Eleven right-handed individuals (4M and 7F; Mage = 55yrs, SD = 5.8), participated in this research study. All participants had normal or corrected-to-normal vision and did not suffer from any known neurological disorders. Informed consent was obtained following the guidelines established by the local ethics committee and all participants received monetary compensation (10€/h) for their participation, regardless of completion. The ethics protocol for this study was approved by the Clinical Research Ethics Committee (CEIC-Parc Salut MAR) of the Pompeu Fabra University-Hospital del Mar, Ref. #2015/6085/I. All methods were carried out in accordance with relevant guidelines and regulations.
Experimental Setup and Apparatus
Participants were situated in an electrically shielded room of the neuroscience laboratory (Center for Brain & Cognition, Pompeu Fabra University), where the task was performed and ocular and electro-encephalography (EEG) was registered. The experimental began after electrode preparation. The participants were seated in a comfortable chair, facing the experimental table, with their chest approximately 10cm from the table edge and both lower arms resting on its surface. The table defined the plane where reaching movements were to be performed. On the same table, approximately 60cm away from the participant’s sitting position, we placed a vertically-oriented, 24” Acer G245HQ computer screen (1920x1080). This monitor was connected to an Intel i5 (3.20GHz, 64-bit OS, 4 GB RAM) computer that ran custom-made scripts which controlled task flow, programmed using OpenFrameworks v.0.9.8 software. The screen was used to show the geometrical arrangements and related stimuli on each trial. A small cross (1x1cm), whose position was synchronized with the planar coordinates of the end-point as it slid along the horizontal plane (table), was used to show the participant’s corresponding movement in the vertical plane on the screen.
As part of the experiment, subjects had to respond by performing overt movements with their arm along the table plane. Their movements were recorded with an Optitrak motion tracking system (Optitrak, Inc; Corvallis, OR, USA) sampling at 100Hz, which tracked the position of a spherical marker placed on the nail of the right-hand index finger, as it slid on the table plane. We performed a spatial calibration to synchronize the axes and units for movements on the table Tracked by the Optitrak system with those of our custom-made software. The marker position was synchronized by the custom-made software controlling the task flow. Subjects were instructed to maintain end-point contact with the table surface at all times, and to apply minimal pressure. They were permitted to minimally lift their elbow off the table to diminish this effect. Given that the monitor was placed vertically and movements were performed horizontally (along the table surface), the movement component along the sagittal plane was rotated 90 degrees, to the frontal plane, to show displacements of the end-point on the screen. The transverse movement component was shown unaltered. Data analyses were performed with custom-built MATLAB scripts (The Mathworks, Natick, MA), licensed to the Pompeu Fabra University.
The subject was required to maintain posture at a fixed distance from the table and to place his/her chin on the chinrest. Eye gaze and pupil diameter from both eyes were tracked and recorded with an EyeTribe oculometer (Oculus, Menlo Park, CA, USA), sampling at 60Hz. We used a chinrest to stabilize posture and fix the head position at a predetermined distance from the screen and from the oculometer. The signals delivered by the oculometer were recorded by the OpenFrameworks custom-made code, along with the movement trajectories and other behavioural data. Behavioural, ocular and electro-encephalographic data from each session were transferred to a MySQL community server database (Oracle, Redwood Shores, CA, USA) for further analysis using custom-designed MATLAB scripts (Mathworks, Natick, MA, USA).
The electro-encephalogram (EEG) data were recorded from 60 points on the scalp by using a high-density 64 active electrode actiCap Brain Product, amplified by a Brain Product amplifier (Brain Products BmbH, Gilching Germany). An appropriate head-sized cap was selected, and conductive gel was carefully applied to obtain electrode impedances <5 kW. Impedances were checked every other block to maintain this condition. Eye movements were measured with electrodes attached to the infraorbital ridge and on the outer canthus of the right and left eyes. Electrode impedances were kept <5 kW, and we used a central electrode voltage reference. The electrophysiological signals were filtered with a bandpass of 0.1–100 Hz and digitized at a rate of 500 Hz. EEG recordings were performed by an independent computer (Intel i5; 3.20GHz, 64-bit OS, 8 GB RAM) running Brain Vision on Windows XP. External pulses, generated by the custom made OpenFrameworks code, were used to synchronize the recordings from both computers on a single trial level.
The participants performed a decision-making task between two reaching movements, seeking precision at target arrival. Maximum precision is attained when arriving at the centre of the wide side of one of both rectangle targets. To provide a positive metric of precision other than the error, we defined a metric of reward that decreased linearly with the error (FIG 1B). During the experiment, we varied three factors: motivation, biomechanics, and the requirement of stopping at the target (FIG 1B-C). The parametrization of the two first factors is explained in the sections Manipulation of Internal Motivation and in the section Manipulation of Biomechanical Factors. The constraint of stopping was implemented by defining two motor control regimes: stop-in, instructing the subject to stop at the target, and cross-over, instructing the subject to punch through it and to stop whenever afterwards.
