This multicentre retrospective cohort study integrated data from three ICU databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV v3.1), the eICU Collaborative Research Database (eICU-CRD v2.0), and the Shenzhen People's Hospital ICU (SZPH-ICU; a tertiary academic medical centre in China). The MIMIC-IV database contains data on over 90,000 ICU admissions at Beth Israel Deaconess Medical Centre, Boston, Massachusetts, USA, between 2008 and 2022 (19). The eICU-CRD is a multicentre database comprising over 200,000 ICU admissions from 208 hospitals across the United States between 2014 and 2015 (20). The SZPH-ICU database includes over 8,000 ICU admissions recorded between 2014 and 2023. All databases provided demographic characteristics, vital signs, treatment measures, nursing records, and laboratory data. The MIMIC-IV database was fully anonymized, with no personally identifiable information retained, and ethical approval was obtained from the Institutional Review Boards (IRBs) of Beth Israel Deaconess Medical Centre and the Massachusetts Institute of Technology. The eICU-CRD is a publicly available, de-identified database (HIPAA certification number 1031219-2) and was exempt from IRB review. The use of the SZPH-ICU dataset was approved by the local ethics committee (approval number: LL-KY-2025288-01), and the requirement for informed consent was waived because of the retrospective study design. This study strictly adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis guidelines (21).

Patients with cirrhosis who met the Sepsis-3 criteria on the first day of ICU admission were included. Cirrhosis was identified using International Classification of Diseases (ICD-9 or ICD-10) codes from patient discharge records (Supplementary Table 1). Patients aged < 18 years or with an ICU stay of < 24 h were excluded. For patients with multiple ICU admissions, only data from the first admission were included.

The primary outcome was all-cause in-hospital mortality. Out-of-hospital follow-up data were unavailable in both the eICU-CRD and SZPH-ICU databases, and only 28-day mortality was available in the MIMIC-IV database.

Based on prior literature and data availability, 25 candidate variables were prespecified (13–18). These included demographic characteristics (age, sex, and body mass index); vital signs (mean arterial pressure, respiratory rate, body temperature, oxygen saturation, and heart rate); laboratory parameters [total bilirubin, albumin, lactate, creatinine, white blood cell count, sodium, and international normalized ratio (INR)]; clinical interventions (renal replacement therapy, vasopressor use, and invasive mechanical ventilation); emergency admission status; and cirrhosis-related complications (ascites, hepatic encephalopathy, viral hepatitis, and alcohol abuse). The Glasgow Coma Scale (GCS) score and urine output were also included. All variables were extracted from the first 24 h following ICU admission. Vital signs were averaged over this period, and laboratory values were recorded as the most extreme values reflecting disease severity (e.g., maximum or minimum values).

For continuous variables (e.g., laboratory measurements and vital signs), outliers were defined as values below the first percentile or above the 99th percentile and were winsorized to the corresponding percentile thresholds. Missing values were imputed using multiple imputation via the “missForest” package in R, which implements a random forest-based algorithm (22).

The MIMIC-IV cohort was chronologically divided into a training set (2008–2016) and a temporal validation set (2017–2022). External validation was performed using the eICU-CRD and SZPH-ICU cohorts. Variable selection was conducted exclusively in the MIMIC-IV training set. To simplify the model and enhance its applicability, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was applied to identify the most predictive variables. A multivariable logistic regression model was then constructed using the LASSO-selected variables and subsequently transformed into a nomogram for clinical use. Model performance was evaluated using AUROC values, the Brier score, calibration curves, and decision curve analysis. To convert the model's predicted probabilities into binary outcomes (high vs. low risk), the optimal probability threshold was determined using the Youden index. The Youden index (J = sensitivity + specificity – 1) identifies the point on the ROC curve that maximizes the combined sensitivity and specificity, thereby providing an optimal balance (23). A web-based risk calculator was developed to facilitate bedside application.

