Learning and visualizing chronic latent representations using electronic health records.

Area of Science:

• * Computational Health Informatics
• * Machine Learning in Medicine

Background:

• * Rising global prevalence of chronic diseases like diabetes and hypertension presents significant healthcare challenges.
• * Electronic Health Records (EHRs) offer vast data but are complex, hindering traditional analysis.
• * Need for advanced methods to extract meaningful clinical knowledge from noisy, high-dimensional EHR data.

Purpose of the Study:

• * To apply Machine Learning (ML), specifically Denoising Autoencoders (DAE), for extracting knowledge from complex EHR data.
• * To develop a visualization method using latent representations (LRs) for mapping patient data in a 2D space.
• * To characterize patient health status evolution for clinical decision support.

Main Methods:

• * Utilized Denoising Autoencoders (DAE) to transform high-dimensional EHR data into lower-dimensional latent representations (LRs).
• * Combined LRs with a visualization technique to map patient data onto a two-dimensional space.
• * Analyzed real-world EHR data from the University Hospital of Fuenlabrada, Spain.

Main Results:

• * DAEs successfully identified distinct clinical patterns associated with hypertension, diabetes, and multimorbidity.
• * Differentiated patient subgroups, such as insulin-dependent vs. non-insulin-dependent diabetics, and identified a group of women with gestational diabetes and hypertension.
• * Demonstrated a proof of concept for visualizing synthetic patient health status evolution.

Conclusions:

• * ML techniques, particularly DAEs, are valuable for extracting clinical knowledge from EHRs.
• * The 2D patient representation aids clinicians in characterizing health conditions and identifying key clinical codes.
• * Visualizing patient health progression can enhance clinical understanding and decision-making.