This study utilized data from two retrospective cohorts. The internal training cohort data were extracted from the publicly available Medical Information Mart for Intensive Care (MIMIC-IV) database, maintained by the Massachusetts Institute of Technology. A detailed description of this database can be found in previous literature (22). The external validation cohort included critically ill patients with urosepsis admitted to Bijie Hospital of Zhejiang Provincial People’s Hospital between January 2020 and January 2025. The use of this cohort was approved by the hospital’s Ethics Committee.

This study focused on patients admitted to the ICU with urosepsis. Patient diagnosis was based on criteria established in previous literature (23, 24). Sepsis was defined according to the Sepsis-3 consensus criteria, requiring an increase in the Sequential Organ Failure Assessment (SOFA) score by ≥2 points from baseline. The diagnostic criteria for UTI differed between the two cohorts: the internal cohort was identified using International Classification of Diseases (ICD) diagnostic codes from the MIMIC-IV database, while the external cohort was confirmed by typical clinical presentation (pyuria and/or bacteriuria on urinalysis) and a positive urine culture. Only patients with an initial admission diagnosis of UTI who also met the sepsis-3 criteria were enrolled in the urosepsis cohort. All enrolled patients met the following criteria: (1) age ≥18 years; (2) a hospital and ICU stay of at least 24 h; and (3) the availability of complete serum lactate and albumin measurements taken at ICU admission.

Data extraction was performed using PostgreSQL (v13.7.2), Navicat Premium (v16.0), and Structured Query Language (SQL). The extracted variables were categorized into six groups: (1) demographics; (2) comorbidities; (3) vital signs; (4) laboratory parameters; (5) disease severity scores; and (6) administered treatments (see Supplementary Tables S1, S2 for a detailed list). The primary endpoint of the study was 28-day ICU mortality, with 28-day in-hospital all-cause mortality as the secondary endpoint. For variable imputation, we only included variables with less than 30% missing data. Multiple imputation was conducted using the R “mice” package (v3.16.0). It should be noted that the key exposure variables (LAR and its components, serum lactate and albumin) had no missing data in the final analysis cohort due to the inclusion criteria requiring complete measurements. For other covariates meeting the missingness threshold, we generated 5 complete datasets through multiple imputation with 50 iterations. The imputation model incorporated all variables used in the primary analysis: continuous variables requiring imputation, the complete primary exposure variables (LAR and its components), the complete outcome variable (28-day mortality), and all other complete categorical and continuous covariates. Descriptive statistics of the pooled imputed datasets were consistent with those from the original data, indicating no substantial bias was introduced by the imputation process. This approach strengthens the plausibility of the missing at random assumption by leveraging all available information to predict missing values.

After conducting hypothesis testing, Cox proportional hazards regression model was used to evaluate the association between LAR and clinical endpoints. To control for potential confounding factors, three progressively adjusted models were constructed: Model 1 was unadjusted; Model 2 was adjusted for demographic characteristics (age, sex, ethnicity, and weight) and baseline comorbidities showing intergroup differences; Model 3 was further adjusted for clinically relevant variables demonstrating significant differences between survivors and non-survivors, including severity-of-illness scores, key laboratory parameters, and treatment measures (detailed in Supplementary Tables S1–S3). This comprehensive adjustment approach follows a well-validated strategy consistently employed in previous studies. To mitigate multicollinearity affecting model stability, the variance inflation factor (VIF) was calculated for all variables included in Model 3, and variables with VIF >5 were excluded. Subsequently, restricted cubic splines (RCS) with four knots (at the 5th, 35th, 65th, and 95th percentiles) were used to explore potential nonlinear relationships between LAR and the endpoints. Kaplan–Meier (KM) survival curves were used for supplementary visualization.

