Speed dependency in α-motoneuron activity and locomotor modules in human locomotion: indirect evidence for phylogenetically conserved spinal circuits

Abstract

Coordinated locomotor muscle activity is generated by the spinal central pattern generators (CPGs). Vertebrate studies have demonstrated the following two characteristics of the speed control mechanisms of the spinal CPGs: (i) rostral segment activation is indispensable for achieving high-speed locomotion; and (ii) specific combinations between spinal interneuronal modules and motoneuron (MN) pools are sequentially activated with increasing speed. Here, to investigate whether similar control mechanisms exist in humans, we examined spinal neural activity during varied-speed locomotion by mapping the distribution of MN activity in the spinal cord and extracting locomotor modules, which generate basic MN activation patterns. The MN activation patterns and the locomotor modules were analysed from multi-muscle electromyographic recordings. The reconstructed MN activity patterns were divided into the following three patterns depending on the speed of locomotion: slow walking, fast walking and running. During these three activation patterns, the proportion of the activity in rostral segments to that in caudal segments increased as locomotion speed increased. Additionally, the different MN activation patterns were generated by distinct combinations of locomotor modules. These results are consistent with the speed control mechanisms observed in vertebrates, suggesting phylogenetically conserved spinal mechanisms of neural control of locomotion.

1. Introduction

Locomotor muscle activity is generated by a vast number of α-motoneurons (MNs) in the spinal cord. Spinal central pattern generators (CPGs) play a critical role in producing the coordinated MN activity during locomotion [1,2]. Recently, evidence for the existence of CPGs consisting of spinal interneurons has been demonstrated by experimental studies using animal models based on electrophysiological, genetic and neurochemical techniques [1,2]. Also, in humans, some studies on human spinal cord-injured (SCI) patients [3] and on healthy participants [4] have demonstrated the ability of the human spinal cord to generate rhythmic and synergistic lower limb muscle activity.

In our previous study [5], we reported that different combinations of locomotor modules, which generate specific combinations of muscle activity (i.e. spatially fixed locomotor muscle synergies), are activated depending on speed and mode of locomotion in humans. The locomotor modules are considered to be encoded in spatial pattern formation networks, which activate multiple MN pools, in the spinal CPGs [6]. Thus, our previous study [5] suggests that the CPG output changes in a manner dependent on locomotion speed and mode. Recently, to estimate how the CPG output is directed to the MN pools within each spinal segment, electromyographic (EMG) signals were used to map the spatio-temporal MN activity in the lumbosacral enlargement (segments L2–S2) along the rostrocaudal direction (figure 1a; see ‘Material and methods’ for further details) [7,8]. Using this method, previous studies have shown some changes in spatio-temporal MN activity depending on speed and mode in humans [7,8]. For example, at non-preferred-speed walking and running, the MN activation patterns demonstrated additional locus of activation compared with those at comfortable speed [8]. Additionally, mode-dependent MN activation patterns were observed during multiple locomotor modes (walking, running, backward stepping and skipping) [8]. In these previous studies [7,8], relationships between the MN activity and timing pattern of locomotor modules have been extensively investigated. However, relationships between the MN activation patterns and the muscle weightings of modules (i.e. spatially fixed muscle synergies) remain unclear.

Figure 1.

Procedures of reconstruction of spatio-temporal activation patterns of motoneurons (MNs) along the rostrocaudal axis of the lumbosacral enlargement from EMGs (a) and locomotor modules (b–d). (a) Activation patterns of MNs in each segment (L2–S2, left vertical scale) were reconstructed by mapping recorded muscle activities (EMGs) based on Kendall's myotomal charts (electronic supplementary material, table S1). (b) Locomotor modules were extracted using non-negative matrix factorization (NMF). The output of each module is explained by the product of the muscle weighting component (bars; specifying activation level of each muscle) and the temporal pattern component (waveforms). The sum of outputs from the modules is approximately equivalent to those of the EMGs. (c) MN activity generated by each module was reconstructed from the output of each module. (d) Activity patterns generated from individual modules were summed into a whole activation pattern. The bar underneath denotes the stance phase (black) and swing phase (grey) in a gait cycle.

