Gait analysis is commonly performed to characterize healthy walking and quantify deviations that may exist due to pathology or injury [1
]. The gait cycle is defined as the time interval between two successive occurrences of one of the repetitive events of walking [10
], typically beginning from initial contact (IC) of the foot to the successive IC. The timing of IC and terminal contact (TC), referred to as heel strike and toe off in healthy populations, is necessary to mark the transition between stance and swing phases of gait. These two events are essential to analyze temporal gait parameters, such as stride time, periods of single and double support [11
] and to compare joint angles, forces, and moments across multiple strides [12
]. The division of the gait cycle also allows clinicians to evaluate deviations in pathologic gait, and improvements achieved with rehabilitation, by providing a clear description of the typical behavior of the lower extremity during each phase of the gait cycle.
A gait or motion capture laboratory, equipped with force plates or instrumented treadmills and cameras, is the “gold standard” [5
] for providing complete biomechanical analysis of the spatiotemporal, kinematic and kinetic parameters of gait, of which, gait event detection (GED) is a necessary component. Gait laboratories, however, require sophisticated motion capture systems, a well-trained team, abundant time and resources for analysis [10
]. Although considered the standard for gait biomechanical analysis, laboratory-based analysis is not fully representative of walking in daily life situations, particularly for patient populations [18
]. Recently, there is a boom in the development of video and wearable sensor-based techniques for gait analysis in ‘real world’ scenarios outside of the typical instrumented motion analysis laboratory [21
]. Such systems do not use kinetic data from force plates in their algorithms to determine gait events such as IC and TC. Rather, they make estimations of gait events from video data or sensor signals, such as those from gyroscopes [21
], accelerometers [28
], EMG [32
], and force sensitive resisters [34
Gait event detection from sensor data has also been leveraged in sensor-based rehabilitation systems. Detection of gait events such as IC and TC are crucial when considering use of orthotic or therapeutic interventions, especially in functional electrical stimulation (FES) [29
] and rehabilitation robotic systems [37
], that use gait events to synchronize stimulation delivery or actuator activation to particular gait phases. Although detection delays and timing estimation errors are inherent in kinematic-based gait phase detection methods when compared to the “gold standard” method (gait events from force plate data), the method used to estimate gait events themselves may impact the timing differences observed. Quantifying gait event delays [5
] and subsequently providing compensation algorithms for gait event timing errors [40
] are crucial for applications in which gait events serve to trigger assistive applications. This approach allows for appropriate correction and compensation of gait event detection errors to minimize timing errors for the assistive applications [40
Kinematic methods use different variables to estimate gait events, therefore, each method may introduce their own systematic characteristic errors in estimating timing of gait events when compared to events determined from force plate data. Thus, analysis of various kinematic methods in detecting gait events is necessary for determining repeatability, accuracy, and reliability to implement in non-laboratory-based gait analysis systems, such that they can be used as standard tools for assessment or to provide gait events as inputs to rehabilitation systems. Therefore, we utilize the gold standard technique (motion capture laboratory and force plates) of gait event detection to evaluate the differences that three different commonly used kinematic methods have on gait event detection (GED) timing. The purpose of this manuscript is to (1) quantify timing accuracy of three kinematic methods; the coordinate-based treadmill algorithm (CBTA) [25
], the shank angular velocity algorithm (SK) [24
] and the foot velocity algorithm (FVA) [11
]; to determine IC and TC gait events compared to the established ‘gold standard’ kinetic-based method (GS), and (2) through statistical modeling, demonstrate that kinematic and therefore sensor-based methods for detecting IC and TC events are insensitive to covariate influences, i.e., subject group differences, person-specific variability, side-to-side gait differences, and the kinematic-based GEDM used to derive the gait events. Finally, we make the case for one particular kinematic-based method that can best be used for sensor-based gait event detection.
Three kinematic gait event detection methods (GEDM) for detecting gait events (IC and TC) during walking were evaluated in three populations (AD, TD, and CP). Comparisons of gait event detection (GED) time and gait detection reliability (GDR) of each kinematic GEDM were made to the gold standard (GS) of using force plates (kinetic) data to detect events. Our approach helps to characterize the errors in GED among different kinematic methods such that, with compensatory algorithms for timing errors, non-laboratory-based gait analysis systems using such techniques can be equivalent to gold standard laboratory systems. A novelty of our work is the advantage of applying the four GEDMs to the same dataset for the direct comparison of GEDM accuracies rather than comparing across different studies and validation techniques. Thus, the present work assists in the validation of various kinematic methods in detecting gait events, and to our knowledge, is the first to use a statistical model demonstrated insensitivity of the GEDMs to variations in group, side, and individuals.
