This study is a secondary analysis of cross-sectional data derived from a large, intensive care cohort. We used the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 3.1) database, a publicly accessible critical care dataset developed by the Massachusetts Institute of Technology Laboratory for Computational Physiology in collaboration with the Beth Israel Deaconess Medical Center in Boston, Massachusetts (https://mimic.mit.edu).19 MIMIC-IV provides de-identified, high-resolution clinical data for patients admitted to intensive care units between 2008 and 2022, and reflects real-world ICU practice within a tertiary care setting. The database includes comprehensive records, comprising patient demographics, vital signs, laboratory measurements, comorbidities, administered therapies, and outcome data. For this analysis, we selected a cohort of adult patients (≥18 years) who experienced their first ICU admission, with a minimum ICU stay of 48h, between 2017 and 2022. This temporal window was selected to ensure consistency with updated diagnostic and therapeutic practices.

The analytical dataset included baseline demographic variables, pre-existing comorbid conditions, and organ system dysfunction through a broad spectrum of laboratory measurements and physiologic markers representative of cardiovascular, respiratory, renal, hepatic, and hematologic function. Additionally, data on medication exposure – specifically the administration of vasoactive agents and diuretics – were incorporated to reflect treatment intentions and hemodynamic support. The predictor variables were extracted from the first 24h following ICU admission to represent the baseline clinical values. Due to missing and inconsistent urine output documentation across patient records, the urine output criteria for AKI classification as defined by guidelines were not applied.20 Consequently, AKI stage was determined based on serum creatinine (Cr) measurements.20 Stage 0 (no AKI) included those with no Cr increase meeting AKI criteria and no AKI-related international classification of diseases (ICD) diagnostic codes. This dual criterion was applied to minimize misclassification and ensure a true non-AKI reference group. Patients with an AKI ICD code but no creatinine rise were excluded from Stage 0. Stage 1 was defined by an increase in Cr to 1.5–1.9 times the baseline or an absolute rise of ≥0.3mg/dL within 48h. Stage 2 involved an increase in Cr to 2.0–2.9 times the baseline, while Stage 3 was characterized by an increase in Cr to ≥3.0 times baseline or an absolute Cr of ≥4.0mg/dL. Baseline serum creatinine was defined as the first available measurement at ICU admission. Pre-existing chronic kidney disease was identified based on ICD diagnostic codes. To enhance cohort validity and decrease bias introduced by incomplete data, only patients with less than 25% missing data across study variables were included. No additional patients were excluded from analysis. This strategy ensured data robustness while preserving clinical heterogeneity. The final analytical sample included patients across all four AKI severity strata (Stages 0–3). To ensure balanced representation and facilitate unbiased phenotypic discovery across the full spectrum of AKI severity (total n=2281), we used an equal sampling strategy by selecting an identical number of patients from each AKI stage subgroup (AKI stages 0–3) for clustering analysis. Given that the smallest subgroup, AKI Stage 2, comprised 343 patients, we randomly selected 343 patients from each AKI stage, resulting in a total analytical cohort of 1372 patients.

The raw, de-identified data used in this study are publicly accessible through the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, available at https://mimic.mit.edu.19 MIMIC-IV is an open-access, relational database constructed from the electronic health records of patients admitted to the Beth Israel Deaconess Medical Center in Boston, Massachusetts. Access to the MIMIC-IV database requires completion of appropriate training and approval under a data use agreement to ensure responsible and ethical utilization of sensitive health information.

This study was conducted using data from the publicly available MIMIC-IV database, which is fully anonymized and did not involve direct interaction with human subjects, the collection of personal information, or any procedures requiring informed consent. The analysis was performed following ethical standards governing the use of publicly available clinical data. The use of the MIMIC-IV database requires completion of ethics training, ensuring responsible handling of sensitive health information. No raw data were republished or redistributed in this study, and all reported findings are based solely on aggregated analyses of the data. Our study followed the ethical guidelines specified in the Declaration of Helsinki and received ethics approval from the institutional review board of our university with the reference number of IR.SBMU.RETECH.REC.1403.896. In order to acquire access to data, we successfully completed the CITI “Data or Specimens Only Research” course (record ID: 53106331) and signed PhysioNet credentialed health data use agreement 1.5.0 on March 17, 2025.

