In hospitalized cirrhosis patients, acute kidney injury (AKI) is a common and severe complication, affecting between 22 % and 53 % of cases [1–3]. Notably, patients admitted with AKI exhibited a 3.75-fold higher mortality risk compared to those without AKI [1]. Generally, AKI is classified into three stages (stage 1 to stage 3) based on serum creatinine (sCr) levels.
The progression and severity of AKI in cirrhosis are closely associated with disease staging, with more advanced stages correlating with worse clinical outcomes. A retrospective analysis of 4654 hospitalized cirrhosis patients with stage 1 AKI (Cerner Health Facts database, U.S., 2009–2017) revealed significant differences between AKI stages 1A (sCr <1.5 mg/dL) and 1B (sCr ≥1.5 mg/dL). Patients with stage 1B exhibited a significantly higher 90-day mortality rate (27.2 % vs. 19.7 %, p < 0.001), a 52 % increased mortality risk, a 24 % reduced likelihood of AKI recovery, and a 42 % higher risk of AKI progression compared to stage 1A [4]. These findings are consistent with data from a multi-hospital network study in the U.S., which showed increased mortality risks in more advanced AKI stages. Stage II and III patients had higher 90-day mortality than those with stage I, while those who achieved complete AKI recovery had the lowest mortality risk [5]. Collectively, these studies emphasize the critical importance of early AKI detection and intervention in patients with cirrhosis.
The significant burden and prognostic implications of AKI in patients with cirrhosis underscore the urgent need for accurate risk stratification [5–7]. A comprehensive study based on the Premier Healthcare Database revealed that among 4.06 million hospitalized patients with chronic liver disease, 1.22 million (30.1 %) were diagnosed with AKI, and 90,235 (2.2 %) were specifically diagnosed with hepatorenal syndrome (HRS). These findings not only confirm the high prevalence of renal dysfunction in this patient population but also show a steadily rising trend in the proportion of hospitalized chronic liver disease patients who develop AKI or HRS during the study period from 2018 to 2023 [8].
While sCr remains the cornerstone of AKI diagnosis, its utility is hampered by inherent limitations: sCr changes lag actual kidney injury, requiring time to reach equilibrium [9]. Moreover, increased ascites, reduced hepatic creatinine synthesis and skeletal muscle loss, leading to falsely low sCr levels even in moderate-to-severe renal impairment, delaying the identification of at-risk individuals [10–12]. Additionally, estimated glomerular filtration rate (eGFR), derived from sCr and/or cystatin C (CysC), is essential for evaluating renal function. However, current eGFR equations exhibit significant limitations in accurately assessing renal function in patients with cirrhosis [13]. Emerging biomarkers like CysC, urinary neutrophil gelatinase-associated lipocalin (uNGAL), N-acetyl-β-D-glucosaminidase (NAG), interleukin-8 (IL-8), and kidney injury molecule 1 (KIM-1) hold promise for enhancing risk prediction by detecting injury at earlier stages. However, their clinical adoption remains limited due to high costs and limited availability [14,15]. This gap highlights the critical unmet need for accessible, cost-effective tools to predict AKI in routine hospital settings. A prediction model utilizing readily available clinical parameters would facilitate timely interventions, such as optimizing fluid balance and avoiding nephrotoxins.
Thus, this study aimed to develop a simple, objective model using readily available clinical data collected within the first 24 hours of admission to predict the risk of AKI in patients with cirrhosis. Given the tendency of clinicians to overlook patients with normal sCr levels upon admission, this study specifically targeted this group. Few studies have compared different eGFR equations in cirrhotic patients, and this study also compared these equations for predicting AKI risk.
2Patients and methods2.1Study designA predictive model for in-hospital AKI, utilizing routine laboratory indices, was developed using a retrospective cohort of patients with decompensated cirrhosis (DC), initially admitted to Mengchao Hepatobiliary Hospital of Fujian Medical University between December 2016 and December 2022. This institution represents the largest referral and academic tertiary hepatobiliary center in Southeast China. External validation of the model was subsequently conducted using the publicly available Medical Information Mart for Intensive Care (MIMIC)-III database (version 1.4), comprising approximately 40,000 patients admitted to intensive care units (ICUs) at Beth Israel Deaconess Medical Center between 2001 and 2012 [16]. The primary endpoint was the development of AKI during hospitalization in cirrhotic patients with a normal baseline sCr level.
