Abstract
Common in-lab, marker-based gait analyses may not represent daily, real-world gait. Real-world gait analyses may be feasible using inertial measurement units (IMUs), especially with recent advancements in open-source methods (e.g., OpenSense). Before using OpenSense to study real-world gait, we must determine whether these methods: (1) estimate joint kinematics similarly to traditional marker-based motion capture (MoCap) and (2) differentiate groups with clinically different gait mechanics.
Healthy young and older adults and older adults with knee osteoarthritis completed this study. We captured MoCap and IMU data during overground walking at participants’ self-selected and faster speeds. MoCap and IMU kinematics were computed with appropriate OpenSim workflows. We tested whether sagittal kinematics differed between MoCap- and IMU-derived data, whether tools detected between-group differences similarly, and whether kinematics differed between tools by speed.
MoCap data showed more flexion than IMU data (hip: 0-47 and 65-100% stride, knee: 0-38 and 58-91% stride, ankle: 18-100% stride). Group kinematics differed at the hip (young extension > knee osteoarthritis at 30-47% stride) and ankle (young plantar flexion > older healthy at 62-65% stride). Group-by-tool interactions occurred at the hip (61-63% stride). Significant tool-by-speed interactions were found, with hip and knee flexion increasing more for MoCap than IMU data with speed (hip: 12-15% stride, knee: 60-63% stride).
While MoCap- and IMU-derived kinematics differed, our results suggested that the tools similarly detected clinically meaningful differences in gait. Results of the current study suggest that IMU-derived kinematics with OpenSense may enable the valid and reliable evaluation of gait in real-world, unobserved settings.
1. Introduction
Gait analyses are a common research tool in biomechanics to detect changes or differences in gait kinematics, which are used as markers of pathology, rehabilitation status, or function (Boyer et al., 2012; Curran et al., 2018; Rodda et al., 2004). While typical laboratory-based gait analyses provide precise, controlled kinematic measurements, they may not accurately represent how people move in uncontrolled, free-living environments. In real-world settings, individuals often walk slower and with shorter stride lengths compared to laboratory settings (Foucher et al., 2010; Hutchinson et al., 2019; Kim et al., 2020; Takayanagi et al., 2019). A mismatch between laboratory-measured gait and real-world gait could lead to incomplete or even spurious conclusions about disease status, intervention success, or function. Thus, accurately and reliably evaluating gait kinematics in real-world, daily life settings is critical.
While traditional, marker-based optical motion capture (MoCap) is challenging to perform outside a laboratory, collecting gait data in free-living environments may be feasible using wearable devices such as inertial measurement units (IMUs). IMUs can be worn by participants as they engage in real-world, daily activities, and many have battery and data storage capabilities that allow for days of continuous data collection. However, the use of IMUs for gait analyses outside the lab is rare. The rarity of real-world IMU gait analysis is partially because the processing needed to transform IMU data (e.g., sensor-frame angular velocity and linear acceleration) into familiar kinematic outcomes (e.g., joint angles) requires data processing and analysis methods that are substantially different from those of optical motion capture gait analyses [e.g., (Dorschky et al., 2019; Seel et al., 2014; Vitali et al., 2017)]. Consequently, many studies use commercially available automated software packages that estimate spatiotemporal outcomes and joint kinematics from IMU data [e.g., (Al-Amri et al., 2018; Monda et al., 2015; Nüesch et al., 2017; Washabaugh et al., 2017)]. While these systems are relatively easy to use, the algorithms are proprietary, and thus a study using one system cannot be replicated or compared against a study using a different system.
Recently, developers at OpenSim released an open-source, freely available software package to estimate kinematics from IMU data (Al Borno et al., 2021; Seth et al., 2018). This software (OpenSense) combines sensor orientation estimates from IMU data with global optimization methods [i.e., inverse kinematics (Lu and O’Connor, 1999)] and musculoskeletal models with anatomical constraints to solve for traditional kinematic data like joint or segment angles. This approach may provide the IMU community an accessible, standardized way to report gait kinematics. Recent studies using OpenSense showed reasonable agreement between IMU and MoCap data for young adult participants (Al Borno et al., 2021; Bailey et al., 2021; Slade et al., 2021). Specifically, these studies demonstrated that MoCap and IMU kinematics were similar in young adult cohorts when offsets between MoCap and IMU model poses were corrected for before running the inverse kinematics procedure. Before OpenSense can be used to estimate gait kinematics for clinical populations in real-world settings, we must demonstrate its ability to detect valid and clinically meaningful differences similarly to standard marker-based optical motion capture without relying on adjusting model poses to data that cannot be captured in real-world settings.
