Reinhard Gentner, Susanne Gorges, David Weise, Kristin aufm Kampe, Mathias Buttmann, and Joseph Classen

Transcription

1 1 Current Biology, Volume 20 Supplemental Information Encoding of Motor Skill in the Corticomuscular System of Musicians Reinhard Gentner, Susanne Gorges, David Weise, Kristin aufm Kampe, Mathias Buttmann, and Joseph Classen Supplemental Data Comparison of Movement Spaces between Violin Playing, Piano Playing, and Grasping We used principal component analysis (PCA) of hand postures to analyze the underlying movement patterns of violin playing, grasping and TMS evoked movements. For violin playing and grasping we selected static finger postures at randomly chosen time points to obtain a set of hand postures. For TMS-evoked movements, we used the posture at the time point of maximum movement amplitude. The extracted principal component loadings (PCs) describe multidimensional joint correlation patterns between the recorded joints motions (10 sensors, see Fig. S1A and S1B for examples) and can be linearly re-combined to reconstruct the original set of postures. The number of PCs needed to account for >90% of the variability of violin (4.4 ± 0.9, Fig. S1C) and piano playing (5.5 ± 1.2, Fig. S1C) exceeded the number of PCs to describe grasping movements of either violinists (3.4 ± 0.7, P = 0.034) or non-musicians (3.2 ± 0.7, P = 0.002, Fig. S1D) indicating a higher complexity inherent in violin and piano playing movements. In subsequent analyses we focused on four PCs ( GRASP-PC4, VIOLIN-PC4, PIANO-PC4 ) because we found these to describe instrumental playing, and grasping movements to a sufficient level of detail. To assess the spatial similarity between violin playing, piano playing, and grasping PCs, we compared the similarity of the subspaces spanned by the first four PCs within and between groups and conditions. We computed the principal angles (a generalization of the notion of angles between lines) between PCs of two different subjects and counted the number of principal angles >0.9. This number was termed shared subspace similarity (SSD) [1]. The SSD of all possible pairs of subjects (intra-individual comparisons were excluded) was determined. Two movement classes are expected to display high SSD values (maximum overlap: 4 dimensions) if the PCs span overlapping subspaces. We selected the SSD measure, because it is invariant against linear transformations (such as rotations) of the PCs and therefore increases the robustness of the results (for a detailed discussion of the SSD, see [1]). The results are displayed in Fig. S1E.

