, FDR‐corrected), but minor positive connectivity (+0.05 Hz): from SMG to AG (see Figure 4a). Interestingly, this connection was strongly weakened by the manipulation task, resulting in reduced effective connectivity from SMG to AG (–0.45 Hz = A‐matrix + B‐matrix) during the task. Furthermore, the manipulation task weakened the connectivity from the semantic node FFG to the action node LOTC (–0.45 Hz, see Figure 4b).

None of the couplings between action nodes and semantic nodes were significantly modulated by the function or the value task according to the pre‐defined FDR‐corrected significance level of p < .05.

All mean coupling parameters across subjects for the intrinsic connections and their task‐specific modulations of the winning model F4M2 are summarized in Table 3.

4. DISCUSSION

The current behavioral and functional imaging results confirm but also clearly extend previous findings on how action knowledge (here: manipulation knowledge) and semantic knowledge (here: function knowledge and value knowledge) about tools is processed in the human brain by exploring the effective connectivity of the underlying neural networks with the help of DCM (Grefkes & Fink, 2011). DCM analyses rest upon the results of the GLM analyses. The current GLM results are well in line with previous imaging studies on action and semantic knowledge about tools. Therefore, we here concentrate on the GLM results with direct relevance for the DCM analysis, and thereby focus the discussion on the novel effective connectivity results (please, see Supporting Information for a more detailed discussion of the activation patterns).

4.1. Brain regions processing action versus semantic knowledge about tools

As expected, the two tasks addressing semantic tool knowledge (function and value knowledge) led to similar activation patterns (see Section 2.1 and Supporting Information). Contrasting these semantic knowledge tasks with the action knowledge task revealed differential activations in the FFG and AG bilaterally and in the mPFC.

In the framework of the initial two stream‐hypothesis (vision‐for‐perception [ventral stream] and vision‐for‐action [dorsal stream; Goodale et al., 1992]), the FFG represents a region within the ventral visuo‐motor stream (Chao, Haxby, & Martin, 1999; Hutchison et al., 2014; Mahon et al., 2007), which, as a whole, supports the recognition of tools and their functions. Recently, Martin and colleagues associated lesions to areas within the ventral stream (here: ATL) with poorer performance in a tool selection task, similar to our tool function task (Martin et al., 2016a). Moreover, (bilateral) activations in the AG and FFG have been shown previously in numerous neuroimaging studies during semantic processing (Binder, Desai, Graves, & Conant, 2009; Chao et al., 1999; Chen et al., 2016; Price, Bonner, Peelle, & Grossman, 2015; Rumiati et al., 2004; Seghier, 2013; Zhang, et al., 2016). Note that the AG was activated by the semantic tasks and the SMG by the manipulation task, suggesting that these subdivisions of the (large) IPL presumably play different roles in the processing of tool knowledge or tool use (Martin et al., 2016a; Randerath, Valyear, Philip, & Frey, 2017).

Finally, note that activation of the mPFC is in line with previous studies on valuation in healthy subjects (Domenech, Redoute, Koechlin, & Dreher, 2017) and particularly in estimating the monetary value of objects (Smith & Milner, 1984).

In contrast, the manipulation task led to activations in the left IPS, left SMG, left IFG and bilateral LOTC. The left IPS, SMG and IFG are parts of a left fronto‐parietal network, hosting action representations. These regions are also activated in healthy subjects during actual tool‐use (Binkofski et al., 1999; Brandi, Wohlschlager, Sorg, & Hermsdorfer, 2014; Yoon, Humphreys, Kumar, & Rotshtein, 2012). The SMG and (a) IPS represent core regions of the ventro‐dorsal stream (Binkofski et al., 2013). Moreover, they are the putative site of parietal action representations (including the skilled actions associated with familiar objects). Lesions to these regions compromise those action representations and lead to impaired tool‐use (Goldenberg, & Spatt, 2009; Martin et al., 2016a; Martin et al., 2016c) and thus tool‐use apraxia (Buxbaum et al., 2000).

