Hierarchical SDG Reproduces CAMHS Clinical Data

This retrospective cohort study (2026) used electronic medical records from 6,924 CAMHS patients in Stavanger, Norway to train a hierarchical synthetic data generator to reproduce 7,730 referral periods and 58,524 episodes of care. The synthetic dataset achieved high statistical similarity (KSC/TVC 0.92, CS 0.77, CSS 0.92), low reidentification risk (singling-out 0.39%, multivariate 5%), and comparable predictive utility (TSTR PRAUC 0.40 vs 0.43). The results support privacy-preserving data sharing for CAMHS research.