Discovery of postoperative clinical states linked to cardiac surgery-associated acute kidney injury Authors: Motahare Shabestari, Vinod Kumar Chauhan, Mohammadtaghi Sarebanhassanabadi, Seyed Jalil Mirhosseini Published: 2026 ABSTRACT AbstractBackgroundCardiac surgery-associated acute kidney injury (AKI) remains a common postoperative complication and is associated with adverse short- and long-term outcomes. Conventional risk scores may have limited performance in complex and dynamically evolving postoperative courses. This study aimed to identify data-driven postoperative clinical states from time-series electronic medical record (EMR) data and characterize how state occupancy evolves before AKI onset.MethodsWe conducted a retrospective cohort study of adults undergoing major open-heart surgery at a tertiary cardiovascular center. AKI was defined using the modified Kidney Disease Improving Global Outcomes criteria. Each postoperative day was represented as a patient-day vector comprising laboratory and static clinical features. An internal benchmark guided preprocessing decisions, including feature normalization and selection. Principal component analysis was used for dimensionality reduction, and multiple clustering algorithms were evaluated using validity, resampling-based stability, and reproducibility metrics.ResultsAmong 3,526 patients, 567 developed AKI. Five clinically interpretable postoperative states were identified. C5 represented an evolving or near-established AKI state, characterized by older age, hyperglycemia, elevated urea and creatinine, hyperkalemia, and hypocalcemia. C4 represented a near-term alarm state, marked by relatively higher hemoglobin, hematocrit, RBC count, calcium, and sodium. C3 reflected a microcytic-hypochromic erythrocyte phenotype with lower MCV, MCH, and MCHC. C1 was characterized by relatively prolonged PT and elevated INR, whereas C2 reflected an overall stable postoperative profile. In patients who developed AKI, state occupancy shifted toward C4 in the immediate pre-onset period and increased further in C5 around the day of AKI diagnosis. By contrast, patients without AKI showed increasing occupancy of the stable C2 state and the relatively protective C1 state as the endpoint approached.ConclusionsPhenotyping of postoperative EMR data identified five interpretable states with distinct laboratory profiles and temporal relationships to cardiac surgery-associated AKI. The AKI trajectory was characterized by transition from an alarm state to an evolving AKI state, whereas the Non-AKI trajectory converged toward stable or protective postoperative states. These findings suggest that dynamic phenotyping may provide a framework for postoperative AKI risk stratification and trajectory monitoring.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13040-026-00572-y.Keywords: Cardiac surgery, Acute kidney injury, Electronic medical records, Unsupervised learning, Phenotyping, Clustering, Principal component analysis, Postoperative trajectories Introduction Each year, nearly 2 million patients undergo cardiac surgery worldwide, and approximately 20%–30% develop cardiac surgery-associated acute kidney injury (AKI) [1]. Over the past two decades, the incidence of this complication has increased, contributing to prolonged intensive care unit (ICU) and hospital stays [2, 3]. Cardiac surgery-associated AKI is also associated with substantially higher mortality, with reported risks increasing by two- to eightfold among affected patients [4]. Clinically, AKI typically develops within hours to days after surgery and is defined by an increase in serum creatinine, a reduction in urine output, or both. Beyond the immediate postoperative period, AKI is associated with adverse long-term outcomes, including chronic kidney disease [5], cardiovascular events [6], reduced long-term survival [2]. impaired quality of life, and the need for ongoing renal monitoring. These short- and long-term consequences highlight the importance of early detection and timely intervention. Several statistical risk tools, including the Mehta score [7], Cleveland Clinic score [8], and Simplified Renal Index [9], have been developed to support early identification of patients at risk for cardiac surgery-associated AKI. Although these tools are attractive because of their simplicity and ease of implementation, they are limited in their ability to capture complex and dynamic postoperative changes, including evolving clinical status and time-varying risk factors [10]. Consequently, their predictive performance may be suboptimal, particularly in patients with heterogeneous or rapidly changing clinical trajectories. In some settings, these models may identify only severe AKI or detect risk in a limited subset of