A comprehensive dataset on biomechanics and motor control during human walking with discrete mechanical perturbations

Introduction

Human walking is inherently unstable, and the feedback control mechanisms to maintain balance are complex, including spinal reflexes as well as vestibular and visual feedback. Reflexes play an essential role in ensuring that corrective actions are taken quickly after an unexpected perturbation due to uneven ground or external forces (Nielsen & Sinkjaer, 2002). Absence or insufficiency of reflexes would increase the risk of falling and injury. On the other hand, hyperactive reflexes can cause abnormal movements with certain neurological disorders, such as cerebral palsy and stroke (Flux et al., 2021; Li & Francisco, 2015). Clinically, reflexes can be evaluated with the familiar tendon tap tests. During gait, the Hoffmann reflex (H-reflex) can be used, which involves electrical stimulation of a sensory nerve, which excites the stretch reflex pathway and causes a muscle activation response that can be detected with electromyography (EMG) (Belanger & Patla, 1984; Zehr, Komiyama & Stein, 1997; Zehr, Stein & Komiyama, 1998).

However, reactive responses to perturbations require mechanisms beyond reflexes. Strategies to recover from perturbations can involve longer time scales and complex neural mechanisms, such as altered foot placement in the contralateral limb (Bruijn & van Dieën, 2018). Electrical stimulation or clinical reflex tests will provide limited insight into the capability of a patient or research participant to maintain balance, and the neural control mechanisms that are responsible.

A more ecological approach is the use of mechanical perturbations to elicit reactive responses during a functional activity such as gait. Wearable exoskeletons have been used to induce stretch reflexes during walking (Stevenson et al., 2015), which is suitable for basic research, but the weight and discomfort of this approach are important limitations. Recently, it has been demonstrated that rapid changes in belt speed can induce reflexes during treadmill walking (Sloot et al., 2015). Such perturbations are similar to those that occur naturally but with controlled timing and magnitude. In Sloot et al.’s (2015) work, perturbations were applied only at one time in the gait cycle, but it is known from other studies that reflex gains are modulated throughout the gait cycle (Zehr & Stein, 1999). To better understand reflexes and the mechanisms behind their control we need to look at responses to perturbations that occur at multiple points in the gait cycle. It is also important to look at the entire perturbation response, including non-reflex responses on a longer time scale, and using full body biomechanical measures in addition to electromyography.

In recent years, it has been emphasized that open datasets are of vital importance to aid scientific advancements by inspiring new analyses, reproducibility, and good data practices (McKiernan et al., 2016). One of the most widely used datasets for human gait is a set of normative data from David Winter (Winter, 2009). It is only for a small number of participants over a few gait cycles, yet it has enabled many studies, like control design for prosthetics (Sup, Bohara & Goldfarb, 2008). More extensive datasets, including gait during various speeds and conditions, have since been published (Fukuchi, Fukuchi & Duarte, 2018), often including motion capture, ground reaction forces (GRF), and EMG. Datasets that include data from inertial measurement units (IMUs), together with conventional instrumentation, are especially valuable because this allows the validation of IMU processing methods (Camargo et al., 2021). However, there are limited publicly available data on human gait during mechanical perturbations.

Moore, Hnat & van den Bogert (2015) published data from fifteen participants walking at three different speeds, during normal walking and during walking with pseudo-random belt speed fluctuations. However, the pseudo-random perturbations, though destabilizing, were small, the standard deviation of belt speed variations was 0.06–0.21 m/s, and may not have exceeded the sensory threshold for reflex responses (Moore, Hnat & van den Bogert, 2015). Liu, Park & Finley (2022) shared data from eleven participants walking at their preferred walking speed with six different perturbation magnitudes with the objective of determining if the reference point influences how dynamic balance is controlled during perturbed walking. Leestma et al. (2023) shared data from eleven participants walking with perturbations caused by platform movement in the transverse plane. However, these datasets do not include EMG data, which can provide information about neuromuscular mechanisms (Huang et al., 2011).

This article presents able-bodied human walking data from 10 participants. The participants walked on a treadmill while belt speed changes were generated at controlled times during randomly selected gait cycles. Motion, ground reaction force, and EMG data were recorded. Motion capture was performed with IMUs in addition to cameras. The data has been organized and made available following best data-sharing practices for other researchers (White et al., 2013). We expect that this work will assist data-driven analyses for the development of models and control systems for robotics and lay the groundwork for clinical applications of neural control identification.

