Subjects
Twenty-two healthy right-handed subjects, 10 males and 12 females ranging in age between 42 and 84 years with a median age of 59, were included in the study. The handedness score according to the Edinburgh Handedness Questionnaire [24] ranged between 50 and 100 with median 100. The subjects had no prior history of psychological disorders, achieved normal Mini-Mental State Examination (MMSE) scores, and exhibited no pathological findings in the T1-MRI brain scans. Demographic data are shown in Table 1. The study received ethical approval from the Kantonale Ethikkommission Bern (KEK), 3010 Bern, Switzerland. Prior to the study all participants gave written informed consent before enrolment, according to the Declaration of Helsinki [25].
Table 1
Demographic subject data.
ID | gender | age (years) | LQ | MMSE |
1 | m | 74 | 90 | 27 |
2 | f | 73 | 60 | 29 |
3 | m | 42 | 100 | 30 |
4 | f | 48 | 60 | 29 |
5 | f | 65 | 50 | 27 |
6 | f | 71 | 100 | 28 |
7 | m | 47 | 100 | 30 |
8 | f | 52 | 100 | 30 |
9 | f | 59 | 90 | 29 |
10 | m | 53 | 100 | 26 |
11 | m | 54 | 70 | 29 |
12 | f | 47 | 100 | 30 |
13 | f | 51 | 100 | 29 |
14 | f | 56 | 100 | 29 |
15 | f | 59 | 80 | 30 |
16 | f | 69 | 100 | 28 |
17 | m | 84 | 100 | 28 |
18 | m | 83 | 80 | 29 |
19 | f | 69 | 90 | 29 |
20 | m | 75 | 100 | 28 |
21 | m | 68 | 100 | 27 |
22 | m | 71 | 100 | 29 |
N or Median | 10 m / 12 f | 62 | 100 | 29 |
Range | | 42–84 | 50–100 | 26–30 |
(m, male; f, female; LQ, laterality quotient; MMSE, Mini-Mental State Examination)
Sensori-motor assessment
Sensori-motor function was assessed with five measurements for both left and right hands: (1) Power grip was calculated from the average of three power grips using a Jamar hydraulic hand dynamometer [26]; (2) Averaged over three trials, precision grip was measured while applying pinch force between thumb and index finger at the groove of a Jamar hydraulic pinch gauge [26]; (3) Motor hand skill of each hand was determined using one of the seven timed subtests comprising the Jebsen-Taylor Test (JTT), namely, “Picking Small Objects” (PSO) in which subjects grasp six small common objects (two each of paper clips, bottle caps and coins) and drop them into an empty can; (4) Two-point discrimination (2PD) was measured using a graded caliper [2-point Discriminator, Medwork Instruments, Vancouver, Canada] on the index fingertip [27]; and (5) tactile object recognition (TOR) was tested using a standardized protocol employing 30 everyday objects as previously described [28]. The assessments were intended to confirm normal sensori-motor abilities in the subjects; they were not incorporated in analyses of the task.
Data glove instrumentation and calibration
We employed the VMG 30™ data glove from Virtual Motion Labs [Virtual Motion Labs, LLC., 3010 LBJ Freeway, Dallas, Texas 75234 (see www.virtualmotionlabs.com)]. The glove is equipped with 29 sensors of which 16 are bend sensors less than 0.35 mm in thickness. Two finger bend sensors per finger for measure the movement extent in the metacarpo-phalangeal (MCP) and proximal interphalangeal (IP) joints, and two finger bend sensors at the thumb measure movement extent in the MCP and IP joints. Four sensors between the fingers measure abduction. One palm arch sensor detects spatial configuration related to the proximal and distal transverse arch of the hand described by Hertling and Kessler [29]. One thumb cross sensor detects the complex movement of the thumb during finger opposition at carpo-metacarpal (CMC) joint. Five sensors situated at the finger tips measure pressure and eight sensors measure hand and wrist orientation (Figs. 1A, 1B). Calibration of the data glove consisted of seven calibration stages: (1) maximal simultaneous flexion and extension of all fingers and thumb simultaneously at a frequency of 1 Hz, (2) alternating maximum adduction and abduction of all fingers at a frequency of 1 Hz, (3) maximal transaxial extension of the thumb (including an associated inward rotation) related to CMC joint to the little finger, and maximal flexion of MCP and PIP joints in all fingers. Finally, one opposing movement of the thumb to the index (4), middle (5), ring (6) and little (7) finger pad, forming an “O” between the thumb and fingers were carried out. Bend sensors were calibrated between a value of 1000 (flexion in the MCP and PIP joints, adduction of fingers, transaxial extension of the thumb and forming the palm arch) and of 0 (maximal extension in the finger joints, abduction of fingers, resting position of the palm arch and CMC joint of thumb). Finger pressure sensors were calibrated between a value of 1000 for no pressure and 0 for maximum pressure.
