Clustering Based on Laboratory Data in Patients With Heart Failure Admitted to the Intensive Care Unit

ABSTRACT

Background

Heart failure (HF) is a common condition that imposes a significant burden on healthcare systems. We aimed to identify subgroups of patients with heart failure admitted to the ICU using routinely measured laboratory biomarkers.

Methods

A large dataset (N = 1176) of patients with heart failure admitted to the ICU at the Beth Israel Deaconess Medical Center in Boston, USA, between June 1, 2001, and October 31, 2012, was analyzed. We clustered patients to identify laboratory phenotypes. Cluster profiling was then performed to characterize each cluster, using a binary logistic model.

Results

Two distinct clusters of patients were identified (N = 679 and 497). There was a significant difference in the mortality rate between Clusters 1 and 2 (50 [7.4%] vs. 109 [21.9%], respectively, p < 0.001). Patients in the Cluster 2 were significantly older (mean [SD] age = 72.35 [14.40] and 76.37 [11.61] years, p < 0.001) with a higher percentage of chronic kidney disease (167 [24.6%] vs. 262 [52.7%], respectively, p < 0.001). The logistic model was significant (Log‐likelihood ratio p < 0.001, pseudo R ² = 0.746) with an area under the curve of 0.905. The odds ratio for leucocyte count, mean corpuscular volume (MCV), red blood cell (RBC) distribution width, hematocrit (HcT), lactic acid, blood urea nitrogen (BUN), serum potassium, magnesium, and sodium were significant (all p < 0.05).

Conclusion

Laboratory data revealed two phenotypes of ICU‐admitted patients with heart failure. The two phenotypes are of prognostic importance in terms of mortality rate. They can be differentiated using blood cell count, kidney function status, and serum electrolyte concentrations.

Heart failure is a common condition that imposes a significant burden on healthcare systems. We aimed to identify subgroups of patients with heart failure admitted to the ICU using routinely measured laboratory biomarkers. Laboratory data revealed two phenotypes of ICU‐admitted patients with heart failure. The two phenotypes are of prognostic importance in term of mortality rate. They can be differentiated using blood cell count, kidney function status, and serum electrolytes concentrations.

Summary.

• Two clusters of ICU‐admitted HF patients were identified based on routine lab data.

• The two clusters are different in mortality rate.

• The high‐risk cluster was older with a higher rate of kidney dysfunction.

• The clusters were not different concerning sex ratio.

• They are different in leucocyte count, MCH, RDW, Hct, and various electrolytes.

1. Introduction

Heart failure (HF) is a common and disabling cardiovascular condition that affects more than 64 million people worldwide and imposes a significant burden on healthcare systems [1]. Despite progress in treatment options, the prevalence of HF is increasing due to the aging population and improved treatment [1]. Laboratory investigations are essential for evaluating and managing HF. They offer valuable insights into the underlying pathophysiology, aid in diagnosis, risk assessment, monitoring disease progression, and guiding treatment decisions [2]. Research studies have been conducted to identify key biomarkers implicated in disease pathophysiology and prognosis, such as neurohormonal activation, inflammation, oxidative stress, and organ dysfunction [3, 4]. Overall, laboratory studies suggest complex relationships between biochemical imbalances and clinical outcomes in HF, underscoring the need for comprehensive laboratory evaluations in HF patients [5, 6].

Despite significant advancements in cardiovascular research, there are still knowledge gaps in our understanding of HF, particularly when comprehensively characterizing patients based on their laboratory profiles in intensive care units (ICU) [5, 7]. Diversity in biomarker assays and lack of standardization may be addressed by integrating data from multiple biomarkers into decision‐making [5, 8]. Existing literature often offers fragmented insights into the laboratory aspects of HF, with limited studies focusing on the laboratory profiles of HF patients admitted to the ICU [8, 9]. Moreover, HF is a heterogeneous syndrome and comprehensive analyses are required to differentiate specific subgroups of patients with HF and clarify their prognostic implications [6, 9, 10, 11]. Novel biomarkers of inflammation, oxidative stress, and organ dysfunction show promise, but their role in comprehensively phenotyping ICU patients with HF is unclear [12, 13, 14].

