A generalized additive model using kinematic and demographic data provides an accurate and objective anesthesia...

To analyze kinematic data extracted from an inertial measuring unit to assess quality of recovery in horses emerging from general anesthesia.

Horses undergoing anesthesia from February 2024 through March 2026 were enrolled in this study. Collected data from the inertial measuring unit included 3-axis acceleration, angular velocity, and orientation. Recoveries were evaluated by 3 scorers using a modified composite grading scale. The mean of the scores was used as the dependent variable. A Python pipeline was written to detect attempts to stand using the first derivative of the acceleration signal. Various kinematic features were calculated from each attempt. These features, as well as some demographic data, were used as independent variables. Four competing regression frameworks were evaluated: a multiple linear model, a polynomial model, a generalized additive model, and a random forest model. Model recommendation was based on 4 metrics: Akaike information criterion, Bayesian information criterion, root mean squared error, and R2 value.

102 mixed-breed horses were enrolled in this study. The generalized additive model achieved the lowest Akaike information criterion (658.24), Bayesian information criterion (737.96), and root mean squared error (5.21) and the highest R2 value (0.8). In this model, the independent variables that showed the highest impact included cumulative fall intensity, time to stand, and duration of anesthesia (linear relationship) as well as mean impact severity (nonlinear).

The generalized additive model was selected as our final model architecture.

Recovery scores based on kinematic and demographic data can be used to obtain objective recovery quality scores.