Feedback control of upright seating with functional neuromuscular stimulation during a reaching task after spinal cord injury: a feasibility study

Five individuals (S1, S2, . . ., S5) with SCI at varying thoracic and cervical levels participated in this study. Each participant had previously received an implanted neuroprosthesis for other studies intended to restore standing, walking, or postural balance. The neuroprosthesis is composed of a stimulator-telemeter [23, 24] connected to intramuscular, epimysial, or nerve cuff electrodes surgically placed to activate nerves serving the muscle groups spanning the trunk and hips. Each subject completed reconditioning exercises at home with the system for several months before the experiments. Table 1 lists the anthropometric and neurological characteristics of each subject at the time of testing, along with the muscle groups targeted by the controller. Supplemental surface stimulation was used in two of the subjects (S3 and S5) to augment the actions of the implanted system. The specific muscle groups activated with surface stimulation are marked with an asterisk (*) in Table 1. Subject S5 had a complete injury at T10 and only a small section of his trunk was paralyzed. We therefore anticipated that the controller’s effect would be reduced on him compared to those with higher injury levels and less volitional trunk control. All subjects were informed of all aspects of the experiment and subsequently signed consent forms approved by the local institutional review board (IRB: VA Northeast Ohio Healthcare System, Protocol Number: 07101-H36, Approval Date: 9/7/2010 or IRB: VA Northeast Ohio Healthcare System, Protocol Number: 18037-H20, Approval Date: 11/2/2018).

The feedback control system (Fig. 1) was composed of a small wireless tri-axial accelerometer, PID control law, and an external control unit to communicate with the implanted neuroprosthesis. The 3-axis accelerometer (CMA3000- D01, VTI Technologies, Vantaa, Finland) was placed on the sternum aligned approximately with the body anatomic axes and its output signal was sampled at 40 Hz. The sensor predominately measured trunk tilt using variation in the component of acceleration due to gravity along one of its axes. As the subject leaned away from the erect position the gravity vector projection on the local accelerometer coordinate frame changed such that variation in the x-axis value corresponded to lateral bending of the trunk, while changes in the z-axis (Fig. 1) corresponded to trunk flexion/extension [20, 21, 25]. Baseline x and z values were set at the beginning of each trial while the subject assumed an upright seated posture. Deviation from baseline was routed to independent PID controllers for sagittal and coronal pane movements governed by Eq. 1 [19, 26].where C(t\(_{\textrm{k}}\)) is the controller gain and the variables P, I, D are the proportional, integral, and derivative terms respectively, T\(_{\textrm{s}}\) is the sample time (0.025 s), and N is the derivative filter coefficient and refers to the bandwidth of the lowpass filter applied to the derivative term (N=100 for all experiments). The error at each time point (e(t\(_{\textrm{k}}\))), was determined by subtracting the baseline x and z acceleration values from the measured accelerometer values. Equation 1 defined the control laws for both flexion/extension and lateral bending movements. During flexion movements all muscle groups on both sides of the body were activated (lumbar erector spinae, quadratus lumborum, posterior portion of adductor magnus, gluteus maximus, gluteus medius, semimembranosus). The iliopsoas was only activated if the trunk extended past the baseline z acceleration value. Surface stimulation of the rectus abdominus was also considered, however in initial testing activation of this muscle did not cause significant flexion movement from an upright position. Additionally, our group has found that activation of the abdominals can be uncomfortable and cause breathing difficulties [27]. In practical use, the wheelchair’s backrest will provide support against backward perturbations. During lateral bending movements only the erector spinae (ES) and quadratus lumborum (QL) opposite to the direction of movement were activated. For example, if the subject leaned left, only the right ES and QL were activated to control the movement. Stimulation parameters sent to the external control unit to activate the subject’s muscles were determined by Equation 2.Where p\(^{i}_{\textrm{applied}}\) is the pulse width (PW) applied to each stimulating channel i of the neuroprosthesis, p\(^{i}_{\textrm{max}}\) is the maximum allowable PW of each channel i determined by increasing stimulation until it became uncomfortable for the subject or hardware limits were reached (250 µs), and p\(^{i}_{\textrm{min}}\) is the minimum PW similarly determined as the lowest value at which visible movement was produced by the target muscle. Stimulation amplitude was kept at a constant 20 mA unless the subject reported discomfort, in which case the amplitude was decreased to a more comfortable level. In cases where both flexion and lateral bending occurred, a muscle group was controlled by both controllers causing redundancies. These were resolved by choosing the larger PW value that resulted from the two controllers.