Overall, there is still a need to identify subgroups of patients with HF admitted to the ICU using routinely measured biomarkers. We aimed to investigate the inherent heterogeneity of patients in the ICU, specifically focusing on laboratory characteristics related to the risk of mortality. We performed a cluster analysis of laboratory data from ICU‐admitted patients with HF. This promotes the development of personalized management strategies in the ICU [9, 10]. We hypothesized that there would be distinct patterns of laboratory findings in HF patients admitted to the ICU.

2. Patients and Methods

2.1. Data Source

Li et al. conducted a study to create and validate a predictive model for in‐hospital mortality among patients with HF admitted to the ICU. They used data from the Medical Information Mart for Intensive Care (MIMIC‐III) database, a publicly accessible critical care database [15]. This database contains anonymized information from patients admitted to the ICU at the Beth Israel Deaconess Medical Center in Boston, USA, between June 1, 2001 and October 31, 2012. For their research, Li et al. [16] included a total of 1,177 patients who were diagnosed with HF (identified through ICD‐9 codes) and were at least 15 years old at the time of admission. Using machine learning models, they developed a nomogram to predict in‐hospital mortality for HF patients admitted to the ICU. Additionally, the authors made their data publicly available as supporting information in their publication, which is stored in a public data repository [17]. To gain further insight into the laboratory data of HF patients admitted to the ICU, we conducted a secondary analysis of this dataset, aiming to uncover any underlying patterns.

2.2. Ethical Considerations

Our study adhered to the principles of the Declaration of Helsinki since we did not directly measure participants' characteristics. Instead, we conducted a secondary analysis using open‐access anonymized data. We credited the primary authors and did not republish raw data. Only the results of our analysis are reported in this article.

2.3. Candidate Variables

Throughout the ICU stay, laboratory test results were collected, and demographic and clinical characteristics were recorded within the first 24 h of hospital admission. The laboratory variables included hematological indices like hematocrit (Hct), red blood cell (RBC) count, mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), mean corpuscular volume (MCV), red blood cell distribution width (RDW), and platelet (Plt) count, white blood cell (WBC) count along with differentials such as neutrophils, basophils, and lymphocytes, coagulation factors like prothrombin time (PT) and international normalized ratio (INR), as well as cardiac biomarkers like N‐terminal pro‐b‐type natriuretic peptide (NT‐proBNP) and renal function markers such as creatinine (Cr) and blood urea nitrogen (BUN). Additionally, we included serum glucose (sGlc), potassium (sK), sodium (sNa), calcium (sCa), chloride (sCl), and magnesium (sMg). We also analyzed acid–base balance indicators like anion gap (AG), bicarbonate, lactate, hydrogen ion concentration (pH), and partial pressure of CO₂ in arterial blood (PCO₂). Other variables were age, sex, saturation pulse oxygen (SPO₂), and urine output (UO) within the first 24 h. We also analyzed data on comorbidities including hypertension (HTN), diabetes mellitus (DM), depression, hyperlipidemia, chronic kidney disease (CKD), and chronic obstructive pulmonary disease (COPD), based on the International Classification of Diseases, Ninth Revision (ICD‐9) codes. The dataset included the mean value for variables with multiple measurements. The vital status at hospital discharge was recorded for survivors and deceased patients. All primary study screening variables had less than 25% missing values.

