BEHAVE: Behavioral Ethology for Human Assessment via Variational Encoding

Quantifying spontaneous human behavior remains a major challenge in psychiatry and neuroscience. We present BEHAVE (Behavioral Ethology for Human Assessment via Variational Encoding), a framework that combines computer vision and unsupervised latent-variable models to capture fine-scale, naturalistic behaviors. BEHAVE segments continuous motion into interpretable motifs and introduces novel metrics of temporal structure, repertoire diversity, and stereotypy. In a naturalistic open-field assay of individuals with euthymic bipolar disorder (BD) and healthy controls, these metrics revealed subtle yet robust BD-associated differences, including reduced exploratory transitions and repertoire narrowing. Compared to clinical scales and standard action-recognition models, BEHAVE achieved superior classification of BD. This approach offers a scalable, bias-resistant path to decoding neuropsychiatric states from natural behavior and lays the foundation for translational biomarkers.