Optimizing the synthesis of clinical trial data using sequential trees.

Area of Science:

• Health Informatics
• Biostatistics
• Data Science

Background:

• Sharing clinical trial data is crucial but requires privacy-preserving methods.
• Synthetic data generation, using sequential trees, is a promising approach.
• The impact of variable order on synthetic data utility was previously unevaluated.

Purpose of the Study:

• To assess the impact of variable order on synthetic clinical trial data utility.
• To develop an optimization algorithm for determining optimal variable order.
• To evaluate the variability in synthetic data utility across different variable orders.

Main Methods:

• Simulated six oncology clinical trial datasets.
• Employed three utility metrics: univariate similarity, multivariate prediction accuracy, and distinguishability.
• Utilized particle swarm optimization and curriculum learning for variable ordering.

Main Results:

• Increased dataset complexity led to higher utility variability with random order.
• Particle swarm optimization with a distinguishability hinge loss improved utility consistency across datasets.
• The optimization method outperformed curriculum learning in maintaining data utility.

Conclusions:

• The proposed optimization approach reliably synthesizes high-utility clinical trial data.
• This method addresses the challenge of variable order dependency in synthetic data.
• Enables scalable, privacy-protective sharing of valuable clinical trial information.