Evaluation of accelerometer fitting position, sampling intervals and data editing techniques for measuring equine lying postures when stabled

The aim of this study was to determine whether a leg-mounted accelerometer could be used to reliably predict the time spent by horses in sternal and lateral recumbency. Six stabled horses were used for the study and each observation period lasted 11.5 hrs (2000–0730). Each evening, three horses were fitted with an accelerometer logger to the front-left (metacarpal) and back-left (metatarsal) legs, and each logger was configured to sample an acceleration value at the same rate (60 s, 20 s or 10 s) on all three axes (X, Y, Z). Observations were repeated for each horse at each sampling interval resulting in six datasets per horse. Ground-truth information was simultaneously recorded using a camera located in each stable and each accelerometer dataset was subsequently annotated with the ground-truth observations. Optimal thresholds for standing and lying (Y-axis) and variants of lying (sternal and lateral recumbency: Z-axis) were determined across all horses using the highest sum of sensitivity, specificity, positive predictive value, and negative predictive value across sampling intervals and accelerometer fitting sites. For standing and combined lying postures, mean predictive performance across all sampling intervals and fitting sites was > 98% for all performance measures. However, the best estimates of lying variants were achieved using data from the front leg where mean predictive performance for all measures was > 86% across all sampling intervals. The effect of short and potentially erroneous readings on the predictive ability of the thresholds was also examined using two event filters. The event filters converted predictions that occurred individually or in consecutive runs of ≤ 2 to the behaviour preceding them, and results were compared to unfiltered predictions. Linear regressions were used to evaluate predicted values with ground-truth observations for total lying time (overall lying time as well as total time in sternal and lateral recumbency), mean lying bout duration (combined lying bouts), and total lying bouts (switches between sternal and lateral recumbency). Herein we make recommendations on the use of event filters for generating reliable estimates (no statistical difference (P > 0.05) from 1 and 0 for slope and intercept, respectively, and R2 ≥ 0.90) of equine standing and lying postures across the tested sampling intervals and accelerometer fitting sites. This is the first study to demonstrate that leg-mounted accelerometers can be used to reliably predict the time spent by horses in sternal and lateral recumbency.