Regarding locomotor speed control by the spinal CPGs, previous studies in animal models demonstrated two interesting characteristics. The first is the rostrocaudal gradient of rhythmogenic capacity in the spinal CPGs. Although rhythmogenic capability is widely distributed along the lumbosacral enlargement, the upper segments have a higher rhythmogenic capability [9–11]. The second is the specific combination of individual classes of interneurons and MNs called a ‘locomotor spinal microcircuit’ [12]. Ampatzis et al. [12] showed that three distinct types of these microcircuits are sequentially activated, from slow to intermediate and then fast, with increasing speed. So far, these speed control mechanisms in spinal circuits have not been confirmed in humans.

It has been suggested that CPGs in legged vertebrates emerged during evolution from a common ancestral circuit [13], and a recent EMG-based study showed that the locomotor modules of humans and those of other mammals and birds evolved from similar circuitry [14]. As spinal CPG mechanisms are phylogenetically conserved at a cellular level across different species, and even between fish and rodents [1], the speed control mechanisms of CPGs are probably conserved in humans. Based on the hypothesis that common speed control mechanisms are shared by humans and other vertebrates, we established working hypotheses as follows. (i) MN activity in rostral segments becomes higher compared with in caudal segments with increasing locomotion speed. Additionally, (ii) if the MN activation patterns are changed with increasing speed, the distinct MN activation patterns are generated by different locomotor modules. To test the hypothesis, here we mapped the recorded EMG patterns onto the approximate corresponding segments of the MN pools in the spinal cord [7,8] and extracted locomotor modules using non-negative matrix factorization (NMF) [5,14] during walking and running over a wide speed range. The acceptance of this working hypothesis would provide indirect evidence that the speed control mechanisms of spinal locomotor circuits are conserved in humans.

2. Material and methods

(a). Participants

Seventeen healthy male volunteers (ages 19–31 years) participated in the study. EMGs and ground reaction force (GRF) data of 16 participants (eight non-runners and eight runners) out of 17 participants were taken from our previous study [5]. EMGs and GRF data of 16 participants (eight non-runners and eight runners) out of 17 participants were taken from our previous study, which focused on a different topic (namely, on the recruitment of motor modules for different locomotion modes and speeds, in contrast with the present topic: the proportion of the MN activity in rostral and caudal segments). Each participant gave written informed consent for his participation in the study.

(b). Experimental set-up and design

Participants walked or ran on a treadmill (Bertec, Columbus, OH, USA). The belt speed was linearly increased from 0.3 m s⁻¹ to 4.3 m s⁻¹ with an acceleration of 0.01 m s⁻² (runners' data were additionally recorded at 4.3–5.0 m s⁻¹; not included in this study). This speed range was set as fast as possible within the safe limits checked in our previous study [5]. The participants were asked to change their locomotion mode (from walk to run) on the basis of their preference under the given speed. The participants' observed walk–run transition speed ranged from 1.9 to 2.3 m s⁻¹.

(c). Data collection

Three-dimensional GRF data were recorded from force plates under the right and left belts of the treadmill (1000 Hz). Surface EMGs were recorded from the following 14 muscles on the right leg (EMGs were recorded from two additional trunk muscles from participants in the previous study [5]; data not included here) for analysing MN activity patterns and locomotor modules: tibialis anterior (TA), gastrocnemius lateralis (LG), gastrocnemius medialis (MG), soleus (SOL), peroneus longus (PL), vastus lateralis (VL), vastus medialis (VM), rectus femoris (RF), biceps femoris (BF), semitendinosus (ST), adductor magnus (AM), tensor fascia latae (TFL), gluteus maximus (GM) and gluteus medius (Gmed). We targeted muscles innervated by MNs located in the lumbosacral enlargement, where locomotor CPGs are known to exist [3]. The EMGs were recorded with a wireless EMG system (Trigno Wireless System; Delsys, Boston, MA, USA). The EMG signals were band-pass filtered (20–450 Hz) and sampled at 1000 Hz. GRF data were smoothed by a low-pass filter (a zero-lag Butterworth filter, 5 Hz cut-off). The timings of heel-contact (HC) and toe-off (TO) were determined based on the vertical component of GRF. Then the stance time, swing time and stride time were calculated. Locomotion mode (walk or run) was defined by the presence and absence of double support time.

