Temporal Connectomics for actigraphy-based psychiatric stratification: a comparative study of circadian, entropy, graph, and sequence representations

Wrist actigraphy is increasingly used for digital phenotyping in psychiatry, yet most analytical approaches compress week-long recordings into static summary statistics that overlook higher-order temporal organization. We introduce Temporal Connectomics, an interpretable representation that characterizes not only how much individuals move, but how their activity is organized across circadian time. The framework combines multiscale entropy with shrinkage-regularized clock-hour coordination graphs to capture complementary aspects of behavioral complexity and temporal coordination. Using a standardized cohort of 117 participants from the public OBF-Psychiatric dataset (ADHD, depression, schizophrenia, and healthy controls), we compared Temporal Connectomics against conventional activity statistics, classical circadian features, graph-only representations, handcrafted feature fusion, and convolutional neural networks under subject-level cross-validation. Temporal Connectomics significantly outperformed conventional static activity summaries while achieving performance comparable to graph-based and deep-learning baselines, demonstrating that temporal organization provides diagnostically relevant information beyond activity magnitude alone. Although graph-only features achieved the highest internal classification performance, the proposed framework offers a unified, interpretable representation that integrates temporal complexity and circadian coordination within a reproducible analytical pipeline. These findings suggest that behavioral organization across time is an informative dimension of digital phenotyping while emphasizing that external validation is required before clinical application.