The bilateral activation of the LOTC during the manipulation task calls for further discussion. Overall, its activation concurs with the proposed role of the LOTC in action processing (Lingnau & Downing, 2015), and in particular, tool‐knowledge and tool‐associated hand actions (Bracci, et al., 2012; Perini, Caramazza, & Peelen, 2014; Vingerhoets, 2008). Thus, LOTC lesions are associated with impaired action recognition (Tarhan, Watson, & Buxbaum, 2015) and deficits in imitation and pantomime of tool‐use (Hoeren et al., 2014). Nevertheless, it cannot be ruled out that the differential LOTC activation (encompassing EBA) may have—in part—been triggered by differences in visual stimuli (here: hands versus objects/money), since visual presentation of body parts and/or tools activates partially overlapping regions in the LOTC (Bracci, Cavina‐Pratesi, Connolly, & Ietswaart, 2016; Bracci et al., 2012; Gallivan, McLean, Valyear, & Culham, 2013). Within the LOTC, the “body‐selective” EBA and the “tool‐selective” posterior MTG are both also involved in decoding the ‘intention’ to perform a motor act, that is, the planning of hand‐ and tool‐related actions (Gallivan, Johnsrude, & Flanagan, 2016; Gallivan et al., 2013; Zimmermann et al., 2018). Such planning strategy of hand‐ and tool‐related actions was presumably required to resolve the applied manipulation task. Consequently, the most parsimonious explanation for the current (large) bilateral LOTC activation encompassing MTG and EBA is a combination of differential stimulus‐related activation (here: hands) and activation related to action processing (also see limitations).

4.2. Effective connectivity during the processing of action and semantic knowledge about tools

The current results shed light on the effective connectivity during processing of action knowledge and semantic knowledge about tools in the left hemisphere. In the DCM analysis, the dorsal regions (LOTC, IPS, SMG, and IFG) that were activated in the action knowledge task were included as “action nodes,” and the ventral regions (FFG, AG, and mPFC) activated in the semantic knowledge tasks were included as “semantic nodes.”

Note that the presence of similar activation patterns in the GLM analysis does not necessarily imply similar interregional connectivity. Thus, the V task was chosen as a further semantic task in addition to the F task to investigate whether the patterns of effective connectivity generalize across different aspects of semantic tool knowledge.

The winning model F4M2 revealed just one significant modulatory effect (p < .05, FDR‐corrected) for the V and the F tasks (see Figure 4c,d). Nevertheless, this finding is important as such a modulatory effect was found for both semantic tasks (V and F) in the same direction (from V1/2 to the FFG), same quality (positive, strengthening), and with similar strength (V: +0.5 Hz and F: +0.76 Hz). This indicates that different aspects of semantic knowledge about tools not only engage similar brain regions as reflected in activation patterns but also show coherent directed influences (effective connectivity) that one region exerts upon another. Further research is warranted to investigate the effective connectivity patterns for other semantic aspects of tool knowledge.

Moreover, we were specifically interested in the interactions between semantic nodes and action nodes that may reflect an information exchange during semantic and action processing. The finding of interregional bidirectional connections between action nodes and semantic nodes as observed in the winning model F4M2 (see Supporting Information, Figure S1e) corroborates the notion that areas processing semantic aspects of tools interact with the areas mediating action representations. In particular, the winning model featured bidirectional couplings between the semantic and action nodes (notably between FFG and LOTC as well as between AG and SMG). These effective connectivity findings are consistent with previous structural and functional connectivity studies. Anatomic studies revealed bidirectional connections between ventral and dorsal regions in the macaque (Borra et al., 2008) and in humans (Takemura et al., 2016). fMRI studies revealed functional connectivity between ventral areas and parietal action representations (Garcea, & Mahon, 2014; Hutchison et al., 2014).