(d). Electromyographic processing

The recoded EMGs were divided into 0.1 m s⁻¹ bins based on the treadmill speed. Thus, the EMGs data of each participant were divided into 40 speed ranges. As the treadmill speed was accelerated at 0.01 m s⁻², each bin contained 10 s of data. Then, the EMG data were rectified and low-pass filtered with a zero-lag Butterworth filter. As the cut-off frequency of the low-pass filter had to be adjusted to each movement frequency, to smoothen the EMG signals to the same extent for different speeds, the low-pass cut-off frequency was adjusted to each speed condition according to the following formula: 10 × stride frequency (Hz) [5]. Subsequently, the smoothed envelopes were time-interpolated so that they had 200 points for each gait cycle.

(e). Spatio-temporal activation patterns of motoneurons along a rostrocaudal direction within the spinal cord

To characterize the spatio-temporal patterns of the spinal MN activity, the processed EMGs were mapped onto the estimated rostrocaudal location of the MN pools in the spinal cord from the L2 to S2 segments (figure 1a) based on Kendall's myotomal charts [15] as in previous studies [7,8,16,17] (see the electronic supplementary material for further details).

To compare the spatio-temporal activation patterns of MNs among speeds, the patterns across all 40 speed ranges were grouped by hierarchical clustering (Ward's method, correlation distance, see the electronic supplementary material for further details). The MN activation patterns were divided into three speed ranges: slow walking (0.3–1.1 m s⁻¹), fast walking (1.1–1.9 m s⁻¹) and running (2.2–4.3 m s⁻¹; detailed in ‘Results’). Data in subsequent analyses described below were compared among the middle speeds of each speed range (i.e. 0.7–0.8, 1.5–1.6 and 3.2–3.3 m s⁻¹ for slow walking, fast walking and running, respectively). Hereinafter, these speeds are referred to as ‘representative speeds’.

(f). Activation ratio between lumbar segments and sacral segments

To evaluate the relative activation between the lumbar and sacral segments, we calculated the ratio between mean MN activity in the main part of the lumbar (sum of activity from L3 to L4) and the sacral segments (sum of activity from S1 to S2) [17]. The activation ratios were compared among the representative speeds for each type of MN activation pattern.

(g). Spatio-temporal activation patterns of motoneurons of each locomotor module

Locomotor modules (figure 1b) were extracted from the processed EMGs using the NMF [14,18] in the representative speeds for each type of MN activation pattern (see the electronic supplementary material for further detail). Then, spatio-temporal activation patterns of MNs were reconstructed from each extracted module. First, the motor output of each module was calculated from the product of the muscle weightings and the corresponding temporal activation patterns (figure 1b). Then, spatio-temporal activation patterns of MNs generated from individual modules were reconstructed from the motor output of each module (figure 1c). These activation patterns of individual modules were summed into a whole activation pattern of MNs over a gait cycle at each representative speed (figure 1d).

(h). Statistics

Differences in the activation ratio between the sacral and lumbar segments among the three representative speeds for each MN activation pattern (slow walking, fast walking and running) were compared using one-way ANOVA with post hoc Holm's test (an updated version of Bonferroni's test). In addition, we compared the number of modules among the three representative speeds by using the non-parametric Kruskal–Wallis one-way ANOVA with the post hoc Steel–Dwass test (non-parametric Tukey's test). Data are presented as the mean and standard error of the mean (mean ± s.e.). Statistical significance was accepted at p < 0.05.

3. Results

(a). Spatio-temporal activity patterns of motoneurons along a rostrocaudal direction in the spinal cord

Using recorded EMGs (typical examples shown in electronic supplementary material, figure S1) and the myotomal charts, we reconstructed the spatio-temporal activation patterns of MNs along the rostrocaudal direction in the spinal cord over the step cycle. The averaged activation patterns across participants from slow walking to fast running are presented in figure 2. These patterns were grouped by hierarchical clustering to evaluate the speed-dependent changes (figure 3a). The activation patterns were grouped into three speed ranges: slow walking (0.3–1.1 m s⁻¹), fast walking (1.1–1.9 m s⁻¹) and running (2.2–4.3 m s⁻¹) (figure 3a).

Figure 2.

Averaged activation patterns of MNs across all participants at all speeds. The speed range and the maximum value of each map are displayed. Colour scale denotes amplitude normalized to the maximum value in each activation pattern. Light blue, deep blue and orange sections indicate three speed ranges (‘slow walking’, ‘fast walking’ and ‘running’, respectively) divided by cluster analysis based on the spatio-temporal activation patterns of the MNs shown in figure 3.

Figure 3.

