Simulating complex patient populations with hierarchical learning effects to support methods development for...

Area of Science:

• Biostatistics
• Health Informatics
• Machine Learning

Background:

• Validating algorithms requires ground truth, which is unavailable in real-world clinical data.
• Simulation studies with synthetic datasets are essential for evaluating methods that disentangle treatment risk from experiential learning.
• Existing simulation methods lack the ability to incorporate hierarchical learning effects.

Purpose of the Study:

• To present a generalizable framework for generating synthetic clinical data with hierarchical learning effects.
• To enable rigorous testing of algorithms designed to differentiate intrinsic treatment risk from learning-associated risk.
• To improve the development and validation of algorithms for clinical data analysis.

Main Methods:

• A multi-step data generation process with customizable modules was developed.
• Synthetic patients with complex features were generated and assigned to case series.
• Hierarchical learning effects were injected at various speeds and magnitudes, with options for missing values and omitted variables.

Main Results:

• The framework successfully generated synthetic data reflecting specified characteristics.
• Deviations in treatment effects and feature distributions were observed in smaller datasets due to random variability.
• Simulated datasets demonstrated changes in adverse outcome probabilities influenced by hierarchical learning effects.

Conclusions:

• The framework extends clinical data simulation by incorporating hierarchical learning effects.
• This enables robust validation of algorithms for disentangling treatment safety signals from experiential learning.
• The work supports the identification of training opportunities and hastens medical treatment improvements.