A novel missing data imputation approach based on clinical conditional Generative Adversarial Networks applied to EHR...

Area of Science:

• Machine Learning
• Biomedical Informatics
• Data Science

Background:

• Missing data is a significant challenge in Electronic Health Records (EHR), leading to spatiotemporal sparsity.
• Existing data imputation methods often lack model integration, are not optimized for EHR specifics, and utilize limited feature information.

Purpose of the Study:

• To propose a novel data imputation strategy for EHR data using a clinical conditional Generative Adversarial Network (ccGAN).
• To address the limitations of current methods by exploiting non-linear, multivariate information and handling high missingness rates in EHR.

Main Methods:

• Developed a clinical conditional Generative Adversarial Network (ccGAN) for data imputation.
• Conditioned the imputation strategy on observable and fully-annotated data to manage high missingness.
• Evaluated performance on a real multi-diabetic centers EHR dataset and a benchmark EHR dataset.

Main Results:

• The ccGAN achieved a 19.79% gain in imputation performance and a 1.60% gain in predictive performance compared to state-of-the-art methods.
• Demonstrated robustness across various missingness rates, with up to a 1.61% performance gain in high missingness conditions.
• Statistical significance confirmed against existing approaches.

Conclusions:

• The proposed ccGAN offers a superior data imputation strategy for EHR data, outperforming current methods.
• This approach effectively handles the inherent challenges of missing data in EHR, improving both data quality and predictive accuracy.