Task performance
The sensori-motor task consisted of regular single motor acts at a frequency of 1 Hz in which the opposed thumb and fingers of one hand surround the cuboid in a continuous and regular action. The almost identical axes of the cuboid avoid the distraction and reorientation induced by significantly different axes. Consecutive steps of the motor act as displayed in the instruction video are depicted in Fig. 1C, which shows successive phases of the thumb and finger trajectories. Each phase begins with a transaxial movement of the thumb versus the ring finger. During the concerted action of thumb and fingers in the workspace, the thumb exerts tangential forces that produce a marked rotation of the object, anticlockwise in the right hand, and clockwise in the left. In the terminology of Bullock et al. [30], 1) the action is prehensile, 2) the stabilizing fingers change continuously during one motor act, 3) the cuboid moves, guided by the tip of the thumb, relative to the contact points of the virtual fingers [19], 4) thumb and fingers move relative to the reference frame defined by the hand base, and 5) the motor sequence of fingers and thumb is repeated at the given frequency.
The cuboid was made of granite with a weight of 29.9 gram and side lengths of 22.54 × 22.54 × 22.57 millimeters resulting in a total volume of 11.5 cm3, comparable with those of the aluminum cube used in [14]. The density of 2.6 g/cm3 was also comparable to that of aluminum, 2.75 g/cm3. A video was filmed to instruct the subjects how to perform the task. This video consisted of three, 20 seconds long, consecutive segments: (1) fixation, (2) observation, (3) active manipulation, each announced by written instruction on a blank white screen for 4 seconds. “Fixation” showed a hand holding the cube; “Observation” showed the same hand manipulating the cuboid at the prescribed 1 Hz; and upon “Active manipulation” the subjects were given the cube by the study physician and requested to manipulate the cube at the required speed as shown in the video sequence displayed during the segment “Observation” on the screen. A right hand was shown for the right hand sensori-motor task and a left hand for the left hand sensori-motor task. The 3 segments were repeated six times showing 3 male and 3 female hands and resulting in a total video length of 7.2 minutes. In- house software recorded the sensor data only during the 20 seconds of active manipulation.
During task performance, subjects were seated at a desk on which was placed a computer screen with their hands supine on the desktop. The motor task was explained by the study physician, and subjects requested to manipulate the cuboid with the left and right hand without the data glove for about 10 seconds as shown by the physician. This procedure ensured that subjects understood the task. Then the calibrated data glove was put on the non-dominant left hand of the subject and checked for fit by the physician. The video was started when the subject's hand was relaxed on the table top. When the instruction “Active manipulation” appeared on the screen, the physician placed the cuboid in the subject's hand; after completion of the segment, the cuboid was removed. The glove calibration procedure required a break of about 2 minutes between acquisitions with the left and right hand.
Data sampling
Data were acquired with software programmed in house and based on the Software Development Kit provided by Virtual Motion Labs. Pre-study testing of the signals produced by the task indicated that they could be most efficiently encoded at a frequency of 50 Hz, implying time frames of 20 msec. This frequency appears sufficient in reference to the published critical thresholds of about 20 Hz for steady visual perception and 10 Hz for visual parsing [31]. One action of consecutive manipulations is denoted a run and consisted of 1000 time frames. In order to i) exclude irregularities as the subject adjusted to prescribed frequency of manipulation observed in the instruction video and ii) to impose a standard number for subsequent analysis, only the last 800 time frames, i.e 16 sec, of each run were analysed.
