Nonstationary multivariate Gaussian processes for electronic health records

Area of Science:

• Computational statistics
• Health informatics
• Machine learning for healthcare

Background:

• Electronic health records (EHR) offer vast potential for improving patient risk stratification and outcomes.
• Modeling complex temporal dynamics and time-varying correlations within EHR data remains a significant challenge.
• Current methods often lack flexibility or are too complex for clinical interpretation.

Purpose of the Study:

• To propose a novel nonstationary multivariate Gaussian process model for analyzing electronic health records (EHR).
• To address limitations of existing models in capturing time-varying relationships among multiple clinical variables.
• To enhance the interpretability and flexibility of EHR data analysis.

Main Methods:

• Development of a nonstationary multivariate Gaussian process model capable of capturing time-varying scale, correlation, and smoothness.
• Implementation of Maximum a posteriori and Hamilton Monte Carlo inference approaches.
• Validation using synthetic datasets and real-world EHR data from Kaiser Permanente Division of Research (KPDOR).

Main Results:

• The proposed model effectively captures time-varying properties of multiple clinical variables in EHR data.
• Validation on synthetic data confirmed model performance.
• Demonstrated statistically significant correlations between a clinical patient risk metric and the model's latent processes using KPDOR data.

Conclusions:

• The novel nonstationary multivariate Gaussian process model offers a flexible and interpretable approach for EHR analysis.
• The model enhances the ability to understand temporal relationships within patient data.
• Findings suggest potential for improved patient risk stratification and clinical decision-making.