A weighted jar was placed on a table marked with four stations, a central ’home’ station surrounded by three radially located target stations labeled ’one’, ’two’, and ’three’ (Fig. 2). Each subject completed the task with only their dominant hand. In the remainder of the report, we will refer to the numbered stations as “ipsilateral” (station closest the dominant hand), “midline” (station two), and “contralateral” (station farthest from the dominant hand). Only subject S2 completed the tasks with her left hand, the other subjects used their right hand. The weight of the jar was varied from 0.9 to 2.26 kg depending on estimated shoulder and arm strength and comfort [17]. Kinematics of the trunk movements were obtained with a 16-camera motion capture system sampling at 100 Hz (Vicon Motion Systems Ltd., Oxford, UK). Reflective markers were placed on the C7 vertebrae, sacrum, and bilaterally on the acromion of the scapula, greater trochanter of the femur, middle of the upper arm, the lateral and medial epicondyle of the elbow, middle of the lower arm, the lateral and medial wrists, and the anterior superior iliac spine of the pelvis. Markers were also placed at each of the stations and on the weighted jar.

The prescribed task was designed to reflect typical ADLs that would occur in the home. Thus, subjects sat in their own wheelchairs in front of the testing table (Fig. 2a), with their non-active hand resting unconstrained across their chests or lower abdomens. Participants were allowed to catch themselves with either extremity if they felt a fall could occur due to trunk destabilization. Subjects were asked to move a weighted jar to and from the target stations (Fig. 2b) in predetermined random order. Distance to the table was customized for each individual to ensure that the jar could be transferred to each of the target stations and returned to home without fear of falling. Subjects were instructed to (1) lift the jar from the home station, (2) move it to the selected target station on cue, (3) release the jar, (4) return their empty hand to home, and (5) resume an upright seated position. Upon receipt of a second cue, the subject reached for the jar at the target station, acquired it, and returned it to the home station. The subjects rested between every 9 reaches to reduce the effects of fatigue, and completed as many of these reaching tasks within a single session as possible, depending on the amount of time required for controller tuning and their own personal time constraints. All data were initially included in the analysis. The number of reaches each subject performed under each condition is shown in Table 2. Subject S2 had a low level of constant stimulation applied to aid in completion of the reaching task while the controller was inactive. This was done at her request since she was a regular user of low level constant stimulation to add stiffness to her sitting posture at home on a daily basis.

Isolating reaching movements: Trunk pitch (extension and flexion motion) and trunk roll (lateral bending) angles were derived from the motion capture data offline during post-processing with MATLAB 2020 (MathWorks, Natick, MA). These measures differ from the tilt information derived from the mounted accelerometer that was only used for real-time control of the trunk. The trunk pitch and roll angles were determined by the angle between the global reference frame and the line defined between the sacrum and C7 marker [28]. Figure 3a shows the plot from a typical trial and the parameters used to process the data. We identified each of the reaching movements automatically based on when the jar marker reached 10% of its maximum velocity. Once each movement was isolated, we centered the data for analysis around the most significant landmark, the peak pitch angle. The data were windowed to 1.5 s before and 1.25 s after the peak pitch angle for subjects S2, S3, S4, S5. Subject S1 completed the task at a slower pace due to reduced hand function; therefore, his data were windowed to 3 s before and 2.5 s after the peak pitch. This process was applied to both the deployment of the jar to the target station and returning it to the home station. The data were normalized to 100% of the movement cycle with the first 50% defining the deployment and the second 50% the return.

Prediction ellipse area: We assessed the trunk kinematics with the area of the ellipse defined by the trunk pitch vs roll angles of each maneuver. The ellipse area (Fig. 3b) quantifies postural sway and represents a subject’s ability to control trunk movements and preserve stability. Larger ellipse areas indicate poorer seated stability [29, 30]. The ellipse area was determined by fitting an ellipse to a plot of the trunk pitch and trunk roll angles with the prediction ellipse method [31]. The prediction ellipse estimates the area in which future observations will occur with 95% probability.

Mean effective reach: We also examined the mean effective reach of the active arm relative to the body axis and was defined as the distance between the shoulder and wrist markers. The mean effective reach is a measure of combined shoulder elevation and elbow extension and reflects the ability of the subjects to hold objects further away from the trunk [16].

Return to upright time: Return to upright time was determined separately for deployment and return motions. It was calculated as the difference between the peak trunk angle and the following trough defined as the minimum trunk angle directly following the peak (Fig. 3a). Quicker return to upright motions enable the subject to reduce the length of time in a potentially unstable leaning position.

The previous analyses examined specific outcome measures. To explore any additional unanticipated impacts of the controller on trunk movements, we further examined the movement of the trunk along the major axis of movement (identified by the prediction ellipse in section E) according to the methods established by Duhamel et al. [32]. These methods allow comparison of the same population under two conditions (no/low constant stimulation and feedback control stimulation) throughout the entire movement cycle to identify effects not captured with the above outcome measures. This is done by characterizing each subject with a single curve by censoring outlier curves and averaging the remaining. Then, 95% confidence bands determined whether the two conditions were similar throughout the entire movement cycle.

