Reasons for performing study: Assessing back movement is an important part of clinical examination in the horse and objective assessment tools allow for evaluating success of treatment.
Objectives: Accuracy and consistency of inertial sensor measurements for quantification of back movement and movement symmetry during over ground locomotion were assessed; sensor measurements were compared to optical motion capture (mocap) and consistency of measurements focusing on movement symmetry was measured.
Methods: Six nonlame horses were trotted in hand with synchronised mocap and inertial sensor data collection (landmarks: T6, T10, T13, L1 and S3). Inertial sensor data were processed using published methods and symmetry of dorsoventral displacement was assessed based on energy ratio, a Fourier based symmetry measure. Limits of agreement were calculated and visualised to compare mocap and sensor data. Consistency of sensor measurements was assessed using Pearson correlation coefficients and linear regression to investigate the effect of speed on movement symmetry.
Results: Dorsoventral and mediolateral sensor displacement was observed to lie within ± 4–5 mm (± 2 s.d., 9–28% of movement amplitude) and energy ratio to lie within ± 0.03 of mocap data. High levels of correlation were found between strides and trials (0.86–1.0) for each horse and each sensor and variability of symmetry was lowest for T13 followed by T10, T6, L1 and S3 with no significant effect of speed at T6, T10 and T13.
Conclusions: Inertial sensor displacement and symmetry data showed acceptable accuracy and good levels of consistency for back movement. The small mediolateral movement amplitude means that changes of <25% in mediolateral amplitude (also unlikely to be detected by visual assessment) may go undetected. New sensor generations with improved sensor sensitivity and ease of use of equipment indicate good potential for use in a field situation.
Back injury in horses poses a diagnostic challenge. Many diagnostic and monitoring techniques including visual observation of asymmetrical movement patterns during clinical examination, palpation and clinical history are all subjective and vary according to experience, tradition and personal bias (Jeffcott 1979; Wennerstrand et al. 2004; Parkes et al. 2009). Symmetrical back movement is a sign of a sound horse. However, the small range of motion and the difficult accessibility of the equine back make routine diagnostic procedures currently used in the limbs unsuitable and insensitive for the back (Faber et al. 2001a, 2002). Optical motion capture has been used to demonstrate symmetry of back movement in healthy horses (Audigiéet al. 1999), to show that chiropractic manipulation results in more symmetrical pelvic motion (Gómez Álvarez et al. 2008) and that small asymmetries might be related to subtle lameness (Gómez Álvarez et al. 2007). Thus a practical and objective method to assess back movement could contribute to elimination of evaluation bias in clinical trials and would lead to more accurate assessment providing continuous rather than discrete (subjective) grading data and, in particular, contribute to distinguishing between back problems and limb lameness.
Optical motion capture (mocap) is the current gold standard to assess 3-dimensional movement and orientation with a high degree of accuracy; however, the constriction of camera calibration means that in practice use is restricted to a gait laboratory and/or a treadmill that requires habituation and is known to cause kinematic changes (Buchner et al. 1994). Finally, the equipment cost restricts its suitability for routine diagnostic use in general practice. It is, however, still the best method to reliably capture back kinematics with high repeatability (Faber et al. 2001b, 2002).
An alternative technique for collecting kinematic data is the use of (wireless) inertial sensors. Small and lightweight sensors mounted on the subject allow data to be collected during unrestricted movement, giving up to 6 degree of freedom (orientation and displacement) information (Keegan et al. 2004; Pfau et al. 2005, 2006, 2007). This allows investigation of locomotion outside the laboratory and thus increases the scope of experiments to include, amongst others, lameness assessment (Keegan et al. 2002; Pfau et al. 2007; Church et al. 2009) and horse–rider interaction (Pfau et al. 2009), potentially combining this with simultaneous force measurements under the saddle (von Peinen et al. 2009).
By double integration, displacement data can be calculated from the acceleration data (Pfau et al. 2005) and symmetry-related parameters (e.g. energy ratio, ER) can then be extracted for each stride (Keegan et al. 2004; Pfau et al. 2007). Validation for withers movement on the treadmill (Pfau et al. 2005) showed good agreement for all movement planes.
The present study aimed to further investigate the accuracy of inertial sensors by direct comparison to the gold standard mocap, to assess repeatability of measurements and to assess validity of ER in measuring back movement symmetry during over ground locomotion at the trot.
It is hypothesised that during trot over ground, displacement data along the equine back will be accurate enough to be of use in a clinical setting and that repeatability of inertial sensors at vertebrae T6, T10, T13, L1 and S3 will be adequate for quantification of movement symmetry.
