Discovering patient groups in sequential electronic healthcare data using unsupervised representation learning.

Area of Science:

• Computational biology
• Medical informatics

Background:

• Natural Language Processing (NLP) inspired unsupervised feature learning methods can derive patient-specific features from longitudinal Electronic Health Records (EHR).
• These methods offer a novel approach to analyzing complex patient data.

Purpose of the Study:

• To apply document embedding algorithms to Paediatric Intensive Care Unit (PICU) EHR data.
• To extract patient-specific features from 1853 patients' ICU journeys.
• To evaluate the clinical utility of these features using K-means clustering.

Main Methods:

• Document embedding algorithms were applied to EHR data comprising lab tests and medication events.
• A document embedding model was trained and used to generate latent patient feature vectors.
• K-means clustering was performed on patient vectors for downstream analysis.

Main Results:

• Latent patient feature vectors were successfully generated for all 1853 patients.
• Unsupervised clustering identified 5 distinct patient clusters.
• Significant variations (p<0.0001) in patient characteristics, surgical interventions, and diagnostic profiles were observed across clusters.

Conclusions:

• K-means clustering confirmed the clinical utility of patient-specific features derived from embedding algorithms.
• The learned latent patient features are directly applicable to other machine learning algorithms.
• Future research will incorporate temporal EHR information and extend embedding algorithms for personalized patient journey predictions.