patients, that restrict their value for early postoperative surveillance [11]. This limited and variable performance has led to the absence of endorsement of these tools in major clinical guidelines, including those from Kidney Disease Improving Global Outcomes (KDIGO). More recently, deep learning models have shown promising performance for AKI risk prediction across several cohorts [12, 13]. However, differences in case mix, data quality, and clinical workflows across institutions may limit model transportability, while reliance on large numbers of input variables can constrain practical implementation [14]. Despite advances in AKI prediction, the clinical evolution of postoperative AKI and its stage-specific manifestations remain incompletely understood. AKI does not progress through a uniform or linear pathway; rather, its trajectory varies across patients and is shaped by factors such as age, baseline comorbidities, perioperative exposures, and postoperative physiological stress [11]. Identifying phenotypic differences across the course of AKI and characterizing transitions between clinical states may improve the recognition of evolving renal injury and support more reliable postoperative risk stratification. Data-driven phenotyping approaches, including methods that combine principal component analysis (PCA) with clustering, offer an alternative to traditional risk models by identifying patient subgroups with distinct clinical profiles and differential risk patterns. For example, Meijer and colleagues applied PCA followed by clustering to baseline data from patients with atrial fibrillation before treatment and identified distinct clusters associated with recurrence-related risk factors [15]. This phenotyping framework can be further strengthened by incorporating time-series patient data, which enables not only the identification of postoperative clinical states but also the characterization of transitions between them. Electronic medical records (EMRs) contain rich longitudinal information on patients’ clinical phenotypes, including laboratory measurements and diagnostic outcomes. These data provide an opportunity to define postoperative disease trajectories and identify meaningful patterns preceding adverse events. In this study, we developed a standardized EMR-based framework for patients undergoing cardiac surgery that included data preprocessing, identification of AKI-related postoperative clinical states, and analysis of state-transition patterns. The overall framework is illustrated in Fig. 1. By mapping each patient-day to a clinical state, this approach facilitates analysis of evolving postoperative profiles and may help distinguish stable recovery trajectories from patterns associated with increasing AKI susceptibility. The first stage of the framework included feature normalization, feature selection, and dimensionality reduction. The second stage identified the optimal clustering approach for detecting latent postoperative states and evaluating their temporal relationship with cardiac surgery-associated AKI. A schematic illustration of the proposed workflow Method Sample selection In this retrospective cohort study, data were collected from the Afshar Hospital, a tertiary referral cardiovascular hospital affiliated with Shahid Sadoughi University of Medical Sciences, between March 1, 2017, and December 31, 2024. Adults aged 18 years or older who underwent major open-heart surgery, including coronary artery bypass grafting, valve repair or replacement, combined valve and coronary procedures, aortic surgery, or other major cardiac operations, were included. Patients who were receiving long-term dialysis and required preoperative renal replacement therapy, those with stage 3 acute kidney injury at the first preoperative assessment, and those without either preoperative or postoperative serum creatinine measurements were excluded. Data were extracted from the EMR system, which contains demographic information, coded diagnoses, laboratory results, and treatment details. Data from each patient’s hospitalization were included in the analysis. Candidate predictors comprised both static variables, such as demographics, and time-varying variables that could change daily, including laboratory test results. To restrict the analysis to routinely available postoperative measurements, laboratory variables were retained only if they had been measured at least once in more than 50% of patients during the first seven postoperative hospital days. This threshold was used to exclude sparsely collected laboratory tests that were ordered only for selected clinical indications rather than as part of routine postoperative monitoring. Among the retained laboratory variables, missingness was minimal, with