Materials and methods
Animals and procedures
Six ponies (age range: 5–20 years, height range: 1.25–1.5 m) were used with no clinical signs of back problems or lameness.
Ten motion analysis cameras (Oqus series300/500)1 were set up in sequence along an 18 m track. Calibration at regular intervals resulted in errors of 1.24–1.57 mm over 3 days.
Six MTx inertial sensors (MTx)2 were synchronised via a wireless transmitter (Xbus). Each inertial sensor contains a 3 axis accelerometer (full scale ± 100 m/s2), a 3 axis gyroscope (full scale ± 1200°/s) and a 3 axis magnetometer (full scale ± 750 mGauss) resulting in a static accuracy of <0.5° for roll and pitch and <1° for heading measurements. All 6 sensors were attached to a wireless transmitter unit (Xbus, Xsens) via serial connections (RS485) and then attached to the horses with hair extension glue at the following landmarks using custom made pads: T6, T10, T13, L1, S3 (identified by palpation of the respective dorsal spinous processes) and left tuber coxae. The wireless transmitter was attached to the horse using an elasticised surcingle and transmitted data of each inertial sensor at a sample rate of 100 Hz per individual channel (3x acceleration, 3x rate of turn and 3x magnetic field) to a nearby laptop computer running MT Manager software2.
Reflective hemispherical markers were secured to each sensor as well as on the lateral aspect of the 2 left hoofs (LF and LH) to record footfall timings of the diagonal pairs.
Each pony was trotted in hand until at least 3 trials of synchronised data collection were successfully recorded.
3D kinematics of the vertebral column: A standard right-handed orthogonal Cartesian coordinate system was used for both mocap and inertial sensor data (craniocaudal or x: positive axis directed along the line of progression; dorsoventral or z: axis vertical [aligned with gravitational field] and positive in the upward direction; mediolateral or y: axis perpendicular to the first 2 axes). Inertial sensor and mocap reference systems were aligned by aligning sensors with the motion capture reference frame and resetting heading to zero. In addition, craniocaudal (x), mediolateral (y) and dorsoventral (z) displacement data in the horse based reference system were calculated following published methods (Pfau et al. 2005, 2006) with modified highpass filter frequencies chosen as 1.5 Hz for dorsoventral and 0.75 Hz for mediolateral movement. Energy ratio was calculated from dorsoventral displacement in order to assess asymmetry of movement (Audigiéet al. 2002). Energy ratio is a Fourier analysis–based measure of asymmetry to indicate lameness. In short, Fourier analysis is used to decompose the dorsoventral displacement data of individual strides into 2 components (amplitudes of the first 2 Fourier waves, the first matching the stride period and the second with half the period). The first component (A1) thus represents the amount of asymmetry and the second one (A2) the amount of symmetry. ER is then defined as A22/A12+A22. Perfect symmetry is thus indicated by an ER value of 1; if asymmetry is present the value decreases nonlinearly with the amplitude difference between A1 and A2.
Stride extraction: Foot contact times (stance times) were determined using 3D marker velocity of the hoof markers. A velocity threshold of 1 mm/frame was used as a cut-off to distinguish between stance and swing phase and strides were defined to start (and end) with foot on the left fore (LF) hoof. For the sensor validation only the middle 3 strides of each trial were used in order to avoid filtering (in particular of mocap data) affecting the comparison. For repeatability calculations all extracted strides were used. Custom scripts were written in MATLAB3 to extract movement data (motion capture and sensor) for each stride and interpolate stride data to 100% stride time.
Agreement between inertial sensors and mocap: Limits of agreement for displacement data calculated from mocap and inertial sensors were assessed in Excel4 and scatter plots created in MATLAB showing differences between methods against their mean (Bland and Altman 1986).
Statistical analysis for consistency of inertial sensor measurements: For each horse, there were 5 sensor locations and 6 trials, as well as 5–7 strides per trial. To assess ‘stride-to-stride’ and ‘trial-to-trial’ consistency, we compared dorsoventral displacement of all strides within a trial (for each horse and sensor combination) and then all trials of each anatomical landmark per horse. A one-tailed Pearson's coefficient of correlation was calculated in SPSS5 for individual strides for separate horse, sensor location and trial combinations to compare all strides within a trial. Displacement values using the average stride cycle (values at 1% intervals using the average stride interpolated to 100% stride time) within a trial were then calculated and again a one-tailed Pearson's coefficient of correlation was calculated to compare between trials of each horse (for each sensor location) and to test the null hypothesis that the pattern of dorsoventral movement is not repeatable ‘stride-to-stride’ or ‘trial-to-trial’ for each horse.
