Estimating military working dog core temperature change with machine learning: A simulation study

Abstract

Military working dogs play a critical role in supporting the United States Military across various missions. Many missions occur in hot environments and predispose military working dogs to hyperthermia, a leading cause of their death. A previous effort created a physics-based model to estimate military working dog core temperature change based on dog attributes, metabolic activity, and environmental conditions. We hypothesize that a machine learning model that builds on the physics-based model may offer additional and complementary benefits. This includes the ability to infer relationships from real-world data, easy modification of features, and compatibility with existing techniques to explain predictions. These benefits would enhance model usability and applicability, especially for scenarios where dogs’ core temperature change patterns may be atypical. As a first step towards a machine learning-based modeling framework, we approximated the physics-based model with three machine learning models, showing feasibility of applying machine learning to this type of data. Of these models, the random forest model, which we dub the “K9-TempML” model, provides the closest core temperature change estimates to the physics-based model and offers well-established methods for feature analyses. We performed feature analysis and augmentation studies on this model to determine the impact of each feature on core temperature change estimates, demonstrating that removing features that are difficult to collect could retain model accuracy but improve usability. Future work includes training a machine learning-based model using real-world data, applying the model to individual military working dogs, and integrating the model into a real-time core temperature change estimation tool and pre-deployment planning aid.