Enhanced selectivity of transcutaneous spinal cord stimulation by multielectrode configuration

We recruited 19 neurologically intact participants (9 female, 10 male, average age 23.12 ± 4.34 years old, demographics in supplementary table 1) who gave their consent to take part in this study that was reviewed and approved by Washington University in St. Louis' Institutional Review Board. Two participants (1 male and 1 female) withdrew from the study due to discomfort with the stimulation. One participant recruited early in the study was excluded from the analysis due to the use of monophasic, rather than biphasic stimulation, for a total of 16 participants included in the analysis for all figure.

2.2.1. Vertebral segment identification and electrode placement

2.2.2. Data acquisition

2.2.3. Transcutaneous spinal cord stimulation

2.2.4. Recruitment curves

Offline data analysis was performed in custom-built software written in Python and MATLAB^® (The MathWorks Ltd, USA). Evoked responses were averaged across the four repetitions, with one average response waveform for each of the ten stimulation amplitudes. Evoked responses for each muscle were normalized to the maximum averaged evoked response in both electrode configurations (one normalization for conventional tSCS and one normalization across the six round electrodes). Data was normalized separately for each condition since the maximum attainable response amplitude for each muscle was determined by either saturation or discomfort at that electrode configuration. To validate the robustness of our results to the normalization process, data was re-analyzed with one normalization across all conditions.

2.3.1. Selectivity index

To quantify the degree of muscle recruitment selectivity enabled by each electrode position, we first generated a recruitment curve for each muscle. This curve represented the muscle's peak-to-peak response as a function of stimulation amplitude. We then quantified the recruitment for each muscle as the area under the curve (AUC) of its recruitment curve normalized to the maximum possible recruitment i.e. maximal recruitment at all stimulation amplitudes. Last, we computed the selectivity SI of muscle m as the normalized recruitment REC_m of that muscle minus the average normalized recruitment of M—1 muscles, excluding the considered muscle, m (figure 2(E)). Conceptually, the selectivity index reflects the normalized recruitment of a given muscle compared to the average activation of all other muscles (Badi et al 2021):

$$\begin{equation}{\text{S}}{{\text{I}}_m} = {\text{ RE}}{{\text{C}}_m} - \frac{{\sum\nolimits_{n \ne m}^M {\text{R}} {\text{E}}{{\text{C}}_n}}}{{M - 1}}.\end{equation} \tag{ 1 }$$

Zoom In Zoom Out Reset image size

The selectivity index varies from −1 to 1 and reflects the activation of a targeted muscle compared to the average activation of all other muscles, with a value of 0 indicating that a given muscle is recruited equally to the average of all other muscles. This computation was performed for each muscle, electrode configuration, and participant.

To identify the target electrode for each muscle, we first compared the selectivity index enabled by three electrode positions ipsilateral to the target muscle (rostral: T10/T11, middle: T11/T12, caudal: T12/L1) and selected the electrode with the highest median selectivity index as the target electrode for that muscle. We then tested whether this target electrode enabled a significantly higher selectivity index for the targeted muscle compared to the conventional tSCS electrode and whether the selectivity index by the target electrode was higher than by its contralateral counterpart i.e. the small-diameter electrode at the same rostrocaudal level on the contralateral side.

2.3.2. Paired pulse suppression

2.3.3. Recruitment probability and spinal activation maps

The probability of recruiting a given muscle was computed for each electrode configuration as the normalized AUC of its recruitment curve at that electrode position. AUC values for each muscle were normalized to the AUC in the electrode that achieved the highest level of recruitment. We then computed the recruitment probability for each muscle as the average normalized recruitment across participants. In this computation, a recruitment probability of 1 would indicate that that muscle was maximally recruited at that electrode position for all participants.

