TMS-based neurofeedback facilitates motor imagery of different hand actions
Poster No:
2064
Submission Type:
Abstract Submission
Authors:
Hsiao-ju Cheng1, Niccolò Voster1, Chantal Wunderlin1, Daryl Chong1, Ingrid Odermatt2, Nicole Wenderoth2
Institutions:
1Singapore-ETH Centre, Singapore, Singapore, 2ETH Zürich, Zürich, Zürich
First Author:
Co-Author(s):
Introduction:
Non-invasive brain-computer interfaces (BCIs) allow the user to modulate brain activity patterns in a goal-directed manner [1, 2]. To date, most non-invasive BCIs can only decode gross movements while many essential daily-life activities require much finer finger and hand control [3]. We have developed a novel BCI using motor imagery (MI) and transcranial magnetic stimulation (TMS)-based neurofeedback (NF) training with the aim to reinforce representations of complex hand actions in the brain. In this proof-of-concept study, we aim to investigate the utility of such a new BCI for daily-life hand function training via MI.
Methods:
We designed 2 experiments to investigate the effect of the TMS-based BCI on 3 hand actions (Figure 1). In experiment 1, the desired corticomotor excitability pattern was derived from the muscle-specific mean MEPs of 6 right-handed healthy adults (data not shown). NF was provided by displaying the target lines at 3 different heights (high, medium, low) with boundary boxes indicating the desired extent of corticomotor excitability. In experiment 2, we used a more adaptive ensemble-based support vector machine (SVM) learning approach. All motor execution (ME) data were used to train the first classifier and each block of MI data was used to train a new classifier in the ensemble. NF was provided based on the group decision of classifiers in the ensemble. To assess the participants' performance as to data separability, an SVM classifier was used to decode target hand actions by normalized MEP amplitudes of 3 finger muscles for each block. Leave-one-trial-out cross-validation was used within blocks ("Cross-validation test"). Accuracy values for each fold were averaged for the block. To understand the relationship between ME and MI, the first classifier trained on ME data in each ensemble was tested on each MI block ("Cross-condition test"). Here we report data from 8 participants (age 24±3 years, 5 females) in Exp. 1 and preliminary data from 4 participants (age 32±5 years, 2 females) in Exp. 2.
·Figure 1. Schematic setup of TMS-based NF training.
Results:
We first examined the ME data to understand whether the evoked MEPs in APB, FDI, and ADM allow us to discriminate 3 hand actions. The results showed 97±1 % (mean±SD) cross-validated classification accuracy for Exp. 1 and 84±2 % for Exp. 2. This indicates that healthy adults could generate comparable corticomotor excitability patterns within a hand action and distinguishable patterns between hand actions, suggesting that ME data could be used as "ground truth" for TMS-based NF training. We then probed the MI-NF data to understand whether participants could modulate their corticomotor excitability patterns with the provided feedback. For Exp. 1, the results showed 40±2% of cross-validated classification accuracy; whereas for Exp. 2, the accuracy was significantly increased to 57±6% (t10=-7.622, p<0.001). This supports that the adaptive approach worked better than the deterministic approach for training self-modulation of corticomotor excitability patterns of 3 finger muscles.
In Exp. 2, we investigated the generalizability between ME and MI (Figure 2). The cross-condition findings suggested a progressive increase in median accuracy with training. We also inspected the unnormalized MEP amplitudes of ME and MI-noNF blocks for every participant. Most participants managed to regulate the muscle-specific MEP with 3 hand actions. With training, the muscle-specific MEP patterns of MI became more similar to those of ME. This demonstrates that NF training could promote the modulation of sensorimotor activities in the brain.
In Exp. 2, we investigated the generalizability between ME and MI (Figure 2). The cross-condition findings suggested a progressive increase in median accuracy with training. We also inspected the unnormalized MEP amplitudes of ME and MI-noNF blocks for every participant. Most participants managed to regulate the muscle-specific MEP with 3 hand actions. With training, the muscle-specific MEP patterns of MI became more similar to those of ME. This demonstrates that NF training could promote the modulation of sensorimotor activities in the brain.
·Figure 2. Cross-condition classification accuracy across MI blocks.
Conclusions:
A novel, personalized, and adaptive MI and TMS-based NF training for complex hand actions was developed and tested. Our findings suggest that healthy adults could simultaneously and selectively modulate brain activities for multiple fingers with the guidance of NF. This demonstrates that TMS-based BCI could be used for hand function training in individuals that are not able to produce overt motor output.
Brain Stimulation:
Non-invasive Magnetic/TMS 2
Higher Cognitive Functions:
Imagery
Learning and Memory:
Neural Plasticity and Recovery of Function
Modeling and Analysis Methods:
Classification and Predictive Modeling
Motor Behavior:
Motor Planning and Execution 1
Keywords:
ELECTROPHYSIOLOGY
Learning
Machine Learning
Motor
Plasticity
Transcranial Magnetic Stimulation (TMS)
1|2Indicates the priority used for review
Provide references using author date format
[2] Hu M, Cheng HJ, Ji F, et al. (2021), ‘Brain functional changes in stroke following rehabilitation using brain-computer interface-assisted motor imagery with and without tDCS: A pilot study’, Frontiers in Human Neuroscience, vol. 15, pp. 692304.
[3] Mihelj E, Bächinger M, Kikkert S, et al. (2021). ‘Mental individuation of imagined finger movements can be achieved using TMS-based neurofeedback’, Neuroimage, vol. 242, pp. 118463.
[4] Blankertz B, Sannelli C, Halder S, et al. (2010). ‘Neurophysiological predictor of SMR-based BCI performance’, Neuroimage, vol. 51, pp. 1303-1309.