Seven horses of different breeds were used in this study: four research Thoroughbreds, two client-owned Warmblood horses and one client-owned pony of unknown breed. There were five geldings and two mares with a mean age of 4.3 years (range 2 to 7 years), mean height of 1.56 m (range 1.35 to 1.69 m). Six horses had front limb shoes, one had hind limb shoes and one had no shoes. Four to six experts subjected the horses to a neurologic and lameness assessment at the Equine Referral Hospital at the Royal Veterinary College (RVC). Three of the horses were mildly lame and three were mildly to moderately ataxic. All procedures were carried out at the RVC Structure and Motion gait laboratory, approved by the ethics and welfare committee at the RVC and complied with the European Animal (Scientific Procedures) Act 1986.

A customised boot (SMBII, Professional's Choice Sports Medicine Products, El Cajon, CA, USA) was placed on each limb. The boots were modified with Velcro on the outside of the boots' tightening straps. An 18 g IMU (MTx, Xsens Technologies B.V, Enschede, The Netherlands) was fixed in a tight pocket with Velcro on both sides and strapped snuggly to the lateral side of the boot using Velcro. The IMU was located at the level of the distal end of the 4th metacarpal bone (MCIV). The right and left IMU were aligned to the same height from the ground using a laser distance measurement device (Disto D3, Leica Geosystems A/S, Herlev, Denmark).

Five 10 g IMUs were placed on the dorsal midline over: (1) the withers (2) the 4th lumbar dorsal spinous process (3) the most dorsal point on the spinous processes of the sacrum and (4 + 5) bilaterally over each tuber coxae. In one horse, four 10 g IMUs were used on the limbs. The dorsum sensors were fixed to the skin with custom made pockets and self-adhesive plaster (Animal Polster, Snøgg Industri AS, Kristiansand, Norway). A cable from each IMU was attached to a data streaming and controlling Xbus Master (Xsens) with a maximum of 5 IMUs per Xbus Master. The two Xbus Masters were connected with a RS232 cable to the USB port of a solid-state drive adapted 10.1” laptop (S10, Lenovo Technology, Hook, UK). The laptop was collecting data at 200 Hz using MT Manager software (Xsens) and remote-controlled via WIFI using dedicated software (TeamViewer GmbH, Goppingen, Germany). The setup is depicted in Figure 1; data collection was initiated and ended with 5 seconds of the horse standing still for each trial, to optimise performance of IMU orientation estimation.

An experienced handler walked the horses at their preferred walking speed along a 25 m runway with a centrally and seamlessly embedded 4.8 m × 0.9 m force plate array (8 × Kistler type 9287BA, Kistler Instrumente AG, Winterthur, Switzerland) where the force data were collected through a custom built 10-plate Interface Rack based on a Data Acquisition unit (NI-6225, National Instruments, Austin, TX, USA) filtered through a low-pass filter (−6 dB point of 100 Hz) and acquired in LabView (v. 8.6, National Instruments).

A 12-camera optical motion capture system (Qualisys Oqus 300 and 500 series, Qualisys AB, Gothenburg, Sweden) was synchronized to the force plates and used to collect 3D kinematic data. The force plates were sampling at 500 Hz and were down sampled using MATLAB (R2011a, The MathWorks Inc., Natick, MA, USA) to the 200 Hz-sampling rate of motion capture and IMUs.

For synchronisation with the force plates, all limb mounted IMU's had a 36 mm reflective marker placed over the centre of the sensor. A 26 mm reflective marker was attached to the proximal dorsal and lateral hoof wall on each leg. The left front leg IMU was tapped with a stick on which a 36 mm reflective marker was glued. The peak in the sensor's z-axis (approximately aligned with the horse's latero-medial axis) acceleration was used to synchronise the IMUs to the force plates with the start of the movement of the reflective marker.

IMU data were processed with automated custom written MATLAB scripts. A horse-based global right-handed Cartesian coordinate system was defined with the positive x-axis pointing in the walking direction of the horse, the positive y-axis pointing to the left side of the horse and the positive z-axis pointing upwards. IMU data consisted of three-dimensional (3D) acceleration, 3D angular velocity, 3D magnetic registration and orientation. Acceleration and angular velocity were rotated into the global horse coordinate system and integrated as described by Pfau et al. [19] with the modification of a 0.5 Hz cut-off for the high pass 4th order Butterworth filter applied before rotation into the global horse coordinate system (hereafter referred to as rotation). Data streams in the local sensor coordinate system included in the analyses were: 3D acceleration before and after filtering, magnetic sensing and orientation (roll, pitch and yaw). Data streams in the global horse coordinate system included in the analyses were: 3D acceleration; 3D angular velocity; 3D velocity and 3D displacement.

Force plate and kinematic data were analysed using a semi-automated custom written MATLAB script. Reflective hoof marker position was used to correctly classify which limb the force plate events related to. The beginning and end of stance was obtained applying a threshold of 10 N to the vertical force signal defining hoof-on/off.

The first two horses were used as a template for the development of prototype gait event algorithms. Based on a frequency analysis, low pass Butterworth filters as well as wavelet symelets decomposition filters applied to each of the data streams were set as listed in Supplementary Table 1. Figure 2 is an example of how the local maxima and minima in the sinusoidal horizontal displacement of the DMC IMU were used as a guide for extraction of features in each of the signals.

