TMS-based neurofeedback facilitates motor imagery of different hand actions
Poster No:
1697
Submission Type:
Abstract Submission
Authors:
Hsiao-ju Cheng1, Olivia Hochstrasser1, Eunice Tai1, Daryl Chong1, Nicole Wenderoth2
Institutions:
1Singapore-ETH Centre, Singapore, Singapore, 2ETH Zurich, Zurich, Zurich
First Author:
Co-Author(s):
Introduction:
Non-invasive brain-computer interfaces (BCIs) enable users to modulate brain activity in a goal-directed manner. Most non-invasive BCIs can decode only gross movements but many daily tasks require finer finger and hand control. We developed a novel BCI using motor imagery (MI) and transcranial magnetic stimulation (TMS)-based neurofeedback (NF) training to reinforce representations of complex hand actions in the brain. This proof-of-concept study investigates the utility of this BCI for training hand function via MI.
Methods:
12 participants (6 males, mean age 32.0 ± 2.7 [SD] years) completed 4 sessions of TMS-based NF training focusing on 3 hand actions (holding a bottle, turning a key, and opening the hand). The first session involved motor execution (ME), followed by 3 motor imagery (MI) sessions (Figure 1). The ME session comprised 4 blocks (2 blocks per hand, counterbalanced across participants). The MI session comprised 4 blocks: 1 without NF and 3 with NF using the right hand. An additional block without NF was added at the end of the 4th session to assess the learning effect. A personalized, adaptive support vector machine (SVM) ensemble was used for NF during the training, in which classifiers were trained using right-hand ME, left-hand ME, and every available MI block data to classify upcoming MI trials and provide feedback accordingly.
We chose average classification accuracy as an outcome measure to gauge the training effect. Block-wise average classification accuracy was derived from an SVM classifier with leave-one trial-out cross-validation.
We chose average classification accuracy as an outcome measure to gauge the training effect. Block-wise average classification accuracy was derived from an SVM classifier with leave-one trial-out cross-validation.
·TMS-based NF training experimental setup and personalized, adaptive algorithm
Results:
We first explore whether average classification accuracy across nine MI+NF blocks exceeded the chance level to understand the feasibility of our approach via a one-sample t-test (Figure 2). We found that accuracy across 9 MI+NF blocks (57.00%±4.32%) was significantly higher than the upper bound of the empirical chance level (45.08%; t11=5.663, p < 0.001). We then evaluated the learning effect using a linear mixed-effect model with MI without NF data and noted that the accuracy of the final MI without NF block (60.5%) showed nearly significant improvement compared to the first block (53.3%; β = 0.091, t33 = 1.974, p = 0.057). We further investigated if the average classification accuracy of 9 MI+NF blocks increased over time using another linear mixed-effect model and found that Session 3 accuracy (59.3%) was significantly higher than Session 2 (53.3%, β = -0.080, t88 = -3.208, p = 0.006) and Session 4 accuracy (58.2%) was marginally significantly higher than Session 2 (β = -0.060, t88 = -2.421, p = 0.053).
·Average cross validation (CV) accuracy of all motor imagery (MI) blocks
Conclusions:
We developed and tested a novel, personalized, and adaptive MI and TMS-based BCI for complex hand actions. Our findings suggest that healthy adults could modulate brain activities for complex hand actions with the guidance of NF. This demonstrates that TMS-based BCI could be used for hand function training in individuals who are unable to produce overt motor output.
Brain Stimulation:
Non-invasive Magnetic/TMS 2
Higher Cognitive Functions:
Imagery
Learning and Memory:
Skill Learning
Motor Behavior:
Motor Planning and Execution 1
Keywords:
Machine Learning
Motor
Physical Therapy
Plasticity
Therapy
Transcranial Magnetic Stimulation (TMS)
Treatment
1|2Indicates the priority used for review