Improving Ovine Behavioral Pain Diagnosis by Implementing Statistical Weightings Based on Logistic Regression and Random Forest Algorithms.

Abstract

Recently, the Unesp-Botucatu sheep acute pain scale (USAPS) was created, refined, and psychometrically validated as a tool that offers fast, robust, and simple application. Evidence points to an improvement in pain diagnosis when the importance of the behavioral items of an instrument is statistically weighted; however, this has not yet been investigated in animals. The objective was to investigate whether the implementation of statistical weightings using machine learning algorithms improves the USAPS discriminatory capacity. A behavioral database, previously collected for USAPS validation, of 48 sheep in the perioperative period of laparoscopy was used. A multilevel binomial logistic regression algorithm and a random forest algorithm were used to determine the statistical weights and classify the sheep as to whether they needed analgesia or not. The quality of the classification, estimated by the area under the curve (AUC) and its 95% confidence interval (CI), was compared between the USAPS versions. The USAPS AUCs weighted by multilevel binomial logistic regression (96.59 CI: [95.02-98.15]; p = 0.0004) and random forest algorithms (96.28 CI: [94.17-97.85]; p = 0.0067) were higher than the original USAPS AUC (94.87 CI: [92.94-96.80]). We conclude that the implementation of statistical weights by the two machine learning algorithms improved the USAPS discriminatory ability.