The development and validation of the predictive model adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines [17]. The study protocol received ethical approval from the Institutional Review Board of the Mengchao Hepatobiliary Hospital of Fujian Medical University (approval number: 2025_014_01). Author S.L. Lin completed the Collaborative Institutional Training Initiative program (CITI; record ID: 54,218,526) and secured authorization for access to the MIMIC-III database. Given that all patient data were de-identified and did not compromise patient privacy, the requirement for informed consent was waived.
2.2The training cohortPatients diagnosed with DC were screened for inclusion in this study by retrieving data from the electronic medical record system. Only first-admission patients were considered. The diagnosis of cirrhosis was established based on clinical parameters, including laboratory tests, endoscopic or radiologic evidence of cirrhosis, historical or present evidence of decompensation [e.g., hepatic encephalopathy (HE), ascites, gastric variceal hemorrhage (GVH), or infection], and liver biopsy results when available. Inclusion criteria required patients to be above 18 years of age, a baseline sCr level recorded at admission, and routine laboratory results completed within 24 hours of admission, such as complete blood count, liver function test, renal function test, electrolytes, and coagulation function. An expected hospital stay of at least seven days was also required to reduce the probability of including patients with chronic kidney disease (CKD). Patients were excluded if they presented with elevated admission sCr values (≥1.2 mg/dL for males or ≥1.1 mg/dL for females), a history of liver or kidney transplantation, AKI upon admission, acute or chronic hemodialysis requirements at the time of admission, missing critical laboratory data within the first 24 hours or with malignant tumors.
Data collection for the training cohort included demographic characteristics, routine laboratory results as previously described, complications at admission (e.g., ascites, HE, or GVH), and comorbidities such as hypertension (HT) and type 2 diabetes mellitus (DM2). Notably, the diagnosis of infection at admission in the training cohort was predominantly based on clinical empirical judgment, owing to the absence of definitive evidence such as pathogen culture results or radiological findings. Consequently, infection-related events were not recorded for the training cohort in final analysis. An overview of the patient selection process is illustrated in the flowchart (Fig. 1A).
2.3The definition of AKI
AKI was defined according to the revised criteria recommended by the International Club of Ascites for patients with cirrhosis: an increase in sCr of ≥ 0.3 mg/dl from baseline within 48 hours of hospital admission, or an increase in sCr of ≥ 50 % from an existing or inferred baseline value within 7 days (any sCr measurement from the past 3 months is applicable as the baseline). The progression stages of AKI are delineated as follows: stage 1 involves an absolute increase in sCr of ≥ 0.3 mg/dL, or an increase in sCr to 1.5 to 2.0 times the baseline; stage 2 includes an increase in sCr to 2.0 to 3.0 times the baseline; stage 3 is characterized by an increase in sCr to more than 3.0 times the baseline, or sCr ≥ 4 mg/dL with an acute increase of ≥ 0.3 mg/dL, or the initiation of renal replacement therapy [18–20].
2.4The equations of eGFR- (1)
aMDRD [21]
175×(sCr−1.154)×(Age−0.203)×1.212(ifblack)×0.742(iffemale)
- (2)
CKD-EPI (2021) [22]
Used sCr only (CKD-EPI1):
142×min(sCr/0.7,1)−0.219×max(sCr/0.7,1)1.200×0.9938Age×1.012(iffemale)Used sCr and CysC (CKD-EPI2):
CKD-EPI (2021) used sCr and CysC
135×min(sCr/0.7,1)−0.219×max(sCr/0.7,1)−1.200×min(CysC/0.8,1)−0.323×max(CysC/0.8,1)−0.778×0.9961Age×0.963(iffemale)135×min(sCr/0.9,1)−0.144×max(sCr/0.9,1)−1.200×min(CysC/0.8,1)−0.323×max(CysC/0.8,1)−0.778×0.9961Age(ifmale)
- (3)
China-sCr-CysC [23]
173.9×(sCr−0.184)×(CysC−0.725)×(Age−0.193)×0.890(iffemale)
- (4)
Mindikoglu-2016 [24]
105.49×(sCr−0.712)×(CysC−0.285)×(0.993Age)×0.864(iffemale)×1.014(ifblack)
- (5)
Kalafateli-2017 [25]
45.9×(sCr/0.011)−0.836×(BUN/2.8)−0.229×(INR−0.113)×(Age−0.129)×(Sodium0.972)×0.809(iffemale)×0.92(ifmoderate/severeascites)
- (6)
GRAIL [26]
GFR Assessment in Liver disease model
For validation, we accessed and deployed the MIMIC-III database following the instructions provided on the PhysioNet project page (https://physionet.org/content/mimiciii/). Patient data were extracted using the same inclusion and exclusion criteria as those applied to the training cohort. Patients were identified as having cirrhosis based on the ICD-9 codes 5712, 5715, and 5716. Geographic characteristics were retrieved from the “patients” table. Routine laboratory results were extracted from the “labevents” table. We identified complications and morbidities from the “noteevents” table. Only patients admitted to the ICU during their first hospital admission were included. The patient selection process is detailed in Fig. 1B Any ambiguity or discrepancies in patient information were resolved through discussion and consensus among all co-authors.