the highest variable-level missingness being 0.82% (Supplementary Table 1). Ethical approval This retrospective study was approved by the institutional review board of the Shahid Sadoughi University of Medical Sciences, and informed consent was waived (IR.SSU.REC.1404.188). Outcome definition The primary outcome was postoperative AKI occurrence (stage 1, 2, or 3). Cardiac surgery-associated AKI was defined using the KDIGO criteria based on serum creatinine trajectories, with AKI onset considered before the seventh day of hospitalization [16]. To ensure accurate diagnosis, clinician-diagnosed AKI, identified through the center’s diagnostic coding system, was cross-checked against the modified KDIGO criteria. The most recent preoperative serum creatinine value was used as the baseline. Data preprocessing Each ICU stay was represented as an ordered sequence of daily snapshots. Each snapshot comprised a 27-dimensional feature vector, including 18 laboratory variables and 9 static variables. Laboratory variables included urea, creatinine, fasting blood glucose (FBS), red blood cell count (RBC), hemoglobin, hematocrit, mean corpuscular volume (MCV), mean corpuscular hemoglobin concentration (MCHC), mean corpuscular hemoglobin (MCH), white blood cell count (WBC), lymphocyte count, neutrophil count, platelet count, prothrombin time (PT), international normalized ratio (INR), calcium, sodium, and potassium. Static variables included sex, age, and surgical procedure indicators, including three-vessel CABG, four-vessel CABG, Bentall procedure, mitral valve repair or replacement, aortic valve repair or replacement, pulmonary valve repair or replacement, and tricuspid valve repair or replacement. The laboratory measurements were summarized using the daily mean to address the irregular timing of ICU measurements and make patient-days comparable. To ensure consistency, the data were regularized to a daily basis, with all repeated same-day lab values averaged. Missing values were not primarily due to sparsely collected variables, as non-routine laboratory tests had been excluded during variable selection. To preserve complete daily trajectories, missing values were imputed within each patient’s longitudinal record using chained-equation imputation. This patient-wise approach was chosen because approximately 78% of missing values occurred in patients hospitalized for more than 5 days, which provided multiple observed patient-days within the same individual’s postoperative course. The goal was to preserve within-patient temporal continuity and avoid borrowing information across clinically heterogeneous patients before dimensionality reduction and clustering. A semi-supervised validation framework was developed to guide decisions on feature normalization and feature selection. A benchmark was constructed by defining two distinct patient-day groups: an AKI index patient-day, defined as the patient-day on which AKI first occurred among patients who developed AKI, and a Non-AKI reference group, defined as the last two available patient-days among patients who did not develop AKI, restricted to patient-days occurring before hospital day 7. This benchmark served as an internal verification target, assuming that an adequate preprocessing strategy should separate these two clinically distinct contexts. We used the silhouette coefficient to quantify separation on this benchmark and selected the preprocessing settings that optimized this metric, including choices for feature normalization and feature selection. Feature normalization To identify a feature normalization approach that preserved physiologically meaningful variability across patient-day feature vectors, we benchmarked the untransformed (raw) data against three scaling strategies: robust scaling, standard scaling, and min-max scaling. Each normalization option was evaluated by its ability to separate AKI index patient-days from Non-AKI index patient-days using the silhouette coefficient. To quantify the sensitivity of this separability, we used a stratified bootstrap procedure (500 replicates, 95% percentile confidence intervals) in which patient-day rows were resampled with replacement within each outcome group. Across bootstrap replicates, robust scaling consistently yielded the highest silhouette coefficient (Fig. 2), which shows the silhouette scores for each normalization strategy across all replicates. As a result, robust scaling was selected as the normalization strategy for the subsequent workflow. Benchmarking data normalization. The performance of three normalization methods was benchmarked against the untransformed (raw) data. Each normalization option was evaluated by its ability to separate AKI index patient-days from Non-AKI index patient-days using the