The purpose of this study was to improve stimulation-induced cycling performance through feedback control of stimulus levels. We found that enforcing a submaximal but higher than typical steady-state exercise intensity through feedback control can significantly increase power maintained and work performed by participants with paralysis. We further found that cadence feedback control of stimulus levels reduced between- and within-stroke power fluctuations, enabling practical implementation of duty cycle reducing paradigms that can further extend exercise durations prior to fatigue. These results may be applied in the development of enhanced stimulation systems and the formation of optimal exercise training regimens for participants with paralysis.
Cycling endurance and intensity
Cadence-controlled stimulation significantly improved endurance in six out of seven participants, as evident by higher Pend values maintained by at least one controlled condition. Participants (P03, P04) who showed remarkable improvements (~ 180% to ~ 900%) in Pend with cadence-control initially output high peak powers but often could not cycle 90 s continuously with S-Max, resulting in near zero Pend values. Both controlled conditions consistently enabled them to maintain power throughout their entire 90 s trial durations, leading to the large percent increases. For participants who produce lower initial peak powers (P02, P06), Pend improvements with cadence control were not always significant. Training should first focus on increasing absolute strength to support cycling against meaningful resistances for these participants. Enabling users to extend exercise durations prior to lower extremity muscular fatigue by increasing baseline strength and/or applying closed-loop control of stimulation will ultimately increase benefits to those muscles and provide a better cardiovascular workout.
Chosen target cadences demanded power outputs greater than a participant’s typical steady state ability and could not always be sustained throughout an entire trial duration. Nevertheless, even when controllers reached maximal stimulus levels and were unable to recruit any additional fibers to maintain a target cadence, end powers still consistently remained above those of conventional stimulation. Cadence control eliminated the high initial peak powers that occur with conventional stimulation by recruiting fewer fibers with submaximal stimulus levels. Submaximal stimulus levels avoid electric field overlap and can prevent fibers that would be in the overlapped regions from experiencing excitation–contraction decoupling due to unintentionally high firing frequencies. Fewer fibers initially contracting can cause a lower accumulation of hydrogen ions and other metabolic byproducts that contribute to force decline. Together these potential mechanisms may account for the improvements in end power achieved with the controllers.
Total work performed within a set cycling duration was also strongly affected by target cadence choice and the resultant power output. We could not ensure a priori that a chosen target output could be maintained long enough to produce equal or greater work within the same cycling duration. Still, the fact that work did accumulate to a significantly greater degree in three participants is encouraging. Extending trial durations would likely further increase differences in total work between S-Max and both controlled conditions for all participants.
Consistently performing more work and maintaining higher steady state powers within each exercise session is crucial to maximizing physiological benefits. Competitive able-bodied cyclists track a functional threshold power (FTP), the maximum steady state power they can maintain for 1 h . Training programs that vary workout durations and intensities based on % FTP have been shown to significantly improve FTP and provide physiological benefits for recreational and competitive cyclists . Such FTP measurements and fine-tuned programs have historically been unavailable to persons with paralysis, as conventional stimulation systems do not sustain a steady muscle power output. The cadence-controlled stimulation schemes presented here could now be used to establish and track a modified FTP in this population. Long-term training regimens based on the FTP equivalent for stimulated cycling could be explored in future studies.
Power fluctuation and controller performances
Uneven force production among independently activated fiber groups was a significant practical limitation of duty cycle reducing stimulation patterns in prior studies [13, 24]. In this study, the C-Cont pattern successfully reduced PFI to the point of no significant difference with S-Max in three of four tested participants and to significantly lower values than both S-Max and S-Cont in the fourth participant. PFI improvements over S-Max are likely due to controllers enforcing more even outputs between the left and right legs, while C-Cont improvements over S-Cont may be due to more precise adjustments by each contact’s individual PI controller. S-Cont modulates and delivers the same PW through all active knee extensor contacts at once, thus operating along a combined recruitment curve (RC) from all activated MUPs. Outputs from independent MUPs activated at the same time will sum approximately linearly when stimulated fiber overlap is low [25, 26]. The combined RC then is likely very steep, especially at lower stimulus values. Small changes in PW could result in large relative changes in muscle output, and the S-Cont controller may have not been as well tuned to account for this, resulting in higher PFIs and RMSE than C-Cont. Conversely, C-Cont modulates PW through the single active knee extensor contact with its own dedicated PI controller. Adjustments in PW thus only depend on what the one knee-extensor MUP needs to meet the desired cadence, removing the potential of over- or under-stimulating through other simultaneously active contacts as is possible with S-Cont. This study thus demonstrates a practical way to employ duty cycle reducing stimulation patterns to better improve cycling ability without the limitations of open-loop implementation.