2.6Missing and outlier dataTo preserve the reliability of the data, included patients must have the baseline total bilirubin (TBIL), prothrombin time (PT) or international normalized ratio (INR). Variables with more than 30 % missing values were removed. The last observation carried forward method was used to fill specific time-point missing values with the last observed value [27]. Additionally, the multivariate imputation by chained equations (MICE) algorithm was employed to address missing values. The MICE algorithm can impute a mix of continuous, binary, unordered categorical, and ordered categorical data [28]. Outliers in continuous variables were managed using the interquartile range (IQR) method, where extreme values were replaced with boundary thresholds to preserve the dataset's integrity and sample size. This approach minimizes distortion from outliers while maintaining statistical power.
2.7Statistical analysisData were analyzed using R software, version 4.2.2 (http://www.r-project.org/). Continuous variables were compared using either a T-test or a Wilcoxon rank-sum test, depending on whether the data was normally distributed. The Chi-square test was used for comparing categorical variables. A predictive model for AKI was built using the training cohort. Although CysC is a valuable serum marker for kidney injury, it is scarce in resource-limited settings [11,29]. Therefore, the current study aimed to construct a new predictive model using routine clinical indices, and CysC was excluded from the model development process. Key risk factors for AKI were identified using the least absolute shrinkage and selection operator (LASSO) regression method, which shrinks less important variables. To balance model complexity and accuracy, predictors were selected based on the one-standard-error (1se) criterion from the LASSO regression. These selected variables were then used to build a logistic regression model, calculating odds ratios (ORs) and their 95 % confidence intervals (95 % CIs). The influence of individual data points was checked using Cook’s distance; values below 0.1 were considered acceptable. Multicollinearity among variables was assessed using the variance inflation factor (VIF). Model performance was evaluated based on discrimination and calibration. Discrimination, measured by the area under the receiver operating characteristic curve (AUROC), indicates the new model’s ability to differentiate high-risk from low-risk individuals. The performance of each eGFR equations in predicting AKI risk was also evaluated and compared using AUROC. The model's calibration, or its ability to accurately predict absolute risk, was assessed with the mean absolute error and calibration curves. Decision curve analysis (DCA) was performed to evaluate the clinical benefit of the new model. The validity of the new model was confirmed through internal validation using enhanced bootstrapping. To address the class imbalance in the dataset, an oversampling method that randomly duplicated instances from the minority class was applied [30]. Additionally, external validation was conducted using the MIMIC-III database. Finally, a nomogram was developed to enhance the usability of the new model. Survival analysis was conducted using the Kaplan-Meier curve and the log-rank test. Statistical significance was set at P < 0.05.