The discriminatory performance of the proposed model was compared, using AUROC values, with the following established scoring systems: MELD, MELD-Na, SOFA, Albumin-Bilirubin (ALBI) score, Child-Pugh score, Chronic Liver Failure–Organ Failure (CLIF-OF) score, CLIF Consortium Acute-on-Chronic Liver Failure (CLIF-C ACLF) score, and CLIF-SOFA score (24–26). DeLong tests were used to compare the AUROC values between the different scoring systems. To further quantify the improvement in risk stratification capability of the new model compared to the traditional scoring systems, net reclassification improvement (NRI) was calculated (27).

Categorical variables are presented as frequencies (percentages), and continuous variables as medians (interquartile ranges). Group comparisons were conducted using the chi-square test for categorical variables and the t-test or Mann–Whitney U-test for continuous variables, as appropriate. All analyses were conducted using R software (version 4.4.0). A two-sided p-value < 0.05 was considered statistically significant.

A total of 2,052 adult ICU patients with cirrhosis or sepsis from the MIMIC-IV database were included in the analysis (Figure 1). The cohort was divided into a training set (2008–2016; n = 1,328) and a temporal validation set (2017–2022; n = 724). The overall median age was 60.0 years [interquartile range (IQR): 53.0–67.0 years], and 65.1% of patients were men. In the training set, 319 deaths were recorded, corresponding to an in-hospital mortality rate of 24.0%, whereas the temporal validation set comprised 260 deaths (mortality rate: 35.9%).

As Table 1 shows, patients who died during hospitalization exhibited significantly worse clinical indicators and higher disease severity scores than survivors in the training set. Non-survivors demonstrated greater haemodynamic instability, reflected by lower mean arterial pressure and higher heart rate. Respiratory impairment was also more pronounced among non-survivors, with elevated respiratory rates and reduced peripheral oxygen saturation. Laboratory indices indicated more severe hepatic and renal dysfunction as well as heightened systemic inflammation among non-survivors, including significantly higher levels of total bilirubin, lactate, creatinine, white blood cell counts, and INR, along with a higher proportion of patients with 24 h urine output < 400 ml. Therapeutic interventions, such as renal replacement therapy, vasopressor use, and invasive mechanical ventilation, were significantly more frequent in the non-survivor group. Similar patterns were observed in the temporal validation set (Table 1). In the eICU-CRD external validation cohort (n = 657), 170 deaths occurred (mortality rate: 25.9%); the median age was 59 years, and 61.6% of patients were men. The SZPH-ICU validation cohort (n = 131; 74.8% men) had a median age of 61 years and included 50 deaths (mortality rate: 38.2%). In both external validation cohorts, non-survivors consistently exhibited more severe organ dysfunction and higher disease severity scores (Supplementary Tables 2, 3). These findings are consistent with clinical expectations, whereby greater physiological derangement and more advanced organ dysfunction are typically associated with an increased risk of mortality.

Variable selection was performed using LASSO regression in the training set to identify predictors most strongly associated with in-hospital mortality (Figure 2). The optimal regularization parameter (λ_0.1se = 0.024) was determined via 10-fold cross-validation, yielding 13 variables with non-zero coefficients: age, respiratory rate, body temperature, oxygen saturation, heart rate, total bilirubin, lactate, creatinine, white blood cell count, INR, vasopressor use, urine output, and GCS score.

A multivariable logistic regression model was constructed using the 13 predictors identified by LASSO (coefficients presented in Table 2). For clinical applicability, the model was visualized as a nomogram (Figure 3). This nomogram serves as a graphical computational tool, enabling clinicians to assign scores to each variable (e.g., age, vital signs, laboratory values) according to an individual patient's characteristics. The scores are then summed to obtain a total score, which corresponds to the predicted probability of mortality on the total score axis. This approach visually quantifies the relative contribution of each predictor to mortality risk and facilitates rapid, individualized risk assessment.