A transformed version of ER, tER, was introduced to ensure that it followed a normal distribution (tested via Kolmogorov-Smirnov analysis) using the following equation (Pallant 2007):
Mean, s.d. and coefficient of variation (CoV) were calculated for tER for all horse and sensor location combinations. Data were also pooled for all horses and mean, s.d. and CoV calculated to assess the variability of tER at different vertebrae. To investigate the effect of speed on asymmetry of back movement, tER from all horses was combined and the relationship between speed and tER was determined for each anatomical landmark using linear regression.
Speed (first derivative of x position of T6 from mocap) at the trot ranged from 2.60–3.69 m/s for all horses.
The dorsoventral (z) displacement showed 2 peaks per stride and amplitude of movement ranged from 40–47 mm. In the lateral (y) direction only one peak per stride was observed and a range of 16–37 mm was recorded for the different sensor locations. In general, dorsoventral movement amplitude increased towards the middle of the back with reduced amplitude cranially and caudally, whereas with the lateral displacement range decreased from cranial to caudal and then increased in the sacral region (see Table 1, movement amplitude).
Table 1. Mean, s.d. and limits of agreement of difference between optical motion capture and inertial sensor displacement data for 5 anatomical landmarks along the spine of a horse
Mean difference (mm)
Mean + 2 s.d. (mm)
Mean − 2 s.d. (mm)
Limits of agreement (LA) (mm)
Movement amplitude (A) (mm)
LA/A × 100 (%)
Mean difference (mm)
Mean + 2 s.d. (mm)
Mean − 2 s.d. (mm)
Limits of agreement (LA) (mm)
Movement amplitude (A) (mm)
LA/A × 100 (%)
Comparison of inertial sensor and motion capture data
A total of 54 strides (mean stride frequency 1.59 Hz) were analysed at the trot. Subjectively, displacement data from inertial sensors followed the mocap data well over a series of strides for all landmarks and in both movement directions (see Fig 1 for T6, T13 and L1).
Figure 2 shows a scatter plot of the difference between the 2 methods over the mean of the 2 methods (Bland and Altman 1986). Both dorsoventral and mediolateral data are shown for 3 sensor locations (T6, T13 and L1) and mean, s.d. and limits of agreement (LA) are given in Table 1. Limits of agreement (Table 1) are generally in the range of ± 5 mm in both movement directions; as a consequence of the reduced mediolateral movement amplitude (A, Table 1) higher percentage differences of 18–28% are observed compared to 9–21% for dorsoventral movement (Table 1).
In addition, mean and s.d. of differences (as well as limits of agreement) for ER values calculated from mocap and from inertial sensors are given in Table 2. In general, inertial sensor data are slightly overestimating ER values (negative difference) and standard deviations of the differences are observed to fall between 0.012 and 0.023.
Table 2. Mean and s.d. and limits of agreement of the difference between energy ratio values calculated from optical motion capture and inertial sensor dorsoventral (z) displacement data for 5 anatomical landmarks along the spine of a horse. Values presented are based on n = 54 strides (9 strides each of 6 nonlame horses)
Mean + 2 s.d.
Mean − 2 s.d.
Width of limits of agreement (LA)
Consistency of inertial sensor data
A total of 205 strides were analysed (31–41 strides per horse) to assess consistency of sensor based measurements.
Pearson's coefficient of correlation ranged from 0.86–1.0 when applied to assess consistency of strides within a trial per sensor and when calculated between the mean strides of each of 6 trials per horse. Correlation values were found to be highly significant (P<0.0001) concluding that displacement was repeatable ‘stride-to-stride’ and ‘trial-to-trial’.
A Kolmogorov-Smirnov test confirmed normal distribution (P values: 0.164–0.996). Mean, s.d. and CoV of tER are shown for each horse and each sensor location in Table 3; Figure 3 presents mean ± s.d. for all 5 sensor locations of all horses. Values of tER generally varied little across the different sensor locations and also between horses. When tER values were combined for all horses (n = 205), the sensor at T13 showed the lowest CoV value, followed by T10, T6, L1 and S3.
