Abstract
Objective
Cirrhosis with acute kidney injury is associated with a high mortality rate, particularly in patients receiving continuous renal replacement therapy. This study aimed to develop a nomogram to predict in-hospital mortality in patients with cirrhosis and acute kidney injury receiving continuous renal replacement therapy.
Methods
A retrospective study was conducted using the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Patients with cirrhosis with acute kidney injury who underwent continuous renal replacement therapy were identified. Predictors were selected using least absolute shrinkage and selection operator regression, and a multivariable logistic regression model was developed. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis. Internal validation was performed via bootstrapping.
Results
The following three independent predictors of in-hospital mortality were identified in the 452 eligible patients: international normalized ratio at continuous renal replacement therapy initiation, presence of sepsis, and vasoactive drug use. The nomogram showed strong discrimination (AUC = 0.828, 95% confidence interval: 0.790–0.866) and consistent internal validation (AUC = 0.825, 95% confidence interval: 0.781–0.870). Calibration and decision curve analyses indicated good agreement and clinical usefulness.
Conclusions
A nomogram incorporating international normalized ratio, sepsis, and vasoactive drug use may help estimate short-term mortality risk in patients with cirrhosis and acute kidney injury receiving continuous renal replacement therapy. As the analysis was based on single-center data, and no external validation was performed, the findings should be interpreted cautiously and verified in future studies.
Keywords
Introduction
Cirrhosis is a progressive hepatic disorder characterized by the replacement of healthy liver parenchyma with fibrotic tissue, resulting in compromised liver function and a spectrum of complications. This condition represents a significant global health challenge, contributing to increased morbidity and mortality rates and imposing a considerable economic burden on healthcare systems because of frequent hospitalizations and complex management needs.1–3 Current therapeutic approaches primarily include lifestyle modifications, pharmacological treatments, and liver transplantation. However, challenges such as delayed diagnosis and limited therapeutic options persist, necessitating further research to enhance the understanding and management of cirrhosis. 4
Acute kidney injury (AKI) is one of the most common and serious complications in patients with cirrhosis. It occurs in 20%–50% of hospitalized patients, with the incidence rising to 30%–80% among those admitted to intensive care units (ICUs).5–7 This underscores the critical importance of AKI in the management of cirrhosis. Continuous renal replacement therapy (CRRT) is frequently used to mitigate the adverse effects of AKI, particularly in patients with severe complications such as hyperkalemia, pulmonary edema, and metabolic acidosis. 8 It is estimated that approximately 6% of patients with AKI in ICUs require CRRT, 9 with mortality rates ranging from 50% to70%.10,11 Furthermore, some studies suggest that individuals undergoing CRRT have an increased risk of adverse outcomes, including mortality, compared with those who do not receive this treatment.12,13 CRRT is a blood purification technique primarily used in the management of AKI, systemic inflammatory response syndrome, and multiple organ dysfunction syndrome, among other critical conditions. In patients with AKI receiving CRRT, mortality rates are particularly high during the early stages of treatment. 14 Consequently, the early identification and management of patients with cirrhosis and AKI who are at an increased risk of mortality during CRRT are crucial for improving survival outcomes.
Although general severity scoring systems, such as the model for end-stage liver disease (MELD) 15 and acute physiology and chronic health evaluation (APACHE II), have been used to predict prognosis in patients with advanced liver disease or critical illnesses, they were not specifically designed for patients with cirrhosis and AKI undergoing continuous CRRT. The clinical profile of this subgroup is unique, as the combination of severe liver dysfunction, AKI, and hemodynamic instability significantly elevates the risk of in-hospital mortality. These factors are inadequately represented in traditional scoring models, thereby limiting their predictive accuracy in this context. Consequently, there is a distinct need for a straightforward and practical model that can accurately identify patients with high-risk cirrhosis requiring CRRT. This study aimed to develop and internally validate a nomogram based on routinely available clinical variables to predict in-hospital mortality in patients with cirrhosis and AKI undergoing CRRT, thereby facilitating timely and individualized clinical decision-making.
