Ongoing research

The use of an artificial neural network to classify gait in Icelandic horsesF.M. Serra Bragança, M. Tijssen, V. Gunnarsson, S. Björnsdóttir, E. Persson-Sjodin, P.R. Van Weeren, J. Voskamp, W. Backand M. Rhodin

The Icelandic horse is a versatile horse remarkable for its five gaits of walk, trot, pace, tölt and canter. Gait classificationis commonly performed visually, but this can be challenging. We hypothesised that an artificial neural network(ANN) could successfully classify all gaits. A group of 26 Icelandic horses were equipped with IMU sensors (samplingfrequency 500 Hz, EquiMoves®). These sensors were attached to each metacarpal/metatarsal bone and one sensor toeach hoof. A high-speed video camera was synchronised with the IMU sensors. Reference classification of each gaitwas performed from the video data by an Icelandic gait specialist. The classification was performed from steady-stategait using an ANN with one hidden layer of 45 neurons and with an input of 23 features. A total of 3,442 strides(walk=715, trot=516, tölt=1,218, left canter=381, right canter=396 and pace=216) were collected. Nineteen horseswere used for training and the remaining horses for cross-validation. Overall, the ANN classified the gait correctlyin 99.2% of the strides, only 0.8% were misclassified compared with the reference. The worst performance was inclassifying pace and tölt, with 3 and 1.3% misclassification respectively. These findings demonstrate the excellentperformance of objective gait analysis combined with an ANN for the classification of 5-gaited Icelandic horses.The technique will be very useful when evaluating performance, phenotyping of gait characteristics for geneticstudies, or as a first step to develop algorithms for objective lameness assessment in the Icelandic horse for which correct classification of the gait is essential.

Objective pain assessment during rest and locomotion in horses with two types of induced lamenessK. Ask, J.P.A.M. Van Loon, F.M. Serra Bragança, E. Hernlund, M. Rhodin and P. Haubro Andersen

Study objectives were to assess the validity of the composite pain scale (CPS) and the facial assessment of pain (FAP)scale in horses with induced orthopaedic pain, as well as to investigate the performance of the FAP scale duringwalk and trot. Lameness was induced by application of sole pressure and intra-articular lipopolysaccharide (LPS)administration (2.5 ng) in eight Warmblood mares, including a two-week washout period between the two inductionmodels. The horses were evaluated before and after lameness induction with inertial sensor (EquiMoves) and opticalmotion capture (Qualisys AB) systems. Direct pain assessment was performed in the stables by two independentnon-blinded observers using the CPS and the FAP scale, and from the side during straight-line walk and trot on thehard and soft surface using the latter scale. Both scales showed excellent inter-observer reliability with a Cronbach’sα of 0.99 for CPS and 0.93 for FAP (P<0.001). Differences in CPS scores were seen between baseline and 3-8.5 hoursafter LPS administration, and in FAP scores between baseline and 3-5 hours after LPS administration (Wilcoxonsigned rank test, P<0.05). An increase in FAP scores was present in both gaits and models when comparing baselinewith induced lameness (mixed models, P<0.0001). Both scales proved very useful in assessing induced orthopaedicpain in horses. Also, relevant FAP parameters could significantly describe the presence of alterations in facial expression during locomotion in lame horses.

Vertical movement of head and pelvis in the Icelandic horse at walk, trot, pace and töltM. Rhodin, F.M. Serra Bragança, E. Persson-Sjodin, T. Pfau, V. Gunnarsson, S. Björnsdóttir and E. Hernlund

Quantitative lameness assessment, utilising upper body symmetry measurements, is becoming popular in equinepractice but the current systems are only validated for trot. Lameness assessment in Icelandic horses is challengingdue to high stride frequency and gait transitions, making it difficult to identify the timing of footfalls visually.The association between the vertical movements of head and pelvis relative to the loading of the limbs in horsesperforming other gaits than trot is poorly understood. This particular knowledge is a prerequisite to identify thelame limb in gaited horses when head and pelvis movement asymmetry is evaluated. Twenty-six Icelandic horseswere equipped with 12 IMU sensors (sampling frequency 500 Hz, EquiMoves®) and measured during walk, trot,pace and tölt while ridden. Stance and swing phase for each limb and the lowest/highest vertical position of thehead (Hmin/Hmax) and pelvis (Pmin/Pmax) were calculated. Hmin/Pmin events occurred at 55.0%/17.9% (walk), 46.5%/65.9% (trot), 44.4%/66.7% (pace) and 52.1%/60.3% (tölt) of the stance phase of the forelimb/hindlimb. All Hmax/Pmax events occurred within the last 10% of the stance phase, during the suspension phase, or during the first 10%of the stance phase of the next limb, except for Pmax at walk (75.7% of stance phase). To conclude, the Hmin andPmin were closely related to midstance of the fore and hindlimb respectively in all gaits, except for the Pmin at walk.Therefore, changes in vertical movement symmetry for Hmin/Pmin are probably good indicators of weight-bearinglameness. Pmax is probably a good indicator of push-off lameness, except at the walk.