silhouette coefficient. Robust scaling yielded the highest silhouette coefficient. (Raw: 0.35 (95% CI 0.18–0.57); MinMax scaler: 0.16 (95% CI 0.08–0.25); Standard scaler: 0.36 (95% CI 0.23–0.5); Robust scaler: 0.56 (95% CI 0.46–0.68)) Feature selection and dimensionality reduction An incremental PCA-based diagnostic was used to assess how the latent structure of the data evolved as feature sets were introduced sequentially in groups. After adding each feature set (Supplementary File), we refitted PCA and examined whether the newly added variables contributed meaningfully to separating AKI index patient-days from Non-AKI index patient-days. The number of statistically supported principal components was determined using Horn’s parallel analysis. To distinguish informative signal from added noise, we compared the dominant contributors to the significant components across successive PCA fits. A newly added feature set was considered noisy when its inclusion did not alter the leading feature contributions of the significant components (Fig. 3 and Supplementary Fig. 1), which shows how the principal components changed with the addition of each feature group. Using this criterion, seven feature sets were identified as informative, and five principal components were observed. Selected features included all the laboratory variables and age. A low-dimensional embedding space learned from laboratory variables and age captures the diversity of data. Explained variance in data vs. their randomization was compared to determine the number of significant PCs. The proportion of variance explained by each PC is shown for both the original and shuffled data (left). On the right, the highest feature loadings for the significant PCs are displayed. Negative loadings are shown in blue, and positive loadings are shown in red. Increased color intensity reflects a higher magnitude clustering Patient-day vectors were embedded into a five-dimensional space defined by the top five principal components (PCs) learned from the selected laboratory features. To identify recurrent physiological states, four clustering methods were evaluated, including K-means, Gaussian mixture models (GMM), agglomerative hierarchical clustering, and partitioning around medoids (PAM), across a range of cluster numbers between 2 and 20. Model selection was based on a combined assessment under resampling. Bootstrap resampling was performed at the patient-day level (n = 500 replicates, 95% percentile confidence intervals) to assess the stability of the clustering solutions. The Silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin index were then computed. The Adjusted Rand Index was also calculated to compare the similarity of different cluster partitions (Supplementary Fig. 2 and Supplementary Fig. 3). Statistical analysis and coding environment Age was compared using Welch’s t-test based on the result of Levene’s test for homogeneity of variance. Categorical variables were analyzed using Fisher’s exact test. All statistical tests were two-tailed, and a p-value below 0.05 was regarded as statistically significant. All analyses and visualizations were carried out in Python 3.10.16. Result Population characteristics A total of 3,526 patients met the initial inclusion criteria, with a median age of 59 years (IQR, 50–66); 2,397 (67%) were male. AKI occurred in 567 patients (16%), of whom 375 (66%) were male, and the median age in the AKI cohort was 62 years (IQR, 55–69). Among patients without AKI, the median age was 58 years (IQR, 49 − 66), and 2,022 (68%) were male. Age differences between the AKI and non-AKI cohorts were statistically significant (Table 1). Categorical variables were compared and tested for significance, as shown in Table 1. Table 1patients’ characteristics (Data are represented in n (%) or median (IQR))Static featuresAKI (n = 567)Non-AKI (n = 2959)P-valueAge62 (55–69)58 (49–66)< 0.001Male375 (66.13%)2022 (68.33%)0.30Bentall surgery9 (1.58%)26 (0.87%)0.16Three-graft CABG365 (64.37%)1811 (61.2%)0.051Four-graft CABG110 (19.4%)627 (21.18%)0.36Aortic valve repair/replace47 (8.28%)204 (6.89%)0.24Pulmonary valve repair/replace1 (0.17%)14 (0.47%)0.49Mitral valve repair/replace29 (5.11%)210 (7.09%)0.10Tricuspid valve repair/replace7 (1.23%)21 (0.7%)0.19 patients’ characteristics (Data are represented in n (%) or median (IQR)) Among the 567 patients who developed AKI, 433 patients (76.36%) had KDIGO stage 1 AKI, and 134 patients (23.63%) developed KDIGO stage 2 or 3 AKI. Renal replacement therapy was required in 32 AKI patients (5.64%). In-hospital mortality was higher in the AKI group than in the Non-AKI group, occurring in 46 patients with AKI (8.11%) compared with 95 patients without AKI (3.21%). K-Means identified five clinical states