Methods
This retrospective cohort study is reported in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis statement.16,17 This study was conducted in accordance with the principles of the Declaration of Helsinki of 1975, as revised in 2024.
Database
This study utilized data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (version 2.2), a comprehensive critical care repository that is publicly accessible and based in the United States. 18 The MIMIC-IV database aggregated clinical information from over 190,000 patients and 450,000 hospital admissions at Beth Israel Deaconess Medical Center between 2008 and 2019. It includes a substantial cohort of critically ill patients, with a particular emphasis on individuals with cirrhosis and AKI undergoing CRRT. This database is exceptionally well-suited to our study objectives, as it provides extensive and detailed clinical data, including patient demographics, laboratory assessments, medication records, vital signs, surgical interventions, disease diagnoses, pharmacological treatments, and follow-up survival. Such comprehensive data facilitate an in-depth analysis of the factors associated with in-hospital mortality. The large and diverse sample size of the database, combined with the richness of the clinical data, makes this database an ideal resource for developing a robust and generalizable predictive nomogram to assess mortality risk in this specific and high-risk patient population.
The institutional review boards of Beth Israel Deaconess Medical Center (Boston, MA) and the Massachusetts Institute of Technology (Cambridge, MA) authorized the creation of the MIMIC-IV database, resulting in a waiver of written informed consent for this investigation. Author Lu-Huai Feng completed the National Institutes of Health online training course (Certification Number: 35897462) to gain access to the MIMIC-IV database.
Participants
All adult patients with a diagnosis of cirrhosis and AKI who received CRRT during ICU admission were consecutively identified from the MIMIC-IV database (version 2.2). The eligibility criteria were as follows: (a) patients diagnosed with cirrhosis who underwent CRRT during their hospital stay, including those who also received intermittent hemodialysis (IHD) or peritoneal dialysis (PD) before or after CRRT and (b) serum creatinine fluctuations consistent with an AKI diagnosis. Exclusion criteria were as follows: (a) age <18 years and (b) preexisting chronic kidney disease (CKD) stage 5, including patients with nondialysis CKD stage 5 and those receiving maintenance hemodialysis, as identified according to International Classification of Diseases, Ninth and Tenth Revision (ICD-9/10) codes recorded in the MIMIC-IV database.
Data extraction
Data extraction was performed utilizing structured query language query tools (V.1.13.1). With consideration for clinical significance, established scientific insights, and predictive variables delineated in previously published literature,19–21 information about the following parameters was systematically collected at the time closest to CRRT initiation.
Demographic variables. These variables included race, age, sex, mean diastolic blood pressure (DBP), mean systolic blood pressure (SBP), and mean blood pressure (MBP). Comorbid conditions. Comorbid conditions were assessed, including CKD, acute coronary syndrome (ACS), ascites, diabetes, chronic obstructive pulmonary disease (COPD), heart failure, shock, malignancy, hepatic encephalopathy, bacterial infections, sepsis, baseline liver diseases, AKI stage, hyponatremia, and MELD score. All comorbid conditions, including cirrhosis, were identified according to ICD-9/10 codes recorded in the MIMIC-IV database. Medical history. Medical history included the use of diuretic medications, such as spironolactone, furosemide, and hydrochlorothiazide; vasoactive agents, including phenylephrine, norepinephrine, epinephrine, dopamine, and dobutamine; and CRRT, as per patient documentation. Laboratory findings. Laboratory data were collected at the time closest to the onset of AKI and included albumin (ALB), aspartate aminotransferase (AST), alanine aminotransferase (ALT), total bilirubin (TB), serum creatinine, hemoglobin, osmotic pressure, blood urea nitrogen (BUN), international normalized ratio (INR), prothrombin time (PT), and glucose levels.