However, to date, only few studies have addressed how context‐dependent information about tools is exchanged between regions of the semantic and action system. Such studies have mostly focused on the putative role of ventral regions in contributing information about tools to action representations. To this end, different methods were employed. For instance, some fMRI studies utilized the fact that parvocellular channels project principally to the ventral visual stream and not the dorsal visual stream and hence presented stimuli only ‘visible’ to ventral visual stream areas. Utilizing this method, these studies found subsequent transfer of information to parietal action representations—particularly between ventral tool‐selective areas (left FFG and MTG) and the left IPL and IPS (Almeida et al., 2013; Kristensen et al., 2016; Mahon et al., 2013). Further support for this notion was provided by Garcea and colleagues, who showed tool‐selective left IPL responses irrespective of whether the tool stimuli were presented in the right or left visual field. This tool‐selective parietal response that was resilient to contralateral visual field biases suggests that the left IPL received tool information from ventral regions (Garcea et al., 2016). Recently, Chen and colleagues revealed with PPI (psychophysical interactions) and DCM that “toolness” itself exhibited a strong modulation of connectivity between ventral areas (left LOTC extending into MTG) and parietal action representations (left IPS; Chen, Snow, Culham, & Goodale, 2017a). Interestingly, Zimmermann and colleagues lately revealed strong functional and structural connectivity of the EBA with the parietal cortex, suggesting a functional role of the EBA in the planning of goal‐directed actions, possibly contributing information about adequate postural configurations for tool use (Zimmermann et al., 2018).

In line with these studies, processing action knowledge in the context of our manipulation task led to a strengthening of connectivity from the left LOTC to the left parietal cortex (SMG), presumably hosting action representations (see Figure 4b). However, as some ambiguity exists about the causes of the current LOTC activations (encompassing posterior MTG and EBA, see Sections 4.1 and 5), we refrain from making any specific inferences about the particular type of information exchange between left LOTC and left parietal cortex (SMG).

With respect to the functional role of the above described connections between ventral areas and parietal action representations, previous studies suggest that the retrieval of action knowledge (needed for adequate manipulation) relies—at least in part—on the identification of the tool (Garcea et al., 2016; Hutchison et al., 2014; Kristensen et al., 2016; Mahon et al., 2013). Alternatively, the particular contribution of the LOTC to the parietal action representations may involve action planning (Gallivan et al., 2016; Gallivan et al., 2013; Zimmermann et al., 2018).

While overall, models featuring couplings between semantic nodes and action nodes were superior to the other models, the model parameters indicated a weakening of effective connectivity between FFG and LOTC during our manipulation task. In the same vein, connectivity from SMG (action node) to AG (semantic node) was also weakened. This relative weakening of the effective connectivity between action and semantic nodes in the context of our manipulation task might have facilitated the evaluation of different feasible hand configurations for interacting with a given tool (supported by the action nodes) irrespective of the identification of the tool (processed by the semantic nodes). Note that a recent fMRI study (Hutchison & Gallivan, 2016) did not reveal significant functional connectivity between ventral areas and action representations during sensori‐motor and visual‐perceptual tasks. At first sight, these findings seem at odds with the theory that action representations are contingent on semantic processing (Almeida et al., 2013; Chen et al., 2017a; Garcea et al., 2016; Kristensen et al., 2016; Mahon et al., 2013). However, an alternative hypothesis claims that—to some extent—object perception can also be mediated via the dorsal pathway (for a detailed review on this notion see Freud, Plaut, & Behrmann, 2016). In a similar vein, neural activity patterns related to motor‐relevant object properties were found in ventral stream areas (Gallivan et al., 2013; Mahon et al., 2007), providing evidence that the ventral stream is modulated by motor attributes contributing to object recognition (Sim, Helbig, Graf, & Kiefer, 2015). These findings clearly challenge a strict functional segregation of processing semantic aspects and action representations of tools.

Interestingly, while connectivity between action and semantic nodes was weakened, a relative strengthening of information transmission from more caudal to more rostral action nodes was revealed (V1/2→LOTC→IPS→SMG→IFG), while the connections were weakened in the reverse direction (IFG→SMG→IPS→LOTC) suggesting a predominantly caudo‐rostral information flow within the fronto‐parietal tool network.