Data analysis
All nineteen sensor time courses of each run and subject reflecting prehensile in-hand manipulation were submitted to principal component analysis (PCA, see results): all ten finger bend sensors, all four ab/adduction sensors, and the two sensors describing the deformation of the palm (palm arch, thumb cross); additionally three pressure sensors at the finger tips 1 to 3 mainly involved in the manipulation task.
Separate analyses were performed for each hand. PCA was performed using in house software written in Matlab [The Mathworks, Inc., Natick, MA] based on the algorithm described by Alexander and Moeller [32]. The sensor amplitudes for each sensor in the 800 time frames were entered in a matrix. The rows corresponded to the 800 time frames and columns to the 19 relevant sensors of a run. PCA is applied to a residual matrix. To compute this matrix, (i) the mean of sensor amplitudes for each time frame and (ii) the mean amplitude for each sensor of all time frames are subtracted from each matrix element, and (iii) the grand mean of sensor amplitudes in the original matrices added. The row, column and grand means of the resulting residual matrices vanish. Using the singular value decomposition implemented in Matlab, the residual matrix was then decomposed into 19 principal components (PC). Each PC consisted of a sensor expression pattern, a time course and an eigenvalue. The sensor expression coefficients describe the amount each sensor contributes to the component. The time course represents the variation of the component with time and the eigenvalue characterizes the fraction of variance described by each component. The sensor expression coefficients and time courses of a PC are orthonormal and range between − 1 and 1; the orthogonality reflects the lack of statistical correlation among the principal components.
Preliminary analysis showed that the first three PCs of each run and subject explained about 75% of the variance, a number consistent with the Guttman-Kaiser criteria for salient PC [33]. Further analysis was therefore restricted to these first three PCs.
Spatial sensor patterns
Statistical analysis of the sensor expression coefficients must take into account the indeterminacy of the signs associated with multilinear models such as PCA [34] i.e. two different sets of coefficients expressing the same pattern might differ only in the signs of the sensor contributions. Before analysis of the subject cohort, alignment of the expression coefficients is therefore necessary. Alignment was performed in two stages. First, pairwise correlations of the expression coefficients were computed for the six runs of each subject and PC and the signs adjusted to yield the highest positive correlation. Second, the realigned expression coefficients of the 22 subjects were submitted to a second pairwise correlation analysis using the most favorable alignment among subjects, i.e. highest correlation, as a standard to determine the relative signs among subjects. Based on the two steps of calibration procedure and preliminary analyses of the principal components, we then assigned a positive sign to highest correlations as reflecting increased bending of the thumb cross, MCP and/or PIP finger sensors. Thus, sensors yielding prominent positive signals indicate bending movements or pressure synchronous with the selected finger sensors. Sensors yielding prominent negative signals indicate that the bending or pressure are out of phase compared to sensors exhibiting a positive sign, but with the same time course.
In order to assure the homogeneity of the component expression coefficients for the complete cohort, k-means clustering was applied to the 3 PCs in all 132 runs and subjects, i.e. 6 runs for the 22 subjects. An iterative method for partitioning data, k-mean clustering yields mutually exclusive clusters after determining their central members. Therefore, each expression coefficient is assigned to a cluster and its distance to the central member, denoted centroid, is computed. Homogeneity of the coefficients would imply that the clusters should correspond to the rank of the PC in explaining the variance of the coefficients, i.e. the PCs explaining the greatest variance would compose one cluster, the PCs explaining the second greatest variance a second cluster, and so on. To be consistent with the number of PCs considered in each run, we limited the number of clusters to three. We implemented the clustering using the program k-mean of Matlab. The distance between centroid and cluster member was computed using the option “correlation”, as suggested by the alignment procedure.
In order to evaluate the salience of the individual sensors in the task, medians, percentiles and confidence levels for the correctly identified component expression coefficients were computed and compared with the centroid. Correctly identified coefficients are those for which the PC is labeled as belonging to its corresponding cluster, i.e. the dominant PC, PC1, of a particular run and subject is correctly identified if it is labeled as belonging to the cluster characterized by a predominance of PC1's. To confirm the salience of individual sensors, a Kruskal-Wallis test of the sensor distributions, corrected for multiple comparisons of ranks, was performed using the Matlab programs, kruskalwallis and post-hoc multicompare.