Missing data handling
The MIMIC-IV database contained missing data across various laboratory parameters. Within our cohort, missing values were observed exclusively for continuous variables, whereas the records of all categorical variables, including comorbidities and demographic information, were complete. The extent of missing data for individual variables ranged from 0.3% to 4.3% and was specifically as follows: AST, 0.3%; ALT, 0.5%; INR, 4.3%; glucose, 2.6%; BUN, 3.5%; and creatinine, 1.6%. To address this issue, we employed mean imputation for continuous variables with a normal distribution and median imputation for variables with a skewed distribution. This methodology is corroborated by previous studies, 22 which indicate that when the proportion of missing data is low (<10%), simple imputation techniques such as mean or median imputation yield results comparable to those obtained using more sophisticated methods, including multiple imputation.
Definitions and outcomes
The primary objective of this study was to assess mortality following CRRT in patients with cirrhosis who developed AKI during ICU admission. AKI was defined based on the Kidney Disease: Improving Global Outcomes (KDIGO) criteria. 23 Although urine output data were available for 91% of patients and could be used to determine the AKI status according to the KDIGO criteria, urine volume may not accurately reflect glomerular filtration because of the potential influence of diuretic use. Therefore, AKI was determined based on changes in serum creatinine levels. In patients who did not receive diuretics, AKI classification using serum creatinine alone was identical to that obtained using the combined serum creatinine and urine output criteria. Serum creatinine level at ICU admission was defined as the reference value, as the true baseline creatinine level prior to ICU admission was unavailable in the MIMIC-IV database. The estimated glomerular filtration rate (eGFR) at ICU admission was calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) creatinine equation, as recommended by the KDIGO guideline. 23 The administration of vasoactive drugs, diuretics, and aminoglycosides was recorded as any instance of their use for any indication during the patient’s ICU stay.
Statistical analyses
Statistical analyses were performed using Statistical Package for Social Sciences (SPSS) software version 26.0 (IBM, Armonk, NY, USA) and R software version 4.2.1. During the analysis, two-sided p-values were applied, with statistical significance determined at a threshold of p < 0.05. Categorical variables were presented as percentages, whereas continuous variables were described as mean ± SD, median, or ranges, depending on the normality of their distribution. The chi-square test was employed for the evaluation of categorical variables, whereas continuous variables were analyzed using either the unpaired t-test or Wilcoxon rank-sum test, according to their distributional properties.
To improve forecast accuracy and enhance result interpretability, this study utilized the least absolute shrinkage and selection operator (LASSO) regression analysis for variable selection. The LASSO model was tuned using 10-fold cross-validation to determine the optimal lambda (λ) value that minimized the mean cross-validated error, thereby preventing overfitting and stabilizing the estimates. All variables presented in Table 1 were incorporated into the LASSO model. This approach effectively identifies the most informative predictors, mitigates the effects of multicollinearity, and reduces the likelihood of overfitting. Variables retained by the LASSO model were subsequently subjected to univariate logistic regression analysis to assess their association with in-hospital mortality. Variables demonstrating a p-value <0.05 in these preliminary univariate logistic analyses were further evaluated using multivariable logistic regression, employing a backward stepwise selection method. Following the development of a predictive model through multivariable logistic regression, a clinical prediction nomogram was constructed.
Clinical and demographic characteristics of patients.
The p-value represents a pairwise comparison between the survivor and in-hospital death groups.
Time from diagnosis of AKI to initiation of CRRT therapy
AKI: acute kidney injury; ACS: acute coronary syndrome; COPD: chronic obstructive pulmonary disease; CKD: chronic kidney disease; ALB: albumin; AST: aspartate aminotransferase; ALT: alanine aminotransferase; TB: total bilirubin; INR: international normalized ratio; PT: prothrombin time; MELD: model for end-stage liver disease; SBP: systolic blood pressure; DBP: diastolic blood pressure; MBP: mean blood pressure; eGFR: estimated glomerular filtration rate; ICU: intensive care unit; CRRT: continuous renal replacement therapy; T2DM: type 2 diabetes mellitus; MELD-Na: model for end-stage liver disease–sodium.
