Predicting missed alerts in detection dogs.

Abstract

Detection dogs are widely used in tasks ranging from explosives detection to medical diagnostics. Yet even well-trained dogs occasionally fail to indicate target odors. Such missed alerts can have severe operational consequences. Here we show that missed alerts are preceded by measurable changes in the dog's behavioral and physiological state. Using a controlled olfacto-treadmill system with precisely timed odor presentation, we analyzed movement dynamics and physiological signals during the seconds preceding each trial. A machine learning model trained on pre-odor movement patterns predicted subsequent missed indications with above-chance accuracy (AUC = 0.81). Adding heart rate, short-term variation in the intervals between heartbeats, and core body temperature to the model increased the identification of missed alerts from 77 to 85%. As these signals emerged seconds before odor exposure, they may reflect variation in a broader performance state rather than isolated trial-specific events. Combining predictions across consecutive trials thus provided a running estimate of miss risk across the session. Together, these findings suggest that detection performance reflects a dynamic biological process and provide a foundation for anticipatory monitoring of working dog readiness.