Development and validation of potential phenotypes of serum electrolyte disturbances in critically ill patients and a Web-based application

Background: Electrolyte disturbances are highly heterogeneous and severely affect the prognosis of critically ill patients. Our study was to determine whether data-driven phenotypes of seven electrolytes have prognostic relevance in critically ill patients.

Methods: We extracted patient information from three large independent public databases, and clustered the electrolyte distribution of ICU patients based on the extreme value, median value and coefficient of variation of electrolytes. Three plausible clinical phenotypes were calculated using K-means clustering algorithm as the basic clustering method. MIMIC-IV was considered a training set, and two others have been designated as verification set. The robustness of the model was then validated from different angles, providing dynamic and interactive visual charts for more detailed characterization of phenotypes.

Results: 15,340, 12,445 and 2147 ICU patients with electrolyte records during early ICU stay in MIMIC-IV, eICU-CRD and AmsterdamUMCdb were enrolled. After clustering, three reasonable and interpretable phenotypes are defined as α, β and γ according to the order of clusters. The α and γ phenotype, with significant differences in electrolyte distribution and clinical variables, higher 28-day mortality and longer length of ICU stay (p < 0.001), was further demonstrated by robustness analysis. The α phenotype has significant kidney injury, while the β phenotype has the best prognosis. In addition, the assignment methods of the three phenotypes were developed into a web-based tool for further verification and application.

Conclusions: Three different clinical phenotypes were identified that correlated with electrolyte distribution and clinical outcomes. Further validation and characterization of these phenotypes is warranted.