There were twelve experimental conditions, as a function of three main experimental factors: two geometrical arrangements (T1-Major, T1-Minor; FIG 1C), three motivated states (0-Solo, 1-Easy, 2-Hard; FIG 1D), and two endpoint control conditions (Cross-Over/Stop-In; FIG 1C). According to these, we designed 10 blocks of 108 trials each; blocks 2, 4, 6, 8 and 10 were Cross-Over, and blocks 1, 3, 5, 7 and 9 were Stop-In. Furthermore, blocks 1 and 2 were of type Solo, 3, 4, 7 and 8 were of type Easy, and 5, 6, 9 and 10 were of type Hard. At each block, the number of T1-Major vs T1-Minor trials within block was the same (50 each); we incorporated 8 single-target trials per block to encourage the participants to equally experience all movements. For each participant, we interspersed the 8 non-Solo blocks into two sessions, maintaining 2 Cross-Over and 2 Stop-In per session, and including blocks 1 and 2 in both sessions to have a baseline motivated state for each session. Both sessions were performed in consecutive days for each subject.
Real-time, visual feedback of hand position was provided throughout the trial by a 1cm cross displayed on the screen. Its position was synchronized with that of the participant’s right index finger sliding on the table. The time-course of a typical trial is shown in FIG 1A. Within each trial, both potential trajectories were defined by the origin cue (pale blue dot; radius 1cm), a via-point (pale-blue dot; radius 1cm), and a target (dark-blue square; side 10cm, depth 1cm --- FIG 1A). Each trial began when the origin cue was shown on the screen (FIG 1A) --- the Time Stimulus Onset (tSO). Then, the subject slid his/her right-hand index finger into the origin cue. After holding position there for a 500ms --- Origin Hold Time (OHT), the geometrical arrangement (via-points and targets) defining one or two potential trajectories was shown. After a 1000ms Observation Time (OT), a GO-signal was given (origin cue disappeared). The instruction was to react as fast as possible, to freely choose an action, and to be as precise as possible by sliding over the via-point and towards the target while maintaining the index fingertip in contact with the table surface at all times. 100ms after entering the target, the participant was informed of the precision attained by a green bar. The bar length ranges between 0 and 900 screen pixels, and its size is scaled between zero if the error is equal or larger than the side of the rectangle, and 900 if the error is zero. Any intermediate value scales the bar down linearly. If accompanied by a simulated partner, a red bar is also shown, indicating the performance of your partner in the same trial. Finally, the ranking of the participant vs the partner is shown each nine trials during 500ms, after the precision bars. Subjects gave informed consent to be photographed for this purpose only. Their photograph was shown alongside that their partner on a vertical axis of precision. No photographs were kept post-experiment.
Manipulation of Motivation
Our manipulation of motivation was performed by means of a social hierarchy defined within the experiment, as a function of the participant’s aiming skill with regard to that of simulated partners. We informed the participant that he/she would perform within a community of participants, and that at each block, he/she would have a different partner from this community to perform alongside. To reinforce the belief of a real partner, at the end of each trial we showed a horizontal green bar displaying the accuracy attained, normalized between 0 and 100%, alongside a red bar displaying the accuracy of their partner. Each nine trials we also showed a photograph of the participant and of their partner, ranked in a top-down fashion, at a height proportionally to their average end-point accuracy. Since the goal of the simulated partner was to introduce a subliminal bias that would modulate the participant’s intrinsic motivation, we informed each participant this was not a task of competition, that the partner was for companionship purposes, that the ranking was only to keep track of your performance, and that the participant should focus only on performing the task and not on outperforming their partner. To parameterize the bias introduced by the presence of the virtual partner, we classified partners of two types, as a function of their worse of better aiming accuracy with respect to the participant (FIG 1D). We also matched the participant and the partner gender to avoid cross-gender effects.
Despite the instruction, we predicted that participants will be concerned by their accuracy with respect to their partner, and that they will adjust their process of selection of motor parameters and/or their policy to select between movements. However, this could be attained by simply trading-off speed by accuracy, or by varying one of these two metrics in a cost-benefit scenario. The influence of the motor cost on the choices was varied in a random fashion on a single trial basis, as we used one of two geometrical arrangements (FIG 1A). By contrast, the social position and the control regime were maintained constant during each block and varied across blocks (10 block types), presented in a counter-balanced fashion. Each participant performed two sessions of six blocks of one-hundred and eight trials each. Each session included six blocks, three per control regime; one block solo, and two blocks alongside both kinds of partners. Subjects were given real-time visual feedback on their trajectories.