5. LIMITATIONS

It might be argued that the differences between the manipulation knowledge task on the one side and the function and value knowledge tasks on the other side are related to increased difficulty in processing the hand postures. However, previous studies analogously reported longer RTs for manipulation tasks compared to function tasks (Evans et al., 2016)—some even without the presentation of hands (Garcea, & Mahon, 2012)—suggesting different cognitive processing strategies rather than mere differences in task difficulty. In the same vein, similar activation patterns were found in imaging studies of action knowledge, which did not use hand pictures as visual stimuli (Assmus, Giessing, Weiss, & Fink, 2007; Chen et al., 2016; Kellenbach et al., 2003; Vingerhoets, 2008). Furthermore, computation of term‐based automatic meta‐analyses with the database Neurosynth (http://neurosynth.org) revealed highly similar bilateral activations in the LOTC for the terms “body” or “hands” (based on 443 studies for “body” and 104 studies for ‘hands’). Importantly, however, the term “actions” (currently based on 501 studies) also revealed bilateral LOTC activations. Moreover, the current large bilateral LOTC activations (for the contrast 2 × M > (F + V)) fully encompassed the Neurosynth results for “hands” and “body” as well as those for “actions.” Thus, the mere fact that there are bilateral activations in the LOTC does not necessarily imply that these are driven by visual stimuli (here: body parts/hands), since action processing also leads to bilateral LOTC activations. This renders it unlikely that the current activations for processing of action knowledge can be solely attributed to differences in visual stimuli.

Taking into account that dysfunction of the anterior temporal lobe (ATL) caused by stroke (Martin et al., 2016a) or neurodegeneration (Hodges et al., 2000) leads to pronounced deficits in function knowledge about tools, and that rTMS over the ATL led to longer RTs for functional judgments (Ishibashi et al., 2011), the absence of significant ATL activation in the current study during the function knowledge task may be considered unexpected. However, neuroimaging studies do not consistently show ATL activation during semantic tasks (Rice et al., 2015; Visser, Jefferies, & Lambon Ralph, 2010). The absence of ATL activation in our study might be due to a methodical limitation, since optimal BOLD imaging of the ATL requires the dual gradient‐echo method (Halai, Welbourne, Embleton, & Parkes, 2014). Adopting this method, Jackson and colleagues could recently show that within the semantic network, the ATL showed functional connectivity with the mPFC and the AG, which were significantly activated in the current function and value knowledge tasks tapping on semantic tool knowledge (Jackson, Hoffman, Pobric, & Lambon Ralph, 2016). Further studies combining the current DCM approach with dual gradient‐echo imaging are warranted to shed further light on the effective connectivity of the ATL within the network supporting conceptual knowledge about tools.

6. CONCLUSION

Effective connectivity analyses with DCM revealed interregional bidirectional connectivity in the networks processing action and semantic knowledge about tools as well as relevant couplings between action and semantic nodes.

The two semantic tasks (function and value knowledge) both significantly strengthened the connectivity from the visual cortex to the fusiform gyrus. As this modulation occurred in a coherent manner for both semantic tasks, these findings suggest that different aspects of semantic tool knowledge are not only processed in similar brain regions, but that these brain regions also show similar interregional effective connectivity during this processing.

The most pronounced modulatory effects of effective connectivity patterns were observed for the action knowledge task with a predominantly caudo‐rostral information flow within the parieto‐frontal action network.

While our effective connectivity results certainly contribute to the understanding of how tool knowledge is processed in the human brain, further research is warranted to characterize the effective connectivity during actual tool use and disturbed connectivity patterns between semantic and action nodes in patients suffering from apraxia or semantic dementia.

CONFLICTS OF INTEREST

All authors state that there are no conflicts of interest.

Supporting information

Additional Supporting Information may be found online in the supporting information tab for this article.

Supporting Information Figure_S1a

Supporting Information Figure_S1b

Supporting Information Figure_S1c

Supporting Information Figure_S1d

Supporting Information Figure_S1e

Supporting Information Table_SI

Supporting Information Table_SII

Supporting Information Table_SIII

Supporting Information Table_SIV

Supporting Information Material_Tool

Kleineberg NN, Dovern A, Binder E, et al. Action and semantic tool knowledge – Effective connectivity in the underlying neural networks. Hum Brain Mapp. 2018;39:3473–3486. 10.1002/hbm.24188