Gait event detection timing of CBTA, SK, and FVA were compared to previously reported accuracies and highlight the benefit of using a standardized dataset. Although GEDM algorithms were the same, variations in validation techniques resulted in different GED accuracies. Zeni et al. compared gait event times detected by CBTA between motion capture data collected at 60 Hz vs. force plate data collected at 600 Hz [12
]. Our study performed the same comparison of gait events detected by CBTA vs. force but data were collected at higher sampling rates for both kinematic (128 Hz) and kinetic (3200 Hz) signals. Zeni reported differences of −17.36 ms (right IC), −15.03 ms (left IC), −0.37 ms (right TC), and 11.69 ms (left TC) [12
] while our analysis resulted in average differences of −26.78 ms (IC) and −2.49 ms (TC). While results differed in absolute value, the direction matched their previous validation demonstrating that CBTA detected right/left IC and right TC before GS. In another study, IC and TC RMSEs were 26 and 25 ms, respectively, when SK was applied to gyroscope signals and compared to gait event times detected by footswitches [27
], whereas, IC and TC RMSEs were 26.53 and 62.44 ms, respectively, when SK GEDM was validated with a motion capture system and force plates in our work. Lastly, two populations (TD and CP) were included in the comparison of gait event times detected by FVA vs. force plate to the literature. Similar to the data collection for CBTA validation, our kinematic and kinetic dataset was collected at higher sampling rates. O’connor et al. demonstrates that IC and TC were detected by FVA 16 ± 15 ms and 9 ± 15 ms later than determined by force plate data, respectively, in TD while IC and TC were detected by FVA 3 ± 9 ms and 6 ± 26 ms earlier in CP [11
]. We report that FVA detects IC later in both TD (39.9 ± 22.67 ms) and CP (77.9 ± 164.9 ms) and detects TC earlier in both groups (TD: −3.19 ± 15.53 ms, CP: −20.68 ± 58.48 ms) compared to force plate data. In general, we report gait event timing errors of similar magnitudes and directions as previous investigations. The direct comparison of GED timing of multiple GEDMs in the same dataset illustrates the variation in amount of delay compensation needed and may facilitate sensor selection.
Statistical modeling was employed to systematically assess the influence of covariates such as gait event, subject group, side, and subject-specific differences on the differences observed between the kinematic GEDMs and the GS. Individual models were created for each of the kinematic GEDMs and incorporated adjustments for the additive covariates (group, gait event, side, and subject). The full model performed well for each of the kinematic GEDMs and provided values for comparison of RMSEs for the reduced and final models. Final models containing only the principal predictor (CBTA, SK, or FVA) and random coefficients grew <10 ms RMSE of the full models. The small RMSE difference demonstrates the large contribution that GEDM has on gait event timing differences compared to GS. The RMSE differences of 5 ms or less between the marginal model (excludes subject-specific adjustments) and the conditional model (includes subject-specific adjustments) at each level of model progression illustrate the robustness of each GEDM to subject differences. The final model shows that only the principal predictor and random coefficients are needed to have the best approximate gait events compared to GS.
Using the RMSE for model comparison indicates little difference between the full model with all predictors and a reduced model with a principal predictor +IC, however, given the relative small effect size for IC, one might question the practical advantage of that model over a simple linear regression using the marginal model coefficients for each of CBTA, SK, and FVA. We recommend the more parsimonious simple linear models with the statistically significant intercepts and slopes appearing as the final marginal model for each measure in Table 4
. Results from our models, inclusive of adults and children with and without CP, suggest that timing differences of our data and that of the literature can be accounted for by the RMSEs of the method employed. A conditional RMSE that improves the marginal RMSE less than
ms suggests that these models can be confidently applied to a wide range of gait characteristics without tuning to subject population characteristics. From a statistical perspective, the tradeoffs between models are minimal. Depending on the desired precision of gait event detection timing, however, the error differences need to be evaluated to determine if they are clinically relevant when applied to gait analysis or rehabilitation applications. Thus, the choice of model and variables included in gait event detection timing compensations may have a small, but meaningful impact upon application.