Across validity metrics, K-means with five clusters was the most consistent model. In particular, the Silhouette coefficient peaked for K-means at K = 5 and declined with larger K. The Calinski–Harabasz index reached the highest value for K-means around K = 5, and the Davies–Bouldin index reached its minimum at K = 5, indicating the best within-cluster compactness and between-cluster separation at this resolution. Other clustering methods showed weaker coherent results across these criteria. Stability analysis showed that K-means achieved its strongest bootstrap reproducibility at K = 5 with comparatively high ARI values. Moreover, agreement between methods was moderate and tended to decrease as K increased. However, at K = 5, K-means provided the best-separated partition among the other approaches evaluated (Supplementary Fig. 2 and Supplementary Fig. 3). Transitions between states To validate the clinical relevance of the five states and investigate AKI progression, state occupancy was quantified as the proportion of patient-days assigned to each state (C1–C5) in both the AKI and Non-AKI groups across 0, 1, 2, 3, and > = 4 days before the endpoint. As Fig. 4 shows in the AKI cohort, patient-days increasingly occupied the C4 state when there was more than one day prior to AKI onset, suggesting that the features of this cluster may act as alarm signs of near-onset AKI. On the day of AKI onset, more patient-days occupied the C5 state, indicating that this state captures the near-onset or evolving AKI phenotype rather than a preclinical risk state. Proportion of patient-day vectors classified into five clinical states, stratified by AKI and Non-AKI outcome. Clusters C1 and C2 are strongly associated with patients who did not develop AKI, while cluster C5 is primarily associated with early-onset AKI In the Non-AKI cohort, as the endpoint approached, the population showed a greater representation of C1 and C2 states. More specifically, there was less concentration in C4 compared to the AKI cohort when there were two or fewer days remaining until the endpoint. This suggests that as patients in the Non-AKI group neared the seventh day, they avoided the alarm state, C4, for AKI. The concentration of the C5 state remained low and constant, at approximately 0.1, across all time points (Fig. 4). Clinical characterization of states The median and interquartile range of each laboratory variable were computed across all patient-days assigned to each state to interpret the clinical states (Supplementary Fig. 4). Furthermore, to compare the states in terms of enrichment and depletion patterns across all the features, a heatmap of state medians standardized within each feature was constructed. For each variable, state medians were z-scored across the states (Fig. 5). Heatmap of state medians standardized within each feature, displaying enrichment and depletion patterns across all features. For each variable, the state medians were z-scored across the states to facilitate comparison The heatmap demonstrated distinct laboratory phenotypes across states. The highest-risk state, C5, was characterized by older age, hyperglycemia, elevated urea and creatinine levels, hyperkalemia, and hypocalcemia. The alarm state, C4, showed higher RBC count, hemoglobin, hematocrit, lymphocyte percentage, platelet count, calcium, and sodium levels, together with lower neutrophil percentage and WBC count. C1 was distinguished by prolonged PT and elevated INR. C3 was marked by reduced red-cell indices, including MCV, MCH, and MCHC, whereas C2 showed intermediate values across most variables. Discussion In this retrospective study, we used an unsupervised learning approach to identify five postoperative clinical states from structured EMR data of patients undergoing cardiac surgery. These states captured distinct laboratory phenotypes and temporal trajectories during the first postoperative week. After initial dimensionality reduction and feature-screening procedures, sex and surgery-related variables showed limited contribution to the dominant postoperative variation and were therefore excluded from the clustering analysis. The resulting states were clinically interpretable: C1 was characterized by prolonged PT and elevated INR, with most other variables remaining relatively unremarkable; C2 reflected an overall stable postoperative profile; C3 was defined by lower red-cell indices, including MCV, MCH, and MCHC; C4 showed relatively higher hemoglobin, hematocrit, and RBC count, together with lower WBC count and neutrophil percentage; and C5 was distinguished by elevated urea, creatinine, and potassium, along with older age and hyperglycemia. When considered in relation to subsequent AKI trajectories, C5 was best interpreted as an evolving