Masked Clinical Modelling: A Framework for Synthetic and Augmented Survival Data Generation

Area of Science:

• Health Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Clinical data access is limited by privacy concerns, hindering healthcare research and development.
• Synthetic data generation offers a privacy-preserving alternative but often lacks utility for clinical insights.
• Existing methods prioritize data realism over the clinical utility of synthetic datasets.

Purpose of the Study:

• To introduce Masked Clinical Modelling (MCM), a novel framework for synthetic data generation and augmentation.
• To evaluate MCM's effectiveness in preserving clinical utility, specifically hazard ratios in survival analysis.
• To demonstrate MCM's potential for advancing privacy-preserving healthcare research.

Main Methods:

• Developed the Masked Clinical Modelling (MCM) framework, inspired by masked language modelling.
• Applied MCM for synthetic data generation and conditional data augmentation.
• Evaluated MCM on the WHAS500 dataset using Cox Proportional Hazards models.

Main Results:

• MCM-generated data improved discrimination and calibration in survival analysis.
• The framework demonstrated superior performance compared to existing synthetic data methods.
• Preservation of hazard ratios, a key clinical metric, was successfully achieved.

Conclusions:

• Masked Clinical Modelling (MCM) offers a robust solution for generating clinically useful synthetic survival data.
• The framework enhances the utility of synthetic datasets for survival analysis and other healthcare applications.
• MCM shows significant potential to overcome data access barriers in medical research.