A new BMJ analysis argues synthetic data can preserve most of the performance teams care about—reporting 89% utility for clinical decision support—while reducing reliance on sensitive patient records. The catch: the paper treats “safe synthetic” as an engineering and governance problem, not a procurement checkbox.
BMJ analysis: synthetic data retains 89% utility for clinical decision support
BMJ Evidence-Based Medicine published an analysis on synthetic data use in healthcare, reporting that synthetic datasets can retain 89% utility in clinical decision support contexts. The paper positions synthetic data as a practical response to common healthcare data constraints—especially limited access to patient-level data due to privacy, compliance, and operational friction—while still enabling model development and analytics.
The analysis outlines nine healthcare uses for synthetic data, includes four hospital case studies describing real-world deployments, and provides safe-use guidelines aimed at reducing re-identification risk and supporting responsible adoption. In other words, it’s not just a performance claim; it’s a playbook framing synthetic data as part of a broader clinical data lifecycle that must be validated, monitored, and governed.
- Data teams get a measurable bar for “good enough.” An 89% utility figure (in clinical decision support) helps teams justify synthetic data for specific workloads—training, testing, and exploratory analytics—when access to real patient data is slow, restricted, or legally risky.
- Privacy and compliance work shifts from “can we use it?” to “how do we prove it’s safe?” Safe-use guidance signals that synthetic data still needs documented risk assessment (e.g., re-identification exposure), controls, and auditability—especially once it feeds downstream models and clinical workflows.
- Procurement should demand evaluation protocols, not vendor promises. The hospital case studies and guidelines imply buyers should require repeatable utility testing, privacy-risk testing, and clear statements of intended use (training vs. decision support vs. sharing) before scaling deployment.