The performance of the nomogram was evaluated using discrimination and calibration metrics, with internal validation conducted using the bootstrap method involving 1000 resamples.17,24 Discrimination was quantified by the concordance area under the receiver operating characteristic (ROC) curve (AUC), which ranges from 0.5, indicating no discriminative ability, to 1.0, indicating perfect predictive accuracy. Calibration was evaluated through a visual calibration plot that compared the predicted probabilities with the actual occurrences of in-hospital mortality. Additionally, the clinical utility of the nomogram was assessed using decision curve analysis (DCA), which facilitates the evaluation of the net benefit derived from the predictors and models. 25
Results
Characteristics of patients
As shown in Figure 1, the flowchart illustrates the patient selection process from the MIMIC-IV database. A total of 452 patients with cirrhosis and AKI who received CRRT were included in the final analysis, comprising 214 patients in the survivor group and 238 in the in-hospital death group. Table 1 presents a comprehensive overview of the baseline demographic and clinical characteristics of the two groups. The median age and sex distributions were comparable between the two groups. Alcoholic liver disease (either alone or in combination with viral hepatitis) and viral hepatitis were the predominant etiologies of cirrhosis in both groups, with no statistically significant differences observed. Most patients in both groups progressed to AKI grade 3 during their ICU stay. The in-hospital death group demonstrated a higher incidence of sepsis and shock than the survivor group, whereas CKD was more prevalent in the survivor group. The use of vasoactive drugs was significantly more common in the in-hospital death group. Laboratory assessments indicated that the in-hospital death group exhibited elevated INR and PT, increased bilirubin levels, and decreased hemoglobin and glucose levels. Additionally, the in-hospital death group experienced severe liver disease, as evidenced by significantly higher MELD and MELD-sodium (MELD-Na) scores than those of the survivor group. No significant differences were noted between the groups in terms of other comorbidities and laboratory parameters. Across the overall cohort, diuretics were administered to 314 patients (69.5%), whereas 57 patients (30.5%) did not receive diuretics during their ICU stay.
Flowchart of patient selection from the MIMIC-IV database. MIMIC-IV: Medical Information Mart for Intensive Care IV.
Selected predictors for the model
Using LASSO regression analysis (Figure 2(a) and (b)), we determined an optimal λ value of 0.02824472, which enabled the reduction of 35 variables (Table 1) to 7 potential predictors of in-hospital mortality. These predictors were subsequently examined through univariate and multivariate logistic regression models, as detailed in Table 2. The multivariate analysis revealed that INR (odds ratio (OR): 2.82, 95% confidence interval (CI): 1.94–4.10, p < 0.001), sepsis (OR: 2.68, 95% CI: 1.69–4.24, p = 0.003), and use of vasoactive drugs (OR: 7.62, 95% CI: 4.24–13.68, p < 0.001) were independently correlated with in-hospital mortality.
Predictor selection using the LASSO binary logistic regression model. LASSO: least absolute shrinkage and selection operator.
Logistic analysis showing the association of variables with in-hospital mortality.
OR: odds ratio; CI: confidence interval; INR: international normalized ratio; PT: prothrombin time; ACS: acute coronary syndrome; N/A: not applicable.
Prediction curve of the probability of in-hospital death
Following the final multivariate logistic regression analysis, a nomogram was constructed to estimate the risk of in-hospital mortality (Figure 3). This nomogram integrated three significant predictors: the INR measured closest to CRRT initiation, sepsis, and the use of vasoactive drugs. Each predictor was assigned a score on the upper points scale based on its relative contribution to the outcome. The sum of these individual scores yielded a total score, which was subsequently mapped onto the bottom probability scale to determine the estimated probability of in-hospital mortality. This predictive model facilitates personalized risk assessment for clinical use, thereby enhancing decision-making processes in patient management.