The aim of this study was twofold. First, we compared IMU- and MoCap-based gait kinematics as calculated using inverse kinematics with OpenSim. In keeping with the goal of using OpenSense to estimate kinematics from unobserved, real-world IMU data, we applied IMU calibration procedures without adjusting for differences between IMU model poses and MoCap model poses, as MoCap data would not be available in unobserved settings. Because this approach will ultimately be applied to monitor clinically relevant changes within an individual or differences in gait kinematics across groups, we also sought to determine whether OpenSense was consistent with optical motion capture methods in identifying clinically relevant differences across multiple cohorts. Thus, the second aim of this study was to determine whether kinematics calculated from IMU-based and MoCap-based kinematics differed between young adults, older adults, and older adults with knee osteoarthritis (i.e., whether there was a tool-by-group interaction). Finally, because IMU kinematics may be more sensitive to increased speed than MoCap data (Potter et al., 2019), we compared kinematics between preferred and faster-than-preferred walking speeds and tested for a tool-by-speed interaction.
2. Methods
2.1 Protocol summary
Participants included 9 young adults (28.7±4.2 yr), 9 asymptomatic older adults (72.1±3.4 yr), and 9 older adults with knee osteoarthritis (69.2±4.6 yr) who were recruited as part of a previous study (Hafer et al., 2020). Each group included 4 females and 5 males. All participants completed IRB-approved informed consent procedures before completing any study tasks. All participants completed overground walking trials in a laboratory.
We collected gait data from participants’ right leg (for young and asymptomatic older adults) or the leg with the greater radiographic osteoarthritis severity [for older adults with knee osteoarthritis; severity evaluated by Kellgren-Lawrence grade (Kellgren and Lawrence, 1957)]. We recorded MoCap and IMU data simultaneously as participants walked overground. Data included 10 trials each at preferred and faster-than-preferred (verbal instruction: “walking to catch the bus”) speeds. Participant characteristics and basic spatiotemporal measures are shown in Table 1.
Participant characteristics and walking speeds
2.2 Inertial Measurement Unit Methods
We affixed IMU sensors (Opal v2, APDM Inc., Portland, OR USA) to each participant’s sacrum, lateral thigh, lateral shank, and dorsal foot. Thigh and shank sensors were placed on participants’ skin with double-sided tape and then secured with elastic wrap. Thigh and shank marker clusters were placed and wrapped on top of (superficial to) the sensors.
IMU sensors captured data at 128 Hz. Sensors were synchronized to each other and collected data continuously for the duration of the session. IMU data collection occurred simultaneously with MoCap data capture, but the two systems were not synchronized in real time. We recorded the beginning and end of walking trials using an IMU event trigger and a data log to match MoCap trials to corresponding IMU trials (one trial = one walk across the MoCap volume). Using trigger events, we split continuous IMU data into walking trials in MATLAB. We extracted one stride from each walking trial based on gait events identified using a one-dimensional continuous wavelet transform of the foot vertical acceleration data (Baroudi et al., 2020; Hafer et al., 2020). Finally, we used a Kalman filter to estimate sensor orientations in a global reference frame (Holmstrom, 2016) and exported the IMU orientation data in quaternion format.
We used the OpenSim “gait2354” model (Delp et al., 2007) to calculate kinematics. The model had virtual IMU sensors placed such that the axes for each sensor approximated the orientation of the native axes of each sensor when a participant stood in a neutral position. For each walking trial, we created a new calibrated model based on a frame of static data that immediately preceded the walking trial. This frame of static data was identified based on raw sacrum sensor accelerometer and gyroscope data, with the assumption that, immediately prior to beginning to walk, an individual would be standing in an approximately neutral position. The OpenSense calibration procedure registers the relative orientation of a sensor’s data in the global reference frame to an assumed model pose, effectively giving a sensor-to-model offset. For this study, we assumed that participants were in a neutral posture immediately before beginning to walk, and we did not attempt to reduce model posture offsets between IMU and MoCap model calibrations.
