Synthesis and quality assessment of combined time-series and static medical data using a real-world time-series generative adversarial network

Jaewon Kim ¹, Hyunwoo Choo ¹, Soo-Yong Shin ¹

⁰Department of Digital Health, Samsung Advanced Institute for Health Sciences and Technology, Sungkyunkwan University, Seoul, Republic of Korea.

Scientific Reports +

|

August 17, 2024

View abstract on PubMed

Summary

This study synthesizes medical data using real-world time-series generative adversarial networks (RTSGAN), ensuring patient privacy. The generated synthetic data accurately reflects real patient information and maintains predictive performance for colorectal cancer survival rates.

Area Of Science

Medical Informatics
Artificial Intelligence
Data Science

Background

Privacy concerns hinder the use of sensitive medical data.
Developing methods for secure utilization of patient information is crucial.
Real-world time-series generative adversarial networks (RTSGAN) offer a potential solution for data synthesis.

Purpose Of The Study

To synthesize high-quality, privacy-preserving medical data using RTSGAN.
To evaluate the fidelity and utility of synthesized time-series medical data.
To assess the privacy risks associated with the synthetic data generation process.

Main Methods

Utilized RTSGAN to synthesize 53,005 data points from 15,799 colorectal cancer patients.
Performed quantitative evaluations including Hellinger distance, TSTR/TRTS AUC, and propensity mean squared error.
Conducted qualitative analyses (t-SNE, histogram) and privacy assessments (distance to closest records, membership inference test).

Main Results

Quantitative metrics demonstrated high data quality (Hellinger distance 0-0.25, AUC 0.98-0.99, PMSE 0.223).
Qualitative analyses confirmed similarity between synthetic and real data.
Synthetic data achieved comparable performance to real data in predicting colorectal cancer survival.
Privacy assessments indicated minimal risk of data exposure.

Conclusions

RTSGAN is effective for synthesizing realistic and privacy-preserving medical time-series data.
Synthesized data accurately represents real-world patient characteristics.
The approach enables secure utilization of medical data for AI model development and clinical insights.

Keywords:

Data quality Generative adversarial networks Medical data Privacy Sequential data Synthetic data