UNSUPERVISED MAPPING OF BEHAVIORAL CHANGES BY INTEGRATING 3D POSE ESTIMATION WITH OPTICAL FLOW DYNAMICS

Conventional behavioral phenotyping in addiction research often relies on pre-defined behavioral metrics and 2D still images (video frames), which overlook animal’s perspective and time-axis component. To overcome these limitations, this study aimed to discover unbiased behavioral markers by integrating 3D pose estimation with spatiotemporal dynamics modeling. We recorded mice from a synchronized multi-view setup (top and side), extracting body keypoints via DeepLabCut and reconstructing them into 3D space using Direct Linear Transformation (DLT). Crucially, we incorporated optical flow concepts into the feature extraction phase to quantify motion dynamics, capturing the continuous flow of movement along the time axis rather than relying on static positions. These high-dimensional features were embedded using UMAP and segmented via unsupervised clustering to construct comprehensive behavioral landscapes independent of human bias. Quantitative analysis of cluster occupancy, transition probabilities, and spatial distributions in cocaine sensitization and conditioned place preference (CPP) models revealed distinct behavioral signatures in cocaine-treated groups compared to controls. Distinct from traditional markers constrained by experimenter bias, our AI-driven approach evaluates the data across the full temporal continuum to classify behavioral patterns objectively. Consequently, this method identifies latent, highly granular, and unbiased behavioral signatures specific to the addiction phenotype, offering a precision framework for characterizing addiction pathology.