Recruitment probability estimates were used to model the activation of motoneuron pools in each spinal segment S_i as a linear combination of the normalized leg muscle recruitment M_j, and W_i,j , the expected segmental distributions of motoneuron pools innervating muscle j at spinal segment S_i, (Wagner et al 2018):

$$\begin{equation}{S_i} = {\text{ }}\frac{{{{\mathop \sum \nolimits }_{{\text{muscles}}}}{W_{i,j}}{M_j}}}{{{{\mathop \sum \nolimits }_{{\text{muscles}}}}{W_{i,j}}}}{\text{ }}.\end{equation} \tag{ 2 }$$

The coefficients W_i,j were obtained from an innervation probability map constructed by Hofstoetter et al, which included qualitative and quantitative data from thousands of participants in anatomical textbooks and electrophysiological studies (Hofstoetter et al 2021b). Probability estimates for the VL muscle were derived from a separate anatomical study showing similar innervation in quadriceps muscles within the L2-L4 spinal segments (Sharrard 1964, Wagner et al 2018). The resulting spinal activation values were interpolated and superimposed onto a 2D image of the human lumbosacral spinal cord.

Statistical analyses were performed using the SciPy statistics toolbox (v1.8.1) for Python. Data for left and right leg muscles were pooled together for analysis. Normal distribution was tested using the Kolmogorov-Smirnov test. As a great majority of datasets were not normally distributed, non-parametric statistical tests were used. A Friedman test for repeated measures was performed to determine the effect of the multielectrode rostrocaudal vertebral level (rostral, middle, and caudal) on selectivity index. A Wilcoxon signed-rank test with a Bonferroni correction for multiple comparisons was then used to determine significant differences in selectivity index across electrode positions. Separate Wilcoxon signed-rank tests were used to compare the selectivity index enabled by the target electrode against the conventional tSCS, as well as against its contralateral counterpart. A Wilcoxon rank sum test was used to test for significant suppression in each muscle by its target electrode.

To study muscle recruitment enabled by conventional and multielectrode configuration electrodes, we recorded leg muscle responses over a broad range of stimulation amplitudes. The motor thresholds for the conventional and multielectrode configurations are shown in supplementary table 2. Overlaid EMG responses to the first stimulation pulse at increasing amplitudes for the conventional and right caudal electrodes are shown for a representative participant in figures 2(A) and (B). Notably, response latencies were progressively higher for caudally innervated muscles in both electrode configurations. This increase in latency is in agreement with previous observations (Courtine et al 2007, Minassian et al 2007a), suggesting that the increase in latency is due to the distance between motor neuron pools and the innervated muscle rather than the electrode location. Identification of the nature of the responses based on their latencies is not always definitive (Ertekin et al 1996, Troni et al 1996). Therefore, we applied existing neurophysiological methods to further analyze these responses.

To understand the relationship between muscle recruitment and stimulation amplitude, we calculated the peak-to-peak amplitude of the averaged responses for each recorded leg muscle (figures 2(C) and (D)). The degree of muscle recruitment was proportional to the stimulation amplitude. However, the degree of recruitment for a given muscle at a given amplitude differed across electrode configurations. Stimulation by the conventional tSCS electrode configuration resulted in broad non-selective muscle recruitment, as reflected by similar motor thresholds across rostral, caudal, and bilateral muscles (figure 2(C)). In contrast, motor thresholds were lower in a subset of muscles recruited by the right caudal electrode, and their recruitment level was close to saturation by the motor threshold amplitude of the others (figure 2(D), note the differences in recruitment level at ∼40 mA). The left leg distal muscles (TA, MG, SL) were also recruited at low amplitudes by the right caudal electrode in this participant. A common observation was that the targeted recruitment of a particular muscle group by the multielectrode configuration was usually accompanied by some undesired recruitment of the other unilateral muscle group (recruitment of right leg proximal muscles in the targeting of right leg distal muscles or vice versa), or the contralateral muscle group, as in this example. Therefore, we sought to quantify recruitment selectivity across muscles and electrode configurations.

To better interpret the degree of muscle recruitment selectivity enabled by each electrode position, we computed the selectivity index (Raspopovic et al 2011, Badi et al 2021) for all muscles (figure 2(E)). Selectivity indices for the conventional and right caudal electrodes in this participant are shown in figure 2(F). The conventional tSCS electrode primarily enabled the recruitment of bilateral proximal muscles (RF, ST, VL), as indicated by a positive selectivity index. The AUC of the recruitment curve for distal muscles (TA, MG, SL) was lower than all others, resulting in a low and negative selectivity index for these muscles. In contrast, for the multielectrode configuration, the selectivity indices of distal muscles of the targeted (right) and non-targeted (left) legs were higher than most other muscles. Moreover, the relative activation of proximal muscles in both legs was lower than that for all other muscles, resulting in a low selectivity index for the non-targeted proximal muscles in this participant. To understand trends in recruitment selectivity that were common across participants, we performed group-level analyses.