(a) Cluster analysis for activity pattern of MNs across all speeds and (b) the representative data from each cluster. (a) Dendrograms represent the results of cluster analysis applied to the activation patterns of MNs across all speeds shown in figure 2. The optimal number of clusters was determined by gap statistics [19]. Three distinct clusters are indicated by light blue, deep blue and orange (‘slow walking’, ‘fast walking’ and ‘running’, respectively). (b) MN activation maps at the middle speed of each speed range for the three clusters are shown as representative data for the clusters. The bars underneath denote the stance phase (black) and swing phase (grey) in a gait cycle.

Specifically, in slow walking (all data shown in figure 2, light blue section; representative pattern shown in figure 3b, left), the patterns showed long-lasting synchronous activation of the lower part of the lumbar (L4 and L5) and sacral (S1 and S2) segments during most parts of the stance phase of the cycle (0–50% of the gait cycle). In addition, weak activation in segments from L4 to S1 occurred during most parts of the swing phase (60–85% of the gait cycle). At the end of the swing phase (90–100% of the gait cycle), preceding heel-contact, segments from L4 to S1 were re-activated with a stronger intensity.

However, in fast walking (all data shown in figure 2, deep blue section; representative pattern shown in figure 3b, middle), the stance phase activation was separated into distinct bursts of the whole lumbar (L2 to L5, at around foot contact, 0–20% of the gait cycle) and the sacral segments (S1 and S2, at around toe-off, 40–55% of the gait cycle). Weak activation of segments from L4 to S1 occurred in the first half of the swing phase (60–80% of the gait cycle). At the end of the swing phase (90–100% of the gait cycle), segments from L4 to S1 were also activated as in slow walking.

In running (all data shown in figure 2, orange section; representative pattern shown in figure 3b, right), the activity of the lumbar and the sacral segments were synchronous in the stance phase (0–30% of the gait cycle). During the initial swing (30–45% of the gait cycle), weak activity of the rostral lumbar segment (mainly in L3) occurred. Following this narrow segment activation, weak activity in the wide lumbar segments (L2 to L5) occurred in the middle of the swing phase (45–70% of the gait cycle). During the end of the swing phase (90–100% of the gait cycle), the activity of the segments from L4 to S1 occurred similarly to that in slow walking and fast walking. Inter-subject variability in EMGs during running (reported in [20]) may raise a question about the averaged data of the reconstructed MN activation patterns obtained from individuals. As seen by visual inspection, inter-subject differences in EMG patterns were quite small at the representative running speed (all individual data are shown in the electronic supplementary material, figure S3). This similarity is supported by high correlation coefficients between individual and averaged data (0.95 ± 0.024, mean ± s.d.; electronic supplementary material, figure S3). Therefore, the averaged running data represent the general trends in the individual data.

(b). Activation ratio between lumbar and sacral segments

As locomotion changes from slow walking to fast walking or running, MN activity in the rostral segments was increased relative to that in the caudal segments. This was statistically confirmed by comparing the activation ratio between the lumbar (L3 + L4) and sacral (S1 + S2) segments (figure 4; ANOVA: F_2,48 = 18.3, p < 0.001; post hoc Holm's test: p < 0.01).

Figure 4.

Mean activity ratios of lumbar versus sacral segments at three representative speeds. Error bars indicate the s.e. (Online version in colour.)

(c). Spatio-temporal activation patterns of motoneurons generated from individual modules

To examine the relationships between the locomotor modules and the activation patterns of the MNs, the modules were extracted from representative speeds lying within the three clusters. The number of extracted locomotor modules was significantly lower during slow walking compared with running (Kruskal–Wallis one-way ANOVA: H2 = 7.40, d.f. = 2, p < 0.05; post hoc Steel–Dwass test: p < 0.05; electronic supplementary material, figure S2). Six types of module were extracted from the combined dataset of the three representative speeds (M1–M6, table 1; bars in figure 5 upper row). From these six types of module, different combinations of modules were used among the three representative speed datasets (table 1).

Table 1.

Characteristics of extracted locomotor modules.

			extracted modules
module type	major muscles	timing	slow walking (0.7–0.8 m s−1)	fast walking (1.5–1.6 m s−1)	running (3.2–3.3 m s−1)	peak activation segment
M1	TFL, Gmed	early stance	○	○		L4
M2	MG, LG, SOL, PL	mid–late stance	○	○	○	S1
M3	TA	initial stance/mid swing	○	○	○	L4
M4	BF, SM	late swing–early stance		○	○	L5
M5	VL, VM, GM, Gmed	early stance			○	L3
M6	AM, RF	initial swing			○	L3

Figure 5.