2.4. Data Analysis

We analyzed the laboratory data obtained from patients with HF admitted to the ICU to identify inherent patterns and groupings. Initially, the dataset was investigated for duplicate entries. Pairwise correlations among the recorded laboratory variables were computed. Variables exhibiting absolute correlations ≥ 0.5 were identified. Then, one variable from each pair was considered for further analysis. Subsequently, missing values were identified and filled using random forest imputation to capture complex interactions between variables and provide robustness to outliers and nonlinear relationships. Following preprocessing, data were standardized to ensure uniformity across variables, regardless of the units of measurement. Then, we used a clustering algorithm to analyze data. Clustering is a machine‐learning technique that organizes similar, unlabeled patients into groups [18]. The clustering algorithm used to group patients based on their laboratory profiles was partitioning around medoids. The partitioning around medoids algorithm offers several advantages including robustness to outliers, flexibility in distance measures (including non‐Euclidean distances), better handling of unevenly sized clusters, and improved performance with high‐dimensional data. Particularly, partitioning around medoids does not make any assumptions about Gaussian distribution. This helped us in managing a large number of laboratory characteristics. Of course, it should be noted that the need to specify the number of clusters beforehand is a limitation of partitioning around medoids. To determine the optimal number of clusters, we used the NbClust R package, which integrates a consensus‐based approach to identify the most appropriate number of clusters. This package evaluates 30 different indices, such as the Elbow method, the Silhouette method, and the Gap Statistic, each providing insights into cluster validity. The Elbow method identifies where the within‐cluster sum of squares begins to level off, suggesting an optimal number of clusters. The Silhouette method measures how similar an object is to its cluster compared to other clusters, with higher values indicating better‐defined clusters. The Gap Statistic compares the log of within‐cluster dispersion for different numbers of clusters to that expected under null reference distribution, thus highlighting significant deviations from randomness. Additionally, NbClust performs stability analysis and aggregation of these indices to offer a recommendation. The final number of clusters was selected based on the consensus of these indices and visual inspection of the clustering results, ensuring that the clusters were both interpretable and meaningful. We used Gower distance to measure dissimilarity between data points providing robustness to outliers and accommodating mixed‐type data. Evaluation metrics, including silhouette score, Davies‐Bouldin index, and Calinski‐Harabasz index were computed to provide insight into the robustness and separation of the identified patient subgroups. The silhouette score quantifies the degree of similarity within clusters versus between clusters, the Davies‐Bouldin index measures the average similarity ratio of each cluster with its most similar cluster, and the Calinski‐Harabasz index evaluates the ratio of between‐cluster dispersion to within‐cluster dispersion. For visualizing the result of clustering, a scatter plot of two principal components was used. Cluster profiling was then performed to characterize each cluster, using a binary logistic model. The performance of the logistic regression was evaluated using the pseudo‐R ² value, which measures the proportion of variance in the outcome variable explained by the model relative to a null model. A higher pseudo‐R ² indicates better model fit and explanatory power. The level of significance was set at two‐tailed α = 0.05. Data analysis and visualization were performed using Python version 3.11.5 and R version 4.3.1 for Windows, two free environments known for their statistical, machine learning, and graphics libraries.

3. Results

3.1. Study Sample

In total, 1177 patients were included in the dataset without any duplicate rows. We excluded one patient because of a missing value in the mortality variable. Of the remaining 1176 individuals, 558 (47.4%) were female. The mean (SD) age was 75.0 (13.4) years. The recorded laboratory data were analyzed to cluster patients. We calculated the pair‐wise correlation of the study variables, excluding any missing values. Figure 1 shows the heatmap of the correlations between variables. For each pair with an absolute correlation value ≥ 0.5, only one variable was included in the subsequent analysis. Ease of measurement, commonality, and higher variance were taken into consideration when selecting each variable to be included. There were large correlations between Hct and RBC; MCH and MCHC or MCV; neutrophil and lymphocyte counts; PT and INH; Cr and BUN or AG; BUN and AG; sNa and sCl; and Bicarbonate and sCl or Ag or PCO₂ (all correlations > 0.50, all p < 0.001). We excluded RBC, MCHC, MCV, INR, Cr, AG, sCl, Bicarbonate, and lymphocyte count from further analysis.

FIGURE 1.

