Identifying and characterizing high-risk clusters in a heterogeneous ICU population with deep embedded clustering

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Data Science

Background:

• Critically ill patients are a heterogeneous group with varied clinical trajectories and outcomes.
• Existing clustering methods struggle to accurately characterize diverse patient populations in intensive care units (ICUs).
• Predicting mortality and kidney injury risk in ICUs requires nuanced patient stratification.

Purpose of the Study:

• To develop and validate a machine learning methodology for identifying and characterizing patient clusters at high risk of mortality and kidney injury.
• To compare the efficacy of different clustering techniques in stratifying critically ill patients.
• To enhance the understanding of risk factors contributing to adverse outcomes in heterogeneous ICU populations.

Main Methods:

• Analysis of prospectively collected data from 743 ICU patients, including comorbidities, clinical examination, and laboratory parameters.
• Comparison of four clustering methodologies, with a focus on deep embedded clustering.
• Training a classifier to predict cluster membership and utilizing SHapley Additive exPlanations (SHAP) for variable importance analysis.

Main Results:

• Deep embedded clustering outperformed traditional algorithms, identifying 6 clinically recognizable patient clusters.
• Identified two high-risk clusters with significantly increased ICU, 30-day, and 90-day mortality rates.
• Discovered a low-risk cluster exhibiting substantially lower mortality rates and varying incidences of acute kidney injury.

Conclusions:

• The developed machine learning methodology effectively characterizes risk groups within heterogeneous ICU populations.
• This approach offers a potential solution for improving patient stratification and risk assessment in critical care.
• Publicly available methodology facilitates improved understanding and management of critically ill patients.