CLEO closed loop framework for synthesizing medical privacy preserving tabular data.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Data Privacy

Background:

• Sharing patient data for multi-center research is limited by privacy regulations, creating data silos.
• Existing methods struggle to balance data utility and privacy preservation.

Purpose of the Study:

• To introduce CLEO (Clean-Learn-Evaluate-Optimize), a novel closed-loop framework for generating privacy-preserving synthetic medical data.
• To formalize data synthesis as a Markov Decision Process using a Gaussian Mixture Model and Q-learning.

Main Methods:

• CLEO integrates a Gaussian Mixture Model generator with Q-learning optimization.
• The framework was evaluated on a multicenter intracranial aneurysm dataset.
• Downstream tasks included training predictive models and empirical privacy auditing.

Main Results:

• CLEO achieved a combined score of 0.9232 ± 0.0124, outperforming comparison methods (TVAE, Gaussian Copula, CTGAN).
• Models trained on CLEO data showed an average AUC of 0.7376, retaining clinical decision signals.
• Privacy auditing indicated a low re-identification risk (Nearest Neighbor Adversarial Accuracy: 0.4940).

Conclusions:

• CLEO offers a controllable and auditable framework for cross-institutional research using synthetic medical data.
• While effective, minority class prediction accuracy needs further improvement due to data imbalance.
• The framework supports research when direct aggregation of real-world medical data is not feasible.