2 2 Intra-group comparisons of GRASP-PC4 from non-musicians (SSD 2.1 ± 0.6) and violinists (SSD 2.1 ± 0.6) as well as between group comparisons (SSD 2.1 ± 0.6) revealed a similar number of overlapping dimensions (ANOVA, F(2,322) = 0.25, P = 0.776, Fig. S1E, blue). Because features of grasping movements from violinists and non-musicians were similar, we only present statistical data from grasping of violinists (numeric SSD values are shown as dashed lines in Fig. S1E). However, all principal results were similar when GRASP-PC4 from nonmusicians were used. The SSD between different pairs of subjects within-groups (Fig. S1E, red) was different (F(2, 76) = 3.59, P = 0.033, one-way ANOVA). Post-hoc testing showed that this was because the SSD of VIOLIN-PC4 (SSD = 1.8 ± 0.8) was significantly smaller when compared to PIANO- PC4 (SSD = 2.2 ± 0.6, P = 0.012, significant after FDRC) or grasping (SSD = 2.1 ± 0.6, P = 0.042, not significant after FDRC). A second one-way ANOVA revealed different SSD values when comparing the between-groups SSD of VIOLIN-PC4, PIANO-PC4 and GRASP-PC4 (F(2, 171) = 32.21, P < 0.001, Fig. 1SE green). Post-hoc testing revealed that the SSD of PIANO-PC4 with GRASP-PC4 (SSD = 1.8 ± 0.7) was significantly (P < 0.001, significant after FDRC) higher when compared to the SSD of PIANO-PC4 with VIOLIN-PC4 (SSD = 0.9 ± 0.5, P < 0.001, significant after FDRC). Furthermore, the overlap between PIANO-PC4 and VIOLIN-PC4 (SSD = 0.9 ± 0.5) appeared to be small. These results indicate that i) the joint correlation patterns underlying grasping movements of violinists and non-musicians are similar, ii) the joint correlation patterns underlying violin playing are more individual and less task dependent than those of grasping movements or piano playing, iii) the joint correlation patterns underlying piano playing are more similar to grasping than the joint correlation patterns underlying violin playing, and iv) the joint correlation patterns underlying violin playing movements differ from the joint correlation patterns underlying piano playing movements (i.e. they represent different motor skills). Stability of Reconstruction Quality across Factorization Methods To address the concern that the results are an artifact of the extraction procedure, and to exclude the possibility that they depend on the orthogonality of factors, we repeated the analysis for violinists using a different extraction method. We adapted a matrix factorization technique, described previously by Tresch and co-workers [2] in which individual factors are not necessarily orthogonal. This method was modified by allowing the coefficients of the extracted components (the four extracted components were called SYN4) to be positive (indicating extension) and negative (flexion). The results were in line with the results obtained with PCA, with regard to the SSD and reconstruction of violin and grasping movements. The SSD of TMS- SYN4 of violinists and VIOLIN-SYN4 (1.1 ± 0.6) was significantly (P = 0.004) higher than the SSD of TMS-SYN4 of non-musicians (0.9 ± 0.5) and VIOLIN-SYN4. Similarly, the reconstruction quality of violin playing movements using TMS-SYN4 from violinists (R = 0.62 ± 0.10) was significantly (P = 0.001) higher as when using TMS-SYN4 of non-musicians (R = 0.57 ± 0.09).

3 3 Supplemental Experimental Procedures TMS-Mapping TMS-mapping was performed by delivering monophasic magnetic pulses to 36 different scalp sites centered over the right primary motor cortex [3]. At each randomly chosen stimulation site, 15 stimuli were applied at a rate of 0.2 (±10%) Hz while the subjects were completely at rest. The stimulation intensity for recording finger movements was set to 1.3 times resting motor threshold. Resting motor threshold was defined as the minimum stimulator intensity for eliciting motor evoked potentials >0.05 mv in at least 5 out of 10 trials with the coil placed over the optimal scalp site targeting the relaxed abductor pollicis brevis muscle of the left or right hand. The coil position was monitored online and controlled by visual feedback from a neuronavigational system (Brainsight, Rogue Research, Montreal, Canada). The subject s head geometry was registered to template MR-scans from the Montreal Neurologic Institute using anatomical landmarks. Finger movements of the left hand were recorded at a resolution of 0.2 by a -Glove at 2 khz sampling rate [4]. In a subset of 4 violinists, we also recorded in a different session finger movements from the right hand after TMS over the left hemisphere. Sensors considered for analysis were located over the proximal interphalangeal (PIP) and metacarpal-interphalangeal (MCP) joints. For the thumb, the metacarpal-interphalangeal joint sensor and the sensor measuring abduction (ABD) between the index finger and the thumb were considered. Raw data from the sensor glove was offline low-pass filtered at 10 Hz (using a 2 nd order Butterworth filter) and normalized to baseline. Extension movements were indicated as positive and flexion movements as negative excursions [3]. Recording of Voluntary Movements After TMS mapping, grasping movements were recorded by the sensor glove, both in musicians and non-musicians, while reaching (1 trial) with the left hand towards 51 different visualized objects presented in random order on a computer screen in front of them [3]. In musicians, finger movements of the left hand were recorded during playing instrument specific compositions. Eight of the nine violinists played (on their own violin) 20 s of a composition from C. Flesch [5] and a composition from R. Kreutzer [6], typical training exercises for violinists, at a metronome paced speed of 1 Hz. Pianists played segments taken from W. A. Mozart s "Ah, vous dirai-je, Maman", A. Cortot s Grundbegriffe der Klaviertechnik, (p. 68, exercise 1C) [7], and M. Moszkowski s Études de Virtuosité (No. 10) [8] on a stationary piano. Before playing, each subject was allowed to briefly familiarize with the compositions. However, none of them explicitly trained the compositions. While all violinists had at least once played them, none of the pianists recollected having played these particular compositions before. Raw data from the sensor glove was preprocessed as the TMS-evoked movements. Finger tapping of the left index finger was assessed in all groups by counting the number of taps during a 20 s time period. The sum out of two blocks was taken and divided by 40 s to obtain the tapping frequency. Principal Component Analysis Principal component analysis (PCA [9]) was used to represent a set of n, 10-dimensional (number of sensors) static finger postures (10xn matrix P), obtained from grasping movements,

