Unsupervised Categorization of Animal Behavioral Patterns Using Deep Neural Networks

Automated quantification of animal behavior can reveal behavioral patterns that are difficult to identify through manual scoring. RAINSTORM was recently developed in our lab as a tool for analyzing rodent exploratory behavior based on pose estimation. This project extends RAINSTORM to support reproducible unsupervised behavioral quantification. This approach reduces reliance on predefined human categories while enabling the detection of subtle differences in behavioral organization that may not be evident through visual observation. The original workflow was reorganized to enable systematic data preparation, model integration, evaluation, and reuse. Unsupervised models inspired by VAME and Keypoint-MoSeq were implemented and compared using either pose data alone (pose_ego) or enriched with contextual information from regions of interest (ego_roi). Models were evaluated based on internal segmentation, behavioral state usage, label stability, and agreement with a reserved human reference. Native VAME with pose_ego achieved the best internal motif separation (silhouette score: 0.28 vs. −0.13), whereas official VAME with ego_roi showed greater agreement with the human reference (ARI: 0.083 vs. 0.058). The latter was selected as the operational configuration for subsequent analyses. Overall, this work provides a reproducible framework for training, comparing, validating, and reusing unsupervised behavioral quantification models, facilitating automated analysis of animal behavior.