The nomogram demonstrated strong predictive performance in both the training and validation cohorts (Table 3 and Figure 4). In the training set, an AUROC of 0.822 (95% CI: 0.797–0.847) and a Brier score of 0.135 indicated excellent discrimination and calibration. Despite differences in baseline characteristics between the temporal validation set and the training set (Table 1), performance remained stable in the temporal validation cohort (AUROC = 0.810, 95% CI: 0.777–0.843; Brier score = 0.169). External validation in the eICU-CRD cohort yielded an AUROC of 0.777 (95% CI: 0.734–0.821; Brier score = 0.150), while validation in the SZPH-ICU cohort yielded an AUROC of 0.763 (95% CI: 0.680–0.846; Brier score = 0.200). Figure 5 presents the calibration curves for all datasets, illustrating the concordance between the model's predicted probabilities and the actual observed mortality rates. Across all datasets, the calibration curves closely aligned with the diagonal line, signifying a high degree of calibration in the model's predictions. In the training dataset, the Youden index was computed across a range of probability thresholds. The maximum value (0.48), corresponding to a sensitivity of 73% and a specificity of 75%, was observed at a threshold of 0.238. This value was therefore selected as the optimal cut-off for distinguishing high-risk and low-risk patients.

The nomogram was systematically compared with eight established scoring systems: SOFA, ALBI, Child–Pugh, CLIF-OF, CLIF-C ACLF, CLIF-SOFA, MELD, and MELD-Na (Figure 4). In the training set, the nomogram (AUROC = 0.822) significantly outperformed all comparator scores (p < 0.05), with the best-performing conventional score (CLIF-SOFA) achieving an AUROC of 0.759. Similar superiority was observed in the temporal validation set (AUROC = 0.810; p < 0.05 for all comparisons).

In the eICU-CRD cohort, the nomogram achieved an AUROC of 0.777, which was significantly higher than those of SOFA, ALBI, Child–Pugh, CLIF-C ACLF, MELD, and MELD-Na (all p < 0.05). Although the differences compared with CLIF-OF (p = 0.105) and CLIF-SOFA (p = 0.107) were not statistically significant, the nomogram still exhibited numerical superiority. In the SZPH-ICU cohort, the nomogram (AUROC = 0.763) did not differ significantly from SOFA (p = 0.062), Child–Pugh (p = 0.093), CLIF-OF (p = 0.158), CLIF-C ACLF (p = 0.194), MELD (p = 0.218), or MELD-Na (p = 0.223), but significantly outperformed ALBI (p < 0.05) and consistently yielded the highest AUROC among all scores.

The incremental value of the nomogram over CLIF-SOFA was further assessed using the NRI index. Table 4 shows that in both the training and temporal validation sets, the NRI was significantly greater than zero (training set: NRI = 0.097, p = 0.002; temporal validation set: NRI = 0.092, p = 0.011), indicating that approximately 9–10% of patients were more accurately reclassified by the nomogram. In the external validation cohorts, the eICU-CRD cohort dataset showed a positive but non-significant NRI (NRI = 0.053, p = 0.234), whereas the SZPH-ICU cohort (n = 131) demonstrated no reclassification benefit (NRI = −0.001, p = 0.994). Overall, the NRI findings were consistent with the AUROC comparisons, supporting the nomogram's superior risk reclassification performance over CLIF-SOFA in the primary cohorts.

Decision curve analysis was used to evaluate the clinical net benefit of the model across a range of decision thresholds. As Figure 6 illustrates, over a broad range of threshold probabilities (approximately 0.05–0.78), employing this model for risk stratification—by intervening in patients whose predicted probabilities exceed the selected threshold—yielded a higher net clinical benefit than either treating all patients or treating none. At threshold probabilities above 0.78, the model's net benefit diminished, aligning closely with the “treat-none” strategy. This outcome is anticipated, as such high thresholds are rarely adopted in routine clinical practice when initiating therapy. Overall, these findings underscore the clinical utility of applying the proposed model.

To facilitate clinical translation, an interactive web-based calculator was developed using the nomogram and is accessible at https://szph-icu.shinyapps.io/Cirrhosis-Sepsis-ICU/ (Figure 7). This tool enables clinicians to input patient-specific parameters, including age, vital signs, and laboratory parameters, to generate real-time estimates of in-hospital mortality risk. Consequently, it aids bedside decision-making and facilitates individualized treatment planning.