Integrative analysis and imputation of multiple data streams via deep Gaussian processes.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Critical Care Medicine

Background:

• Healthcare data, especially from critical care, presents challenges: related physiological measurements treated independently, irregular sampling times, and prevalent missing values.
• Existing imputation methods often neglect the temporal nature of data and fail to provide uncertainty estimates for predictions.

Purpose of the Study:

• To address the challenges of missing values in critical care data analysis.
• To introduce a novel method for handling missing data that leverages both longitudinal and cross-sectional information.
• To provide uncertainty estimates for imputed values in time-series healthcare data.

Main Methods:

• Deep Gaussian process emulation with stochastic imputation.
• Leveraging longitudinal and cross-sectional data relationships.
• Quantifying uncertainty in imputed values.

Main Results:

• The proposed method outperforms conventional techniques like Multiple Imputations with Chained Equations (MICE), last-known value imputation, and individual Gaussian Processes (GPs).
• Demonstrated superior performance on a clinical dataset.
• Successfully handled missing values while preserving temporal data characteristics and providing uncertainty.

Conclusions:

• Deep Gaussian process emulation with stochastic imputation is a robust method for analyzing critical care data with missing values.
• The approach enhances the reliability of healthcare data analysis by incorporating temporal dynamics and uncertainty quantification.
• The method offers a significant improvement over existing imputation strategies for complex clinical datasets.