Predictive accuracy and generalizability of a detection dog behavioral evaluation

Abstract

Detection dogs play a vital role in public and environmental security. Demand for highly capable detection dogs continues to rise, yet many dogs bred or trained for these purposes lack the necessary behavioral characteristics to succeed. Thus, efforts have been aimed at developing reliable and efficient methods of identifying dogs suitable for operational detection work. This study assessed the accuracy and generalizability of a previously validated standardized behavioral evaluation for detection dog candidates in an internal cohort of dogs from a detection dog breeding and training program (n = 241) and an external cohort of dogs from novel populations (n = 67). Analyses identified consistent underlying behavioral domains and individual traits that were predictive of outcome across ages and populations. Feature selection identified reduced subsets of behavioral measures that achieved acceptable to strong predictive performance with the internal cohort at 3-mo and 12-mo of age (AUC = 0.75 and 0.82, respectively), with moderate predictive balanced accuracy (60%) when applied to the external cohort. Collectively, results support the validity and generalizability of the behavioral assessment and demonstrate its practical utility for optimizing detection dog selection, though accuracy of prediction of various traits may be population-specific due to differences in the prevalence of behavioral phenotypes.