4 4 musical instrument playing, or TMS-evoked movements as weighted linear combinations of m, 10-dimensional principal component loadings (10xm matrix W), according to P = W*C, where C represented the mxn dimensional weighting coefficient matrix. A TMS-evoked movement was represented as a single static posture at the time of maximum angular excursion [3]. Static finger postures during grasping (20 per grasp) and violin playing (200 per composition) were randomly extracted from the corresponding movement time course. The procedure led to similar results if the extraction procedure was repeated or a different number of postures was selected. Reconstruction of Voluntary Movements Violin playing, piano playing and grasping movements were reconstructed by finding the mxn optimal weighting coefficient matrix A to minimize W TMS *A - M, where the 10xm matrix W TMS contained m = 4 PCs derived from TMS-evoked movements. The 10xn matrix M represented the time course of the voluntary movements (n samples). For solving this problem, a standard least squares algorithm implemented in Matlab (The Mathworks, Natick, MA, USA) was used. The reconstruction quality was expressed by the magnitude of the correlation coefficient [10] between M and W TMS *A. Each movement was reconstructed by four PCs of non-corresponding subjects and the mean correlation coefficient across musical exercises/grasped objects was used as a marker of how well the PCs of a particular subject were suited to reconstruct a class of voluntary movements.

5 Figure S1. (A and B) Principal component analysis was used to represent the recorded set of postures as combinations of principal component loadings (PCs). The first two PCs derived from instrumental playing (A), and grasping (B) movements are displayed as bars for each joint (same assignment of numbers to joints as in Figure 2B) in representative subjects. The PC effect of a signal step (±50 added to baseline position) is visualized below. (C and D) The cumulative variance explained by the 10 PCs, extracted from postures adopted during instrumental playing (C), and grasping (D), is illustrated (mean ± s.e.m.). (E) Overlap (mean ± s.d.) of PC space (4 PCs) within and between different data sets quantified as the shared subspace dimensionality (SSD). 5

6 Supplemental References 1. Cheung, V.C.K., d'avella, A., Tresch, M.C., and Bizzi, E. (2005). Central and sensory contributions to the activation and organization of muscle synergies during natural motor behaviors. Journal of Neuroscience 25, Tresch, M.C., Saltiel, P., and Bizzi, E. (1999). The construction of movement by the spinal cord. Nature Neuroscience 2, Gentner, R., and Classen, J. (2006). Modular organization of finger movements by the human central nervous system. Neuron 52, Gentner, R., and Classen, J. (2009). Development and evaluation of a low-cost sensor glove for assessment of human finger movements in neurophysiological settings. J Neurosci Methods 178, Flesch, C. (1987). Das Skalensystem für Violine - Tonleiterübungen durch alle Dur- und Molltonarten (Berlin: Ries&Erler ). 6. Kreutzer, R. (1958). 42 Etüden oder Capriccen, (Frankfurt: Peters Edition). 7. Cortot, A. (1928). Grundbegriffe der Klaviertechnik., (Paris: Salabert). 8. Moszkowski, M. (2000). Fifteen Études de Virtuosité, Op. 72, (Van Nuys: Alfred Pub Co Inc). 9. Joliffe, I.T. (2004). Principal component analysis, 2 Edition, (New York: Springer). 10. Shadmehr, R., and Mussa-Ivaldi, F.A. (1994). Adaptive representation of dynamics during learning of a motor task. J Neurosci 14,