Spatio-temporal activation patterns of MNs generated by individual locomotor modules at three representative speeds. The output of an individual module is explained by the product of a muscle weighting component (top bars) and its corresponding temporal pattern component (the same colour waveform). Based on the output, MN activity generated by an individual module was reconstructed. The activity patterns generated by the same modules are shown in the same column, while those at the same speeds are shown in the same row. These activity patterns at the same speed were summed into a total activation pattern over a gait cycle (the second column from the left). Correlation between the pattern directly reconstructed from the EMGs (the first column from the left) and the pattern reconstructed from the modules is shown just above these two patterns. An enlarged view of the x-axes of muscle weightings is shown in the upper right. The bars underneath denote the stance phase (black) and swing phase (grey) in a gait cycle.

Figure 5 shows the spatio-temporal activation patterns of MNs reconstructed from individual locomotor modules at the three representative speeds. Each module activated the MN pools in somewhat narrow segments (two to four segments) at a specific timing point in the gait cycle (figure 5, third to the sixth columns from the left). The MN activation patterns of individual modules at each speed were summed into a total activation pattern over a gait cycle at each speed (figure 5, second column from the left). The original patterns were reconstructed with high accuracy by summation of the activation from individual modules (r = 0.91–0.98). The most strongly activated segments by each module were located at rostral segments, in most cases, of the newly recruited module types as speed increased (table 1). Specifically, in slow walking, one of the three modules (M2) innervated the MN pools in a sacral segment (peak segment: S2). In addition, the other two modules (M1 and M3) activated a lower lumbar segment (peak segment: L4). In fast walking, the newly recruited modules (M4) mainly innervated lower lumbar and upper lumbar segments, respectively (peak segment: L4). During running, M5 and M6 were newly recruited and they innervated the MN pools in an upper lumbar segment (peak segment: L3).

4. Discussion

The present results confirmed our working hypotheses and demonstrated that (i) MN activity in the rostral segments increased compared with the caudal segments with increasing locomotion speed, and (ii) the three different MN activation patterns used in a wide range of speed are generated by distinct combinations of locomotor modules. These results are consistent with the characteristics of the speed control of the spinal CPGs observed in other vertebrates [9–12]. Therefore, our results support the hypothesis that similar basic locomotor neural circuits are used among different vertebrate species even though they have significant morphological differences and exhibit different locomotion styles (e.g. aquatic or terrestrial, non-legged or legged) [13,21]. Thus, our results indicate a possibility that the commonality of the spinal locomotor circuits can be extended to humans.

Our results showed that the MN activation patterns were divided into three speed ranges (slow walking, fast walking and running; figure 3a). In the fast walking pattern (figure 3b), two clear bursts occurred at around foot contact (0–20% of the gait cycle) in the lumbar segments and at around toe-off (40–55% of the gait cycle) in the sacral segments. This activation pattern was similar to those presented as normal-speed walking patterns in many studies [7,8,16,17]. Additionally, as natural self-selected walking speed is approximately 1.2–1.5 m s⁻¹ [17,22], this activity pattern would be most frequently used as an ordinary pattern in our daily life. Conversely, in the slow walking pattern (figure 3b), long-lasting synchronous activations of the lower part of the lumbar segments (L4 and L5) and the sacral segments (S1 and S2) were observed during most parts of the stance phase of the cycle (0–50% of the gait cycle). Such MN activation patterns have been found in several studies at slow speeds (0.28 m s⁻¹ [7]) and low step frequencies (40 steps min⁻¹ and 40–80 steps min⁻¹ in young and older adults, respectively [23]). Additionally, the long-lasting synchronous activations of the lower lumbar and sacral segments were also observed in toddlers while stepping [17]. It has been demonstrated that posture control is an important aspect for the control of locomotion as a common characteristic in slow-speed walking [24] and toddler walking [25]. Thus, both may represent similar MN activation patterns. Regarding running, to our knowledge, only one previous study [8] examined the MN activation pattern in slow-speed running (1.38–3.33 m s⁻¹, considered as jogging). In this study, we examined it at a faster speed (–4.3 m s⁻¹). As a result, the MNs activation patterns during faster running were almost the same patterns as those during slower running (figures 2 and 3a).

