Med-Diffusion: Diffusion Model-Based Imputation of Multimodal Sensor Data for Surgical Patients.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Data Science

Background:

• Multimodal medical data completeness is crucial for surgical outcomes.
• Missing data in patient records, due to sensor issues, hinders clinical understanding and predictive modeling.
• Data sparsity is a significant challenge in rare or complex medical cases.

Purpose of the Study:

• To introduce Med-Diffusion, a diffusion-based generative framework for imputing missing multimodal clinical data.
• To enhance sensor data integrity and improve the performance of predictive models using heterogeneous data types.
• To address data sparsity and improve the understanding of patient conditions.

Main Methods:

• Developed Med-Diffusion, a conditional diffusion model framework.
• Integrated one-hot encoding, simulated bit encoding, and feature tokenization for heterogeneous data.
• Utilized diffusion modeling to learn data distributions and synthesize plausible data for incomplete records.

Main Results:

• Med-Diffusion effectively imputes missing multimodal clinical data, including categorical and numerical variables.
• The framework successfully mitigates data sparsity caused by sensor inaccuracies.
• Extensive experiments show Med-Diffusion enhances the performance of downstream predictive models.

Conclusions:

• Med-Diffusion offers a robust solution for reconstructing missing multimodal clinical data.
• The framework improves data integrity, leading to better predictive model performance.
• This approach has the potential to advance the development of algorithms for predicting disease progression.