Synthetic Patient Data Generation for Multi-Treatment Causal Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing synthetic datasets for causal inference in healthcare are limited by small, non-representative populations, dissimilarity to real target populations, and lack of ground truth for multiple treatment effects, posing challenges in validating causal models and preserving patient privacy.
Innovation Solution
A large-scale synthetic dataset is generated using an adapted ADS-GAN model and neural network to mimic a nationwide cohort of hypertension patients, ensuring patient privacy and providing ground truth for multiple treatment effects, while maintaining a nearly identical joint distribution to real data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If de-identification of patient records is used to mitigate legal and ethical risks, then patient privacy is protected, but the data becomes insufficient for avoiding re-identification through linkage with other identifiable datasets
Solution Approach 1:
The patent creates synthetic patient data that copies the statistical properties and relationships of real patient data without containing actual patient information. This synthetic dataset preserves the joint distribution and causal structures needed for model validation while being inherently privacy-protected since it contains no real patient records to re-identify
Solution Approach 2:
The patent introduces synthetic data as an intermediary between real patient data and research models. This intermediary layer allows researchers to validate causal inference models on data with known ground truth effects while maintaining privacy protection, as the synthetic data serves as a mediator that preserves statistical properties without containing identifiable information
2Measurement precision
If existing synthetic datasets are used for causal inference validation, then ground truth can be generated, but the datasets are limited by small non-representative patient populations and dissimilarity to real target populations
Solution Approach 1:
The patent transforms the synthetic data generation process by changing key parameters: creating large-scale datasets (millions of records vs. small existing datasets), preserving complex joint distributions of multiple variables, and maintaining representativeness of target populations through careful modeling of real data relationships and structures
Solution Approach 2:
The patent creates a universal synthetic data framework that can serve multiple functions: validating causal inference models, preserving patient privacy, representing diverse populations, and providing ground truth for multiple treatment effects simultaneously. The synthetic data generation process is designed to be adaptable to different disease areas and population characteristics
3Adaptability or versatility
If observational data is used to infer causal treatment effects, then RCT feasibility constraints are overcome, but confounding variables affect both treatments and outcomes making causal claims difficult
Solution Approach 1:
The patent creates synthetic data that copies the confounding variable structures present in real observational data, allowing researchers to test and validate causal inference models' ability to handle confounding. The synthetic data preserves the relationships between confounders, treatments, and outcomes, enabling rigorous model validation without requiring actual RCT data
Data Source
AI summary
The present disclosure shows a generated, large-scale and realistic patient dataset that mimics real patient data distributions, supports multiple treatments, and provides ground truth for the effects of these treatments. The datasets generated are of synthetic patients modeled on a large nationwide cohort of real patient data including their history of diagnoses, medications, and laboratory values. A data generation process is designed by adapting an Anonymization Through Data Synthesis Using Generative Adversarial Networks (ADS-GAN) model for fictitious patient information generation and using a neural network for treatment outcome generation. The synthetic dataset demonstrates strong similarity to the original dataset as measured by the Wasserstein distance. In addition, privacy of the original patients is preserved so that the dataset can be made available to the research community to evaluate causal inference models.