Among the three different MN activation patterns, MN activity in the rostral segments was increased relative to that in the caudal segments as the speed increased (figure 4). Also, the locomotor modules extracted in faster speed (fast walking and running) innervated rostral segments (table 1 and figure 5). Previous studies demonstrated several functional differences among lumbosacral segments considered to be related to the observed speed-dependent change in MN activation ratio between the lumbar and sacral segments. A previous study showed that vertebrates have multiple rhythmogenic modules in their entire lumbosacral enlargement [26]. However, rostral segments have a higher capacity to generate rhythmic activity of MNs than do caudal segments [9,10]. Therefore, although the function of generation of rhythmic MN activity is widely scattered in the lumbosacral enlargement, the capacity is graded along the rostrocaudal direction. In addition, rostral segments of the lumbar cord play a crucial role in the function of the CPG as a leading oscillator that propagates motor bursts to caudal segments [27]. Also, in humans, the rostral segments have been suggested to work as a leading oscillator [3,4]. From the viewpoint of the rostrocaudal functional gradient, in fast locomotion, high activity of the rostral rhythm generator may be required to achieve a high step frequency. Indeed, non-NMDA receptors, which are responsible for receiving glutamatergic input to the locomotor CPG, exist in more rostral lumbar segments and are indispensable for achieving high-frequency locomotor behaviour [11]. This glutamatergic mechanism for faster locomotion might explain the higher activation in the caudal segments at faster speed locomotion presented in this study. However, slower locomotion may not require strong locomotor drive from the rhythm generator in the upper segments. Indeed, it has been demonstrated that posture control is an important aspect of the control of locomotion at slow speeds [24]. In addition, another study suggested that the locomotor generator in the sacral segments is related to body support through sensory inputs to the foot during walking [28]. Therefore, in contrast with the lumbar segment, sacral segment activity might play an important role for posture control, especially in slow walking. Therefore, it is plausible that the observed speed-dependent change in MN activation ratio between the lumbar and sacral segments reflected these functional differences of the spinal segments.

In this study, focusing on the relationships between the MN activation patterns and extracted locomotor modules, we found that three different MN patterns were generated by distinct combinations of locomotor modules (table 1 and figure 5). It is suggested that the locomotor CPG may consist of a timing structure and a spatial pattern formation network [6]. The locomotor modules are considered to be encoded in the spatial pattern formation networks of the spinal CPGs, which activate multiple MN pools [6]. Recent molecular and genetic techniques have revealed that CPGs consist of spinal interneurons, and each type of interneuron plays a particular role in controlling locomotion [1,2]. In addition, spinalized vertebrate studies have shown that mathematically extracted locomotor modules, like those in this study, are organized in the spinal interneuronal circuits [27,29]. Thus, assuming that the locomotor modules consist of spinal interneurons, our results suggest that each interneuronal locomotor module has specific connectivity with MN pools. Indeed, a recent study in zebrafish revealed the existence of a specific combination of individual classes of interneurons and MNs called a ‘locomotor spinal microcircuit’ [12], which has interesting characteristics regarding speed dependency. The authors identified three distinct microcircuits with separate interneuron types innervating slow, intermediate or fast MNs. Furthermore, the microcircuits are sequentially activated from slow to intermediate and fast with increasing speed [12]. Thus, this principle of spinal circuit organization represents a neural mechanism to modulate the locomotor speed by stepwise recruitment of different microcircuits. In addition, a study in mice [30] showed specific connectivity (i.e. microcircuits) between motoneuron subtypes and V1 interneuron subtypes, which regulate locomotor speed [31]. It is not clear to what extent these findings can be extended to humans. Nevertheless, the greater part of the spinal locomotor networks are conserved across vertebrates [1,19]; thus, it is probable that the recruitment patterns of spinal neurons might also be conserved in humans. If the speed-dependent recruitment mechanisms of the microcircuits for locomotion are phylogenetically preserved in humans, this would explain the present result that the different MN activation patterns were generated by different locomotor modules.