The heatmap of pairwise correlations between study variables was used for clustering patients, with correlation values represented by a gradient of colors to visually indicate the strength and direction of relationships. AG, anion gap; BUN, blood urea nitrogen; CK, creatine kinase; Cr, creatinine; Hct, hematocrit; INR, international normalized ratio; MCH, mean corpuscular hemoglobin; MCHC, mean corpuscular hemoglobin concentration; MCV, mean corpuscular volume; NT ptoBNP, N‐terminal pro B‐type natriuretic peptide; PCO₂, partial pressure of CO₂; pH, hydrogen ion concentration; Plt, platelet count; PT, prothrombin time; RBC, red blood cell; RDW, red blood cell distribution width; sCa, serum calcium; sCl, serum chloride; sGlc, serum glucose; sK, serum potassium; sMg, serum magnesium; sNa, serum sodium; SPO₂, saturation pulse oxygen; UO, urine output.

3.2. Clustering

Figure 2 displays the missing data patterns for the laboratory characteristics that were used for cluster analysis. Only 4.28% of the whole data was missing. To address these missing values, we employed random forest imputation using all the variables in the dataset. Subsequently, the complete laboratory data were scaled to have a mean of 0 and a variance of 1. Next, the best number of clusters was investigated using the NbClust R package. Figure 3 shows that the optimal number of suggested clusters was two. We clustered patients into two subgroups (N = 679 and 497) using partitioning around medoids. Figure 4 illustrates the result of partitioning patients (upper panel) and their mortality (lower panel). Comparing the two panels suggests that patients in cluster two are at a higher risk of mortality. The silhouette score, Davies–Bouldin index, and Calinski–Harabasz index were 0.12, 3.24, and 90.44, respectively. The silhouette score and Davies–Bouldin index indicate a moderate separation between clusters, with some overlap. However, the Calinski–Harabasz index suggests a relatively high ratio of between‐cluster variance to within‐cluster variance, indicating appropriate clustering.

FIGURE 2.

Patterns of missing data in laboratory variables. BUN, blood urea nitrogen; CK, creatine kinase; MCH, mean corpuscular hemoglobin; NT ptoBNP, N‐terminal pro B‐type natriuretic peptide; PCO₂, partial pressure of CO₂; pH, hydrogen ion concentration; Plt, platelet count; PT, prothrombin time; RDW, red blood cell distribution width; sCa, serum calcium; sGlc, serum glucose; sK, serum potassium; sMg, serum magnesium; sNa, serum sodium; SPO₂, saturation pulse oxygen; UO, urine output.

FIGURE 3.

The optimal number of clusters in laboratory data was determined by conducting a majority vote among 30 indices. The majority consensus identified two clusters as the most appropriate grouping.

FIGURE 4.

Scatter diagrams showing the distribution of patients based on two key principal components (PC1 and PC2) derived from laboratory data. The upper panel displays two distinct groups of patients, while the lower panel differentiates between patients who survived and those who did not. The diagrams illustrate how patients are clustered and how their outcomes vary based on these principal components.

3.3. Cluster Profiling

3.3.1. Mortality, Age, Sex, and Comorbidities

There was a significant difference in the risk of mortality between Clusters 1 and 2 (50 [7.4%] vs. 109 [21.9%], respectively, χ ² [1] = 50.85, p < 0.001). Patients in Cluster 2 were at a greater risk for mortality (high‐risk cluster). Overall, patients in cluster 2 were significantly older (mean [SD] age = 72.35 [14.40] and 76.37 [11.61] years, respectively, mean difference = −4.03, t [1174.0] = −5.13, p < 0.001, Cohen's d = −0.3). While more female patients belonged to the high‐risk cluster, there was no statistically significant difference in the sex ratio between the two groups (females 313 [46.1%] vs. 245 [49.3%], respectively, χ ² [1] = 1.050, p = 0.305). Our investigations also showed that there was no significant difference between the two subgroups in hypertension (p = 0.980), diabetes mellitus (p = 0.607), hyperlipidemia (p = 0.296), and COPD (p = 0.999). But, the clusters were significantly different in the percentage of patients with CKD (167 [24.6%] vs. 262 [52.7%], respectively, χ ² [1] = 96.720, p < 0.001).

