Clustering based on renal and inflammatory admission parameters in critically ill patients admitted to the ICU

Abstract

Introduction: The COVID-19 pandemic has been associated with significant variability in acute kidney injury (AKI) incidence, leading to concerns regarding patient heterogeneity. The study's primary objective was a cluster analysis, to identify homogeneous subgroups of patients (clusters) using baseline characteristics, including inflammatory biomarkers. The secondary objectives were the comparisons of MAKE-90 and mortality between the different clusters at three months.

Methods: This retrospective single-center study was conducted in the Medical Intensive Care Unit of the University Hospital of Clermont-Ferrand, France. Baseline data, clinical and biological characteristics on ICU admission, and outcomes at day 90 were recorded. The primary outcome was the risk of major adverse kidney events at 90 days (MAKE-90). Clusters were determined using hierarchical clustering on principal components approach based on admission characteristics, biomarkers and serum values of immune dysfunction and kidney function.

Results: It included consecutive adult patients admitted between March 20, 2020 and February 28, 2021 for severe COVID-19. A total of 149 patients were included in the study. Three clusters were identified of which two were fully described (cluster 3 comprising 2 patients). Cluster 1 comprised 122 patients with fewer organ dysfunctions, moderate immune dysfunction, and was associated with reduced mortality and a lower incidence of MAKE-90. Cluster 2 comprised 25 patients with greater disease severity, immune dysfunction, higher levels of suPAR and L-FABP/U Creat, and greater organ support requirement, incidence of AKI, day-90 mortality and MAKE-90.

Conclusions: This study identified two clusters of severe COVID-19 patients with distinct biological characteristics and renal event risks. Such clusters may help facilitate the identification of targeted populations for future clinical trials. Also, it may help to understand the significant variability in AKI incidence observed in COVID-19 patients.

Fig 1. Hierarchical clustering on principal components using principal component analysis.

A: Graph of variables: projection on the first and second dimensions of the covariates used for the clustering. B: Dendrogram of ascending hierarchical clustering analysis. Dendrogram obtained after application of hierarchical clustering analysis. The vertical axis of the dendrogram represents the distance between clusters. The horizontal vertical axis represents the patients and clusters. Each junction between two clusters is represented on the graph by the split of a vertical line into two vertical lines. The vertical position of the split, shown by the short horizontal bar gives the distance between the two clusters. The red line shows the cut level that determines the number of clusters. C: Representation of the patients into the first two dimensions. Axes correspond to the first and second dimension of MCA. All patients were represented by their individual coordinates in these dimensions. MCA: multiple correspondence analysis, HC: hierarchical clustering, all patients were represented by their individual coordinates in these dimensions.

Fig 2. Prediction to belong to Cluster 1 or Cluster 2 using random forest algorithm.

%Inc MS:, per cent increase in mean squared error, CRP: C reative protein, Il: interleukin,; LFABP: liver fatty acid binding protein, MCA: multiple correspondence analysis, mHLA DR, Monocytic human leukocyte antigen-DR, PCT: procalcitonin, SUPAR: soluble urokinase plasminogen activator receptor, TNF: tumor necrosis factor.

Fig 3. Prediction to belong to Cluster 2 using classification and regression tree analysis.

Tree is built such as 0: should not belong to cluster 2, 1: should belong to cluster 2; and below: left number actual patients in the studied population who did not belong to cluster 2; right number: actual patients in the studied population who belonged to Cluster 2. CRP: C-reactive protein (in mg/L), creat U: urinary creatinine (in mmol/L), Il, interleukin (in pg/L,; LFABP, liver fatty acid binding protein (in ng/mL,; SUPAR, soluble urokinase plasminogen activator receptor (in ng/mL) The binary tree was built in the training set using Breiman methods with Rpart package version 4.1–10, R version 3.1.0. The structure is similar to a real tree, from the bottom up: there is a root, where the first split happens. After each split, two new nodes are created. Each node contains only a subset of the patients. The partitions of the data, which are no longer split, are called terminal nodes or leafs. The second stage of the procedure consists in pruning the tree using cross-validation. Pruning means to shorten the tree, which makes trees more compact and avoids over-fitting to the training data. Each split is examined if it makes a reliable improvement. The six variables used by the binary tree are neutrophils, CRP, IL-6, SUPAR and LFABP/urinary creatinine ratio. The accuracy of the binary tree evaluated in the training dataset is given in Table 4.

Fig 4. Predictors of death and MAKE 90—Random forest.

Contribution of each variable to the risk of death (A) and MAKE 90 (B) %Inc MSE: per cent increase in mean squared error, CV: cardiovascular, Il: interleukin, KDIGO: acute kidney injury classification, LFABP: liver fatty acid binding protein, RAGE: receptor of advanced glycation end product, SUPAR: soluble urokinase plasminogen activator receptor, TNF: tumor necrosis factor, SAPSII: simple acute physiology score II.