Prediction of working outcomes in trainee dogs using the novel Assistance Dog Test Battery (ADTB)

Abstract

Canine behaviour is commonly assessed using test batteries comprising a test protocol and ethogram scoring system. These are particularly valuable for assistance dog organisations as a tool for evaluating trainee dogs’ proficiency in fundamental skills. The goal of this study was to design and validate a new test battery to assess the suitability of trainee dogs for assistance work at different stages of the training programme. The main objective was to develop a machine-learning tool capable of predicting working outcomes. Accordingly, the novel Assistance Dog Test Battery (ADTB) was developed. Trainee assistance dogs participating in this research performed the test at 3 weeks and 10 weeks after starting formal training. The results from the univariate logistic regression analysis were used to select the variables for the reduced feature sets that were used for modelling. The machine learning models were built using the data collected at 3 and 10 weeks separately and predicted working outcomes with an area under the ROC curve of 0.74 and 0.84, respectively. This research demonstrated the relationship between the novel ADTB ethogram measures and working outcomes in assistance dogs. The machine learning model created using the data collected at 3 weeks achieved comparable performance to the state-of-the-art, while the model built using the data collected at 10 weeks substantially outperformed it. These preliminary results suggest that the ADTB is a reliable tool for the prediction of working outcomes in trainee assistance dogs. Hence, assistance dog organisations can reduce the cost of training by using model predictions as a guide for deciding which dogs to withdraw from training. The data collected and the code developed in this research are publicly available on Mendeley Data (https://doi.org/10.17632/5mzfpt455r.1) and GitHub, respectively (https://github.com/mmarcato/dog_ethogram/).