3.3.2. Laboratory Characteristics

The dataset was divided into training (70%) and test sets. To demonstrate the discriminatory ability of the clusters in practice, we developed a binary logistic model using the train set. Table 1 presents the outcomes of the logistic model, facilitating a comparison between the two clusters based on laboratory characteristics. The model was significant (Log‐likelihood ratio p < 0.001, pseudo R ² = 0.746) and calibrated (calibration coefficient = 0.890, Brier Score = 0.098) with an area under the ROC curve = 0.905 (accuracy = 0.881, sensitivity = 0.853, specificity = 0.901, positive likelihood ratio = 8.616) based on the test set. Of the statistically significant characteristics, the odds ratio for leucocyte count, MCH, RDW, Hct, sK, sMg, sNa, lactic acid, and BUN were practically important.

TABLE 1.

Laboratory profiling of Clusters 1 (N = 679) and 2 (N = 497).

Characteristic	Cluster		Logistic model			Adjusted odds ratio
	1	2	Coeff.	SE	p	OR	Lower 95% CI	Upper 95% CI
Constant			−21.492	39.093	0.582	—	—	—
Plt	251.53 (111.92)	227.84 (113.54)	−0.004	0.002	0.031*	0.996	0.993	1.000
NT proBNP	5066.39 (6256.5)	19132.6 (15553.35)	< 0.001	< 0.001	< 0.001*	1.000	1.000	1.000
sGlc	147.01 (51.03)	151.27 (52.08)	−0.002	0.005	0.771	0.998	0.988	1.009
sCa	8.52 (0.54)	8.47 (0.61)	0.170	0.493	0.730	1.186	0.451	3.116
Leuko.	9.74 (3.92)	12.05 (6.38)	0.207	0.052	< 0.001*	1.230	1.112	1.361
MCH	29.23 (2.52)	29.96 (2.7)	0.565	0.078	< 0.001*	1.759	1.509	2.051
PCO2	46.98 (12.92)	43.51 (12.14)	−0.032	0.024	0.181	0.969	0.925	1.015
RDW	15.49 (1.97)	16.59 (2.18)	0.489	0.092	< 0.001*	1.631	1.362	1.954
Hct	32.51 (5.58)	31.09 (4.52)	0.078	0.037	0.034*	1.081	1.006	1.161
UO	2209.84 (1353.09)	1457.5 (993.79)	−0.002	< 0.001	< 0.001*	0.998	0.998	0.999
SPO2	96.14 (2.19)	96.46 (2.43)	0.112	0.123	0.362	1.119	0.879	1.425
PT	16.54 (5.92)	18.77 (8.85)	0.054	0.049	0.272	1.055	0.959	1.161
pH	7.39 (0.06)	7.36 (0.07)	−7.706	4.786	0.107	0.000	0.000	5.334
Baso.	0.44 (0.56)	0.35 (0.29)	1.393	1.150	0.226	4.028	0.423	38.382
sK	4.08 (0.37)	4.30 (0.44)	1.266	0.440	0.004*	3.547	1.497	8.405
sMg	2.06 (0.22)	2.2 (0.27)	1.553	0.656	0.018*	4.725	1.306	17.085
sNa	138.83 (3.77)	138.99 (4.61)	0.141	0.044	0.001*	1.151	1.056	1.254
Lactic acid	1.75 (0.9)	2.0 (1.08)	1.029	0.379	0.007*	2.800	1.332	5.886
BUN	25.29 (12.27)	51.32 (23.12)	0.145	0.016	< 0.001*	1.156	1.119	1.194
CK	201.43 (754.9)	304.83 (2070.48)	< 0.001	< 0.001	0.979	1.000	1.000	1.000
Neutro.	78.57 (11.53)	82.13 (10.27)	0.052	0.037	0.165	1.053	0.979	1.132

Abbreviations: BUN, blood urea nitrogen; CK, creatine kinase; Hct, hematocrit; MCH, mean corpuscular hemoglobin; NT ptoBNP, N‐terminal pro B‐type natriuretic peptide; PCO₂, partial pressure of CO₂; pH, hydrogen ion concentration; Plt, platelet count; PT, prothrombin time; RDW, red blood cell distribution width; sCa, serum calcium; sGlc, serum glucose; sK, serum potassium; sMg, serum magnesium; sNa, serum sodium; SPO₂, saturation pulse oxygen; UO, urine output.

