Predicting equine behavior from small datasets using machine learning with LLM-generated explanations

Understanding horse fearfulness is crucial for building strong human-animal relationships, influencing training methods, task selection, and predicting reactions to new stimuli. This interdisciplinary study aimed to identify key characteristics for predicting a horse’s fearfulness. Using classical machine learning, we analyzed anatomical, kinematic, and housing-related data from 49 horses, with fear scores obtained through a standardized behavioral test. To ensure an unbiased evaluation, the leave-one-out cross-validation method was applied. The study’s main contributions are: (1) an iterative feature selection approach that reduces the number of required measurements while maintaining prediction accuracy; (2) a unique dataset on Lipizzan horses, revealing that head and body anatomical characteristics are critical for assessing fearfulness; (3) identification of the Decision Tree algorithm as the most accurate machine learning method for modeling fearfulness; (4) integration of Large Language Models (LLMs) to generate clear, interpretable textual explanations of the Decision Tree, improving the understanding of key predictive features. This study bridges behavioral science and artificial intelligence, offering a novel AI-driven approach to equine behavior analysis, with practical applications in horse training, selection, and welfare management. • Iterative feature selection cuts measurements while preserving prediction accuracy. • AI can assess horse fearfulness using conformation traits of head and body. • Decision Tree is most accurate for predicting fearfulness in Lipizzan horses. • Leave-one-out validation suits small datasets with many more features than samples. • LLM explains Decision Tree via text, diagram, and feature descriptions.