Methods

Equipment

The data were collected in the Human Motion and Control Laboratory at Cleveland State University. The instrumentation as described by Moore, Hnat & van den Bogert (2015) includes and R-Mill instrumented treadmill (Forcelink, Culemborg, Netherlands) and motion capture system. Additionally, Trigno Legacy EMG Sensors (Delsys Inc., Natick, MA, USA) with analog EMG and three channel Dual Range Accelerometer were used to gather analog EMG data and Trigno IM Sensors (Delsys Inc., Natick, MA, USA) with digital EMG and nine channel Quad range IMU chip were used to gather inertial measurement unit data.

Similarly to Moore, Hnat & van den Bogert (2015), Cortex 5.5.01 software (Motion Analysis, Santa Rosa, CA, USA) received the force plate data, analog sensor (EMG) data, and 3D marker data via a National Instruments USB-6255 data acquisition unit. Cortex sends the data, in real-time to the D-Flow 3.34.0 software (Motek Medical, Amsterdam, Netherlands), which is responsible for controlling the speed of the treadmill belts. The force plate and marker data were recorded at a rate of 100 Hz, while the analog EMG data was recorded at 1,000 Hz.

A Delsys Trigger Module (Delsys Inc., Natick, MA, USA) and Phidget InterfaceKit 8/8/8 1018_2 (Phidgets Inc., Calgary, Canada) were used to synchronize the Trigno IMU sensors with the existing laboratory setup. The trigger module was plugged into the IM sensor base and connected to the Phidget board; which was connected to the D-flow computer via a USB connection. When recording began in the D-Flow program a 5 V signal was sent to the trigger module, beginning recording of IMU sensor data in the EMGworks 4.7.9 software (Delsys Inc., Natick, MA, USA).

The layout of the treadmill and motion capture equipment, similar to Moore, Hnat & van den Bogert (2015) can be seen in Fig. 1. The walking surface of the treadmill is 500 × 1,800 mm, for each belt. The coordinate system of the motion capture system is defined so that the positive z-direction is pointing towards the back of the treadmill, the positive y-direction points upward, and the positive x-direction is to the right. When calibrating the cameras, a point on the treadmill’s surface is set as the origin point for the coordinate system, it was also used as the ground reaction force measuring system origin point (Moore, Hnat & van den Bogert, 2015).

A flow chart showing how the instrumentation interacts with the computers can be seen in Fig. 2. The figure also shows the file names that the protocol produces, along with a brief description of what can be found in each file.

Perturbation signals

A Lua script was created within the ‘Script’ module of the D-Flow software to generate the perturbations and control the speed of the treadmill belts. The script received vertical ground reaction force data from the Mocap module in D-Flow. Once the treadmill was at the normal walking speed of 1.2 m/s, the vertical GRF was used to keep track of the participant’s gait cycles, specifically when they entered and exited the stance phase (heel strike and toe-off, respectively) to determine the average duration of their unperturbed stance phase. At heel strike, a random number generator determines if a perturbation will occur during that gait cycle. There was a 1 in 6 (about 17%) chance that a perturbation will occur for any gait cycle, except for the two gait cycles immediately following a perturbation. If a perturbation is set to occur, another random number generator determines what type of perturbation (10%, 20%, … or 80%) will occur during that gait cycle. A running average of the stance duration was used determine when the command perturbation time should occur. Since we could not know the exact stance duration while the perturbed gait cycle was occurring, we verified that perturbations were delivered at the correct time during post processing. During each perturbation, the right belt speed will increase from 1.2 m/s to 1.95 m/s for 25 ms at a rate of 15 m/s². The speed remains at 1.95 m/s for 75 ms, then decreases to 1.2 m/s for 25 ms, at a rate of 15 m/s². The left belt remains at a constant 1.2 m/s throughout the entire trial. The complete Lua script that generated the perturbations is included in Supplemental File 1. A sample of the ground reaction forces and belt speed can be seen in Fig. 3, the code to generate this figure is available in Supplemental File 2.

Protocol

Experimental setup

The experimental protocol consisted of seven 5-min trials of perturbed walking. Before each participant’s data-gathering session, the motion capture system was calibrated using the manufacture’s recommended procedure. Once participants arrived for their data-gathering session, they changed into athletic shoes, shorts, and a t-shirt. Before changing, participants were given an EMG and instructions on where and how to apply an EMG on their gluteus maximus muscle.