^*Significant at p < 0.05.

4. Discussion

Patients with HF who are admitted to the ICU represent a cohort of critically ill individuals who require intensive monitoring and management. To the best of our knowledge, no previous study has been reported in the literature on the phenotyping of patients with HF admitted to the ICU. We conducted a cluster analysis to uncover inherent patterns and associations in laboratory data from ICU‐admitted patients with HF. We used partitioning around medoids to address challenges inherent in clustering analyses, such as handling unevenly sized clusters, robustness to outliers, and flexibility in distance measures. This allowed us to identify patient subgroups based on their laboratory profiles. Our findings showed the presence of two distinct clusters of patients with HF, each displaying laboratory characteristics. Patients in Cluster 2 were found to be at an increased risk of mortality. The discriminatory ability of the identified clusters was further validated by developing a binary logistic model. This model showed significant predictive power in distinguishing between patient subgroups based on their laboratory characteristics. Significantly, laboratory parameters such as leucocyte count, MCH, RDW, Hct, and various electrolytes were found to be key determinants driving differentiation between clusters. These highlight the clinical relevance of laboratory characteristics as prognostic markers in patients with HF and underscores the significance of categorizing patients with HF according to their laboratory profiles.

4.1. Implications

4.1.1. Mortality, Age, Sex, and Comorbidities

An important finding in our study was a significantly higher mortality rate observed in Cluster 2. This underscores the need for identifying phenotypes based on laboratory profiles, allowing for early risk stratification and targeted interventions. We found that patients in the high‐risk cluster had a higher average age. This finding is plausible and consistent with previous research that has identified advanced age as a significant risk factor for HF and increased mortality risk [19]. Furthermore, it is consistent with the global trend of an aging population and the growing burden of HF among older adults [1]. Although not statistically significant, we also observed a slight predominance of females in our study, which is in agreement with epidemiological data indicating a higher incidence of HF among women, especially in older age groups [20]. The prevalence of CKD was significantly higher in Cluster 2, which is consistent with the relation between HF and kidney dysfunction contributing to the discrepancy in laboratory findings between the two clusters [21]. Overall, these findings showed the clinical plausibility and accuracy of our cluster analysis, regardless of the high performance of the logistic model.