For each trial of IMU data, we used the calibrated model (i.e., one model per trial for IMU data) to calculate the generalized coordinates using the inverse kinematics algorithm in OpenSim (Lu and O’Connor, 1999). We extracted pelvic tilt, hip flexion, knee flexion, and ankle flexion angles from the inverse kinematics analyses for each trial. All kinematic time series were normalized to 0-100% of the gait cycle.
2.3. Optical Motion Capture Methods
To compare our IMU data to MoCap data, we used a standard marker set to model the pelvis and leg. The marker set included anatomical markers on the right and left anterior and posterior superior iliac spines, right and left greater trochanters, medial and lateral femoral epicondyles and malleoli, first and fifth metatarsal heads, and lateral and medial heel (note—foot markers were placed on shoes). Rigid clusters of four non-collinear markers each tracked the thigh and shank. We collected MoCap data with 15 cameras (Vicon, Oxford UK) capturing at 120 Hz and 2 force plates (AMTI, Watertown MA) capturing at 1200 Hz.
We used Visual3D (c-motion, Germantown, MD USA) to perform initial processing of MoCap data. We first low-pass filtered marker data at 8 Hz and then identified gait events (e.g., heel strike and toe off) automatically using force plate contact. Subsequently, we exported marker and force data for a static calibration trial and any walking trial containing a stride of data with clean force plate contact (at least 3 strides per condition for each participant). After exporting from Visual3D, we used a Matlab pipeline to rotate the data to match the OpenSim coordinate system convention and convert files to the OpenSim format.
After converting marker data to OpenSim format, we scaled the gait2354 model (Delp et al., 2007) for each participant using participant body mass and the marker data from the static calibration trial. We then modified each scaled model by locking the subtalar joint, such that the only degree-of-freedom at the ankle was flexion. The pelvis, thigh, shank, and foot were scaled based on anatomical markers on these segments (i.e., bilateral ASIS and PSIS markers, greater trochanter, medial and lateral femoral epicondyles, medial and lateral malleoli, 1st and 5th metatarsal heads and calcaneus).
For each trial, we used the subject-scaled model (one model per subject for marker data) to calculate the generalized coordinates using the inverse kinematics algorithm. We then extracted the same coordinates (pelvic tilt, hip flexion, knee flexion, and ankle flexion angles) for comparison with the IMU data. All kinematic time series were normalized to 0-100% of the gait cycle.
2.4 Statistical Analysis
Because we identified gait events differently for MoCap and IMU data (i.e., a MoCap gait cycle may have been slightly offset from an IMU gait cycle), we corrected for inter-tool timing differences for each speed condition before running statistical analyses. To do this, we first identified the difference in the timing of the peak knee flexion angle between MoCap and IMU trials. Then, we time-shifted IMU data forward or backward such that the peak knee flexion angle occurred at the same percentage of the gait cycle for both IMU and MoCap data. Data were then averaged within each tool and speed for each participant, such that statistical analyses included four time series per participant per joint (i.e., MoCap and IMU at preferred and faster speeds). We compared time-series waveforms for pelvic tilt and hip, knee, and ankle flexion across tools (MoCap vs. IMU), groups (young, healthy older, vs. older with knee OA), and walking speeds (preferred vs. faster) with a continuous 3-way ANOVA implemented using statistical parametric mapping [SPM, (Pataky, 2010)]. Where we found significant group effects or significant interactions, we used t-tests implemented in SPM for post-hoc comparisons.
3. Results
Marker data showed more anterior pelvic tilt than IMU data (0-100% gait cycle, p<0.001) and more flexion than IMU data across joints (hip: 0-47 and 65-100% gait cycle; knee: 0-38 and 58-91% gait cycle; ankle: 18-100% gait cycle; all differences p<0.001) [Figure 1]. Group kinematics differed at the hip (young hip extension > knee OA hip extension at 30-47% gait cycle, p<0.001) and ankle (young plantar flexion > older healthy plantar flexion at 62-65% gait cycle, p=0.023) [Figure 2]. Critically, group by tool interactions were minimal (61-63% gait cycle at hip, p=0.046) and did not overlap with main effects of group or tool.
Sagittal pelvis, hip, ankle, and knee angles across tools (mean±SD). Shaded grey rectangles indicate difference between tools (p<0.05) by statistical parametric mapping.
