Generating high-fidelity synthetic patient data for assessing machine learning healthcare software

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Data Privacy

Background:

• Growing demand for AI in healthcare relies on patient data.
• Patient privacy concerns hinder the use of historical health data.
• Synthetic data offers a privacy-preserving alternative.

Purpose of the Study:

• To develop a method for generating realistic synthetic patient data from UK primary care records.
• To address privacy issues associated with using real patient data for AI models.
• To quantify the risk of patient re-identification from synthetic data.

Main Methods:

• Integration of resampling, probabilistic graphical modelling, latent variable identification, and outlier analysis.
• Focus on handling missing data and complex variable interactions.
• Sensitivity analysis of machine learning classifiers and re-identification risk assessment.

Main Results:

• Synthetic data closely matches original data in feature distributions and dependencies.
• Machine learning classifier performance is comparable between real and synthetic data.
• The risk of generating synthetic data identical or very similar to real patients is low.

Conclusions:

• The integrated approach successfully generates realistic synthetic health data.
• This method enhances data utility for AI while mitigating privacy risks.
• The developed technique offers a viable solution for privacy-preserving healthcare analytics.