Objective evaluation of stride parameters in the five-gaited Icelandic horseV. Gunnarsson, M. Tijssen, S. Björnsdóttir3 J.P. Voskamp, P.R. Van Weeren, W. Back, M. Rhodin, E. Persson-Sjodin and F.M.Serra Bragança

Evaluation of gait quality in the Icelandic horse at breeding shows and in competitions has so far only been basedon subjective judging scales. The aim of the study was to provide quantitative data for temporal stride parametersfor the five gaits of the breed. Twenty-six Icelandic school horses, ridden by experienced riders, were equipped withinertial measurement unit (IMU) sensors (EquiMoves®) that were attached to each metacarpal/metatarsal bone andset to a sampling speed of 500 Hz. A video camera was also synchronised with the IMU sensors. Representativestrides (>200) for each gait were selected from the videos by a qualified judge and stride parameters were calculatedin descriptive statistics in Minitab based on hoof-on and hoof-off IMU data on a stride per stride basis. Mean±SEfor each gait were: Walk (715 strides): duty factor front limbs (DF-front) 62.6±0.09%, duty factor hind limbs (DFhind)59.0±0.06%, lateral advanced placement (LAP) 25.6±0.12%. Trot (516 strides): DF-front 43.9±0.18%, DF-hind45.8±0.23%, suspension 3.5±0.22%, diagonal advanced placement (DAP) 4.9±0.19%. Tölt (1,218 strides): DF-front45.1±0.14%, DF-hind 44.8±0.12%, LAP 19.4±0.21%. Pace (216 strides): DF-front 35.7±0.28%, DF-hind 42.1±0.41%,suspension 9.9±0.56%, LAP 12.5±0.52%. Canter (777 strides): DF-front 39.3±0.14%, DF-hind 43.8±0.19%, suspension7.1±0.22%. Valuable quantitative data for several important stride parameters of all five gaits of the Icelandic horsewere collected in a field setting using IMU sensors. In combination with traditional subjective methods, this objective technique might enhance assessment of gait quality in competitions and breeding shows.

Validation of gait event detection algorithm using hoof-mounted inertial measurement units (IMU)M. Tijssen, M. Rhodin, S. Bosch, J.P. Voskamp, M. Marin-Perianu, M. Nielen, W. Back, P.R. Van Weeren and F.M. SerraBragança

Inertial measurement unit (IMU) sensors are versatile and affordable tools for gait analysis. The objective of thisstudy was to validate a hoof-on/hoof-off detection algorithm for hoof-mounted IMU sensors. Tri-axial acceleration(accel) and rate-of-turn (RoT) were measured with IMU sensors (sampling frequency 200 Hz, Inertia-Technology)attached to the lateral quarter of the right front (F) and hind (H) hooves of seven Warmblood horses. As a goldstandard, horses were walked and trotted over the force plate (sampling frequency 200 Hz). Axes were synchronisedby calculation of the root of the sum of squares resulting in one-directional IMU signals. Algorithms to detect hoofeventsbased on peak detection were developed; a threshold of mean+2.58×SD was used for the vertical force andmean+1.96×SD for the IMU data. The accuracy and precision of these algorithms was calculated as the mean timeof FP minus IMU in milliseconds (ms) and the SD of these differences. At total of 152 steps (36-walk and 40-trotfor both F and H) were analysed. For hoof-on, accuracy in accel/RoT were 11/87 ms (F), -16/79 ms (H) and forprecision 29/66 ms (F), 15/98 ms (H). For hoof-off, accuracy in accel/RoT were -91/-11 ms (F), -142/-15 ms (H) andfor precision 108/14 ms (F), 140/23 ms (H). Hoof-on events were detected too early by both IMU algorithms andhoof-off events too late. These preliminary results show that combining these IMU algorithms is very promisingfor gait classification. Further algorithm development will include break-over phase detection to improve hoof-off accuracy and hoof-event detection on soft surfaces.