Sagittal pelvis, hip, ankle, and knee angles across groups (mean±SD). Shaded grey rectangles indicate significant main effect of group (p<0.05) by statistical parametric mapping with accompanying text indicating post-hoc group differences.
As expected [e.g., (Fukuchi et al., 2019)], there were significant effects of speed on kinematics (Figure 3). At the faster walking speed, the pelvis was more anteriorly tilted (3-53% and 59-94% gait cycle, p<0.01) and the hip was more flexed at 0-22 and 60-100% gait cycle (p<0.01) and more extended at 37-52% gait cycle (p=0.01). At the faster speed, the knee was more flexed from 5-21 and 52-71% gait cycle (p<0.01) and more extended from 81-92% gait cycle (p=0.014). Ankle dorsiflexion was greater at faster speed from 1-30 and 67-100% gait cycle (p<0.001) and ankle plantar flexion was greater at faster speed from 42-64% gait cycle (p<0.001). There were significant tool-by-speed interactions at the pelvis and all three joints, with marker data having greater increases in flexion than IMU data for faster speeds at the hip and knee (hip: 12-15% gait cycle, p=0.002; knee: 60-63% gait cycle, p=0.045). There were no significant group-by-tool-by-speed interactions. Full statistical results can be found in Table 2.
Statistical results for all comparisons. %s indicate ranges of the gait cycle found to be significantly different via SPM. Post-hocs indicate direction of main effect or interaction effect differences. Not sig. indicates a main, interaction, or post-hoc effect was not significant. N/a indicates a post-hoc comparison was not run because a main or interaction effect was not significant.
Sagittal pelvis, hip, ankle, and knee angles across speeds (mean±SD). Shaded grey rectangles indicate significant main effect of speed (p<0.05) by statistical parametric mapping.
4. Discussion
The ability to reliably measure kinematics outside the lab setting with IMUs would expand our understanding of how age, pathology, or interventions affect real-world gait. In Thus, the current study focused on determining whether lower extremity sagittal plane kinematics differed when derived from traditional MoCap vs. wearable IMUs and whether differences across young adults, older healthy adults, and older adults with knee OA could be detected similarly with both tools. We found that lower extremity kinematics were significantly different between MoCap- and IMU-derived data, but that between-group differences were detected similarly by both tools (i.e., minimal group × tool interaction). Our results align with differences we would expect between young and older adults: greater peak hip extension and ankle plantar flexion in young adults compared to older adults with or without knee OA (Boyer et al., 2017). These results suggest that, while kinematics derived from these two systems differ, kinematics derived using IMU data and OpenSense kinematics may enable detection or tracking of valid and clinically relevant gait differences.
Our finding of between-tool differences in kinematics differs from other recent studies that found minimal differences in kinematics between MoCap and IMU data using an OpenSim workflow (Al Borno et al., 2021; Bailey et al., 2021; Slade et al., 2021). These between-study differences are likely the result of different model calibration procedures for the IMU data. In the current study, we used periods of apparently static data as calibration time points, assuming that participants were standing in an approximately neutral posture at these times. Any differences in participant posture from neutral or between assumed IMU alignment and actual anatomical reference frames would therefore be carried forward into inverse kinematics results. In contrast, Al Borno et al. and Bailey et al. corrected for differences between MoCap and IMU model postures and Slade et al. carefully controlled participant posture during calibration data capture (Al Borno et al., 2021; Bailey et al., 2021; Slade et al., 2021). Compared to a measured or carefully controlled IMU calibration posture, our assumed IMU calibration postures would be expected to introduce some differences between our IMU and MoCap-derived kinematics. However, if researchers or clinicians plan to implement an OpenSense workflow in unobserved, real-world data collections, measured or carefully controlled calibration postures would not be available. Thus, our results may provide a more realistic estimate of the accuracy of OpenSense kinematics in unobserved data collections.