In human walking, generation of muscle activity is largely affected by sensory input [32]. In slow walking, as discussed above, large sacral activity in mid-stance is presumably derived from foot-support interactions thorough load feedback. This sacral activity is most probably related to constancy in triceps activity over the speed range and to negative speed dependency in PL activity with regard to load information in slow-speed walking, as discussed in [33]. At higher speeds, changes in locomotor muscle activity are clearly related to reinforcement of sensory feedback depending on speed increases. It is assumed that extensor activity during early stance phase (M5 activity and forward shift of activation timing of M2 in running) was related to reinforcing load feedback (i.e. extensor reinforcing reflex) [34], and hamstring activity at the end of swing (M4 activity) was related to an increase in knee-extension speed (i.e. stretch reflex) [35]. Although there was no doubt about the contributions of these reflex-driven muscle activities to control speed in human locomotion, a modelling study showed that reflex-based muscle activity alone without CPGs cannot control locomotor speed [36]. In addition to reflex control of locomotion, it has been shown that sensory feedback affects lumbar burst generators in CPGs [4,37]. Vibration of leg muscles facilitated locomotor-like muscle activity evoked by spinal electromagnetic stimulation to the lumbar segment [4]. A study in mice demonstrated more direct evidence that ascending afferent pathways from the sacral segments to the lumbar segments enhance locomotor bursts [37]. Assuming that similar mechanisms are shared between mice and humans, strong afferent feedback at higher speeds may powerfully facilitate the lumbar burst generators of the CPGs.

Regarding the organization of locomotor networks, Hof et al. [38] showed that locomotor EMGs in the speed range of 0.75–1.75 m s⁻¹ could be estimated by two simple functions, one constant and one proportionally increasing with walking speed (i.e. fixed CPG networks). Nevertheless, this model cannot explain the large activity in the triceps and PL at mid-stance in slow walking or the rapid increase of extensor activity after the walk–run transition. Instead, as Pearson [39] proposed, CPG networks would be flexibly reorganized by sensory feedback. Reorganization of CPG networks depending on changes in locomotor speed have been revealed in fishes and mice by recent molecular studies [2]. Thus, regarding speed control in locomotion, sensory input would be related to not only reflex-driven muscle activity, but also rhythmic burst generation and reorganization of CPGs, probably contributing to the speed-dependency of MN activity and locomotor modules observed in this study.

There are several limitations regarding the method of reconstructing the spinal MN activation. EMG cross-talk is always a potential issue with recordings of surface EMGs. In a previous study, it was demonstrated by modelling the potential effect of different levels of cross-talk in the EMGs that the cross-talk has little effect on the estimation of MN activity patterns [17]. The study showed that the level of cross-talk from adjacent muscles increased incrementally (from 10% to 100%); nevertheless, the appearance of a new location of activity or notable temporal shifts of the activity did not occur. Additionally, the number of muscles and the muscle type are also important variables. Regarding this point, it has been demonstrated that the activity patterns, analysed with two different sets of muscles (12 muscles and 20 muscles), are relatively robust [16]. Presumably, this result stemmed from the fact that each segment in the spinal cord innervates multiple muscles and each muscle is innervated by several segments to the contrary. The activity patterns in this study were similar to those in previous studies using various sets of muscles [7,8,16,17,23]. However, it should be kept in mind that the sets of muscles analysed would affect the extraction of locomotor modules by NMF [40].

In conclusion, we found the following spinal activation patterns regarding speed control of human locomotion: (i) MN activity in the rostral segments increased compared with the caudal segments with increasing locomotion speed; and (ii) different MN activation patterns are generated by distinct combinations of locomotor modules. These results are consistent with the speed control characteristics of vertebrate CPGs. This commonality supports the hypothesis that basic locomotor neural circuits are highly conserved across significant morphological differences and phylogenetic distances in vertebrates [13,21]. Thus, our results provide important insight into not only human locomotor control but also the evolution of vertebrate locomotion.

Ethics

The study was in accordance with the Declaration of Helsinki and was approved by the local ethics committee of the National Rehabilitation Center for Persons with Disabilities (Tokorozawa, Japan).

Data accessibility

Electronic supplementary material can be found in http://dx.doi.org/10.5061/dryad.4ch98 [41].

Authors' contributions

H.Y. and K.N. designed the experiment. H.Y. performed analyses and collected data. All authors interpreted results, wrote the manuscript and approved the final manuscript.

Competing interests

We have no competing interests.

Funding

This work was supported by a Grant-in-Aid for Japan Society for the Promotion of Science Fellows (15J09583) to H.Y.