Ever since the 1937 publication, the proposed motor homunculus is standard in our textbooks though more recently the evidence accumulates that this picture oversimplifies reality. With my work bundled in this thesis, I added to this evidence and challenged the mere one-to-one mapping between M1 and muscles. In contrast to Penfield and Boldrey (1937), I opted for a non-invasive, secure, and painless brain stimulation method, namely transcranial magnetic stimulation (TMS) combined with neuro-navigation (nTMS). I employed nTMS to map the cortical representations of eight muscles in the hand and the forearm.
Using data of N=20 subjects, I performed test-retest assessments to first establish the reliability of my procedure. I chose the size of excitable areas in M1 and their centroids as primary outcome measures and estimated the corresponding intraclass correlation coefficients. All the ICCs values demonstrated good reliability in most muscles. Random positioning of the TMS, essential to my approach, yields a substantial number of stimulations that do not elicit muscle activity, that is, motor-evoked potentials (MEPs) cannot be observed. To my surprise these non-MEP stimulations have been ignored in the literature potentially causing an overestimate of the (size of the) excitable area in M1.
In the analysis of the overlap between cortical representations of different muscles. For this, I introduced how to warp the subject-specific neuroanatomy and the TMS stimulation parameters to a standardised template. Motivated by previous research that indicated the presence of overlapping cortical representations, I used this (in the field of TMS novel) approach and examined the overlap of representations within functionally relevant muscle groups (within-hand, within-forearm, and between hand and forearm). Using again an ANOVA with repeated measures, the synergistic muscle pairs turned out to have a significantly larger overlap in cortical representation than their non-synergistic counterparts. This agrees with my overall hypothesis that muscle synergies in the upper extremities are partly represented in the cortex.
I devoted Chapter 4 to the quest of predicting the presence of MEPs when merely using the TMS setup parameters, stimulation intensity and position and orientation of the TMS coil. The fairly large number of stimulations in my experiment (more than 13,000) allowed for employed different supervised machine learning approaches. When predicting data within a subject, a subset of the subjects’ data served as a training set with which the machine learner could be (cross)-validated before testing on remaining subset. For the between-subject prediction, I used the data from all subjects for training and validation. The model with highest predictive capacity was a decision tree-based bagging ensemble. In the within-subject design it reached a maximum accuracy of 90% with an average of subject as high as 77%. Possible reasons for this arguably poor performance are the yet limited sample size next to a pronounced between-subject variability that was manifested is substantial differences in the predictor importance between subjects. Still, I believe that machine learning we help simplifying the motor mapping process in experimental settings and will eventually enter clinical assessments, as discussed in the final Chapter 5.