3Results3.1Training cohortA total of 1863 patients were enrolled during the study period. After excluding those who did not meet the inclusion criteria, 1091 patients were included in the analysis, and 122 developed AKI development (10.8 %) (See Table 1). The age of the entire cohort was 54.22 ± 12.47 years, with 74 % being male. The most common etiology of cirrhosis was Hepatitis B virus-related (70.03 %). Other causes included alcohol-related liver disease (n = 82), autoimmune liver diseases (n = 18), hepatitis C virus infection (n = 11), Wilson’s disease (n = 3), and unknown etiology (n = 213). The AKI group experienced a significantly longer hospital stay days (28.02 ± 14.15 vs. 21.05 ± 11.94, P<0.001) and was slightly older (57.22 ± 12.41 vs. 53.86 ± 12.43 years, P=0.006). Patients with AKI exhibited worse liver function, evidenced by higher levels of TBIL (125.05 ± 95.67 vs. 77.87 ± 78.63 µmol/L, P<0.001), direct bilirubin (DBIL) (66.76 ± 55.3 vs. 38.41 ± 45.82 µmol/L, P<0.001), alanine aminotransferase (ALT) (87.47 ± 72.89 vs. 73.07 ± 66.15 U/L, P=0.043), and aspartate aminotransferase (AST) (124.14 ± 94.64 vs. 100.09 ± 83.57 U/L, P=0.009). Furthermore, Child-Turcotte-Pugh scores (CTP) were higher, serum albumin (ALB) levels were lower, and with predominant clinical complications (HE and ascites) in the AKI group than in the non-AKI group, consistent with more severe liver disease (CTP: 9.05±1.08 vs. 8.36±1.17, P<0.001; ALB: 2.97 ± 0.62 vs. 3.12 ± 0.68 g/L, P=0.012). The AKI group also presented with significantly higher white blood cell counts (WBC) (6.26 ± 2.85 vs. 4.85 ± 2.25 × 10⁹/L), higher neutrophils (NE) (4.37 ± 2.35 vs. 3.19 ± 1.83 × 10⁹/L), and abnormal coagulation function (prolonged PT and increased INR). Additionally, the AKI group has a higher rate of HT and DM2. Although baseline sCr and sCr at 48 hours were not different, the maximum sCr was significantly higher in the AKI group, along with elevated CysC (1.1 ± 0.23 vs. 10.82 ± 0.18 mg/L and 1.17 ± 0.33 vs. 1.09 ± 0.3 mg/L, respectively), confirming the development of kidney injury. Moreover, baseline eGFR was generally similar between groups across equations, with the exception of the China-sCr-CysC equation, which estimated a lower eGFR in the AKI group (78.26 ± 18.62 vs 82.83 ± 20.59; P=0.014).
Characteristics of the training cohort.
In the training cohort, sCr alone and several eGFR equations, specifically aMDRD, CKD-EPI1 (using sCr only), CKD-EPI2 (using sCr and CysC), China-sCr-CysC, Mindikoglu-2016, Kalafateli-2017, and GRAIL, were evaluated for their ability to predict AKI. All demonstrated poor discrimination, with AUROC values of 0.52–0.54. These findings suggest that neither sCr alone nor these equations are sufficient to identify patients with cirrhosis at high risk of AKI (Fig. S1).
3.2AKI predictive model developmentLASSO regression was performed to identify variables associated with AKI in cirrhosis (Fig. S2–3). Using the 1se criterion (lambda.1se=0.05), six variables were selected: age, NE, DBIL, CTP, HT, and DM2. The logistic regression model was constructed. The formula of the model was λ=−8.59 + 0.02 × Age + 0.20 × NE + 0.01 × DBIL + 0.43 × CTP + 0.50 × DM2 (Yes=1, No=0) + 0.44 × HT (Yes=1, No=0). The probability of the AKI could be calculated as P=100/(1 + eλ). As presented in Fig. 2 and Table S1, the ORs and 95 % CIs for the risk factors associated with AKI were as follows: age (1.22, 1.10–1.35), DBIL (1.01, 1.00–1.01), CTP (1.53, 1.28–1.83), DM2 (1.65, 1.03–2.64), and HT (1.55, 0.94–2.54). The model indicates that CTP was the strongest predictor of AKI. Although HT did not reach statistical significance, it may still have clinical relevance in assessing AKI risk.
Forest plot of variables in the new model. The forest plot presents ORs and 95 %CIs for risk factors associated with AKI based on multivariate logistic regression analysis. Points represent the adjusted ORs for each variable; horizontal lines indicate the corresponding 95 % CIs; and the vertical dashed line marks the null value (OR = 1.0). P-values are provided for each variable.