Any one of the three kinematic GEDMs can be translated into sensor-based systems for gait event detection [28
] as they all showed high gait detection reliability of IC and TC. However, not all kinematic GEDMs are equally practical for implementation. Despite the GED accuracy of the CBTA, implementation of this GEDM requires a relatively high number of sensors and sophisticated arithmetic processing. Seven inertial measurement units were used in a CBTA-based sensor system to generate the necessary input signals for GED as well as other spatiotemporal parameters [43
]. The number of sensors may be reduced, however, depending the desired output. For example, CBTA-based sensor system may only require three IMUs if only bilateral GED is required. This GEDM may be useful as an alternative for laboratory-based gait analysis, however, it may not be computationally and cost efficient for wearable applications such as orthotics and prosthetics.
Foot velocity (FVA) and shank angular velocity (SK) GEDMs are more easily applied to wearable sensor signals and are commonly used in research laboratories [5
]. Foot velocity can be captured via shoe/foot-attached accelerometers, requiring a minimal sensor setup of one sensor on each side [28
], by integrating acceleration over time. A source of GED timing error is the potential drift introduced with signal integration. Sophisticated techniques, such as zero velocity update [45
], extended Kalman filters [46
] combined with sensor fusion [47
], however, may be used to reduce drift. In addition to timing errors introduced from the signal, previous studies have reported that FVA is not applicable to clinical cases [11
]. Pathologic gait, as demonstrated in CP, does not have a regular heel to toe progression; therefore, difficulty in detecting gait events with non-kinetic based methods is a limitation of this GEDM. If adjustments cannot be made for missing the trough in the foot velocity signal (Figure 1
), it can result in decreased detection reliability.
We have identified SK, out of the three GEDMs evaluated in this study, as the kinematic-based method that can best be used for sensor-based gait event detection. The SK GEDM can be applied to signals collected via shank-attached gyroscope sensors (one sensor on each side). Gyroscopes are easy to use, miniature in size and can be used with FES systems, exoskeletons and other clinical/research training systems [5
]. This minimal sensor set up facilitates implementation (similar to FVA), is of modest cost, and computationally efficient in wearable applications. Unlike CBTA, minimal processing is required to condition the sensor signals for the SK detection algorithm providing improved processing capacity and enabling implementation of more sophisticated control algorithms in the system [50
]. The ample amount of data on reliability and performance of the SK GEDM evaluated in patient populations such as amputees [51
], spinal cord injuries [49
], and post-stroke survivors [52
] demonstrate its robustness to variations in gait. The SK GEDM includes the advantages of the other two GEDMs, such as minimal sensor set (similar to FVA) and comparable detection accuracy and reliability to CBTA, as well as demonstrated GED ability in multiple patient populations.
A limitation of this study is the assumption that GEDM performance in a motion capture system translates to equivalent performance in a sensor-based system. We have demonstrated the advantage of assessing GEDM performance on the same dataset. Previous work reported real-time GED performance using SK against force plates in adults, typically developing children, and children with CP during treadmill walking [5
]. Gait detection reliability was higher when SK was applied to sensor signals (AD: 99.8% [50
], TD: 99.9%, CP: 99.6% [5
]) than to motion capture signals (AD: 96.7% TD: 96.3% CP: 95.2%). Comparisons of SK GEDM accuracies between the sensor system (gyroscope signal) and motion capture system are outlined in Table 5
. Although gait event detection had a greater RMSE in the sensor system, with the exception of TC in AD, the range of delay was similar (sensor: 49 ms, motion capture: 40 ms) and smaller than the FVA range (166.21 ms).
Other study limitations include small group sizes, although a large number of gait cycles were included in analysis, GEDM performance was isolated to treadmill walking, and we did not account for speed variation. The walking speeds were similar for the AD and TD groups (AD: 0.92 ± 0.18 m/s, TD: 0.99 ± 0.17 m/s), while individuals with CP walked slower (CP: 0.72 ± 0.15 m/s). Exclusion of one subject in the CP group from analysis indicates that kinematic GEDM, and wearable sensors, may be challenging to use in patient populations with more severe gait deviations (discussed above). Further investigation is required to fine-tune our model for these patient populations.