Clinical phenotype of ARDS based on K-means cluster analysis: A study from the eICU database

Area of Science:

• Critical Care Medicine
• Data Science
• Pulmonology

Background:

• Acute Respiratory Distress Syndrome (ARDS) is a complex critical illness with varied clinical presentations.
• Current classification systems may not fully capture the heterogeneity of ARDS phenotypes.
• Machine learning offers novel approaches to stratify patients based on complex data patterns.

Purpose of the Study:

• To characterize distinct clinical phenotypes of ARDS using machine learning (K-means clustering).
• To compare the prognostic value of identified ARDS phenotypes against the established Berlin classification.
• To analyze phenotypic conversion over time and its impact on patient outcomes.

Main Methods:

• Utilized the eICU database to screen ARDS cases and collect clinical data at diagnosis, Day 1, Day 3, and Day 7.
• Employed K-means cluster analysis to derive distinct ARDS phenotypes, optimizing cluster number with Calinski-Harabasz, Gap Statistic, and Silhouette Coefficient.
• Compared survival outcomes and phenotypic characteristics across identified clusters and the Berlin classification.

Main Results:

• Identified three distinct ARDS phenotypes (Phenotype-I, II, III) from 5054 cases.
• Phenotypes differed significantly in laboratory indicators, vital signs, and severity scores (e.g., APACHE IV).
• Phenotype-I showed the lowest in-hospital mortality (10%), while Phenotype-II had the highest (31.8%), with statistically significant survival differences (P < 0.05).

Conclusions:

• K-means clustering provides a valuable tool for identifying distinct ARDS clinical phenotypes.
• This machine learning-derived classification offers superior prognostic clarity compared to the Berlin standard.
• The new classification aids in predicting patient prognosis and understanding ARDS heterogeneity.