Characterization of Synthetic Health Data Using Rule-Based Artificial Intelligence Models.

Area of Science:

• Artificial Intelligence
• Health Informatics
• Data Science

Background:

• Synthetic health data generation is crucial for privacy-preserving analysis.
• Evaluating the quality of synthetic data is essential for reliable research.
• Existing methods for synthetic data quality assessment have limitations.

Purpose of the Study:

• To apply and characterize eXplainable AI (XAI) for assessing synthetic health data quality.
• To compare XAI-derived insights with conventional utility metrics.
• To explore the potential of XAI in understanding the mechanisms of generated data.

Main Methods:

• Generated synthetic datasets using conditional Generative Adversarial Networks (GANs).
• Applied a rule-based XAI algorithm (Logic Learning Machine) for analysis.
• Assessed classification performance across different data training/testing scenarios (real vs. synthetic).
• Compared rules extracted from real and synthetic data using a similarity metric.

Main Results:

• XAI successfully assessed synthetic data quality through classification performance analysis.
• Analysis of extracted rules (number, coverage, structure, cut-off values, similarity) provided quality insights.
• Demonstrated significant differences and similarities between rules from real and synthetic data.

Conclusions:

• XAI offers an original and effective approach to synthetic health data quality assessment.
• Rule analysis via XAI can reveal underlying mechanisms of data generation.
• XAI enhances the trustworthiness and utility of synthetic health datasets.