Cook's distances were calculated to assess the influence of each observation on the model. No outliers were detected, as none of the Cook's distances exceeded 1 (Fig. S4). The VIF values for age, NE, DBIL, CTP, DM2, and HT were 1.20, 1.13, 1.26, 1.06, 1.12, and 1.18, respectively. All VIF values were less than 10, indicating no multicollinearity among the risk factors included in the model. The new model achieved an AUROC of 0.755 for predicting AKI risk in the training cohort (Fig. 3). The optimal cut-off value of the model, based on the Youden index of the ROC curve, was -1.89. When this threshold was applied, the model demonstrated balanced performance, yielding specificity, sensitivity, and accuracy values of 0.78, 0.63, and 0.76, respectively. The calibration curve (Fig. S5), after 1000 bootstrap repetitions, demonstrated that the predicted probabilities were highly consistent with the observed probabilities (mean absolute error=0.007). To address the significant class imbalance within the cohort, the new model was trained using bootstrap validation (n=1000 resamples) with upsampling method. The model achieved an AUROC of 0.733, similar to the original AUROC, with sensitivity (0.697) exceeding specificity (0.626). DCA curve confirmed the clinical utility of the new model across threshold probabilities ranging from 0.01 to 0.42, consistently providing a higher net benefit than both 'treat all' and 'treat none' strategies (Fig. S6). This supports the model’s robustness in identifying high-risk patients for AKI. To facilitate clinical application, a nomogram was developed (Fig. 4), allowing clinicians to input routine clinical indexes (age, NE, DBIL, and CTP) and the presence of DM2 and HT to calculate an individualized probability of AKI development.
Among 810 cirrhosis patients admitted to the ICU, 31.6 % developed AKI. Gender distribution (69 % vs 66 % male, P=0.421) and mean age (56.0 ± 11.9 vs 55.6 ± 11.5 years, P=0.611) were comparable between AKI and non-AKI groups (See Table 2). The AKI group demonstrated significantly higher mortality (62 % vs 45 %, P<0.001), elevated liver enzymes (ALT: 56.05 ± 39.47 vs 46.41 ± 33.23 U/L, P<0.001; AST: 105.72 ± 69.7 vs 82.68 ± 59.62 U/L, P<0.001), worse synthetic function (ALB: 2.83 ± 0.65 vs 3.05 ± 0.63 g/dL, P<0.001), higher bilirubin (TBIL: 93.09 ± 74.79 vs 48.2 ± 47.84 mg/dL, P<0.001; DBIL: 61.59 ± 52.19 vs 34.65 ± 30.95 mg/dL, P<0.001), impaired coagulation (INR: 1.86 ± 0.57 vs 1.55 ± 0.43, P<0.001; PT: 18.78 ± 4.51 vs 16.49 ± 3.52 s, P<0.001), systemic inflammation (WBC: 9.36 ± 5.52 vs 8.16 ± 4.71 × 10⁹/L, P=0.003; NE: 7.44 ± 4.93 vs 6.18 ± 4.08 × 10⁹/L, P<0.001; lactate: 2.62 ± 1.43 vs 2.28 ± 1.26 mmol/L, P=0.001), higher ascites rate (61 % vs 6 %, P<0.001), hepatic encephalopathy (32 % vs 19 %, P<0.001), infection (14 % vs 9 %, P=0.049), and more severe CTP (10.94 ± 2.26 vs 8.96 ± 1.54, P<0.001) with predominant class C classification (72 % vs 34 %, P<0.001). Although the incidence of AKI was higher in the validation cohort than in the training cohort, the demographic and clinical characteristics of the two cohorts were similar. This discrepancy occurred because the MIMIC-III cohort comprised critically ill patients who were more severe compared to the local hospital cohort. In the validation cohort, the new model attained an AUROC of 0.729 (Fig. 3), outperforming sCr (0.500), the aMDRD equation (0.494), the CKD-EPI1 equation (0.506), and GRAIL (0.538) (Fig. S7). The enhanced bootstrap validation with 1000 repeats showed the reliability of the new model with the AUROC was 0.752 (sensitivity=0.748 and the specificity=0.677).
Characteristics of the validation cohort.
To assess the prognostic value of the model, a subgroup of 231 patients with complete 96-week follow-up was identified from the MIMIC-III database. Patient characteristics are summarized in the Table S2. Predictive scores were calculated using the model's equation. Patients were divided into two groups: a high-level group (scores above the cut-off value -1.89) and a low-level group (below -1.89). As shown in the survival plot (Fig. 5), the high-level group exhibited a poorer 96-week prognosis compared to the low-level group (log-rank test, P=0.016).
Prognostic value of the new model for 96-week survival in cirrhosis. Kaplan-Meier survival curve compares the high- and low-score groups (red, n=158; blue, n=73) over 96 weeks. Survival probability declined more rapidly in the high-score group, and the between-group difference was significant (log-rank test: P=0.016). The accompanying risk table presents the number of participants at risk at baseline and every 24 weeks.