Table 3. Mean (s.d., CoV) of tER values for each horse and sensor location. tER values can be seen to be comparable between horses (all horses were free of lameness) and between sensor locations with a slight tendency towards increasing asymmetry (decreasing value of tER) for the more caudal sensors
Effect of speed on asymmetry
Figure 4 presents scatter plots of tER values over speed for all 5 anatomical landmarks. Values for the linear regression lines are given in Table 4. R2 values are generally observed to be close to zero (indicating no linear relationship between speed and asymmetry). A significant deviation from a horizontal line is only observed for L1 and S3 (negative linear regression - representing increased asymmetry [decreased tER] with increasing speed).
Table 4. r2 values for linear regression of speed and tER per sensor. P values for assessing deviation from a horizontal line, slope (best fit value ± 95% confidence interval) and intercept of the regression line (best fit value ± 95% confidence interval) are also given
3.81 × 10−4
3.24 × 10−4
2.31 × 10−3
5.63 × 10−2
1.03 × 10−1
−9.55 × 10−4± 3.4 × 10−3
−8.32 × 10−4± 3.24 × 10−3
−2.21 × 10−3± 3.23 × 10−3
−1.29 × 10−2± 3.71 × 10−3
−2.10 × 10−2± 4.34 × 10−3
0.9859 ± 1.08 × 10−2
0.9864 ± 1,02 × 10−2
0.9901 ± 1.01 × 10−2
1.021 ± 1.16 × 10−2
1.041 ± 1.36 × 10−2
Quantification of back movement
Consistent with previous work on equine kinematics of anatomical landmarks along the back (Audigiéet al. 1999; Buchner et al. 2000; Faber et al. 2001a,b) our inertial sensor data showed a double sinusoidal pattern for dorsoventral displacement and an approximately sinusoidal pattern for mediolateral displacement for each stride. The range of displacement for T13 (which was found to correspond most closely to the body centre of mass movement, Buchner et al. 2000) was found to be 54 and 19 mm for dorsoventral and mediolateral movement, respectively, confirming the results from this earlier study. The dorsoventral displacements of the sensors along the back were found to be of similar shape but slightly differing amplitudes (Fig 1); greatest towards the caudal thoracic region and smallest in the craniothoracic and sacrum region. Mediolateral displacement decreased from the cranial thoracic to the lumbar region and then increased in the sacral region. The pattern of displacement amplitudes corresponds to the body centre of mass (most similar to T13 movement, Buchner et al. 2000), the further away from this, the greater the lateral movement. Differences between movement patterns of the different sensors allow for quantification of flexion-extension and lateral bending, which may be important clinically to assess asymmetry in horses with suspected back problems.
Accuracy of inertial sensors
When comparing inertial sensor with mocap data we found that the 2 methods agreed well for all investigated anatomical landmarks. Differences between inertial sensor data and mocap ranged (± 2 s.d.) from 8–16 mm for dorsoventral data and from 7–11 mm for mediolateral data (similar to earlier values obtained for treadmill exercise, Pfau et al. 2005). However, we adapted the high pass filter frequencies employed during the double integration from acceleration to displacement and used a cut-off of 0.75 Hz for mediolateral and 1.5 Hz for dorsoventral movement. The high pass filter cut-off frequency for dorsoventral movement was chosen just below stride frequency thus not affecting signal amplitudes occurring at the stride frequency and hence no information relating to the first Fourier harmonic representing ‘asymmetry’ (Audigiéet al. 2002) is lost by this process. The lower high pass filter cut-off frequency in the mediolateral direction reflects the fact that approximately one left-right movement can be observed for each stride while in the dorsoventral direction 2 similar (at least in the sound horse) up-down movements are observed - one for each stance phase - resulting in a higher movement frequency. The average stride frequency observed here was 1.59 Hz and thus just above the higher of the 2 high pass filter cut-off frequencies used for dorsoventral movement; when choosing a cut-off frequency for different data sets the stride frequency should be taken into account. During over ground locomotion, the number of mocap strides per trial is in practice restricted by the number of cameras and thus filtering effects at the start and end of each data sequence have to be considered. Here, for example, some circular trajectories can be observed in the Bland and Altman style plots presented in Figure 2 as an effect of temporary (but systematical) differences between mocap and inertial sensor data over a series of consecutive samples (e.g. as an effect of filtering or inaccuracies in sensor orientation estimates and subsequent double integration). We aimed to minimise these effects and only compared the 3 ‘central’ strides out of each sequence (usually comprising 5–6 mocap strides and even more inertial sensor strides). As an alternative, we could have performed the comparison on the treadmill similar to an earlier study (Pfau et al. 2005). However, this seemed suboptimal due to the known detrimental effect of changes in the Earth's magnetic field near large amounts of ferromagnetic material (de Vries et al. 2009) but it would have allowed for collection of a longer series of strides, thus minimising filter effects. We considered over ground locomotion to be more relevant for a practical (clinical) application and thus performed the validation for over ground locomotion.
