The IMU measurements were validated using 24 GAIT4Dog tests recorded in this study. For each GAIT4Dog test, four to five strides from each of the four limbs were generally recorded by the GAIT4Dog system. A dog typically took a couple of seconds to go through the active area of the walkway (4.9 m long and 0.6 wide). All strides identified in the GAIT4Dog system were matched (“synchronized) with IMU measurements as explained in the section of “Synchronized IMU and GAIT4Dog tests”. Overall, the number of synchronized swing start events detected was 237 in forelimbs and 239 in hindlimbs in the four dogs. The number of synchronized swing end events was 230 and 239 in forelimbs and hindlimbs, respectively. The four parameters listed in Eq. (1) were then calculated based on these detected events.
Figure 4 shows the Pearson’s correlation results between IMU and Gait4Dog measurements of the four gait parameters (SWD, STD, SDS, SDE) defined in Eq. (1). All results obtained in the two systems were highly correlated (p < 0.001) for both forelimbs and hindlimbs. The correlation coefficient (r) ranged from 0.81 to 0.99 in forelimb and 0.83 to 0.98 in hindlimb. The correlation coefficient was equal or greater than 0.97 in STD, SDS, and SDE in forelimbs, as well as in SDE in hindlimbs. The smaller correlation coefficients in other measures were attributed to slightly greater variations between the two systems as well as the smaller data range.
Pearson’s correlation of the four gait parameters (Eq. (1)) obtained from the IMU (vertical axes) and GAIT4Dog (horizontal axes) systems in the synchronized validation study. The correlation coefficient r and p values are given in each pairwise plot. SWD, swing duration; STD, stance duration; SDS, stride duration based on swing start; SDE, stride duration based on swing end.
The Bland–Altman analysis is a reliable way to evaluate the agreement and bias between two measurement systems32. Figure 5 shows the Bland–Altman plots of the four gait parameters (SWD, STD, SDS, SDE) obtained from the synchronized IMU and GAIT4Dog tests. Each symbol represents a comparison of the same step parameter measured using both IMU and GAIT4Dog systems (indicated using subscript “IMU” and “G4D”, respectively). The horizontal axis is the arithmetic mean of the two measurements, and the vertical axis represents the difference between the two measurements.
Bland and Altman plots of the four gait parameters (SWD, STD, SDS, SDE) to study the agreement between IMU (with subscript IMU) and GAIT4Dog (with subscript G4D) measurements in the synchronized validation study. The horizontal axes are the arithmetic mean of IMU and G4D readings, and the vertical axes are their difference. The measurement unit is second in all axes. The dashed lines are the mean bias between the two systems. The solid lines indicate the 95% confidence interval of the agreement. SWD, swing duration; STD, stance duration; SDS, stride duration based on swing start; SDE, stride duration based on swing end.
In forelimbs, the mean bias (dashed line) between the IMU and Gait4Dog results was − 0.0023 s, 0.0032 s, 0.0008 s, and 0.0014 s for SWD, STD, SDS, and SDE, respectively. The corresponding 95% confidence intervals of the agreement (solid lines) were [− 0.023 s, 0.018 s], [− 0.018 s, 0.025 s], [− 0.012 s, 0.010 s], and [− 0.017 s, 0.019 s]. The disagreement between IMU and GAIT4Dog results was less than 10% of the mean value in more than 92% of the data points (97.4%, 92.2%, 100%, 100% in SWD, STD, SDS, SDE, respectively). The stride duration by swing start (SDS) had the best agreement between the two systems, which was consistent with the high correlation coefficient (r = 0.99) shown in Fig. 4.
In hindlimbs, the mean bias between the IMU and Gait4Dog results was 0.0022 s, − 0.0027 s, 0.0006 s, and − 0.0005 s for SWD, STD, SDS, and SDE, respectively. The corresponding 95% confidence intervals of the agreement were [− 0.022 s, 0.026 s], [− 0.028 s, 0.023 s], [− 0.028 s, 0.029 s], and [− 0.011 s, 0.010 s]. While the mean bias between the two systems remained low, the confidence intervals were wider than those in the forelimb comparison. Overall, the disagreement between the two systems was less than 10% of the mean value in 96.9%, 78.7%, 99.0%, and 100% of the data points in SWD, STD, SDS, and SDE, respectively. A careful examination indicated that the small percentage number of 78.7% in STD could be mainly attributed to the small mean STD values, which were the smallest among all parameters. SDE had the best agreement between the two systems with sub-millisecond mean bias and 10 ms in the confidence interval. This was consistent with the high correlation coefficient (0.98) shown in Fig. 4.
Among the four gait parameters calculated in this study, the SWD and STD values are affected by both the swing start and swing end detection (Eq. (1)). The SDS and SDE values are only affected by the swing start and swing end, respectively. Therefore, the accuracy in SDS and SDE provides a good estimation of the accuracy of swing start and swing end detection, respectively. Because the temporal resolution of the IMU system was 0.01 s (100 Hz data acquisition rate), the small limits in confidence intervals of SDS in forelimb ([− 0.012 s, 0.010 s]) and SDE in hindlimb ([− 0.011 s, 0.010 s]) suggested that the accuracy of swing start detection in forelimbs and swing end detection in hindlimbs may be close to the system limit.