To identify the optimal electrode position to target each recorded leg muscle, we first computed the selectivity index for all muscles and electrode configurations across participants. We then compared the selectivity index enabled by each electrode rostrocaudal position (rostral, middle, or caudal) ipsilateral to the targeted muscle (figure 3(A)). Selectivity indices enabled by the multielectrode configuration depended on an electrode's rostrocaudal position for most recorded muscles: Friedman effect of electrode's rostrocaudal position in RF (χ² = 9.75, p = 0.008), VL (χ² = 15.06, p < 0.001), ST (χ² = 0.063, p = 0.969), TA (χ² = 10.94, p = 0.004), MG (χ² = 6.06, p = 0.048), and SL (χ² = 10.56, p = 0.005). Statistical significance for comparisons across electrode positions is shown above each muscle. The electrode with the highest median selectivity index was taken as the target electrode for that muscle for all subsequent analyses. Results for normality using the KolmogorovSmirnov test can be found in supplementary table 3.

Zoom In Zoom Out Reset image size

We hypothesized that the target ipsilateral electrode in the multielectrode configuration could enhance the recruitment selectivity of a targeted muscle group compared to conventional tSCS (figure 1(D)). To test this hypothesis, we compared the selectivity index for each muscle enabled by the conventional tSCS and the ipsilateral target electrode (figure 3(B)). Overall, the target electrode in the multielectrode configuration significantly enhanced recruitment selectivity compared to the conventional electrode for most muscles: Wilcoxon signed rank effect of RF (W =99, p = 0.002), VL (W =43, p < 0.001), ST (W =228, p = 0.500), TA (W =54, p< 0.001), MG (W =175, p = 0.096), and SL (W =94, p = 0.001).

To test whether the improvements in muscle recruitment selectivity could be achieved by having the small diameter electrode at the right rostrocaudal level regardless of the electrode's lateral position, we compared the selectivity index between the target ipsilateral electrode and its contralateral counterpart (figure 3(C)). As an example, we compared the selectivity index of the right RF when targeted by the right rostral electrode to the selectivity index of the same muscle targeted by the left rostral electrode. Overall, the ipsilateral target electrode significantly enhanced recruitment selectivity compared to its contralateral counterpart: Wilcoxon signed rank effect of RF (W =54, p < 0.001), VL (W =80, p < 0.001), ST (W =55, p < 0.001), TA (W =75, p< 0.001), MG (W =100, p = 0.002), and SL (W =65, p < 0.001).

Although selectivity index values for individual participants slightly varied when normalization of peak responses was performed across all electrode configurations (supplementary figure 1), the general trend across electrodes remained the same. The target electrode in the multielectrode configuration was the same for each muscle and achieved a higher level of selectivity than the conventional electrode and its contralateral counterpart. These results indicate that positioning small-diameter surface electrodes over the predicted locations of a muscle's motoneuron pools in a rostrocaudal direction while maintaining laterality can improve the recruitment selectivity of that muscle.

Studies examining the potential use of tSCS for recovery of motor function below the injury use the posterior root-muscle (PRM) reflex to confirm the effective position of the stimulating electrode over the spinal cord (Hofstoetter et al 2015, Minassian et al 2016, Gad et al 2017, Sayenko et al 2019). To investigate the reflex mechanisms behind spatially selective tSCS, we included in the recruitment curves protocol an additional paired pulse with a conditioning-test interval of 33.3 ms (figure 4(A)). We hypothesized that if elicited responses were indeed mediated by PRM reflexes, compound muscle action potentials to the test (second) pulse would reflect significant depression, a hallmark behavior of reflex responses (Olsen and Diamantopoulos 1967, Hofstoetter et al 2019).

Zoom In Zoom Out Reset image size

Recruitment curves for the first (top panel) and second (bottom panel) pulses delivered through the left rostrocaudal electrode in a representative participant are shown in figure 4(B). The amplitude of the response to the second pulse was largely suppressed in most muscles. However, the amount of suppression was inversely proportional to stimulation amplitude. Therefore, we quantified the level of suppression for each electrode at the highest stimulation amplitude, where the amount of suppression was generally the lowest. A comparison in response amplitudes between the first and the second pulse for the left caudal electrode in the same representative participant is shown in figure 4(C). Overall, there was a significant reduction in response amplitude to the second stimulation by the left caudal electrode in that participant.