Participants were then asked to step into a full-body safety harness. After adjusting the straps to fit, they were asked to take a few steps to ensure that the harness did not affect their gait. Once in the harness, eight wireless EMG sensors were placed on the right leg. The sensors were placed on the belly of the following muscles: tibialis anterior, lateral gastrocnemius, soleus, rectus femoris, vastus lateralis, vastus medialis, biceps femoris, and gluteus maximus. Locations for these sensors/muscles can be seen in Fig. 4. Eight IMUs were placed on the participants in the following locations: on the trunk at T10, the sacrum, left and right thigh, left and right shank, and left and right foot, as seen in Fig. 4. The EMGs and IMUs were wrapped in pre-wrap to reduce the risk of sensors falling off during the trials. Twenty-five reflective markers were then placed on anatomical landmarks on the body according to the lower body marker set, as seen in Fig. 4; see Table 2 for anatomical landmark names and locations.

Table 2:

Descriptions of the names and locations of the 25-marker, marker set used in this study.

Marker number	Marker name	Anatomical location
1	T10	T1O
2	SACR	Sacrum bone
3	NAVE	Navel
4	XYPH	Xiphoid process
5	STRN	Sternum
6	LASIS	Pelvic bone left front
7	RASIS	Pelvic bone right front
8	LPSIS	Pelvic bone left back
9	RPSIS	Pelvic bone right back
10	LGTRO	Left greater trochanter of the femur
11	FLTHI	Left thigh
12	LLEK	Left lateral epicondyle of the knee
13	LATI	Left anterior of the tibia
14	LLM	Left lateral malleolus of the ankle
15	LHEE	Left heel
16	LTOE	Left toe
17	LMT5	Left 5th metatarsal
18	RGTRO	Right greater trochanter of the femur
19	FRTHI	Right thigh
20	RLEK	Right lateral epicondyle of the knee
21	RATI	Right anterior of the tibia
22	RLM	Right lateral malleolus of the ankle
23	RHEE	Right heel
24	RTOE	Right toe
25	RMT5	Right 5th metatarsal

DOI: 10.7717/peerj.17256/table-2

Note:

The column labeled marker name matches the column headers in the motion capture data files at Zenodo, doi: 10.5281/zenodo.10557913.

Testing protocol

Once the instrumentation and markers were placed on the participant, the force plate was zeroed. Participants were then asked to step onto the treadmill, and the safety rope was attached to the harness so that no additional forces were acting on the participant when walking but would prevent a complete fall if their balance was lost at any point during the trials.

After securing the harness, participants were given a verbal countdown to the start of the trial. When the trial begins, the treadmill belts accelerate to 1.2 m/s at a rate of 0.2 m/s², and the recording of EMG, IMU, motion capture, force plate, and treadmill belt speed data begins. There are then 15 s of walking before the perturbations will start occurring. Perturbations occur randomly during the stance phase of the participant’s right leg for the remainder of the trial, as described above. Five minutes from the start of the trial, both belts will decelerate at a rate of 0.2 m/s² until the belt speeds are 0 m/s. Participants were then given at least a 5-min break between trials to rest and for the investigator to ensure none of the sensors became unsecured during the trial.

Results

Here, we present some basic results. A description of the raw data, along with an overview of some computed variables, is given.

Processed data

Basic analysis is shown in this section to show that the methods and data described can provide meaningful results. Data analysis was performed using MATLAB (Mathworks, Natick, MA, USA). The EMG signals were advanced 48 ms to compensate for the known lag in the Delsys system. The raw EMG signal was filtered using a second-order bandpass Butterworth filter (30–300 Hz), rectified, smoothed with a second-order lowpass Butterworth filter (20 Hz) to obtain the EMG envelope. All filters were run forward and backward in time to eliminate time lag.

