Novel application of machine learning to enhance untrained and inexperienced evaluators’ diagnosis of acute pain in pigs

Abstract

Accurately identifying pain is a critical first step required to adequately mitigate pain and improve pig health, welfare and quality of life. The objective of this study was to verify whether random forest and support vector machine algorithms trained utilizing experienced evaluators could improve the accuracy of pain diagnosis in untrained and inexperienced evaluators using the Unesp-Botucatu Pig Composite Acute Pain Scale (UPAPS). Four-minute, pre-recorded video clips of 45 male pigs in pain-free (pre-surgical castration) and painful conditions (post-surgical castration) were used. Previously generated scores from three experienced evaluators using UPAPS on a video database were used to train and test random forest and support vector machine models. Following this, ten inexperienced evaluators were recruited to assess the same video clips using the UPAPS. Scores from inexperienced evaluators were then inputted for machine learning algorithms and pain diagnosis was adjusted accordingly. Both machine learning models performed well based on area under the curve, sensitivity > 90 %, and specificity > 95 %. Area under the curve, specificity, and sensitivity of untrained inexperience evaluators were statistically (p < 0.05) equivalent between the original UPAPS, and UPAPS adjusted by random forest and support vector machine. In conclusion, the random forest and support vector machine algorithms trained using experienced evaluators did not modify the discriminatory diagnostic ability of untrained inexperienced evaluators scoring UPAPS. In future studies, additional machine learning techniques could be implemented to investigate whether they improve the accuracy of pain diagnostic. In addition, further studies are needed to develop a concise and standard training program for inexperienced evaluators and investigate its effects on the accuracy of pain diagnosis.