To analyze this effect across participants, we computed the level of suppression as the ratio between the first and second responses. A ratio smaller than one would indicate response suppression, i.e. that the response to the second stimulation pulse was lower than the response to the first pulse. The amount of suppression for each muscle, when targeted by its optimal electrode, is shown in figure 4(D). There was a significant amount of post-activation depression of the evoked response by the paired-pulse paradigm at each optimal electrode configuration: one-sided 1-sample Wilcoxon signed rank test in RF (W =489, p < 0.001), VL (W =407, p = 0.004), ST (W =527, p < 0.001), TA (W =528, p < 0.001), MG (W =528, p < 0.001), and SL (W =528, p < 0.001). The amount of suppression for all electrode configurations and muscles is shown in supplementary figure 2. A significant amount of suppression was also observed in the conventional tSCS electrode and has been extensively reported by several groups (Hofstoetter et al 2019, 2021a, Oh et al 2022, Dalrymple et al 2023). These results serve as a neurophysiological confirmation that the optimal multielectrode configuration enables the effective stimulation of the respective segmental afferents in the targeted leg muscles.

To understand how different electrode configurations could target the posterior roots projecting to spinal cord segments containing the motor neurons responsible for activating the hip, knee, and ankle joints, we developed a neuroanatomical map of spinal cord activation for each electrode position. We first calculated the probability of activating each muscle at each electrode position by averaging the normalized recruitment of that muscle across participants (see section 2). Recruitment probabilities for electrodes located ipsilaterally and contralaterally to the targeted muscles at different rostrocaudal positions are shown in figure 5(A). Next, we compiled an atlas of the expected anatomical locations for the motor neuron pools associated with the recorded leg muscles (Sharrard 1964, Wagner et al 2018, Hofstoetter et al 2021b) (figure 5(B)). Finally, we projected the recruitment probability of each muscle onto its respective innervation probabilities at each spinal segment (figure 5(C)). The resulting spinal activation maps elicited by each multielectrode position are shown in figure 5(D).

Zoom In Zoom Out Reset image size

Our analysis of recruitment probability across participants revealed a high similarity between the optimal electrode configurations and the recruitment of the posterior roots projecting to the targeted spinal cord regions involved in the activation of the hip and ankle joints. For example, the left rostrocaudal electrode predominantly activated upper left lumbar segments (figure 5(D), top left panel), while the right caudal electrode activated motor neuron pools located around right sacral segments (figure 5(D), bottom right panel). It is important to note that the combination of the segmental innervation probabilities and the muscle recruitment probability was agnostic to the electrode configuration. Nevertheless, the spinal activation maps closely reflect the cathode location for that configuration.

Our study builds upon previous research in non-invasive tSCS, which has shown that stimulating different areas of the lumbar spinal cord can selectively activate motor neuron pools in leg muscles. Specifically, rostral and caudal positions of small electrodes centered over the midline have been shown to activate proximal and distal muscle groups, respectively (Krenn et al 2015, Sayenko et al 2015), and lateral positions can selectively recruit both proximal and distal muscles ipsilateral to the cathode position (Calvert et al 2019).

In our work, we demonstrate that by combining these approaches, we can achieve selective activation of specific muscle groups in the hip, knee, and ankle joints, independently for the right and left legs. We found that different rostrocaudal positions of a small diameter electrode provided unique degrees of selectivity for different muscles. The rostral electrode, positioned lateral to the midline and centered rostrocaudally over the T10/T11 interspinous ligament, was the optimal position for RF and VL muscles, whereas the caudal electrode, centered over the T12/L1 ligament, was the optimal contact for TA, MG, and SL. However, optimal selectivity for the ST was more challenging to achieve. The caudal electrode was marginally better than other rostrocaudal positions, but this difference was non-significant. Despite the caudal segment being the most probable site for ST recruitment, we observed that ST muscle recruitment was highly likely to occur at low stimulation amplitudes across all three segments. This consistent recruitment of ST and other hamstring muscles at different stimulation sites has been previously observed by other groups and has been attributed to their broader segmental innervation than other muscles (Hofstoetter et al 2021b). In their work, Hofstoetter and colleagues argue that PRM reflexes of the hamstring muscles may not provide useful information in intraoperative monitoring to guide electrode placement. Our results similarly suggest that responses from the other leg muscles should be prioritized when selecting the optimal electrode placement for a particular individual.