In this study, we developed and externally validated a predictive model for AKI in hospitalized cirrhotic patients who presented with normal sCr levels at admission. The model achieved an AUROC of 0.755 in the training cohort and 0.729 in an independent validation cohort. Furthermore, higher model scores were associated with poorer prognosis among patients with cirrhosis. A score that integrates routinely available admission variables to estimate AKI risk could serve as a practical tool to stratify patients for preventive measures, monitoring, and early intervention, ultimately improving patient care and outcomes.
The model was constructed with six routinely measured variables: age, NE, DBIL, CTP, and the presence of DM2 and HT. In elderly patients, reduced kidney function and a lower renal reserve increase the risk of AKI and restrict the body’s capacity to heal after such damage [31]. In a large cohort study of 25,230 patients conducted at a single tertiary care academic medical center, 69.2 % of individuals aged ≥55 years developed AKI. Furthermore, the prevalence increased with age, affecting 63.4 % of those aged 55–64 years, 70.1 % of those aged 65–74 years, and 72.3 % of those aged ≥75 years [32]. NE in cirrhosis exhibit paradoxical hyperactivity, generating reactive oxygen species (ROS) and NE extracellular traps (NETs), while also displaying immunosuppressive characteristics that impair bacterial clearance. This interplay fosters systemic inflammation, contributing to renal injury and impeding infection control, thereby accelerating AKI progression [33]. Recent studies have identified uNGAL, originally detected in activated NE, as a significant predictor of AKI. Levels exceeding 220μg/mL uNGAL are linked to increased mortality (OR=1.74) [34]. DBIL was reported as a significant risk factor associated with AKI development in cirrhosis [35]. Increased DBIL reflects cholestatic dysfunction and tubular toxicity, contributing to bile cast nephropathy (BCN). Histologically, even modest DBIL elevations (< 10 mg/dL) can induce bile casts, impairing renal function in patients with cirrhosis [36,37]. CTP is a widely used clinical tool for assessing the severity and prognosis of cirrhosis. Studies have shown that CTP is associated with the renal artery resistive index (RARI), an important non-invasive predictor of pre-renal AKI, with higher CTP values corresponding to higher RARI levels [38]. Consequently, CTP has been identified as a risk factor for AKI development in patients with cirrhosis [35,39]. Although HT and DM2 were not identified as risk factors in a meta-analysis of 14 studies, both were considered important predictors of AKI development in our study. HT and DM2 have been recognized as risk factors for AKI in various diseases [40,41]. This discrepancy could be explained by differences in the studied cohorts, as our study focused on cirrhosis patients with normal sCr. Ascites can exacerbate AKI risk by promoting portal hypertension and increasing intra-abdominal pressure, with approximately 50 % of patients with cirrhosis and ascites developing AKI [20,42]. In our study, ascites was also significantly associated with AKI development. However, because ascites assessment is already incorporated into the CTP score, it was not included as a separate risk factor in the new model. Collectively, these findings underscore the complex, multifactorial pathogenesis of AKI in cirrhosis and demonstrate the utility of our model in integrating routinely measured clinical parameters to enhance risk stratification.
Compared to previous studies, both Zheng et al. and Tian et al. employed machine learning strategies in their respective analyses. Zheng developed an XGBoost model using baseline data from 950 patients with cirrhosis in the MIMIC-III cohort, achieving an AUROC of 83.20 % [43]. Tian's study utilized a larger dataset, comprising 1149 patients from the eICU-CRD cohort as a training set and 789 patients from the MIMIC cohort as a validation set, to construct a new scoring model using advanced machine learning algorithms. This model attained AUROC values of 80.50 % and 77.20 % in the training and validation sets, respectively [44]. Although machine learning models demonstrate promise, their complexity and "black-box" nature often hinder their clinical applicability. In contrast, Patidar et al. adopted a more traditional approach, employing multivariate logistic regression with data from 397 patients with cirrhosis. Their model, which incorporated WBC, sCr, and INR, achieved an AUROC of 77.00 %, decreasing to 70.00 % in an external validation set. This study identified that cirrhosis-related AKI is primarily caused by elevated endotoxin levels resulting from bacterial translocation, leading to vasodilation and glomerular dysfunction; sCr was the most significant predictor of AKI risk [45]. Our study found that while WBC and NE (a component of WBC) are important markers of systemic inflammation, NE provides a more precise reflection of the inflammation. Although baseline INR was higher in patients with AKI, this parameter was excluded from our model after feature selection due to its indirect impact on kidney function. In contrast to cohorts studied in previous studies, our exclusion criteria specifically omitted patients with elevated baseline sCr. We postulated that marked sCr elevations would more likely prompt immediate clinical intervention, whereas individuals with normal baseline sCr levels might be at higher risk of delayed detection of renal dysfunction during hospitalization.
