Dexterous Control of Seven Functional Hand Movements Using Cortically-Controlled Transcutaneous Muscle Stimulation in a Person With Tetraplegia

Abstract

Individuals with tetraplegia identify restoration of hand function as a critical, unmet need to regain their independence and improve quality of life. Brain-Computer Interface (BCI)-controlled Functional Electrical Stimulation (FES) technology addresses this need by reconnecting the brain with paralyzed limbs to restore function. In this study, we quantified performance of an intuitive, cortically-controlled, transcutaneous FES system on standardized object manipulation tasks from the Grasp and Release Test (GRT). We found that a tetraplegic individual could use the system to control up to seven functional hand movements, each with >95% individual accuracy. He was able to select one movement from the possible seven movements available to him and use it to appropriately manipulate all GRT objects in real-time using naturalistic grasps. With the use of the system, the participant not only improved his GRT performance over his baseline, demonstrating an increase in number of transfers for all objects except the Block, but also significantly improved transfer times for the heaviest objects (videocassette (VHS), Can). Analysis of underlying motor cortex neural representations associated with the hand grasp states revealed an overlap or non-separability in neural activation patterns for similarly shaped objects that affected BCI-FES performance. These results suggest that motor cortex neural representations for functional grips are likely more related to hand shape and force required to hold objects, rather than to the objects themselves. These results, demonstrating multiple, naturalistic functional hand movements with the BCI-FES, constitute a further step toward translating BCI-FES technologies from research devices to clinical neuroprosthetics.

Keywords: brain-computer interface; functional electrical stimulation; functional hand grasping; neuro-orthotics; spinal cord injury.

Figures

Figure 1

The BCI-FES system and experimental setup. The participant sits on wheelchair in front of the monitor which shows him the cued hand movement. The participant is required to grasp and transfer the object to the raised platform. (1) Neural activity is recorded from a 96-channel MEA implanted in the motor cortex; (2) A wavelet decomposition is performed on the raw data to extract neural information related to motor intent; (3) Wavelet scales 3 through 6 are used to generate Mean Wavelet Power (MWP)-based neural features; (4) Machine-learning algorithms decode the MWP activity for each attempted hand movement; (5) Hand movement is evoked using targeted transcutaneous FES delivered through cuffs wrapped around the forearm.

Figure 2

MEA location and signal quality over time. (A) Red regions are brain areas active during imagined hand movements. The implanted MEA location from post-op CT is shown in green. (B) MWP data for all channels were collected over a 108 s period at the beginning of periodic test sessions where the participant was instructed to imagine cued hand movements. MWP features were calculated to approximate the power in the multiunit frequency bands a plotted as a function of post-implant days. A 33% decline in the signal quality was observed over time from the MWP data.

Figure 3

Standardized GRT objects and functional grasps. Schematic showing the different GRT objects with associated dimensions and weights. Hand schematics illustrate the grasp/movement enabled by FES for the object. Fingers that were activated and used to perform the grasps/movements are highlighted in blue. *For the Fork object, a 4.4 N force is required to depress the cylinder.

Figure 4

Neural decoder training. Representative plots showing (A) threshold crossing raster plot, (B) the corresponding MWP activity across all channels of the MEA, and (C) neural decoder output as the participant attempts the seven cued hand movements. Solid lines indicate neural decoder output and dotted lines indicate the cue start and stop times. Of the seven possible hand movement states that can be predicted by the decoder, the output score from the one with the highest amplitude greater than zero was used to turn on/off the stimulation; (D) Confusion matrix showing the decoder response probability for each movement cue.

Figure 5

Functional performance evaluation using the Grasp and Release Test (GRT). (A) Sequential snapshots of the participant manipulating the Can object as part of the GRT. The participant starts from a rest state, opens his hand and place it around the Can, grasps the Can, transfers it to the raised platform, and then releases the Can. A new object is then placed in front of the participant to attempt the next transfer. (B) GRT scores showing the mean number of successful transfers with and without the BCI-FES system. With the use of the system, the participant not only improved his GRT scores over his baseline, demonstrating an increase in number of transfers for all objects except the Block, but was also able to grip and transfer two objects (Paperweight, Fork) that he could not manipulate at baseline. (C) Mean transfer times for each object with and without the BCI-FES system. With the use of the system the participant's transfer speed increased for all objects except for the Peg and Block which he was able to transfer faster on his own using adaptive grips. #The participant was never able to transfer these objects without the system. *The Can transfer required two hand movements—Hand Open and Can grasp. **p < 0.05 (paired t-test).

Figure 6

Neural decoder outputs during the GRT. Representative decoder outputs during a GRT test block showing instances of decoder misclassification (black triangles). All seven hand movements are available to the participant as part of the decoder and he has to evoke the correct movement (solid lines) during the 30 s trial period (dotted lines) given to him to complete the GRT for that object. Only decoder outputs above the activation threshold of zero are shown for visual clarity. Successful transfer of Can required the participant to evoke two hand movements—Hand Open and Can grasp (70–100 s). During the Can transfer, the decoder had two misclassifications (one of each for the PEG and VHS grasps). However, the participant was able to evoke the correct hand movements to successfully complete two Can transfers during the trial period. Similarly, during the Block transfer (170–200 s), the participant incorrectly evoked the decoder for Paperweight on four occasions. This did not affect the GRT scores for the Block, however, as the decoder for Paperweight kicked in after the participant had completed the Block transfer.

Figure 7

Neural representation of functional hand movements in the motor cortex. The participant was asked to attempt the cued hand movement. No FES was provided during this task so that neural data can be captured without any stimulation artifact. (A) Principal component analysis (PCA) of MWP activity shows clustering of neural activity for each hand movement during decoder activation for each functional movement. Dotted lines indicate a Gaussian mixture distribution model fit. (B) Heat maps of averaged MWP activity during neural decoder activation overlaid on the physical layout of the electrode array for each attempted hand movement. Corner reference (non-active) electrodes in the electrode array are labeled with white squares. (C) Heat map showing the pairwise Euclidean distances between vectorized MWP spatial patterns and highlights the separability in neural representation between different hand movements. Darker colors indicate that the neural representations are similar while lighter colors indicate that the representations are dissimilar. (D) Correlation between individual decoder accuracy and separation in neural representation (aggregate Euclidean distance for each movement) shows that higher neural discriminability leads to higher decoder accuracy. The trend line indicates a linear fit.