Validation of a handheld smartphone markerless method for quantifying thoracolumbar flexion–extension range of motion in horses at trot

Abstract Background Back pain and poor performance in horses are difficult to assess objectively and thoracolumbar flexion‐extension range of motion (FE ROM ) is not reliably evaluated by subjective observation alone. A handheld, smartphone‐based markerless computer vision system (RealHorse®; RH) may enable objective clinical and field‐based FE ROM assessment but requires validation against an established optical motion capture reference system (Qualisys®; QS). Objectives To compare the accuracy and precision of RH relative to QS for measuring FE ROM in horses trotting on a straight line and on a circle. Study Design Cross‐sectional comparative validation study. Methods Fifty‐nine horses were recorded trotting on a straight line; 23 of these were also recorded on a circle. Data were collected simultaneously using RH and QS, with a light source for manual frame‐level synchronisation. The validation videos of horses were not used to train the RH algorithm. Agreement between RH and QS was analysed separately for straight‐line and circular trot at stride and trial levels using mean signed error (MSE), mean absolute error (MAE), and Bland–Altman limits of agreement (LoA). Results On the straight‐line, stride‐level MSE was −0.08°, MAE 0.96°, and LoA −2.49° to 2.32°. Trial‐level agreement was high, with MSE −0.13°, MAE 0.44°, and LoA −1.22° to 0.94°. On the circle, stride‐level MSE was −0.71°, MAE 1.14°, and LoA −3.17° to 1.74°. At the trial‐level, variability was moderate (MSE −0.62°, MAE 0.78°, LoA −2.18° to 0.93°). Overall agreement was lower on the circle than on the straight‐line and higher at trial‐level than at stride‐level. Conclusions Clinically, RH may support objective clinic and field‐based monitoring of thoracolumbar flexion‐extension range of motion, especially for repeated within‐horse assessments, but decision thresholds remain to be established. Stride‐level agreement between RH and QS was influenced by expected stride‐to‐stride variability, whereas averaging across strides improved agreement at the trial‐level.