Interestingly, we found that the middle electrode, which was centered at the T11/T12 vertebral level, did not provide optimal selectivity for any muscle. This finding is noteworthy because conventional tSCS protocols often center the electrode at this level. We do not suggest that this placement is incorrect or ineffective, as if the objective were to target both proximal and distal muscles simultaneously by one stimulation site, T11/T12 or T12/L1 would be ideal placements. Rather, it may not be the optimal position for achieving selective activation of specific muscle groups.

Despite this, we do not recommend discarding the T11/T12 position in the multielectrode configuration altogether. In fact, it may be crucial in the context of multipolar configurations of tSCS, which have been shown to enhance recruitment selectivity further and prevent the unwanted activation of non-targeted muscles in eSCS (Struijk et al 1993, Wagner et al 2018, Hofstoetter et al 2021b, Rowald et al 2022). Therefore, while T11/T12 may not be the most selective electrode position for targeting specific muscles, it may still have important roles in optimizing tSCS protocols for further personalized improvements.

Our study revealed that achieving optimal recruitment selectivity using tSCS is not universal across individuals and targeted muscles. As shown in figure 3(B), the conventional tSCS electrode placement outperformed the target electrode in some participants' muscles. While our results provide a valuable starting point for improving muscle recruitment selectivity in individuals with neuromotor disorders, it is important to recognize that interindividual differences in neuroanatomy and neurophysiology, particularly in the damaged nervous system, will require personalized approaches. To optimize stimulation protocols for each individual's residual ability and specific needs, we believe it is crucial to test muscle recruitment for each participant and verify that these activations are mediated by posterior root-muscle reflexes. Modifications in stimulation parameters such as amplitude, frequency, and pulse width should then be made and reported based on these initial activation thresholds and selectivity starting points.

Recent advancements in eSCS have demonstrated the potential of machine learning approaches to identify optimal stimulation parameters for targeting motor function rapidly, across subjects and species (Govindarajan et al 2022, Bonizzato et al 2023). Similarly, algorithms for automated calibration of stimulation parameters can be employed to determine clinically suitable tSCS settings across participants with different anatomies and neuromotor disorders (Salchow-Hömmen et al 2021). Therefore, these approaches hold promise for streamlining and optimizing the customization of stimulation parameters in tSCS applications.

The motor-enabling effects of SCS are primarily associated with the recruitment of proprioceptive afferent fibers in the posterior roots (Dimitrijevic et al 1980, Capogrosso et al 2013, Minassian et al 2016, Formento et al 2018). However, because of the complex topographic anatomy in the lumbosacral spinal cord (Wall et al 1990), non-invasive stimulation at different lumbosacral vertebral segments can excite both afferent and efferent pathways to different degrees. The engagement of each pathway is crucial to both the immediate prosthetic effect that may be enabled by tSCS and the rehabilitation effect that may be observed through SCS-assisted therapy (Seáñez et al 2022). Therefore, careful evaluation of evoked responses should be performed.

While direct recruitment of motor axons within the anterior roots, such as in functional electrical stimulation, can indeed produce desired movements in leg muscles Faghri et al 1992, Popovic et al 2011), the artificial recruitment of large-diameter motor fibers makes it technically challenging to produce and sustain large forces (Kirsch and Rymer 1987, Giat et al 1993). Moreover, the post-synaptic activation of motor fibers bypassing spinal and descending circuits limits the types of movements that can be performed to those that can be pre-programmed (Ajiboye et al 2017), and prevents the voluntary modulation and neuroplasticity potential of these prosthetic effects. In contrast, the recruitment of afferent fibers by SCS enables interaction with descending and spinal circuits, which can enhance residual descending inputs (Minassian et al 2016, Guiho et al 2021) and lead to a natural recruitment order that is fatigue-resistant and capable of sustaining the whole body weight for extended periods (Formento et al 2018, Wagner et al 2018). This presynaptic recruitment of primary afferents, in turn, promotes neuroplasticity, which is believed to mediate the neurological recovery observed during neurorehabilitation facilitated by SCS (Nishimura et al 2013, Asboth et al 2018).