Accuracy as a percentage of the range of movement was found in the range of 9–21% (dorsoventral) and 10–24% (mediolateral) and the higher mediolateral values were explained by the lower range of movement. It has been shown that even experienced clinicians are not able to reliably detect movement asymmetries below a level of 25% (Parkes et al. 2009). Therefore, considering the reduced accuracy (due to the range of motion) in the mediolateral direction, the inertial sensor measurements used here are borderline sensitive to detect relevant asymmetries in this direction. However, inertial sensors still provide a valid technique to collect objective back movement data during over ground locomotion and thus should be considered a useful source of evidence complementing results of subjective visual assessment.
Direct comparison of ER values from mocap and inertial sensor data indicate that inertial sensor measurements slightly overestimate ER values and thus slightly underestimate the amount of asymmetry. Energy ratio values generally lie within 0.03 (2 s.d.) of mocap based values with the exception of the most caudal sensor (S3) and this sensor location should thus preferably not be used for assessing movement (a)symmetry. Further investigations should include horses with different grades of lameness in order to assess the influence of the nonlinear relationship between amplitude ratios (e.g. symmetry indices, Buchner et al. 1996) and energy ratio on the accuracy of sensor based ER (or tER) calculations.
Consistency of inertial sensor data
We assessed consistency of inertial sensor data and were able to show that inertial sensors produce repeatable measurements of dorsoventral displacement data. In particular, we observed no measurable effect on dorsoventral displacement correlation between trials (at different speeds), nor between strides within each trial, with highly significant correlation values. Dorsoventral displacement data were then further used to assess movement asymmetry and the influence of sensor location as well as the influence of trotting speed on asymmetry. Energy ratio, a published method based on Fourier decomposition of the movement signal (Audigiéet al. 2002), was used to calculate asymmetry. Here, ER was transformed into tER in order for the data to follow a normal distribution. Then, consistency of asymmetry (tER) measurements was described using CoV and the effect of speed on asymmetry was evaluated using linear regression. Symmetry was generally higher for the thoracic sensors (T6, T10, T13) compared to the sensors in the lumbosacral region (L1, S3) supporting published findings (Audigiéet al. 2002) and thus confirming the calculation of ER (or tER) from inertial sensor data captured during over ground locomotion.
In order to obtain data comparable to an in-field (clinical) situation, the horses were trotted up over ground and not on a treadmill. Since speed was harder to control in this situation (across all trials: 2.60–3.69 m/s), the influence of speed on variability of inertial sensor data was assessed. Individual horses were reasonably consistent with regards to speed and only one horse (Horse 6) had a CoV for speed that was >10%. Since dorsoventral movement amplitude and consistency of stride parameters have been found to be affected by speed (Peham et al. 1998; Minetti et al. 1999; Pfau et al. 2005), we investigated whether back movement asymmetry would also be affected by changes in speed. Linear regression analysis (at a 95% confidence interval) of tER against speed was performed, which resulted in generally very low r2 values. We showed that asymmetry at T6, T10 and T13 was not influenced by speed, whereas movement asymmetry increased with increasing speed for L1 and S3 (despite the very low r2 value a slope significantly different from zero was observed due to the relatively high number of strides [n = 205]). This suggests that in order to avoid speed affecting results adversely, movement asymmetry should be assessed cranial to (and including) T13. In practice, however, this might be difficult to achieve at least in the ridden horse due to the location under the saddle.
Accuracy of dorsoventral and mediolateral displacement data calculated from inertial sensors mounted along the spine of horses trotting over ground lies within ± 4–8 mm (± 2 s.d.) of the gold standard mocap, and ER values varied by ± 0.03 (± 2 s.d.) between the 2 methods. However, due to the relatively small range of motion in the mediolateral direction, subtle mediolateral asymmetries might potentially go undetected. Sensor data were found to be consistent between trials and between strides within individual trials thus allowing for data pooling. Movement symmetry was generally observed to be higher in the thoracic region (T6–T13) and speed had no significant influence on movement asymmetry of these sensors suggesting these as preferred locations.
The ease of use of inertial sensors combined with the unrestricted capture volume allows unobtrusive collection of objective movement data even under the constraints of a clinical work-up. Further studies should apply this technique to horses with clinical signs of back problems allowing a more objective assessment of, for example, treatment success or changes with different surfaces and or different exercise activities.