Although we have conducted an extensive search for better signal markers, the phase detection accuracy of swing end in forelimbs and swing start in hindlimbs remained slightly worse than their aforementioned counterparts. It’s known that the IMU signals may be affected by variations in sensor mounting positions. The 3D printed sensor holder used in this study helped to minimize variations in mounting position. However, the IMU sensors were mounted at higher locations away from the paw to avoid affecting the dog’s movement. Because the swing start and end events were defined based on the contact between the ground and paw, the distance between the sensor and paw may ultimately affect the accuracy of detection paw-ground contact.
It should also be noted that any disagreement between the two systems may also be attributed to possible errors in the GAIT4Dog system. The walkway system is made from a large number of pressure sensors with finite sensor size and response. The paw touch events are determined based on the pressure readings using a proprietary algorithm. It’s expected that pressure readings may be affected by specific paw position, contact size, and contact speed, which can lead to fluctuations in the detection of swing start and end events. We had encountered several occasions when a dog’s steps were detected by the IMU system but not recognized by GAIT4Dog on the walkway.
Nevertheless, the overall good agreement between the IMU and GAIT4Dog systems suggests that the wireless 4-limb IMU system may be able to characterize the gait types of the dogs used in this study. In general, a canine gait can be classified as “walk”, “amble”, “trot”, “canter” etc., based on the on-ground paw patterns. For example, in “walk”, three paws are on-ground, and the other is off-ground. The off-ground paw (in swing phase) may alternate from the forelimb to the hindlimb, and from the left limb to the right limb. In “amble” gait, the on-ground paws may alternate between two limbs on the same side and two limbs in diagonal. “Trot” is a slightly faster gait and is characterized by two diagonal limbs swinging in unison while the other two are on the ground.
In reality, a dog’s gait sequence is often complicated with mixed gait types. The top panel of Fig. 6 shows a continuous temporal “box diagram” of a stride sequence constructed using the swing start and end data measured from a dog’s four limbs. The solid boxes represent swing phases of a limb, and the stance phases are represented by the white spaces between two solid boxes. The pawprint diagrams at the lower panel visualize the on-ground paw patterns at a specific time where off-ground paws are represented in light-gray, and on-ground paws are shown in solid colors.
An example of time-resolved gait type detection using the 4-limb IMU system. The solid-color boxes in the top panel indicate the swing phase of a limb. The white space between two solid-color boxes indicates the stance phase (on-ground paw). In the lower panel, the pawprint in light gray represents off-ground paws, and the pawprint in solid colors represents on-ground paws. LH: left hindlimb, LF: left forelimb, RF: right forelimb, RH: right hindlimb.
A quick overview of the temporal box diagram suggests that many strides can be characterized as “trot” due to the alternating diagonal on-ground paw patterns. This is most obvious at the beginning and toward the end, as evidenced in the paw diagrams around 43 s and 46 s. However, deviations can be observed in the middle part of the stride sequence. The dog appeared to speed up from ~ 44 s to ~ 45 s, which can be corroborated by smaller stance durations (white spaces between boxes in solid colors) in all four limbs. The two paw diagrams close to 44 s and 45 s show only the right hind paw on-ground, and all other three paws are off-ground. Such a paw pattern only exists in the traditional “canter” or “gallop” gait. It is interesting to note that the diagonal “trot” gait can still be observed in the middle part of the stride sequence. Therefore, we can conclude that this dog’s gait type changed rapidly during the speed-up period from the beginning toward 44–45 s and during the slow-down period thereafter.
In conclusion, we developed a 4-limb wireless IMU system for canine gait analysis. The IMU sensors can be easily mounted to a dog’s limbs using a 3D printed sensor holder, which showed no adverse impact on dog’s movement. Through intensive searching and validation, effective IMU signal markers were identified for detecting the swing start and end events in both forelimbs and hindlimbs. The system was tested on four dogs in a synchronized study using a commercially available GAIT4Dog system as a comparison reference. Owing to the novel sensor mounting system and robust signal processing algorithms, we achieved a great agreement between the measurements obtained in the IMU system and the GAIT4Dog walkway. We demonstrated that this 4-limb system was able to reveal complicated gait patterns in detail.
To the best of our knowledge, this is the only IMU based wireless gait system that can autonomically detect stride parameters in all four limbs of a dog with precision comparable with a GAIT4Dog system. With its low cost, lightweight, and wireless capability, this 4-limb IMU based gait analysis system is well positioned for applications on free roaming animals. It enables new research opportunities in studying the sophisticated coordination among all four limbs in dogs. Such a system may also be adapted for gait applications in other quadrupedal animals. Once fully developed, such a tool would be valuable for animal gait analysis at veterinary clinics as well as in translational research involving canine models (and potentially other quadrupedal animal models) of human diseases.
