MediSim: Multi-granular simulation for enriching longitudinal, multi-modal electronic health records.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHRs) are crucial for healthcare but often contain missing or incomplete data across various modalities.
• Augmenting EHRs with realistic, simulated data is essential for robust AI model development and validation.
• Existing methods struggle to effectively simulate multi-modal EHR data, especially in low-data scenarios.

Purpose of the Study:

• To introduce MediSim, a multi-modal generative model designed for simulating and augmenting EHR data.
• To address the challenge of missing modalities and visits within EHRs.
• To improve the quality and completeness of EHR data for downstream AI applications.

Main Methods:

• Developed MediSim, a multi-modal generative model utilizing a multi-granular, autoregressive architecture.
• Employed iterative, reinforcement learning-based training to enhance simulation in low-data settings.
• Integrated encoder-decoder model pairs to effectively handle complex modalities like clinical notes and medical images.

Main Results:

• MediSim demonstrated superior performance over baselines in predicting missing codes (over 74% improvement).
• Achieved up to 65% better downstream predictive performance compared to original deficient records.
• Successfully generated realistic synthetic clinical notes and X-ray images for downstream tasks.

Conclusions:

• MediSim effectively simulates and augments multi-modal EHR data, addressing data deficiencies.
• The model significantly enhances data completeness and improves the performance of downstream predictive tasks.
• MediSim's capability to generate comprehensive, high-dimensional EHR data holds substantial potential for advancing AI in healthcare.