4.1.2. Laboratory Biomarkers

Our study showed that leucocyte count was significantly elevated in Cluster 2, which was associated with a higher mortality risk. This can be explained as the connection between an elevated leucocyte count, heightened systemic inflammation, and a worsened prognosis in patients with HF [22]. Inflammation plays a crucial role in the development and progression of HF, contributing to myocardial remodeling, neurohormonal activation, and endothelial dysfunction. Inflammatory markers, such as the C‐reactive protein‐to‐albumin ratio and the systemic immune‐inflammation index, have been suggested to have significant prognostic value in patients with HF, particularly in those with reduced ejection fraction and implantable cardiac defibrillators. These markers further highlight the role of inflammation in HF progression, supporting the integration of immune‐related biomarkers in risk stratification and management strategies. Elevated leucocyte count has been shown to reflect underlying inflammatory processes and is independently associated with adverse outcomes, including the exacerbation of HF, hospitalization, and mortality [23]. Further characterization of the immune profile, both systemic and cardiac, can contribute to the advancement of precision medicine [22]. Our analysis revealed that Cluster 2 had higher MCH and RDW levels, suggesting changes in the shape and hemoglobin content of RBCs. Elevated RDW has been suggested as a biomarker of adverse outcomes in HF, independent of traditional markers such as hemoglobin levels [24]. Increased RDW has been associated with impaired production of RBCs, chronic inflammation, and oxidative stress in HF patients. It serves as a predictor of mortality, hospitalizations, and disease progression [24]. Due to the large correlations between MCH and MCHC or MCV, we excluded MCHC and MCV to prevent multi‐collinearity in the development of the study model. In our study, a higher mean MCH has been associated with the high‐risk cluster (Cluster 2). We did not find specific evidence in the literature regarding the association of increased MCH with a higher risk of poor outcomes in patients with HF. However, some studies have suggested a potential association between high MCV or MCHC and poor outcomes in cardiovascular disease and CKD [25], although there are controversies surrounding this relationship [26]. A U‐ or J‐shaped association between hemoglobin concentrations and cardiovascular and all‐cause mortality has been reported, even after adjusting for cardiovascular risk factors [27]. Lower Hct and plt count, along with a higher MCH in Cluster 2, may suggest a nutritional deficiency of folate, 25‐hydroxy‐vitamin D [28], and vitamin B12 in the high‐risk cluster. However, since there is a lack of data on nutritional status, we cannot substantiate this assertion. Further studies are needed to investigate the nature and exact mechanisms underlying this association. Commonly, hyponatremia, hyperkalemia, and hypomagnesemia have been reported in HF [29]; however, we are not aware of direct evidence about higher levels of sNa, sK, and sMg in a high‐risk cluster or subgroup of ICU‐admitted patients with HF. However, the higher levels of sK, sMg, sNa, lactic acid, and BUN observed in Cluster 2 patients may result from multifactorial physiological perturbations inherent in the pathophysiology of HF. The elevated sK and sMg levels indicate impaired renal excretion due to decreased renal perfusion and neurohormonal activation leading to dysregulation of aldosterone and antidiuretic hormone. Concurrently, the higher sNa may reflect the activation of compensatory mechanisms, such as the renin‐angiotensin‐aldosterone system, in response to volume depletion or neurohormonal derangements. The elevation in lactic acid levels suggests tissue hypoperfusion and anaerobic metabolism, which are seen in advanced HF states characterized by reduced cardiac output and tissue perfusion [30]. Furthermore, the elevated BUN levels indicate impaired renal function, likely secondary to reduced renal perfusion and neurohormonal dysregulation [21, 31]. These findings collectively highlight the complex interactions between cardiac, renal, and metabolic abnormalities in patients with HF, contributing to the observed laboratory abnormalities in Cluster 2. The observed differences in laboratory profiles between clusters further emphasize the heterogeneity of the HF populations and the importance of personalized risk assessment approaches that integrate multiple biomarkers to guide clinical decision‐making. Our findings further support existing research advocating for the use of multiple biomarkers in the prognostication and management of cardiovascular disease, underscoring the potential for personalized risk assessment strategies. For instance, the Intermountain Risk Score has been developed as a robust tool for predicting short‐ and long‐term mortality in patients with acute coronary syndrome, particularly in ST‐elevation myocardial infarction and cardiogenic shock [32, 33]. This score, similar to our clustering approach, integrates multiple laboratory parameters, including those evaluated in our study, to enhance risk stratification.

Recent studies have advanced our understanding of clustering methods in patients with HF admitted to the ICU, highlighting the use of traditional and novel analytical approaches. For instance, Huang and Huang [34] studied 1176 ICU patients with HF and demonstrated the application of hierarchical clustering combined with Shapely Additive explanations (SHAP) to elucidate predictors of in‐hospital mortality in HF patients, emphasizing the importance of urine output, leukocytes, and bicarbonate levels in differentiating mortality risk. Meng et al. [35] expanded on this by addressing the interpretability and fairness of deep learning models, revealing that while high prediction performance is achievable, there is a critical need to consider model fairness and potential biases related to demographic features. In another study, Huang and Huang further highlighted the utility of bootstrap simulation and SHAP in enhancing model transparency and selection, noting significant variability in model metrics that can influence model choice [36]. Huang and Huang [37] also compared analytically derived distributions and simulation‐based methods for evaluating model accuracy, finding that XGBoost performed optimally, with bootstrap simulation providing more detailed insights into model metric variability. Collectively, these studies underline the importance of integrating transparent, fair, and robust evaluation methods to improve the predictive accuracy and clinical relevance of clustering models in HF management.