A nomogram was developed for predicting the risk of in-hospital mortality in patients with cirrhosis and acute kidney injury receiving continuous renal replacement therapy, based on the results of multivariate logistic regression analysis. Based on the corresponding variable, a score can be obtained, and the scores are added to yield a total score. Finally, the total score can be used to calculate the corresponding in-hospital mortality risk. In our example, the red dots for INR, sepsis, and use of vasoactive drugs represent the values of the patient’s respective variables. A patient with sepsis was treated with vasopressors and had an INR of 1.8. The total score was 21 + 51 + 77 = 114, corresponding to an in-hospital mortality risk of 71.1%. INR: international normalized ratio.
Performance of the nomogram
The performance of the nomogram was rigorously evaluated to ensure its reliability and predictive accuracy. The Hosmer–Lemeshow test yielded a p-value of 0.536, suggesting an absence of significant overfitting within the prediction model. The model demonstrated substantial discriminative capability, with an uncorrected AUC of 0.828 (95% CI: 0.790–0.866) and an internally validated AUC of 0.825 (95% CI: 0.781–0.870). To further compare its predictive performance, ROC analyses were conducted for the nomogram and each individual predictor. The AUC values for INR, sepsis, and use of vasoactive drugs were 0.718 (95% CI: 0.671–0.764), 0.671 (95% CI: 0.627–0.714), and 0.719 (95% CI: 0.682–0.757), respectively, as shown in Figure 4. These metrics underscore the robustness of the predictive capability of the model.
Receiver operating characteristic (ROC) curves for the nomogram and individual predictors of in-hospital mortality. The nomogram demonstrated the highest discriminative ability (AUC = 0.828, 95% CI 0.790–0.866) compared with INR (AUC = 0.718, 95% CI 0.671–0.764), sepsis (AUC = 0.671, 95% CI 0.627–0.714), and use of vasoactive drugs (AUC = 0.719, 95% CI 0.682–0.757). The diagonal dashed line indicates no discrimination. AUC: area under the receiver operating characteristic curve; INR: international normalized ratio; CI: confidence interval.
The calibration curve (Figure 5) corroborated the model’s accuracy, demonstrating a strong concordance between predicted and observed probabilities of in-hospital mortality. The calibration curve’s slope was 0.99, with a mean squared error of 0.008 and a Brier score of 0.16, further emphasizing the model’s reliability and precision in estimating the risk of in-hospital death.
Calibration curves for the nomogram predicting the risk of in-hospital mortality in patients with cirrhosis and acute kidney injury receiving continuous renal replacement therapy. The x-axis represents the nomogram-predicted probability of in-hospital mortality, and the y-axis represents the actual probability of in-hospital mortality. The diagonal dashed line represents a perfect prediction by an ideal model. The solid line shows the performance of the nomogram, where a closer fit to the diagonal line indicates better predictive accuracy.
Subgroup analyses
To assess the robustness and generalizability of the nomogram, we conducted subgroup analyses across a range of clinically relevant patient characteristics, as illustrated in Figure 6. The model consistently exhibited stable discriminative performance across all examined subgroups, with AUC values exceeding 0.75. These subgroups exhibited variations in comorbidities (CKD, diabetes, and heart failure), clinical conditions (sepsis, shock, and the administration of vasoactive drugs or diuretics), demographic factors (age, sex, and race), and baseline liver disease etiologies. No statistically significant interactions were observed between the subgroup variables and model performance (all p-values > 0.05), indicating that the nomogram’s discriminative capacity was not significantly affected by underlying patient characteristics. These findings suggest that the nomogram maintains stable predictive accuracy across diverse populations of critically ill patients with cirrhosis undergoing CRRT.
