Generation of realistic synthetic data using Multimodal Neural Ordinary Differential Equations

Area of Science:

• Artificial Intelligence
• Biostatistics
• Health Informatics

Background:

• Healthcare organizations face challenges collecting representative disease population data due to data silos.
• Sharing patient data across institutions is restricted by legal regulations, hindering comprehensive analysis.
• Existing federated data access solutions are technically and organizationally complex.

Purpose of the Study:

• To introduce a novel AI approach for generating realistic synthetic patient data.
• To overcome limitations in data sharing for improved statistical modeling and clinical study generalization.
• To enable smooth interpolation and extrapolation of clinical study data using continuous time-scale trajectories.

Main Methods:

• Development of Multimodal Neural Ordinary Differential Equations (MultiNODEs), a hybrid AI model.
• Integration of static and longitudinal patient data within the MultiNODEs framework.
• Implicit handling of missing data values during synthetic data generation.

Main Results:

• MultiNODEs successfully generated highly realistic synthetic patient trajectories on a continuous time scale.
• The method demonstrated capabilities in integrating diverse data types (static and longitudinal).
• Application to real-world clinical and simulated epidemiological data validated the approach.

Conclusions:

• MultiNODEs offer a viable alternative to direct data sharing for enhancing medical research.
• The AI approach improves the generalizability of statistical models by providing representative synthetic data.
• This method facilitates smoother data analysis and insights from clinical studies.