or near-established AKI state, C4 as a potential early warning state, C1 and C2 as relatively protective postoperative states, and C3 as a lower-risk state without clear protective enrichment. Unsupervised phenotyping has been increasingly used to identify heterogeneous clinical states among surgical and critically ill patients [17–19]. This approach is well-suited to complex, high-dimensional data because it can reveal latent phenotypes and temporal patterns that may not be apparent using prespecified clinical categories alone. These data-driven phenotypes may, in turn, support more granular postoperative risk stratification. In cardiac surgery-associated AKI, prior unsupervised studies have mainly focused on preoperative variables to define patient subgroups with differential AKI risk [20–22]. Other work has used clustering to identify intraoperative phenotypes associated with postoperative complications in cardiac surgery patients [18]. By contrast, the application of unsupervised methods to postoperative time-series data for identifying alarm and high-risk states related to cardiac surgery-associated AKI remains underexplored. This study showed that when AKI onset was more than two days away, the proportion of patient-days assigned to C4 was broadly similar between the AKI and Non-AKI groups. However, as the interval to AKI narrowed, particularly on the day before AKI onset, C4 occupancy increased in the AKI group relative to the Non-AKI group. This temporal pattern suggests that the C4 phenotype may represent a near-term alarm state. The dominant features of C4 were hematologic, including relatively higher RBC count, hematocrit, and hemoglobin compared with the other states. Importantly, hemoglobin values in C4 were not markedly supraphysiologic but were relatively higher across states, generally ranging from 13.0 to 14.8 g/dL. In the near-onset period, occupancy of C1, C2, and C3 decreased among patients who subsequently developed AKI, whereas C4 showed the opposite pattern. One plausible explanation is RBC transfusion among patients presenting with lower hemoglobin values, particularly because C1, C2, and C3 had lower hemoglobin levels than C4. In our cohort, RBC transfusion was more frequent among patients who developed AKI than among those who did not, occurring in 176 of 567 AKI patients (31.04%) compared with 533 of 2,959 Non-AKI patients (18.01%). The most commonly documented indications for transfusion were intraoperative or early postoperative bleeding and hemodynamic instability. These findings support transfusion exposure as one plausible contributor to the C4 hematologic profile and subsequent AKI risk; however, other mechanisms, including relative hemoconcentration, postoperative intravascular volume shifts, or a combination of transfusion and fluid-balance effects, may also contribute to the C4 phenotype. The relationship among anemia, transfusion, and AKI after cardiac surgery is complex. Prior studies have shown that AKI risk is higher when anemia and RBC transfusion coexist than with either factor alone, independent of the transfusion threshold used [23, 24]. Therefore, the transition from lower-hemoglobin states toward the relatively higher-hemoglobin C4 state may reflect a clinically important postoperative trajectory in which baseline anemia, transfusion exposure, bleeding, hemodynamic instability, and fluid shifts interact to increase AKI susceptibility. C4 was also characterized by a lower WBC count and neutrophil percentage, together with a higher lymphocyte percentage and platelet count relative to the other clusters. However, as shown in Supplementary Fig. 4, these values generally remained within reference ranges. This pattern is more consistent with resolution of the early postoperative inflammatory stress response than with overt leukopenia, neutropenia, or thrombocytosis [25, 26]. Calcium and sodium levels were also relatively higher in C4 than in the other clusters, although they generally remained within reference ranges. Similar postoperative recovery of calcium and sodium concentrations has been reported in previous cardiac-surgery cohorts and does not appear to be independently linked to AKI occurrence [27]. Taken together, C4 appears to represent a clinically mixed state, with several features compatible with postoperative recovery rather than overt organ dysfunction. Nevertheless, its concurrent pattern of relatively higher hemoglobin, hematocrit, and RBC count may identify a subgroup of patients with a hematologic and volume-related profile associated with increased near-term susceptibility to AKI. Another important finding was the progressive increase in C5 occupancy as AKI onset approached. Specifically, the proportion of patient-days assigned to C5 rose