The initial phase of analysis involved rigorous data preprocessing to ensure completeness and reliability. Variables showing large missingness (>20%), high cardinality, or severe binary imbalance (<10% prevalence) were excluded. Redundant features with strong collinearity (Spearman ρ>0.5) were filtered to decrease the risk of multicollinearity. This moderate threshold was chosen to prevent collinearity-driven distortions in distance-based clustering and to promote model parsimony suitable for clinical translation. Missing data were imputed using predictive mean matching to preserve distributional properties. Factorial analysis of mixed data was used to project high-dimensional clinical data into lower-dimensional principal components. Hierarchical clustering on principal components was applied to the transformed data to identify latent patient subgroups. The number of clusters was suggested by the dendrogram and supported further by average silhouette width, which quantifies intra-cluster cohesion and inter-cluster separation. The clustering validity was further supported by the cophenetic correlation coefficient, which measured the efficacy of the dendrogram. Variable contributions and squared cosine (Cos2) metrics were used to identify the most informative features. Cluster differentiation was statistically validated through chi-squared tests comparing AKI stage and mortality distributions. Survival characteristics across clusters were analyzed using Kaplan–Meier survival curves with log-rank testing. To further characterize cluster-defining features, a logistic regression model was trained using high-importance variables identified from the clustering phase. Hyperparameter tuning for model regularization (alpha and lambda) was performed via 10-fold cross-validation. Classification performance with 10-fold cross-validation was evaluated using the area under the ROC curve (AUC), sensitivity, specificity, accuracy, kappa statistic, and positive/negative predictive values. A confusion matrix was used to assess classification agreement.

The sample included 1372 patients with AKI severity evenly distributed across stages 0–3 (n=343 per stage). The median follow-up time was 5.5 days with an interquartile range of 3.1–11.9 days. There were no duplicate rows in the study dataset. Fig. 1 illustrates the initial variables and the patterns of missing data. In total, 11.4% of the initial data were missing. We then excluded variables with >20% missing data from further analysis: cardiac output, C-reactive protein, direct bilirubin, central venous pressure, arterial O2 saturation, serum albumin, and marital status. Furthermore, the variable race was excluded due to its high cardinality (30 levels). Binary features were also evaluated for imbalance, and those with fewer than 10% positive cases were excluded from the analysis. Specifically, among comorbidities, liver disease (3.7% positive), type 1 diabetes mellitus (2.2%), cerebrovascular disease (1.2%), and acute myocardial infarction (0.4%) were excluded. Similarly, medication use variables of dobutamine (4.6%), milrinone (3.1%), dopamine (1.5%), and bumetanide (1.3%) were also excluded due to low prevalence. Fig. 2 illustrates the heatmap of correlations among the study variables using Spearman's correlation coefficients. There were large and statistically significant correlation between ALT and AST (ρ=0.86), serum chloride and sodium (ρ=0.63), norepinephrine and vasopressin (ρ=0.61), systolic and diastolic blood pressure (ρ=0.57), prothrombin time and total bilirubin (ρ=0.55), arterial CO2 pressure and PH (ρ=−0.51), all p<0.01. To reduce redundancy and decrease the risk of multicollinearity, we excluded one variable from each pair of the correlated features. Specifically, we excluded AST, serum chloride, vasopressin, diastolic blood pressure, total bilirubin, and arterial CO2 pressure. These selections were based on clinical relevance, physiological specificity, and the goal of retaining the variable that provides broader or more direct insight into patient status. The remaining data contained only 3.8% missing values, which were imputed using predictive mean matching.

Fig. 1.

Visualization of the patterns and percentages of missing data in the initial study dataset.

Fig. 2.

The heatmap of pairwise Spearman's correlations among study variables.

Fig. 3A presents a scatter plot of the first and second principal components, with deceased patients indicated in black. A greater concentration of deceased patients is observed on the right side of the plot, suggesting an association between the first component and mortality. Similarly, Fig. 3B displays the same scatter plot colored according to AKI stages (0–3). Patients with lower AKI stages are predominantly located on the left side of the plot. Stages 1 and 2 show substantial overlap, particularly along the first principal component. Hierarchical clustering on principal components best suggested the presence of two clusters (Fig. 3C). The average silhouette width was 0.56, indicating moderately well-separated clusters. The cophenetic correlation coefficient was 0.59, suggesting that the hierarchical clustering dendrogram reasonably preserved the original distance structure. Fig. 3D presents a scatter plot illustrating the results of the hierarchical clustering. Fig. 3E and F illustrate the variable contributions and the squared cosine (Cos2) values of the variables, respectively, indicating that the clusters can be effectively distinguished using a subset of patient features. The clear separation between the two cluster centroids, together with the distribution of clinical severity along the first principal component, indicates that the clustering scheme may hold prognostic significance. Cluster 1 included 701 patients, and cluster 2 included 671 patients. Cluster 1 included 314 (44.8%), 159 (22.7%), 148 (21.1%), and 80 (11.4%) patients with AKI stages 0, 1, 2, and 3, respectively, while cluster 2 included 29 (4.3%), 184 (27.4%), 195 (29.1%), and 263 (39.2%) patients across the same stages. There was a significant difference in the distribution of AKI severity between the clusters, χ2(3)=342.210, p<0.001. There was also a statistically significant difference in mortality rates between clusters 1 (145 deaths; 21.9%) and 2 (367 deaths; 54.7%), χ2(1)=154.500, p<0.001, odds ratio (95% CI)=4.28 (3.39, 5.42). Overall, cluster 2 was a high-risk subgroup of patients with greater AKI severity and higher mortality, signifying a poorer prognosis compared with cluster 1. Fig. 4A demonstrates a significant difference in 30-day survival probability between the two clusters, as confirmed by the log-rank test.

