(J-397) Continuous Decoding of Dynamic and Static Force from Motor Cortical Population Activity

Force control at the hand is essential for everyday life. People with impaired upper limb motor control live with reduced quality of life partly due to their limited control in executing force during grasping movements. While there has been many non-human primate studies on static grasping forces and dynamic grasping forces at the level of single-units, the underlying representation of force in motor cortex (M1) population activity is not yet well understood. We set out to answer the following questions: Does the motor cortex encode for force at the population level during dynamic and static force control, and how well are we able to predict force output? This is important to understand especially in the context of using brain-computer interfaces (BCIs) to restore motor function.

Two human participants with tetraplegia, P2 and P3, performed a force-matching task using isometric wrist extension. We chose this behavioral task because our participants have residual motor control over the wrist, and we wanted to examine the neural correlates of executed motor output. Our four force conditions were comprised of static and dynamic periods of force output where we varied either the target force or rate of force output. The participants followed a visual guide (MuJoCo VR environment) and adjusted their rate of force output to match a sequence of the four conditions. The visual guide was converted into time series data for continuous prediction. We will refer to it as the “force trace”, the commanded change in force over time. M1 activity was recorded intracortically with the Utah Array (Blackrock Neurotech Salt Lake City, Utah). We also recorded forearm electromyography (EMG) activity to serve as a representation of participants' true attempts to match the force trace. EMG activity was recorded with the Myo Armband from Thalmic Labs (Waterloo, CA). We performed factor analysis followed by varimax rotation on Gaussian-smoothed firing rates and trained a long short-term memory (LSTM) network on the resulting M1 population activity for continuous prediction of the force trace. Then, we compared actual versus predicted force trace by calculating R-squared (R²) on the testing set and then again after separating static and dynamic periods. We then trained a second LSTM network to predict EMG activity and performed the same analysis.

The R² value for predicted force trace was approximately 0.8. Next, separating the force trace into static and dynamic periods, we found that the static periods had higher R², approximately 0.7, compared to the dynamic periods, approximately 0.4-0.5. This implies that there is force information encoded in M1 during the static periods that can be continuously decoded with an LSTM. Next, we wanted to understand if the lower R² values during the dynamic periods was due to lack of force information in M1 during those periods, or if it was a consequence of fitting the LSTM network to a rigid force trace that is not representative of neural output. So, we looked at the results for predicting EMG.

We found that the R² for predicted EMG was approximately 0.8-0.9, during the static periods it was approximately 0.8, and during the dynamic periods it was approximately 0.7. The higher R² for predicted EMG activity during the dynamic periods compared to the predicted force trace is promising because we see that EMG can be decoded from M1 during this period. Meaning, there is force information encoded in M1 during dynamic force output that can be continuously decoded. In Fig 1, we see that the predicted EMG tends to match the curvature of the actual EMG linear envelopes nicely meanwhile the predicted force trace has a much harder time matching the slope of the actual force trace. When predicting the force trace, there are a couple possibilities for why we see lower performance during the dynamic periods compared to EMG. First, the force trace does not account for trial-by-trial variability in the participant's execution of the task (i.e., participant onset of force output and rate of force output) while EMG inherently does. Second, continuously decoding neural data, an input that is highly variable, to predict a rigid output, force trace, may introduce inaccuracies when using an LSTM network. Thus, the output, force trace, may not be representing the input, neural data, well. Our next steps involve addressing trial-by-trial variability when designing force decoders for restorative BCIs, and we would like to further explore the choice of using an LSTM network on this type of data.