The between-group kinematic differences in the current study agreed with age- or knee OA-related differences in previous studies. This finding, despite overall differences between IMU and MoCap kinematics, supports the use of IMU-derived data for detecting or tracking clinically meaningful differences in sagittal gait kinematics. In addition to older adults having less hip extension than young adults (Boyer et al., 2017), sagittal hip range of motion may also decrease with increasing knee OA severity (Astephen et al., 2008). Both of these findings agreed with the current result of less hip extension during mid-late stance in older adults with knee OA compared to young adults. Our finding of less plantar flexion at 62-65% of the gait cycle in older compared to young adults agrees with the established evidence of decreased peak plantar flexion with increasing age (Boyer et al., 2017). While previous studies suggested that older adults and adults with knee OA may have less sagittal knee range of motion compared to young adults (Boyer et al., 2017; Kaufman et al., 2001; Messier et al., 1992), we did not detect differences in knee kinematics in this study. This may be due to the large standard deviation in the knee kinematics of the knee OA group, which counterbalanced apparent trends towards less knee extension during midstance and less knee flexion during swing in the knee OA group (Figure 2).
Not surprisingly, sagittal plane kinematics differed between preferred- and faster-than-preferred walking speeds. Beyond overall differences in kinematics between speeds, we were interested in whether speed affected kinematics differently for MoCap and IMU data. Integrated IMU data (e.g., linear velocity and displacement, angular orientation) are susceptible to greater error when noise increases with greater movement speed. While this error is generally low at walking speeds (Potter et al., 2019), OpenSense does not directly implement error-reducing procedures such as linear drift correction or stride-by-stride integration resetting. We did find significant tool × speed interactions for every kinematic variable (Table 2). However, these interactions appeared to have minimal clinically relevant significance as post-hoc comparisons of the effect of speed on MoCap or IMU data revealed only that MoCap picked up greater flexion at faster speed where IMU did not detect a similar increase in flexion for 4% of the gait cycle at the hip and 4% of the gait cycle at the knee. Thus, IMU data appeared—overall—to detect similar speed-related differences in kinematics as MoCap data overall.
This study extends the demonstration of OpenSense as a viable open source tool for estimating gait kinematics using IMUs. To date, many validations of IMU kinematic data have relied on proprietary commercial software packages. While proprietary software may provide internally reliable and accurate estimates of kinematics, the analysis methods within these software packages are generally unknown, making comparison across studies or replication of results with different systems difficult. Open source software provides a more consistent analysis approach and increases accessibility of IMU analysis to researchers or clinicians who do not have the bandwidth to develop custom algorithms.
This work has several limitations. Most significantly, demonstration of the validity of this IMU approach in a lab setting does not guarantee similar validity in real-world, unobserved settings. While we attempted to mimic some features of unobserved data collection (e.g., selecting calibration time points based on IMU signal characteristics rather than known, controlled postures; identifying gait events from IMU signals), we did not have to consider other unobserved data challenges, including detection of walking activity or sensor placement variation. While many methods exist for activity classification and gait event detection [e.g., (Attal et al., 2015; Benson et al., 2019; Gurchiek et al., 2020)], the impact of variations in sensor placement remains a relatively understudied area. As part of the development of this study, we examined the effect of variation in sensor placement on IMU kinematic outcomes. We found that modest to moderate differences in sensor placement (i.e., ∼2-4 cm superior/inferior position or ∼90° medial rotated position) could impart differences in peak sagittal joint angles of up to 10° (Supplemental material). This finding may support the use of additional IMU data alignment techniques prior to inputting data to OpenSense to minimize the effect of day-to-day or inter-participant sensor placement differences. In addition to the limitation of a controlled testing approach, we only reported sagittal kinematic results. Kinematic differences imparted by the presence of knee OA may have been missed in this study because we used a 1 degree of freedom knee joint that restricted frontal or transverse plane knee rotation.
Overall, this study demonstrates that an open-source, freely available IMU inverse kinematics approach (i.e., OpenSense) provides reasonable estimates of gait kinematics that can differentiate between clinically relevant groups in controlled collection settings. This work is an early step in determining whether an inverse kinematics approach sufficiently detects or tracks meaningful decrements or improvements in real-world gait function. Real-world demonstration of these methods will be essential to verify that these in-lab findings translate to less-controlled, potentially unobserved data collections.
Data Availability
All data produced in the present study are available upon reasonable request to the authors. Some data and code used in this work are available at https://simtk.org/projects/knee-oa-age-imu.
Data Availability
Sample data and code are available at https://simtk.org/projects/knee-oa-age-imu
Supplemental Material
A supplemental file demonstrating an analysis of the effect of differences in IMU sensor placement on inverse kinematics results is included with this article.
References