Discriminative performance of the nomogram across patient subgroups. AUC values with 95% confidence intervals (CIs) are shown for each subgroup, stratified by key clinical characteristics including comorbidities, liver disease etiology, medication use, demographic variables, and laboratory findings. Horizontal bars represent the 95% CI. p-values indicate whether model performance differed significantly across subgroups; no statistically significant interaction was observed (all p > 0.05), suggesting consistent discrimination of the nomogram across diverse patient profiles. The subgroup with biliary cirrhosis was excluded from analysis due to insufficient sample size. AUC: area under the receiver operating characteristic.
Clinical use of the nomogram
The clinical utility of the nomogram was evaluated using DCA, as illustrated in Figure 7. The decision curve delineates the net benefit associated with employing the nomogram for predicting in-hospital mortality across a range of threshold probabilities. Notably, the nomogram yields substantial clinical benefit when the threshold probability is between 0.05 and 0.58. This finding suggests that, within this range, the nomogram can effectively assist clinicians in identifying patients at an increased risk for in-hospital mortality.
Decision curve analysis for the nomogram predicting the risk of in-hospital mortality in patients with cirrhosis and acute kidney injury receiving continuous renal replacement therapy. The x-axis represents the high-risk threshold for predicting the risk of in-hospital mortality, and the y-axis represents the net benefit.
Discussion
In this study, we developed and internally validated a nomogram designed to predict in-hospital mortality among patients with cirrhosis undergoing CRRT following AKI. The nomogram integrates three significant predictors, including INR, sepsis, and the use of vasoactive drugs, identified through a rigorous selection process using LASSO regression and multivariate logistic regression. Our nomogram exhibited strong discriminatory power, with an AUC of 0.828 (95% CI: 0.790–0.866), and internal validation confirmed similar performance (AUC = 0.825, 95% CI: 0.781–0.870), thereby confirming its robustness in predicting in-hospital mortality. Furthermore, calibration analysis and DCA demonstrated good calibration and suggested clinical utility, supporting its value for risk stratification within relevant probability thresholds. To the best of our knowledge, no existing predictive models specifically address mortality risk in patients with cirrhosis undergoing CRRT. This highlights the novelty and clinical significance of our nomogram, which addresses a critical gap in the literature. The nomogram may provide clinicians a useful and individualized tool to estimate the probability of in-hospital mortality among patients with cirrhosis undergoing CRRT and guide clinical decision-making in critical care settings.
This study contributes to the current understanding of the complex relationship between AKI and cirrhosis, a domain that remains underexplored in the existing literature. Although previous studies have documented the adverse effects of AKI on patients with cirrhosis,5,26,27 our findings add to the existing evidence by highlighting specific predictors of in-hospital mortality among patients with cirrhosis undergoing CRRT. Notably, our study provides evidence that factors such as INR, sepsis, and the need for vasoactive drugs are independent predictors associated with increased in-hospital mortality risk. Coagulation abnormalities, indicated by an elevated INR, are well documented in cirrhosis due to compromised liver function and have been associated with increased mortality rates in hepatic and critically ill populations. Similarly, sepsis is a major contributor to mortality in ICU patients, particularly those with liver cirrhosis and renal failure, where it is often accompanied by systemic inflammation and multiorgan dysfunction. Finally, the use of vasoactive drugs, which often indicates hemodynamic instability, has been associated with adverse outcomes in critically ill patients. These medications are typically administered to manage hypotension and shock, conditions that are closely associated with increased mortality rates. These considerations emphasize the clinical importance of the nomogram in supporting clinical assessment and identifying high-risk patients in the ICU.