Criteria for a trial to be considered faulty were: if the position of the endpoint left the origin before the GO signal, if the reaction time (RT) was shorter than 200ms or longer than 1000ms, or if the stylus reached the target before first crossing over the via-point. Each individual trial was shown to the subject a single time. In other words, even if a trial were faulty, the sequence of trials resumed to present the subject with the next trial. Error trials in which the subject crossed the via-point (where a significant part of the trajectory was however performed) but failed to enter the target or failed to stop at the target during the THT time were disregarded from further analysis. Visual feedback was provided during the movement by showing the end-point position as a small cross in real time. Furthermore, the colour of the via-point and target cues changed to green as tip of the small cross slid over them.
Each end-point trajectory was first low-pass filtered at 10Hz. We then identified and quantified the following kinematic markers for each individual trajectory: Peak Velocity (PV), Time-to-Peak Velocity (TTPV), Movement Time (MT), Peak Acceleration (PA), Time-to-Peak Acceleration (TTPA) both for the stop-in and cross-over regimes, and Peak Deceleration (PD), and Time-To-Peak Deceleration (TTPD) for the stop-in only regime --- FIG 1D, and were converted into z-score for statistical analyses (for each subject individually), and performed a Greenhouse-Geisser correction for non-sphericity. As a function of our hypotheses of control for each control regime, we then fitted a generalized linear model (GLM) on a single subject basis for several of these metrics. For the stop-in regime this comprised the late kinematic markers, involved in the control of stopping: PV, TTPV, PD, TTPD; and for the cross-over regime the early kinematic markers: PA, TTPA, PV, TTPV, as target arrival occurred around peak velocity. The GLM included the following variables: motivated state [0, 1, 2], #Block [1, 2, 3], #Trial (within Block --- [1-108]), Biomechanical Cost (Biomechanical Arrangement --- 1|2 (T1-Major|T1-minor) x Choice --- T1|T2) and interactions between them, as defined by equation 1. Note that the BM×C, was define to group low-cost and high-cost movements within arrangement (0-Low Cost/1-High Cost), and to ultimately quantify the influence of motor cost on our metrics.
For each explaining factor, we identified the regression coefficients b that significantly differ from 0 (as reported in FIG 2A-B and Suppl. FIG 1A-B) by running a Mann-Whitney-Wilcoxon (rank sum) test between the distribution of b pooled over all subjects and a null distribution. The null distribution was built by shuffling the regressed values fk across trials within each subject, then pooling the surrogate coefficients of each type across all subjects together. Note that the z-scoring applied to the kinematic markers allows for the alignment of the distributions across the different subjects. In this way we identified the explaining factors (and interactions thereof) that affected the kinematic markers (as quantified by the regression) to compare the effect of motivation with other expected factors like habituation or fatigue (#Block and #Trial) and biomechanical constraints (BM). We set the threshold for group significance at p<0.05 with the Mann-Whitney-Wilcoxon (rank sum) test to compare the distribution of regression coefficients estimated from the data and the corresponding null/surrogate distribution.
Like the movement markers, we also z-scored our recordings of pupil diameter and regressed this metric with the same GLM used to test behaviour (Eq. 1). We calculated the GLM b-coefficients per subject at each 30ms during two intervals, centred around the tSO and at the tGO events (FIG 4C-D). We determined group significance for each variable by running a t-test across the b-regression coefficients obtained across subjects (FIG 4A). Again, the threshold for statistical significance was set at p<0.05 (without correction for multiple comparison, considering each time step to be independent).
We selected an interval of interest of 1200ms for further analyses, starting 800ms before the first stimulus onset --- the initial cue, and ending 400ms later (see Fig. 1A). We selected this interval for two main reasons, first to seek for baseline changes in the network of motivation, which should not depend on each specific trial but on the block of trials --- the partner was the same during each block of trials, and so should the modulation of motivation. Second, this interval preceded any movement, and was therefore likely to contain fewer artifacts than those following the stimulus presentation.
We applied a 4th order notch filter around 50, 100 and 150Hz to prevent electrical interference from the power supply line and a 4th order bandpass Butterworth filter between 0.1 and 100Hz. Electrodes with EEG level exceeding either 200V or voltage step/sampling 50V within intervals of 200ms were removed off-line. Baseline was corrected, and the datasets were z-scored by using the recordings during trial block types 1 and 2 of each session --- when the participant was playing solo, as a reference for baseline motivation.Eye related artefacts were removed by means of independent component analyses (ICA), implemented in with custom-made Field-trip open-source toolbox (www.fieldtrip.com) and EEGLAB scripts (http://sccn.ucsd.edu/eeglab, UC San Diego, CA, USA). The procedure to identify eye-movement related sources was semi-automatized, first correlating each source obtained with the signal from the electrodes recording eye movements to obtain a first metric of relatedness. Second, we visually inspected all sources to corroborate that their shape and spatial location matched those of ocular artefacts. Eye-related sources were removed and the cleaned signal obtained by inverting the ICA process.