From the available in-laboratory data, cadence-controlled patterns were more efficient than conventional stimulation, producing more work per unit of charge injected. Efficiency even increased when work was not significantly higher, due to lower levels of Q needed to sustain the target output. Higher efficiency may extend battery life of the stimulation systems and provide further assurance that no overstimulation or damage to the neural tissue will result over time, as similar outputs may be achieved with less injected charge . Additionally, prior research shows a correlation between stimulation cost, the inverse of efficiency presented here, and the oxygen cost (the rate of pulmonary oxygen uptake) of an exercise for a given power output . Greater stimulation efficiency with cadence-control may therefore coincide with less oxygen cost, which could make maintaining a greater exercise intensity more feasible for deconditioned persons who may have partial paralysis of the ventilatory muscles. Though not formally measured in this study, P05 anecdotally reported less breathlessness after controlled cycling bouts compared with conventional stimulation.
Cadence only steadily declined below target once controllers were unable to recruit more unfatigued MUPs. For P04, the only participant for whom PW data from both controlled conditions is available, both TmaxPW and Ttarget occur later with C-Cont than S-Cont for comparable target outputs. All contacts required maximum PWs by the end of each S-Cont trial, while four of the six independently controlled contacts do not yet reach maximum with C-Cont (Fig. 6). This agrees with our prior open-loop study  where reduced duty cycles were found to further extend muscle output. By incorporating reduced duty cycles into closed-loop stimulation models, originally recruited fibers fatigue less rapidly and additional fibers do not need to be recruited as frequently, enabling the controller to work below maximum PWs and sustain target output even longer. This also supports the assumption that independent fiber groups are activated by submaximal PWs through independent contacts with low overlap, and are thus able to periodically rest during carousel stimulation, despite lacking formal overlap assessments in this participant.
Both controllers only adjusted stimulation through contacts that activated the knee extensors. Stimulation to the knee flexors, hip extensors, and adductors remained constant at maximal values. Progressive fatigue in these other muscle groups may have influenced controller adjustments, potentially resulting in premature increases in quadriceps activation to make up for declining output of the other muscles. Despite this limitation, basic PI control of only the knee extensor fiber groups yielded average absolute RMSE across all participants of only 2.4 rpm (6.4%). Controlling stimulus levels to all involved muscle groups may further improve target tracking ability and would likely increase the performance benefits of S-Cont and C-Cont stimulation. While direct comparisons to other fatigue-delaying strategies are not made, improvements achieved in this study may also be further enhanced by incorporating asynchronous, interleaved, or other such strategies [14,15,16, 28] for all involved muscle groups during closed-loop cycling. More sophisticated control schemes have been proposed from simulations with musculoskeletal models to optimize performance and more faithfully produce the desired output [18, 29, 30]. However, these proposed control schemes have not yet been successfully deployed in clinical tests with paralyzed participants or without simultaneous, prioritized control of a motor. Future work should seek to incorporate these more advanced controllers into a motorless stimulation system to potentially provide even greater improvements in exercise performance and physiological outcomes.
Recently, a closed-loop control scheme using feedback from IMU sensors to spatiotemporally adjust implanted epidural electrical stimulation was shown to improve power output during cycling in one participant with SCI . The study’s control scheme focused on adjusting the timing and location of epidural stimulation, not on regulating the intensity of resulting muscle contractions for cadence control. The study also only presents data for less than 75 pedal strokes, fewer than what was accomplished in most of the trials reported herein, and it is unclear if the participant could have continued further with this strategy. It is thus difficult to compare the effectiveness of these closed-loop, implanted strategies, but it is probable that the epidural modality would also benefit from similar cadence-control techniques.