Nevertheless, we were unable to obtain comprehensive data on CRRT, including treatment duration, frequency, and dosage. This limitation restricted our ability to assess the prognostic significance of these factors within our cohort. These parameters are clinically relevant, as the intensity of CRRT may be associated with patient outcomes and often reflects the severity of illness. The lack of such data constitutes a fundamental limitation inherent to secondary analyses utilizing the MIMIC-IV database. Regarding the assessment of liver disease severity, MELD and MELD-Na scores were considered as candidate variables but were not retained following LASSO selection. This outcome is likely attributable to multicollinearity with INR, a central component of the MELD score, which emerged as an independent predictor. This finding suggests that individual objective parameters sometimes serve as proxies for composite scores in multivariable models, particularly in critically ill populations where overlapping indices can reduce incremental predictive value. Because MELD and MELD-Na were not included in the final model, a direct comparison of their discrimination performance with that of the nomogram was not performed. The predictive information provided by these composite scores was likely captured by INR and other retained variables in the model. Furthermore, chronic liver failure–sequential organ failure assessment (CLIF-SOFA) and APACHE II scores could not be calculated in this study, as several necessary parameters such as hepatic encephalopathy grade, detailed chronic health data, and specific physiological variables are not consistently available in the MIMIC-IV database. This limitation restricted our ability to compare the nomogram with other established prognostic models for critically ill patients with cirrhosis. Additionally, we recognized that the Child–Pugh classification could not be evaluated in this study. Calculation of the Child–Pugh score requires data on the extent of ascites and hepatic encephalopathy, which are not systematically documented in the MIMIC-IV database. Given that the Child–Pugh score remains a critical prognostic instrument in cirrhosis, the inability to incorporate it constitutes a study limitation. Despite the absence of detailed CRRT treatment data and Child–Pugh scores, our nomogram appears to retain potential clinical utility by relying on objective and routinely available predictors. Future prospective studies incorporating granular CRRT information and standardized liver function assessments, including the Child–Pugh score, can further improve risk stratification and model performance in this high-risk cohort.
The implications of our findings may be meaningful for clinical practice. Through the identification of specific risk factors linked to increased mortality risk in patients with cirrhosis and AKI, the nomogram’s performance was rigorously validated to ensure its clinical applicability. An AUC of 0.828 in the original dataset demonstrated good discriminatory ability, supporting the model’s effectiveness in stratifying in-hospital mortality risk. Internal validation, with a similar AUC of 0.825, suggests that the model’s generalizability and stability differ across patient groups. Further validation was achieved using the calibration curve, which showed near-perfect concordance between predicted and observed outcomes (slope = 0.99 and Brier score = 0.16), indicating good calibration and potential clinical utility in mortality risk prediction and its potential utility in accurate clinical decision-making. Moreover, bootstrapping with 1000 resamples confirmed the nomogram’s reliability, with consistent AUC results across internal datasets. These findings collectively suggest that the model demonstrates robust internal validity and appears applicable for patients with cirrhosis admitted to the ICU undergoing CRRT. Although the model is not intended to guide or dictate CRRT initiation, it may provide clinical value by identifying patients at higher risk of adverse outcomes. This information can assist in determining the appropriate intensity of monitoring and supportive care, facilitate communication with families regarding prognosis, and aid in risk stratification for prognostic assessment and clinical decision-making.
The primary and concluding rationale for utilizing the nomogram is centered on the need to evaluate the specific requirement for additional treatment or care on an individual basis. However, the effectiveness of risk prediction, encompassing aspects of discrimination and calibration, does not fully reflect the clinical implications associated with a particular level of discrimination or degree of miscalibration.16,28,29 To gain a deeper understanding of its practical applicability, we employed DCA to examine the impact of the nomogram on clinical decision-making. DCA provides insights into the net clinical benefit by quantifying the trade-off between true positives and false positives at various threshold probabilities. In this context, a threshold probability represents the predicted risk level at which clinicians might reasonably choose to intensify management, initiate closer monitoring, or reconsider treatment strategies for a patient. This approach helps clinicians assess whether using the nomogram can provide added value compared with default strategies of treating all or none. In this setting, “treating” does not refer to a specific therapeutic intervention but rather to broader clinical management decisions, such as initiating or intensifying supportive care, adjusting CRRT parameters, and implementing closer hemodynamic and laboratory monitoring for patients with a higher predicted mortality risk. Our findings indicate that within a threshold probability range of 0.05–0.58, the nomogram provides substantial net clinical benefit. This suggests that its application may assist clinicians in the early identification of high-risk patients, designing of individualized treatment plans, and optimization of resource allocation in the intensive care setting. By enabling individualized risk assessments, the nomogram supports a more personalized approach to patient evaluation and management. Consequently, the model can help improve risk stratification or clinical decision-making, which might, in turn, benefit outcomes. However, this conclusion requires prospective confirmation.