Curve analysis on trunk movement: The goal of defining the trunk movements of each subject by a single curve for each condition was complicated by the high degree of variation observed in the kinematics of individuals with musculoskeletal disorders [33] that could corrupt a typical mean curve. To remove atypical curves before ensemble averaging, we assessed reliability with intraclass correlation coefficients (ICC). ICCs have been shown to account for both differences in means among measures and parallelism between test results [34]. ICC was calculated for each subject during each condition and target station. ICC estimates and their 95% confidence intervals were calculated based on repeated measurements, absolute-agreement, and a 2-way mixed-effects model [35]. To obtain reliable curves for each subject the ICC was calculated for the whole set of measures (subject, stimulation condition, and target station, 369 curves total) and a minimum level ICC\(_{\textrm{m}}\) was identified based on the probability density function of the ICCs estimated with a Gaussian kernel. An ICC\(_{\textrm{m}}\) = 0.9 was chosen, as this was the main inflection point of the Gaussian kernel plot. Figure 4 shows the process of assessing the trunk movement curves for reliability. First, the curves were grouped based on subject, condition, and target station, similar to Table 2. Second, ICC was computed for each set of curves. Third, if the ICC > ICC\(_{\textrm{m}}\), for a given group then all curves were included in subsequent analysis, and if not the least similar curve was removed. The second and third steps were repeated until ICC > ICC\(_{\textrm{m}}\) for all remaining curves. If ICC remained below ICC\(_{\textrm{m}}\) when only four curves remained in the set, data for that given group was ignored. Each subject was ultimately represented by a single mean curve for each target station taken as the ensemble average of the reliable set of curves.

Confidence bands The 95% confidence band for the difference between the means of the two conditions (feedback control vs no/low constant stimulation) was computed with the bootstrap method [32, 36]. All bootstrap bands were constructed using 1000 iterations. This test allowed us to determine when the feedback control had an impact on trunk movements during the reaching tasks.

Table 3 summarizes the statistical results for each outcome measure. Application of feedback controlled stimulation resulted in reduced postural sway for two of the subjects (S1, S4), allowed three subjects to extend their arms farther to accomplish the tasks (S1, S2, S3), and quickened return to upright motion for four subjects (S1, S2, S3, S4).

Prediction ellipse area: Ellipse area is shown in Fig. 5a. The results are separated into each target station followed by the aggregate data. Statistics were only performed on the aggregate data. Generally, ellipse areas were lowest for the ipsilateral station and greater for the midline and contralateral stations. There was a statistically significant reduction of ellipse area with feedback control for subjects S1 (z = 2.28, p = 0.022) and S4 (z = 4.69, p < 0.001) only.

Mean effective reach: The mean effective reach (Fig. 5b) did not show any consistent trends between the three stations. Statistically significant increases were observed for subjects S1 (z = − 2.14, p = 0.032), S2 (z = − 2.68, p = 0.007), and S3 (z = − 2.47, p = 0.014). A statistically significant decrease was observed for S4 (z = 6.19, p < 0.001). No differences were observed for subject S5.

Return to upright time: The time required for the subject to return to an upright position generally varied from 0.5 to 2 seconds, with S1 taking up to 4 seconds (Fig. 6). The return to upright time after deploying the jar to any target station was statistically faster with the controller active for subjects S1 (0.32s faster, z = 2.11, p = 0.035), S2 (0.18 s faster, z = 4.6, p < 0.001), S3 (0.21 s faster, z = 2.89, p = 0.004), and S4 (0.18 s faster, z = 3.72, p < 0.001) (Fig. 6a). Only S4 saw a significant time to upright reduction when returning the jar home (0.17 s faster, z = 4.02, p < 0.001) (Fig. 6b).

Curve analysis on trunk movement: Figure 7 shows the movement of the trunk along the major axis during both conditions after removal of the unreliable curves. Horizontal areas shaded in grey show where the bootstrap test determined there was a significant difference between the two conditions. These differences occurred both in trunk movements during ipsilateral and midline stations with no differences observed in the contralateral station. The differences are concentrated in two areas: the return to erect motion and the starting posture. The feedback controller resulted in a quickened return to upright motion after deploying the jar to the ipsilateral and midline position. Figure 7a shows average trunk movement curves as a result of reaching to the ipsilateral station. While the peak excursion occurs at 29 cycle percent for both conditions a significant difference occurred between 31 and 38 percent cycle during the return to upright motion. Feedback control resulted in shallower trunk angles at the same point in the cycle and showed a return to upright at 41 percent cycle compared with feedback control to 48 cycle percent with no/low stimulation. A similar effect can be observed during the other return to erect motions; however, this was not found to be significant during reaching to the contralateral station or during motions returning the jar to the home station. The feedback control also resulted in increases in the starting angle of the subjects during these two motions.