Extraction of Neuronal Signals (Source Space)
The pre-processed EEG datasets were transformed from electrode into source space via custom-made MATLAB scripts based on the default EEGLAB ICA algorithm in combination with the electrode spatial location map. The Brain Products Unicap 64 configuration was used to establish a spatial reference between the electrode placement and the sources location, and we assumed a spherical head model.
EEG Analyses in both Electrode and Source spaces
We used cross-validated classification to assess how much information about motivation is present in the EEG signals. Following the task description, we defined three motivated states as a function of the types of partner the participant performed the task with: Solo (M=0), when alone; Easy (M=1), when performing with a partner of lesser skill than the participant; Hard (M=2), when performing with a partner more skilled than the participant (see behavioural task, FIG 1D). The signals, in electrode or source space, were first filtered in three frequency bands: a [8-12Hz], b [15-32Hz], g [40-80Hz]. We cut out our interval of interest from the EEG temporal series, from 800ms before the first stimulus onset, until 400ms after (yielding 1200 time points at a resolution of 500Hz), i.e. at the tShowTarget event (FIG 1A). We applied our classification pipeline to each recording session, type of signal and frequency band. This pipeline relied on two specific metrics to capture complementary aspects of the brain network of motivation: signal power and pairwise correlation between signals (as a proxy for interactions, which is equivalent to spectral coherence averaged over each frequency band used here). The classifiers were trained to predict the Motivation Category with a train set (entailing 80% of the trials) and to yield a performance with a separate test set (the remaining 20% trials). This procedure is repeated 100 times with distinct train/test sets obtained by randomly splitting the 1296 total trials (432 trials / Motivation Category), to comply with the nowadays standards for proper assessment of the generalization capability of classification 67.
The code for classification consists of custom-written (open-source) python scripts with the numpy, scipy and scikit-learn libraries . We used two classifiers: multinomial linear regression (MLR) and 1-nearest-neighbor (kNN with k=1), in a similar manner to a previous study for fMRI data . It followed nowadays standards for proper assessment of the generalization capability of classification procedure . Each classifier performed a distinct mapping from features to motivated category. The MLR calculates a linear weighted sum of the entries, which is then rectified by a sigmoid function, to calculate an output ranging from 0 to 1 for each category. This indicates the confidence of the classification in the predicted category; the decision is made by selecting the category whose confidence is the highest. The training procedure tunes the weights (or regressors) for all features to reduce the prediction error on the train set. By contrast, the 1NN calculates a similarity measure between the sample in the test set and all samples in the train set (using the Pearson correlation between the two vectors of features) and attribute to the test sample the category of the most similar train sample. Here the 1NN does not involve training.
To identify the most informative (or “best”) features for the classification, we used Recursive Feature Elimination (RFE) on the MLR; because the MLR gives a better performance than the 1NN. In brief, its strong weights (in absolute value) correspond to important features (also because the MLR gives a better performance than the 1NN). RFE provides a ranking of features by their importance by iteratively excluding features that weakly contribute to the classification and optimising the remaining weights. Ultimately, the most relevant features survive the pruning process. In practice, we performed RFE for each train set (in each train/test split) and then evaluate the stability of the feature rankings by calculating the Pearson correlation coefficient between all pairs of rankings. In the case of stable ranking (typically Pearson correlation coefficients above 0.6), the mean ranking over all train/test splits is calculated. This gave two sets of best features: electrodes for power and interactions between pairs of electrodes for correlation (FIGS 5-6 and Suppl. FIG 2).
The common sources were calculated for each participant using the power of source signals. Following the heuristic that the eight best sources obtained using RFE were sufficient to attain an over 90% accurate classification (FIG 6B), we pooled the eight best sources over all sessions of all subjects, over both motion types. We used the Girvan-Newman algorithm  to calculate communities of similar sources (spatially distributed over the cortical surface) among the original 25-45 sources obtained per subject. Similarity between pairs of sources was established using the Pearson correlation over the vectorized head maps, after retaining only the more localized part (similar to a spatial rectification to favour narrow bumps). Once the sources were grouped in communities, the centroid head map of each community was calculated by averaging all corresponding sources. We performed this operation independently for each frequency band, yielding six communities of common sources for the a, six for the b-band and five for the g-band.