Fig. 3.

(A) Scatter plot of the first two principal components, with deceased patients marked in black. A higher frequency of deceased patients is visible on the right side of the plot, indicating an association between the first component and mortality. (B) Scatter plot of the first two principal components colored by AKI stages (0–3). Patients with lower AKI stages are represented mainly on the left, with stages 1 and 2 overlapping along the first principal component. (C) Dendrogram from hierarchical clustering performed on the principal components, suggesting two distinct patient clusters. (D) Scatter plot showing patient clusters derived from hierarchical clustering on principal components. (E) Variable contributions to the clustering, highlighting patient features that distinguish the two clusters above the average contribution (indicated by the dashed red line). (F) Squared cosine (Cos2) values for the variables, indicating the quality of variable representation.

Fig. 4.

(A) Kaplan–Meier survival curves showing 30-day survival probability for the two clusters. A significant difference between clusters was observed by the log-rank test, with cluster 2 showing higher mortality. (B) Receiver operating characteristic (ROC) curve illustrating the performance of the logistic model in discriminating between the two clusters. The area under the curve (AUC) demonstrates strong model accuracy and robustness.

We developed a logistic regression model to evaluate the discriminative ability of the two clusters using the features identified as important in the cluster analysis. Hyperparameter tuning via 10-fold cross-validation identified the optimal logistic regression model parameters with an alpha of 0.100 and a lambda of 0.044. Table 1 presents the logistic regression model showing variables that differentiate the two clusters and identify a distinct high-risk phenotype. Variables related to critical illness severity – such as sepsis, vasopressor use, chronic comorbidities (e.g., chronic kidney disease and heart failure), and markers of organ dysfunction – were strongly associated with membership in the high-risk cluster. The regression coefficients and odds ratios indicate that these features are powerful predictors of cluster assignment. Additionally, elevated respiratory and heart rates were statistically significant, further supporting the physiological distinctiveness of this group. In contrast, higher arterial oxygen levels, normal pH, and older age were negatively associated with high-risk cluster membership. These statistical patterns may help clinicians identify ICU patients with phenotypic features compatible with more severe AKI trajectories. Model performance was evaluated using 10-fold cross-validation. Fig. 4B presents the model's performance, as measured by the area under the ROC curve (AUC). The large AUC demonstrated the model's strong performance in distinguishing the two clusters, thereby validating the effectiveness of the clustering scheme. Table 2 summarizes the diagnostic performance of the model in identifying the positive class (class 2) among the clustered patients. The logistic model demonstrated strong and balanced classification performance, effectively distinguishing between the two clusters with high concordance between predicted and actual labels. The diagnostic accuracy indicates minimal systematic bias in misclassification. Overall, these results support the model's reliability and robustness in identifying the two patient profiles.

This study presented evidence that AKI phenotypic profiling provides enhanced risk stratification relative to conventional serum creatinine-based staging by integrating a broader spectrum of patients’ clinical characteristics. We suggested that high-risk patients can be identified using routinely available clinical data, enabling timely risk recognition. The strong discriminative performance of the logistic model supports the feasibility of building real-time clinical decision support tools to stratify ICU patients into phenotypic clusters at admission. These phenotypic profiles could guide tailored interventions; for example, patients in the high-risk cluster may benefit from closer monitoring, early nephrology consultation, and aggressive hemodynamic and metabolic optimization. This two-cluster approach may improve prognostic accuracy, facilitate communication with families, and support shared decision-making, compared with the conventional four-stage severity classification. Phenotype-guided research and trials enable the design of more targeted therapies and personalized treatment strategies in AKI management.

This study had several limitations that warrant consideration. The analysis was based on retrospective data, which may introduce biases inherent to observational studies, such as unmeasured confounding factors. Although the clustering approach was unsupervised, selection bias related to missing data and variable exclusion criteria cannot be fully eliminated. Nevertheless, the data preprocessing ensured statistical integrity by removing highly imbalanced or collinear variables and applying robust imputation techniques. Race can affect the generalizability of research findings to other populations. However, we excluded it due to its high cardinality, sparse distribution, and inconsistent documentation, which could distort factor projections and bias cluster centroids. Our analysis focused primarily on physiological and biochemical variables rather than sociodemographic determinants. Comorbidity variables with fewer than 10% positive cases were also excluded to prevent statistical sparsity and instability in the factorial analysis. Inclusion of rare features can distort covariance structures and compromise cluster reproducibility. The exclusion preserved statistical power, ensured meaningful variable contributions to multidimensional variance, and enhanced clinical interpretability of the resulting phenotypes. We used advanced unsupervised learning for phenotype discovery, provided equal representation of AKI severity stages, and internally validated both clustering and classification performance. While these features support the robustness and translational potential of the findings, external validation with an independent dataset is still needed to confirm the generalizability and reproducibility of the model across diverse populations and clinical settings.