Temporal sensor patterns
To investigate the temporal properties of the PC clusters, frequency spectral analysis was applied to the time courses of correctly identified PCs. In addition, time delays between PCs for each run and subject were computed using the Matlab program finddelay. The sampling frequency of 50 Hz determined the maximum delay of 25 frames in the program, corresponding to one half of a sampling cycle.
In addition to the PCA, the frequencies and time delays among twelve individual sensors for all runs and subjects correctly assigned to Cluster 1 for both hands were also analysed; the sensors comprised the ten finger bends (i.e. related to MCP and PIP joints) and thumb cross (i.e. related to CMC joint and palm arch sensors). To reduce the noise in the time courses due to the discontinuous signal, the time courses were first filtered using a finite impulse response (FIR) filter with low pass cutoff frequency of 10 Hz. To achieve similar gain levels, they were normalized such that the magnitude of the maximum amplitude was unity. This preprocessing was implemented using the Signal Processing Toolbox of Matlab. The frequencies were determined by the time difference between signal maxima using the Matlab program findpeaks, Matlab. The time differences between minima and null positions confirmed the frequencies. The delays were limited to maximum delay of 25 frames as above.
Graph analysis of selected sensor time series
Using the same time series of the 12 MCP/PIP finger bend, palm arch and thumb cross sensors included in the PCA and cluster analyses, we performed graph analysis with GraphVar (Release V2.01) [35] as implemented in Matlab. Restricting to runs for which PC1 was assigned to the associated cluster, the analysis required first calculation of the 12 × 12 Pearson correlation matrices for 98 runs of the right hand and 105 of the left hand. From these were calculated mean matrices yielding a weighted undirected graph with 12 nodes and 66 edges for each hand. Negative weights, corresponding to negative correlations, were retained. To investigate subnetworks, the graphs were thresholded in steps of 0.05 for positive and negative weights. Global efficiencies for both graphs were calculated without thresholds. A null model network consisting of 100 random fully connected weighted graphs generated with 1000 iterations served as basis for comparison using the Mann-Whitney-U-Test. Finally, the graphs were submitted to the Louvain community detection algorithm [36] s implemented in the brain connectivity toolbox [37] using a gamma of one in order to determine the modularity of the graphs.
Temporal evolution of finger movements in space
To complement the group PCA and temporal analysis of individual sensors and underpin their understanding, we acquired 3D data for male subject ID 10 in an additional acquisition. The group cluster analysis showed that PC1 of the subject had been assigned to the corresponding cluster in all runs of the right and in most runs of the left hand (Figs. 2, Fig S2). Software provided by Virtual Realities converts the raw sensor data into the C3D file format (www.c3d.org) used in biomechanics, animations and gait analysis laboratories. This format comprises 23 data points representing a standardized 3D hand model, each consisting of x, y, and z values in millimeters. Because the finger tips play a central role in the task, we focused on the five data points representing the end of the distal phalanges to calculate spatial finger trajectories and average speed. A trajectory was defined as the points between consecutive maximal extensions of the thumb derived from repeated manipulations, as determined by the program findpeaks of Matlab.
As in group acquisitions, the data were acquired as 6 runs of 16 sec each. However, the sampling rate was 36.97 Hz, the prescribed rate for the 3D acquisition mode. Since the time between maximum extensions of the thumb varied, the number of trajectories was less than optimum: 80 for the right hand and 75 for the left. The speed of finger movement were then computed by dividing the path length of the trajectory by its duration. From the ensemble of trajectories for each hand were calculated a mean trajectory and 95% CL (confidence level) region, the latter using an error ellipsoid for each time point with the Matlab program error ellipse (https://ch.mathworks.com/matlabcentral/fileexchange/4705-error_ellipse). For visualization, the trajectories of approximately 37 frames were resampled to 100 frames and the mean trajectories and CL region displayed (Video 1) using the open source software Mokka version 0.6.2 (https://biomechanical-toolkit.github.io/mokka/).