Detection dog performance state estimation from pre-stimulus video and physiological signals using deep learning and Bayesian inference

Detection dogs play a critical role in operational settings ranging from explosives detection to medical diagnostics. Their unmatched olfactory capabilities allow them to locate trace-level targets that remain beyond the reach of current technology. However, detection performance can decline without overt behavioral signs, posing a challenge for timely and informed deployment decisions. Identifying subtle precursors to missed alerts remains an open problem with both practical and computational significance. To address this, we recorded detection dogs on a treadmill-based scent delivery system under tightly controlled conditions, enabling precise timing of odor presentation. We focused on the 4.5 seconds preceding each odor pulse to capture anticipatory movement through video and physiological state via sensors. Using markerless pose estimation, we extracted high-resolution joint trajectories, while concurrent recordings provided heart rate, heart rate variability, and core temperature. We trained a spatiotemporal deep learning model that integrates Graph Attention Networks and Temporal Convolutional Networks to predict missed indications from these short, pre-stimulus windows. The model successfully identified 85% of missed alerts with 81% precision. Heart rate variability emerged as the most informative physiological input, suggesting a strong autonomic component to declining readiness. To move from isolated predictions to actionable insights, we implemented a Bayesian aggregation framework that estimates each dogs latent miss rate over time. This probabilistic formulation enables adaptation to individual baselines and operational risk thresholds, supporting context-sensitive deployment decisions. While data were collected in a laboratory setting, our findings highlight behavioral and physiological signatures that precede performance failure. This work lays the foundation for real-time readiness monitoring systems that integrate wearable sensing with interpretable machine learning, supporting timely and welfare-conscious deployment decisions.