A joint learning method for incomplete and imbalanced data in electronic health record based on generative...

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Electronic health records (EHR) often contain incomplete and imbalanced data, hindering accurate clinical predictions.
• Previous methods addressing these issues separately lead to reduced prediction task performance.

Purpose of the Study:

• To propose a unified framework for simultaneously addressing incomplete and imbalanced data in EHR.
• To develop and evaluate a novel model, MVIIL-GAN, for joint learning of data imputation and generation.

Main Methods:

• Developed the Missing Value Imputation and Imbalanced Learning Generative Adversarial Network (MVIIL-GAN).
• Employed joint learning for high missing rate data imputation and conditional EHR data generation.
• Utilized sample-level and variable-level discriminators for distinguishing generated data.

Main Results:

• MVIIL-GAN integrates missing value imputation and data generation into a single step.
• Achieved improved parameter optimization consistency and enhanced prediction task performance.
• Outperformed existing methods on the MIMIC-IV dataset with high missing and imbalanced data.

Conclusions:

• The proposed MVIIL-GAN framework effectively handles challenges of incomplete and imbalanced EHR data.
• Joint learning in MVIIL-GAN improves clinical prediction accuracy.
• MVIIL-GAN offers a promising approach for leveraging EHR data in clinical research.