4.2. Limitations and Strengths

The generalizability of our results may be influenced by the characteristics of the study population, which are mainly derived from a single‐center database. The dataset spans over a decade, which provides a broad perspective on clinical practices and laboratory techniques. However, changes in these factors over time may have influenced the results, highlighting the importance of considering temporal variations in future studies to enhance the robustness and applicability of the findings. To validate our findings and improve external validity, it is important to conduct future studies that incorporate multi‐center data with prospective designs. One limitation of our study was the absence of albumin measurements, which are used in the prognostication of HF and could have provided additional insights into patient risk stratification. Another limitation of this study was the lack of treatment data, which limited the ability to adjust for treatment effects and understand their influence on cluster membership. Addressing this limitation would require recruiting a very large number of patients. Meanwhile, one strength of the study was that the results were based on objective laboratory assessments. This was evident in the calculation of coefficients and odds ratios, as well as in the significance and high performance of the logistic model we developed. Such objective assessments plausibly demonstrate minimal variability among patients of different nationalities. Another strength of this study was the relatively large sample size, which decreased the variance in the conclusions made.

Researchers have explored phenotypic differences in HF, but they have largely focused on ejection fraction‐based classifications or genetic markers [38]. Our approach, which integrates laboratory data, provides a more granular perspective on how biomarkers influence mortality risk, particularly in ICU settings where clinical decision‐making is based on rapidly changing physiological parameters. Our findings regarding inflammation and electrolyte imbalances expand on existing research, which has highlighted the importance of these markers in ICU‐admitted patients with HF [39]. The identification of two distinct phenotypes among ICU‐admitted HF patients has direct clinical implications. Patients in Cluster 2, with higher mortality rates, elevated leucocyte counts, and kidney dysfunction, would likely benefit from early aggressive intervention, including closer monitoring of renal function and perhaps therapies with anti‐inflammatory properties [40]. Additionally, the observed abnormalities in electrolyte levels in this cluster suggest a need for electrolyte rebalancing and consideration of aldosterone antagonists, which have been shown to improve outcomes in HF patients with renal dysfunction [41, 42]. Given the increased inflammatory markers in Cluster 2, the integration of immune‐modulating therapies may also be beneficial. This aligns with recommendations for utilizing biomarkers like C‐reactive protein to guide inflammation‐targeted therapies in high‐risk HF subgroups [43].

5. Conclusion

In conclusion, our study demonstrated the clinical use of clustering laboratory data to better understand the different phenotypic profiles of patients with HF in the ICU. We provided insights into patient heterogeneity and prognosis and identified two distinct clusters of HF patients admitted to the ICU, each with specific laboratory characteristics and clinical profiles. Of the two identified clusters, one had a higher mortality rate, consisting of older patients with kidney dysfunction and specific blood count and electrolyte features. This emphasized the importance of categorizing HF patients based on their laboratory profiles for prognostic purposes and the potential for implementing personalized management strategies.

Author Contributions

S.N. contributed to the study planning and primary interpretation of the results and carried out the literature review. B.M. developed the research design, analyzed data, and contributed to the primary interpretation of the results. Z.H. contributed to the study idea and helped with the clinical interpretation of the findings. All the authors participated in drafting and its final approval. They critically reviewed and approved the final version of this manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The raw data were licensed under a Universal (CC0 1.0) Public Domain Dedication license [17]. The dataset is openly available in the DRYAD public data repository, accessible via the digital object identifier (DOI) 10.5061/dryad.0p2ngf1zd. This dataset was published on July 5, 2021, and was last accessed on May 2, 2024.