SADI: Similarity-Aware Diffusion Model-Based Imputation for Incomplete Temporal EHR Data.

Area of Science:

• Health Informatics
• Machine Learning
• Biostatistics

Background:

• Missing values in temporal electronic health records (EHR) complicate analysis and bias results.
• Current imputation methods struggle with sparse EHR data common in non-ICU settings.
• Existing models rely on temporal and feature correlations, which are weak in sparse data.

Purpose of the Study:

• To introduce a novel imputation method, Similarity-Aware Diffusion Model-Based Imputation (SADI), for temporal EHR data.
• To address the limitations of current state-of-the-art (SOTA) models in handling sparse EHR data.
• To leverage diffusion models and patient similarity for improved imputation.

Main Methods:

• Developed SADI, a diffusion model-based imputation technique.
• Incorporated a similarity-aware denoising function with a self-attention mechanism.
• Modeled correlations across time points, features, and similar patients.

Main Results:

• SADI demonstrated superior performance over SOTA imputation methods.
• Experiments were conducted on Critical Path For Alzheimer's Disease (CPAD) and PhysioNet Challenge 2012 datasets.
• SADI proved effective under various missing data mechanisms (MCAR, MAR, MNAR).

Conclusions:

• SADI offers a significant advancement in imputing missing temporal EHR data.
• The method's ability to utilize information from similar patients is a key innovation.
• SADI provides a robust solution for EHR data imputation, especially in non-ICU settings.