in the immediate pre-onset period and was highest around the day of AKI diagnosis, indicating that this state captures a near-onset, evolving, or already established AKI phenotype. C5 was characterized by older age, hyperglycemia, elevated urea and creatinine, hyperkalemia, and hypocalcemia. This phenotype represents a biochemical renal-dysfunction state that overlaps with early AKI development and, in some patient-days, with the diagnostic window of AKI itself. Thus, C5 summarizes the laboratory abnormalities accompanying early postoperative AKI evolution. In contrast to C4, which may serve as an earlier alarm state, C5 identifies patients who may already have entered the biochemical phase of AKI and require close monitoring and kidney-protective management. Postoperative hyperglycemia was another defining feature of C5. Poor glycemic control after cardiac surgery has been associated with a higher risk of AKI, potentially through oxidative stress, endothelial dysfunction, and amplification of inflammatory signaling within the renal microvasculature [28]. Therefore, the higher glucose levels observed in C5 may reflect both systemic metabolic stress and a potentially modifiable contributor to early renal injury. Elevated urea and creatinine are established biochemical markers of reduced renal function and are central to AKI recognition. Similarly, hyperkalemia is a common manifestation of impaired kidney function, reflecting reduced glomerular filtration and impaired distal tubular potassium secretion [29]. Calcium levels were lowest in C5 compared with all other clusters, suggesting that hypocalcemia may be part of the biochemical profile accompanying postoperative AKI. Although this association has not been extensively studied specifically in cardiac surgery populations, observational studies in critically ill patients have reported an association between lower serum calcium and AKI [30, 31]. For example, Zhao et al. reported a higher risk of AKI among patients with calcium levels below 8.3 mg/dL [32]. Mechanistically, hypocalcemia may contribute to renal vulnerability through altered calcium-dependent cellular signaling, impaired tubular function, and increased susceptibility to ischemic and inflammatory injury [33]. In the Non-AKI group, the proportion of patient-days assigned to C2 increased as patients approached the endpoint, whereas C4 occupancy decreased over the same interval. C2 was marked by balanced values across most laboratory variables, suggesting that many patients without AKI transitioned toward a stable postoperative state by the end of the first week. This pattern supports the interpretation of C2 as a favorable recovery phenotype and reflects clinical stability rather than progression toward renal dysfunction. In the Non-AKI group, the proportion of patient-days assigned to C1 also increased as patients approached the endpoint. C1 was defined mainly by relatively prolonged PT and elevated INR, together with more controlled potassium levels compared with the other clusters. Although evidence linking postoperative coagulation profiles to cardiac surgery-associated AKI remains limited, preoperative PT prolongation has been associated with postoperative AKI risk in prior studies [34]. In our analysis, however, C1 was enriched among Non-AKI trajectories, which suggests that this coagulation profile may reflect a more protective postoperative state rather than a potential risk for renal dysfunction. One possible explanation is that modest relative hypocoagulability may limit renal microvascular thrombosis after cardiac surgery. Cardiac surgery-associated AKI is increasingly understood as a microvascular disorder involving endothelial activation, platelet aggregation, fibrin deposition, inflammatory injury, and regional renal hypoxia, rather than as a purely hemodynamic phenomenon [35]. Activation of platelets and the coagulation cascade may promote intrarenal microthrombi formation, medullary hypoxia, and tubular injury, even without sustained systemic renal hypoperfusion [36]. In addition, thrombotic microangiopathy has been recognized as an underdiagnosed contributor to severe AKI after cardiovascular and aortic surgery, supporting the pathogenic relevance of renal microvascular obstruction [37]. Within this framework, a state of relative hypocoagulability may reduce intrarenal fibrin formation and platelet-fibrin microaggregate deposition. It preserves renal microcirculatory flow and is consistent with established models of coagulation-driven postoperative AKI [38]. The distribution of patient-days assigned to C3 was relatively stable across postoperative time points and did not differ between patients who did and did not develop AKI. This pattern suggests that C3 represented neither