The statistical analysis was carried out using R version 2.13.1 [20] with the packages lattice [21] for graphics, pastecs [22] for descriptive statistics, psy [23] for Intra Class Correlation (ICC) and a function for agreement analysis with repeated measures [24] where gait events are assumed to be repeated dependent measures on the independent items horses. Agreement analysis corrected for repeated measurements on each horse was performed with accuracy defined as the mean difference (bias) between the methods with upper and lower limits of agreement as described by Bland and Altman [25]. Precision was defined as the global SD of the bias and percentage of error calculated as described by Metzelder et al. [26]. The ICC was calculated according to Schrout and Fleiss [27]. Stance was calculated combining the bias corrected hoof-on/off results. Algorithms detecting fewer than 95% of all steps in all horses were not further investigated.

Biomechanical analysis is increasingly applied as a paraclinical diagnostic tool in human [28] and veterinary medicine [9,10,17,29,30]. Objective recognition of normal repeatable movement patterns in horses requires at least 3–5 consecutive strides [31]. The stride length of a horse and e.g., the limited extent of the calibrated area of 3D motion capture systems and/or the restricted size of standard force platforms means that even the most advanced comparative gait laboratories only allow recording of three to four consecutive strides during over ground walk and trot. The use of an instrumented treadmill increases the number of consecutive strides [32,33]. However surface properties of the treadmill belt are different to “natural” equestrian surfaces and treadmills have been shown to alter gait parameters during walk resulting in a decreased step frequency, increased step length, increased stance time and less vertical movement of the hoofs [14,15]. In humans, treadmill walking and running result in a significantly longer stride time and increased stability of trunk acceleration characterised by Lyapunov exponents of state space representations [16,34]. In addition, at least two habituation sessions are required for trot and nine for walk to obtain reproducible kinematic variables for a horse on the treadmill [35], making the instrumented treadmill comparatively impractical as a paraclinical diagnostic tool. Therefore there is a need for accurate and precise kinematics with the capacity to record many strides for objective assessment of the ataxic horse.

To the authors' knowledge there are no reports concerning the use of gyroscopes, accelerometers or IMUs in other anatomical locations than the hoof for determining gait event timings in the horse. Algorithms using the sacrum, hip, thigh, shank and ankle have successfully been developed in humans [36] and hoof-on detected using IMUs on the sacrum for horses during trot [37]. Gyroscopic sensors were applied in a study of hoof break-over but not correlated to gold standard gait event timings, e.g., derived from force platforms [38]. A study using hoof-mounted accelerometers by Witte et al. [18] obtained a mean absolute error of 2.4 ms for front hoof-on during walk and 3.6 ms for front and hind limb hoof-off. The accuracy reported here, using vertical velocity for front hoof-on (bias: 2.7 ms), and horizontal acceleration for front hoof-off (bias: 0.7 ms) indicates a better accuracy for hoof-off and similar accuracy for hoof-on using distal limb-mounted IMUs compared to uni-axial accelerometers. It is likely that the precision of hoof-mounted accelerometers is better compared to the fetlock-mounted IMUs indicated by the low interquartile range reported by Witte et al. [18], however a direct comparison would require the SD of the difference between the methods. In addition Peham et al. [39] described a method (based on motion capture and the horizontal velocity of a dorsal hoof marker), which was found to have an average error (accuracy) for stance phase duration of 10.8 ms. The DMC/DMT mounted IMUs in this study have a better accuracy for stance phase detection (0.01–1.3 ms bias).

The vibration damping properties of the flexor tendons and the distal foot [40,41] result in a decrease of the impact-accelerations of the DMC/DMT relative to the hoof [41] and could decrease the accuracy and precision using this location for gait event detection. However there is good correlation between the metacarpophalangeal joint angle and the vertical ground reaction force [42] indicating that the resulting forces could be measured at the DMC/DMT level as shown by the high accuracy and precision of both the horizontal and vertical acceleration and velocity to detect hoof-on/off at the DMC/DMT level.

The good accuracy (3 ms), precision (36 ms) and limits of agreement (−11 to 17 ms) for vertical velocity of the sacrum to detect hoof-on was expected based on a sacrum mounted IMU in non-lame and lame horses during trot [37] indicating that we can obtain accurate and precise stride but not hoof-off, step or stance duration from sacrum withers, lumbar or tuber coxae mounted IMUs. A sacrum mounted IMU applied to humans [43,44] found a lower bias for stride step and stance duration (ranging from 2 μs to 2 ms) and similar limits of agreement (−15 ms to 25 ms) [43] compared to the present study.

The inter-individual variation is contributing to a higher SD but a low intra-individual variation, for which the repeated measures agreement analysis corrects [25]. The wider limits of agreement for stance phase can therefore be attributed to an increased intra-horse variation.

The 10 N cut-off for hoof-contact in the vertical ground reaction force is low for a horse, but similar to what has been found in humans [45]; we found that a higher cut-off on average moves the time for hoof-on one frame later (at 200 Hz) and the time for hoof-off one frame earlier increasing the bias. The 10 N cut-off is therefore closer to the kinematic gait events but could increase the SD due to the oscillations in the force trace after impact. Future validation studies could use a percentage of body weight or a percentage of peak vertical ground reaction force to remove the likely effect of different mass of the subjects. Further work should focus on accuracy and precision for 3D position estimate in walk and trot using limb mounted IMUs to add robust spatial parameters of the distal limbs to the set of parameters that can be quantified with IMUs during over ground locomotion in horses.

We demonstrate good accuracy and precision for detection of the gait events front and hind hoof-on/off using DMC/DMT and sacrum mounted IMUs in horses during walk. Stance phase duration is highly accurate but less precise. Based on these findings development of a portable system for spatiotemporal gait analysis in the horse seems promising for objective assessment of ataxia.