Unsupervised Imputation of Non-ignorably Missing Data Using Importance-Weighted Autoencoders.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Deep Learning (DL) methods are increasingly used in biomedical sciences.
• Missing data in biomedical datasets poses significant challenges for DL models.
• Variational Autoencoders (VAEs) are popular for unsupervised learning but struggle with complex missingness patterns.

Purpose of the Study:

• To formally address missing data within VAEs for biomedical applications.
• To propose a novel VAE architecture, NIMIWAE, capable of handling both ignorable and non-ignorable missing data patterns.
• To facilitate downstream analysis of high-dimensional incomplete biomedical datasets through improved imputation.

Main Methods:

• Developed a new VAE architecture, NIMIWAE, designed to flexibly account for missing data during training.
• Implemented a method to draw samples from the approximate posterior distribution for multiple imputation.
• Validated the approach through statistical simulations and a case study on an Electronic Health Record (EHR) dataset.

Main Results:

• NIMIWAE demonstrates superior performance compared to existing methods in unsupervised learning tasks.
• The proposed method achieves higher imputation accuracy on complex, incomplete datasets.
• Successful application to a large EHR dataset of 12,000 ICU patients with partially observed features.

Conclusions:

• NIMIWAE offers a robust solution for handling missing data in VAEs for biomedical research.
• The method enhances the utility of DL for analyzing complex and incomplete health data.
• This work paves the way for more accurate analysis of real-world biomedical datasets.