a protective recovery state nor an early warning state for postoperative AKI in our cohort. The defining laboratory features of C3 were lower MCV, MCH, and MCHC, consistent with a microcytic-hypochromic erythrocyte phenotype. This profile most likely reflects pre-existing iron-deficient erythropoiesis or chronic microcytosis rather than an acute postoperative renal-injury signal [39]. Among patients undergoing cardiac surgery, iron deficiency is common and may result from chronic inflammation, heart failure-related iron dysregulation, malnutrition, and long-term exposure to antiplatelet or anticoagulant therapy [39, 40]. Importantly, the comparable distribution of C3 between the AKI and Non-AKI groups indicates that lower MCV, MCH, and MCHC, when considered in isolation within the early postoperative trajectory, did not identify patients at increased AKI risk. This interpretation is consistent with prior cardiac-surgery data showing that lower preoperative corpuscular indices were not independently associated with postoperative AKI after cardiac surgery [41]. It is also compatible with evidence from elective cardiac surgery, suggesting that non-anemic iron deficiency does not necessarily translate into worse early postoperative outcomes [40]. Therefore, C3 is best interpreted as a low-risk baseline hematologic phenotype, which likely reflects chronic iron status or microcytosis, rather than a marker of cardiac surgery-associated AKI. The primary clinical purpose of identifying these postoperative states is prognostic risk stratification. This state-based model provides an interpretable framework for monitoring dynamic postoperative trajectories and distinguishing patients who are progressing toward renal deterioration from those following a stable recovery pattern. Recognition of a C4-like warning profile, or transition from relatively protective or lower-risk states toward C4, may support earlier clinical awareness before overt renal dysfunction develops. In practice, this could prompt closer reassessment of intravascular volume status, recent transfusion exposure, bleeding or hemodynamic instability, glycemic control, urine output trends, and the frequency of renal laboratory monitoring. By contrast, transition into C5 may indicate near-established or evolving AKI and should prompt intensification of kidney-protective management. In conclusion, using a day-level phenotyping framework, we identified five postoperative clinical states after cardiac surgery. Compared with patients who did not develop AKI, those who developed AKI showed fewer patient-days in the stable recovery state, C2, and a progressive shift toward C4 as AKI onset approached. C4 was characterized by relatively higher hemoglobin, hematocrit, and RBC values, while most other laboratory markers remained within physiologic ranges. This profile was compatible with partial recovery from the early postoperative stress response, but also reflects transfusion exposure, hemoconcentration, intravascular volume shifts, or their combination. Accordingly, C4 represents a near-term warning state for increased AKI susceptibility. In the final 24 h before AKI onset and around the day of AKI diagnosis, occupancy increased in C5, the highest-risk state, which was defined by renal dysfunction markers and electrolyte derangements. Thus, C5 captured a near-established biochemical AKI phenotype. This study has several limitations. First, the cohort was derived from a single tertiary center, which may limit external generalizability. Second, the analysis was restricted to in-hospital patient-days and AKI occurring within the first seven postoperative days, or before discharge if earlier. Therefore, later-onset AKI and post-discharge kidney outcomes were not captured, and the five-state structure may not fully represent the broader spectrum of postoperative renal trajectories. Third, we emphasized population-level state occupancy patterns rather than patient-specific transition pathways. Future studies incorporating individualized longitudinal modeling, external validation, and prospective assessment of state-guided clinical monitoring may better define how these phenotypes can be translated into postoperative AKI risk stratification. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Author contributions Motahare. Sh: Conceptualization, Data curation, Project administration, Validation, Writing original draft, Methodology, Formal analysis, Software, Visualization. Vinod. Ch: Supervision of the methodology and analysis, Writing review, and editing. Mohammadtaghi. S: Data curation, Writing review, and Editing. SeyedJalil. M: Conceptualization, Data curation, Writing review, and Editing. Funding None. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.