Physiological effects of controlled stimulation
Muscle oxygenation data may explain variations in functional performance among select participants. In P05, S-Cont delayed rapid declines in SmO2 and resulted in significant functional improvements compared with S-Max. In the other participants tested, however, such drastic SmO2 declines did not occur with any stimulation pattern. P05 may have had a previously undetected perfusion issue that limited oxygen delivery to the working muscles relative to their oxygen consumption. By delaying the incorporation of some fibers until needed with a controller, oxygen was not depleted immediately in all fibers. This may enable more efficient aerobic respiration to occur in some parts of the muscles longer than with S-Max, and account for the significant improvements in cycling performance. P05 also cycled against greater resistances and produced much greater powers than the other participants, which contributes to more dramatic declines in SmO2  and provides ample opportunity for improvement with the controller. It is possible that, once other participants gain strength to cycle against greater resistances, the controllers will become more beneficial as demand on the muscles increases and better pacing strategies become more valuable. Future work should assess the relative rates of SmO2 decline at various gears and cadences to determine a combination that enables high power outputs with the physiological advantage of slower oxygenation decline.
A significant heart rate increase when cycling with cadence-controlled stimulation is another notable physiological benefit seen in P05. Paralysis, particularly when caused by SCI, hinders appropriate cardiorespiratory responses to stimulated exercise . Loss of quick afferent feedback from the working muscles to the autonomic nervous system due to the spinal cord lesion eliminates the influence of the exercise pressor reflex on the regulation of heart rate, respiration rate, and blood pressure during exercise. We have observed heart rate often changes only negligibly and sometimes even declines in participants with SCI despite cycling to the point of lower extremity exhaustion. The reduction in heart rate is likely due to increased venous return when the typically sedentary lower extremities are activated by stimulation. Specifically, contraction of the paralyzed muscles creates a pumping effect that can greatly increase venous return and stroke volume and subsequently reduce heart rate for any given cardiac output. These factors prevent conventional stimulation-induced cycling from providing a meaningful cardiovascular workout and present obstacles for providing the working muscles with the resources needed to keep moving. The ability of cadence-controlled stimulation to overcome these barriers to elevating heart rate in P05 is extremely promising. Greater work achieved with the controller may explain the elevated heart rate in this participant. Additionally, altered hemodynamics may have also played a role. The controller maintains a submaximal cadence and does not maximally recruit all available fibers from the outset of the exercise, causing only a subsection of the quadriceps to contract and relax at comparatively slower rates. Using only a subsection of available fibers to maintain a submaximal cadence may reduce venous return through decrease muscle pump compared with conventional stimulation that activates all available fibers at higher cadences . This could potentially ease the blood volume-induced heart rate depression, further contributing to the greater heart rates achieved. The statistically significant increase in heart rate, while only 6 beats per minute, may have facilitated greater oxygen delivery to the quadriceps muscles to help P05 sustain power and perform even more work, perpetuating the cycle.
Advantages of motorless exercise control
Prior studies incorporating feedback control during stimulation-induced cycling utilized a motor to assist or resist stimulated muscle contractions to maintain a target cadence and/or power output [18, 35,36,37,38]. These approaches can provide greater resistance in the beginning of a trial to maximize the load against which the muscle must work before it fatigues. Maximizing resistive loads may be the key to achieving greater load-dependent physiologic improvements with these systems, especially in bone density [39, 40]. However, excessive motor resistance may prematurely fatigue the muscles and drastically reduce the duration of exercise. Conversely, assisting the pedaling motion when muscle output becomes insufficient for target maintenance can shield paralyzed musculature from positive stress and decrease required effort that would help them improve. There are mixed opinions as to whether keeping the legs cycling after the muscles can no longer contribute to the motion provides adequate physiological benefits , and the fatigued muscles may be better served resting without continued ineffective stimulation so that they may recover and perform subsequent bouts of meaningful, leg-driven (as opposed to motor-driven) cycling. Additionally, integrating a motor significantly increases the complexity of the control algorithm since care must be taken not to risk harming the participant, and increases the weight and cost of the cycling apparatus. Control methods from this study are easily implemented with no additional hardware or musculoskeletal modeling requirements and only minute increases in computational complexity, making them ideal for practical every day and potentially overground use. Maintaining a mid-level intensity using only the capabilities of the activated muscles may enable meaningful work against resistive loads without undue fatigue or motorized assistance. Future work should determine the relative advantages of the simple control strategies presented here compared to those that employ motors.