Our comparison of different formulas revealed that eGFR is less reliable in predicting cirrhosis-related AKI, a gap inadequately addressed in existing literature. The currently available formulas, such as MDRD and CKD-EPI, were originally developed for general populations and do not account for the patients with cirrhosis. Even equations specifically adjusted for cirrhosis, including Mindikoglu-2016 [24], Kalafateli-2017 [25], and GRAIL [26], demonstrated poor performance in predicting AKI risk among cirrhotic patients in our study. These findings are consistent with the previous study [46]. In contrast, our newly developed model showed superior performance compared to eGFR and sCr, highlighting the need for more precise approaches to assess renal function in cirrhosis. However, the underlying mechanisms of renal dysfunction in cirrhosis remain insufficiently explored and warrant further investigation. The validation cohort from the MIMIC-III database further demonstrated the robustness and external applicability of our model, confirming that it could generalize across diverse clinical settings. This is particularly important given the heterogeneity of cirrhosis etiologies and severities in real-world clinical practice.
This study has several limitations. First, although class imbalance in our retrospective dataset was unavoidable, internal validation using an upsampling approach may partly mitigate this issue and support the model’s reliability. Additionally, the study focused on cirrhosis with normal sCr, which may limit generalizability to broader populations. Given the heterogeneity of cirrhosis, future studies should assess whether specific subgroups (defined by etiology or by severity of liver dysfunction) exhibit differential risk profiles for AKI. AKI is commonly categorized into three subtypes: prerenal AKI, acute tubular necrosis (ATN), and HRS, each with distinct prevalence and mortality rates [5] However, the etiology of AKI was not distinguished in our study, which may introduce bias. To mitigate this limitation, future research should investigate the model's predictive value across AKI subtypes in cirrhosis. Another challenge concerns defining baseline sCr, which is critical for diagnosing AKI in cirrhosis. However, achieving a consensus on this measure is difficult. Although we used admission sCr as the baseline, consistent with clinical guidelines and previous studies [18,47], this choice may introduce bias. Furthermore, we were unable to compare our model with emerging kidney biomarkers for risk assessment, such as uNGAL, NAG and KIM-1. Additionally, we did not analyze patients’ treatment regimens, although differences in the dosage and timing of diuretics, albumin, and terlipressin could influence AKI risk. Nonetheless, all participants had normal baseline sCr, helping to align patients’ baseline status. Moreover, the treatment protocols at our center adhered to clinical guideline [19], allowing the study to accurately reflect the real-world clinical practice. Future prospective studies are needed to confirm and extend our findings.
5ConclusionsIn conclusion, this study presents a robust predictive model for AKI in hospitalized cirrhotic patients with normal sCr at admission. The model's consistent performance across different cohorts underscores its potential utility as a clinical decision-support tool and may help mitigate diagnostic delays and inaccuracies in AKI risk assessment.
Data availableThe training dataset generated and analyzed during the current study is not publicly available due to the inclusion of identifiable patient clinical information but is available from the corresponding author, upon reasonable request and subject to institutional and ethical approvals. The validation dataset used in this study is available from the MIMIC-III database at https://physionet.org/content/mimiciii/ (access requires credentialed approval). The R code used for data extraction from the MIMIC-III database is accessible at https://github.com/AlongLin/Renal_AKI.git.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation processDuring the preparation of this work, the authors used DeepSeek-R1 AI model to improve the clarity, coherence, and fluency of the English writing. After using this tool, the authors thoroughly reviewed and edited the content as needed and take full responsibility for the content of the published article.
CRediT authorship contribution statementShenglong Lin: Conceptualization, Formal analysis, Validation, Writing – review & editing. Wenhai Ke: Formal analysis, Writing – review & editing. Rong Wang: Formal analysis, Writing – review & editing. Huaxi Ma: Data curation, Writing – review & editing. Ziyuan Liao: Data curation, Writing – review & editing. Jianyu Wang: Data curation, Writing – review & editing. Minghua Lin: Writing – review & editing. Haibing Gao: Conceptualization, Writing – review & editing.
None.