Overall, evoked responses by all target electrodes were mediated by PRM reflexes, as evidenced by the suppression of the conditioned response. However, the probability of efferent fiber recruitment differed substantially between L2-L4 innervated (RF, and VL) and L4-S2 innervated (ST, MG, TA, and SL) muscles (figure 4(D)). There were a few cases for proximal muscles in which the response average amplitude to the second pulse was higher than the average amplitude of the first response (values >1 in RF, ST, and VL). This observation has been recently reported by others (Oh et al 2022, Skiadopoulos et al 2022), although, to our knowledge, has not been studied in detail. In contrast, there were no cases suggesting efferent recruitment in TA, MG, or SL, and only one case in ST. This suggests that higher stimulation amplitudes could be employed at caudal segments to target distal muscles with a reduced likelihood of recruiting motoneurons directly.

It should be noted that the degree of suppression was calculated at the maximum stimulation amplitude, which is associated with a greater likelihood of directly recruiting motor neurons. This selection was deliberate to facilitate a conservative analysis of recruitment mechanisms in a 'worst-case' scenario. Routine applications of tSCS use stimulation amplitudes between 0.8 and 1.2 times the motor threshold to enhance motor function after paralysis (Hofstoetter et al 2013, Minassian et al 2016, Inanici et al 2021, Samejima et al 2022). As these amplitudes are considerably lower than saturation, we expect the probability of directly recruiting anterior roots in a therapy setting to be lower than those reported here. Nevertheless, because spine curvature can dramatically alter the probability of directly recruiting the anterior roots (Binder et al 2021), verification of post-activation depression should be carefully confirmed in different therapeutic settings (e.g. sitting, standing, or in the position required to use a rehabilitation device).

Despite the marked effects of stimulation frequency, intensity, and sensory feedback on evoked responses (Sayenko et al 2014, 2015), continuous SCS is commonly used in studies involving SCS-assisted neurorehabilitation (Gerasimenko et al 2015a, Meyer et al 2020, Inanici et al 2021). In this application, stimulation parameters are typically fixed at the onset of therapy and kept constant across different types of movements. This continuous, non-selective SCS largely diverges from the natural spatiotemporal activation patterns observed during movement (Yakovenko et al 2002) and can disrupt the natural feedback from proprioceptive afferents (Formento et al 2018), which is critical to enable the spinal regulation of movements during fine motor control (Barra et al 2022). And although long-term training combined with continuous SCS can indeed induce improvements in motor function that persist even without stimulation, they may take several months to a year of intense rehabilitation to appear (Angeli et al 2018, Gill et al 2018, Seáñez and Capogrosso 2021).

Spatial and temporal control of SCS aims to facilitate movements through the selective activation of specific motor neuron pools at the appropriate phases of movement (Wenger et al 2014, 2016). Recent clinical studies have shown that spatiotemporal eSCS can rapidly enable (within one week) powerful facilitation of walking, cycling, kayaking, and other dexterous activities in people with SCI (Wagner et al 2018, Rowald et al 2022) as well as upper limb movements and grasping ability in people with stroke (Powell et al 2023).

Combining long-term activity-based training with tSCS has been shown to improve standing, balance, and induce functional recovery (Gerasimenko et al 2015a, Gad et al 2017, Sayenko et al 2019) that is comparable to that achieved through eSCS (Harkema et al 2011, Rejc et al 2015, Seáñez and Capogrosso 2021). It is unlikely that non-invasive tSCS will achieve the same level of muscle recruitment selectivity as eSCS. However, using multielectrode configurations of tSCS to selectively target facilitation in certain muscles during movements could help individuals learn to perform complex movements faster, potentially accelerating their recovery process. By speeding up the recovery process, these technologies could then become more accessible to individuals with SCI by reducing the amount of training required and, ultimately, the cost.

4.5.1. Selective targeting of agonist and antagonist muscles

4.5.2. Potential improvements by different electrode configurations and higher electrode density

4.5.3. Robustness across experimental sessions

4.5.4. Translation into clinical applications

4.5.5. Controls vs SCI