The vertical and horizontal force data, EMG envelope, and motion capture data for each participant were divided into gait cycles using the ground reaction forces recorded from the right force plate. For each gait cycle, the angle of the trunk, right/left hip, right/left knee, and right/left ankle are calculated from the sagittal plane marker coordinates (Flux et al., 2021). Then, using the perturbation information recorded from D-Flow, each gait cycle was labeled based on the kind of perturbation that occurred during that gait cycle: no perturbation, 10%, 20%, … 80%. For each perturbation type, the corresponding gait cycles were averaged after removing outliers. EMG was normalized to the peak of the non-perturbed average. The standard error of the mean (SEM) was calculated for all averages. Figure 5 shows average EMG, joint angles and ground reaction forces for each perturbation type. The code to generate this figure can be found in Supplemental File 2.

Typically, for each participant, there are 1,600+ gait cycles without perturbation, and 22–30 gait cycles for each type of perturbation. The averaged belt speed signals in Fig. 5 demonstrate that the perturbations were consistently delivered at the same time, though not perfectly. Some variation in perturbation timing is inevitable because the timing is specified as a fraction of the stance phase. Stance phase duration is estimated in real time, from previous gait cycles and cannot predict the duration of the stance phase during which the perturbation is delivered.

Discussion

This dataset contains motion, ground reaction force, IMU, and EMG data from normal and perturbed gait data, where perturbations were generated at controlled times during randomly selected gait cycles. The EMG and motion data for the unperturbed gait are consistent with what has been seen in previous studies on normal gait (Winter & Yack, 1987; Wang & van den Bogert, 2020). The EMG data for the perturbed gait shows evidence that a reflex occurs (Fig. 5), as shown for a simpler protocol in Sloot et al. (2015). The number of gait cycles with the same perturbation timing (22–30) appears to be large enough to obtain a reliable average of the EMG response, despite the inherent noisiness of EMG.

The EMG data, after averaging, shows evidence that the magnitude of the reflex response is dependent on when the perturbation occurred during the stance phase (Fig. 5). This aligns with what has been seen in H-reflex studies (Belanger & Patla, 1984; Zehr, Komiyama & Stein, 1997; Capaday & Stein, 1986). However, in our study, mechanical perturbations were applied, which is a more ecological approach to induce and study reflexes during gait, and how the reflex gains may be modulated during the gait cycle. It is also possible that the muscle activity that is seen in response to the perturbations could be proportional to other sensory inputs or reactions after sensory integration, and is not only due to a stretch reflex (Safavynia & Ting, 2013). Further analysis on the data should be done to determine if this is the case.

One appealing application of the data would be the development of human-inspired and intuitive reflex control systems for assistive devices, such as prostheses and exoskeletons (Geyer & Herr, 2010; Ferris et al., 2012). Typically, reflex control parameters are tuned or optimized to perform predictive simulations of human gait, using objectives such as minimal metabolic energy expenditure (Geyer & Herr, 2010; Veerkamp et al., 2021). Even when the simulated gait looks good, the controller may not perform well during perturbations, in simulation or in robotic devices. We suggest that feedback controllers for robotic assistive devices can be optimized to generate movement responses after each perturbation that matches the human responses in this dataset.

The provided dataset contains IMU data, which we have not performed any analysis on, but could be useful for evaluating methods for processing IMU data. IMU technology is cheap and increasingly used in clinical motion capture applications. Our data may help validate the use of IMU data for perturbation responses. IMU processing typically uses machine learning methods, trained and validated on normal gait data (Cho et al., 2018), which can introduce bias towards normal gait in clinical applications. This dataset can test the capability of ML to correctly quantify unusual or unexpected features in human gait.

Some technical limitations should be taken into consideration. First, it should be noted that the GRF data, specifically the anterior-posterior COP, is unreliable during belt acceleration and deceleration (Hnat & van den Bogert, 2014). Second, there is not IMU data for all participants. This is from an issue that arose with the IMU sensors dying during the data gathering sessions, leading to incomplete trial sets for participants 2, 15, 90, and 99. Finally, due to the large volume of data, careful curation is needed. Markers were occasionally dropped. Sometimes, a participant might touch the force plate during the swing phase, or take an abnormal step, which then results in an incorrectly timed perturbation. In the code that generated Fig. 5 (Supplemental File 2), such situations were automatically detected and excluded from the gait cycle averaging process.

Conclusion

We have presented a comprehensive dataset of motion, ground reaction loads, and EMG from human participants during perturbed walking. The raw data from this study is available publicly for reuse. Although this data has multiple other uses, we believe it is ideal for identifying control laws. Researchers working on mathematical models for gait control may find this data useful as a way to generate and validate their models.

Supplemental Information