The limitations of this study necessitate thorough consideration. As this was a retrospective observational study, it is inherently subject to potential selection bias, information bias, and unmeasured confounding, which should be considered when interpreting the results. First, the generalizability of our findings may be limited due to the absence of external validation and the single-institution scope of the MIMIC-IV database. Although the MIMIC-IV database encompasses data from multiple ICUs, all patients were treated within a single hospital system, which may restrict the model’s applicability to broader populations. Furthermore, the lack of comparable publicly available datasets prevented us from conducting external validation. Second, the MIMIC-IV database lacks comprehensive standardized information regarding CRRT initiation, such as whether the therapy was initiated according to predefined protocols or based on clinician discretion, which can influence the comparability of the outcomes. Additionally, the database does not provide structured data on cirrhosis stage or AKI etiology, thereby limiting the possibility of subgroup analyses based on these critical prognostic factors. Furthermore, although the overall proportion of missing data was minimal, even small amounts of missing information can introduce bias in studies involving critically ill populations. These limitations may have reduced the specificity of our findings and limited the applicability of the model in clinical settings where treatment criteria or disease staging are crucial for risk stratification. Finally, the absence of novel AKI biomarkers, such as neutrophil gelatinase–associated lipocalin (NGAL), kidney injury molecule-1 (KIM-1), and interleukin-18 (IL-18), in the database limited the evaluation of their potential prognostic significance. Future investigations must aim to include larger and more diverse cohorts, conduct external and prospective validation, and incorporate additional biomarker data to verify and improve the clinical applicability of the nomogram in predicting in-hospital mortality among patients with cirrhosis and AKI undergoing CRRT.
Conclusion
The nomogram developed in this retrospective cohort study may serve as a useful tool for estimating in-hospital mortality among patients with cirrhosis and AKI receiving CRRT. It demonstrated strong internal predictive performance and may assist clinicians in identifying patients at higher risk, thereby supporting more informed clinical decision-making and resource allocation in critical care. Nevertheless, given the retrospective and single-center nature of this study, further external validation is warranted to confirm its generalizability and clinical utility.
Footnotes
Acknowledgments
Not applicable.
Author contributions
Study design: Zhenhua Yang, Lu-Huai Feng, and Tianbao Liao; data collection: Yang Lu, Xuefei Zhou, Tingting Su, and Tianbao Liao; manuscript preparation: Tingting Su and Tianbao Liao; data analysis and interpretation: Zhenhua Yang, Lu-Huai Feng, Tingting Su, and Tianbao Liao. All authors confirm that they contributed to manuscript review, revised it critically for important intellectual content, and have read and approved the final draft for submission.
Consent for publication
Not applicable.
Data availability
The data analyzed during the current study are available from the corresponding author on reasonable request.
Declaration of conflicting interest
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Ethics approval and consent of participate
Approval for the establishment of the MIMIC-IV database was granted by the Institutional Review Boards of Beth Israel Deaconess Medical Center (Boston, MA) and Massachusetts Institute of Technology (Cambridge, MA), leading to a waiver of informed consent participates and ethics approval for this study. All patient data were fully deidentified, and no individual can be identified in any way. Author Lu-Huai Feng obtained certification (certification number 35897462) upon completion of the National Institutes of Health online training course to access the MIMIC-IV database (version 2.2).
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
