Synthetic Clinical Data for Privacy-Preserving Research

Abstract Background Synthetic clinical records and narratives can support research and information-system testing without routine access to identifiable patient files. A reusable generator must coordinate record relationships, encounter context, and configurable documentation conventions. Fluent prose and relational consistency alone do not establish clinical fidelity, privacy protection, or reliable training labels. Objective To describe a configurable framework for linked synthetic clinical records and unstructured notes, evaluate the integrity of archived outputs, and verify reference functions for reproducibility and annotation. We also specify an extension toward a virtual electronic health record and continual, on-demand corpus generation. Methods Five stages construct patient anchors, encounters, operational events, demographics, and coverage-guided clinical notes. We examined an archived census of structured exports and two disjoint narrative sets, reproduced isolated implementation fixtures, and tested deterministic seeds, bounded timing, export checks, exact span compilation, and patient-level partitioning in a separate reference module. Python notebooks combine structured sampling with four-bit MedGemma inference. The author-reported generation environment used virtual machines with dual-GPU cloud accelerators. Historical outputs and revised software were evaluated separately; system extensions are presented as proposals. Results The archived census recorded 50,000 patients, 3,225,625 unique encounters, and 375,000 operational events covering 250,000 visits. All sampled encounter counts agreed with patient anchors, and operational patient-visit mappings were valid. The 300 narrative documents represented 250 patients and visits. Strict encounter ordering held, although 341 intervals were shorter than 60 seconds. In the 200-document narrative set, 175 texts contained Markdown fences and 30 contained standalone missing-value tokens. All 39 reference-module tests passed, including 1,000 constructed timing cases. This establishes selected software contracts; no new language-model corpus, clinical assessment, or trained de-identification benchmark was produced. Conclusions Explicit record contracts and coverage planning provide a reusable basis for synthetic clinical data generation. The evidence establishes selected structural properties and bounded software behaviors in one evaluated configuration. The architecture could supply a virtual EHR and repeated narrative releases, but clinical fidelity, sustained operation, privacy risk, and transfer to other settings require independent evaluation.