The internal cohort was first randomly split into a training set and a validation set in a 7:3 ratio. In the training set, the Boruta algorithm was employed to identify variables associated with short-term ICU mortality. The confidence level (p-value threshold) for this process was set at 0.01, with a maximum of 100 iterations for important source runs, and multiple comparison adjustments were performed using the Bonferroni method. Subsequently, three machine learning algorithms were applied to further identify core variables significantly related to patient outcomes from a broader set. The random forest algorithm was configured with 100 trees, a maximum tree depth of 3, a minimum of 2 samples required to split an internal node, and 1 sample required at a leaf node. The GBM was implemented using the Cox proportional hazards loss function (“coxph”). Key parameters included a learning rate of 0.1, 100 boosting stages, a subsample fraction of 1.0, the “friedman_mse” criterion for evaluating split quality, a minimum of 2 samples required to split an internal node, 1 sample required at a leaf node, and a minimum impurity decrease threshold of 0. LASSO regression (least absolute shrinkage and selection operator) was also applied to simultaneously perform feature selection and regularization. Variables identified as important by all three algorithms were considered significant predictors.

Following the identification of key variables, these were used to construct a final, interpretable multivariate Cox proportional hazards regression model for risk prediction. The regression coefficients (β-values) in the final model formula were directly derived from this Cox model, representing the logarithm of the hazard ratio associated with a one-unit increase in each predictor, after adjusting for all other variables in the model. Each patient’s risk score was calculated using the following formula: Risk Score = (β₁ × Variable₁) + (β₂ × Variable₂) + (βₙ × Variableₙ). The predictive performance of the model was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC). Furthermore, its stability was assessed in an external validation cohort with distinct patient characteristics.

For data visualization and comparative analysis, patients were stratified into low (<0.48718), medium (0.48718–0.84375), and high (>0.84375) LAR groups based on tertiles. It is important to emphasize that this categorization was intended solely for descriptive and exploratory purposes and does not imply the existence of a clinically validated risk threshold. Continuous variables are presented as mean ± standard deviation and were compared using Student’s t-test or analysis of variance (ANOVA) after assessing normality and homogeneity of variances. Categorical variables are expressed as numbers (percentages) and were compared using Pearson’s chi-square test or Fisher’s exact test, as appropriate. All statistical analyses were performed using R software (version 4.5.1). A two-sided p-value <0.05 was considered statistically significant.

According to the strict inclusion criteria, 1,055 critically ill patients with urosepsis were included in the analysis. The mean age of the cohort was 70.5 years, and 439 (41.6%) were male. Baseline characteristics stratified by the LAR are presented in Table 1. Compared to patients with lower LAR, those with higher LAR had a higher burden of baseline comorbidities, required more intensive therapeutic interventions, and exhibited less stable vital signs, characterized by an increased respiratory rate and decreased systolic, diastolic, and mean arterial pressures. Furthermore, multiple laboratory parameters differed significantly between the groups. Higher LAR was associated with elevated levels of RDW, WBC, neutrophil count, anion gap, serum glucose, lactate, INR, PT, PTT, ALT, AST, total bilirubin (TB), and lactate dehydrogenase. Conversely, albumin, total serum calcium, total carbon dioxide combining capacity, ionized calcium, pH, and partial pressure of carbon dioxide (PCO₂) were lower in the high LAR group. Clinical severity scores were positively correlated with LAR levels. Mortality was significantly higher in the high LAR group compared to the low LAR group for both 28-day ICU mortality (29.2% vs. 10.4%, p < 0.001) and in-hospital mortality (27.2% vs. 9.83%, p < 0.001).

As shown in Table 2, a higher LAR was significantly associated with increased 28-day ICU mortality across all three Cox regression models: the unadjusted Model 1 (HR = 1.59, 95% CI: 1.39–1.81, p < 0.001), Model 2 adjusted for demographics and baseline comorbidities (HR = 1.48, 95% CI: 1.29–1.70, p < 0.001), and Model 3 further adjusted for intergroup differential variables (HR = 1.35, 95% CI: 1.08–1.68, p = 0.008). Similarly, compared to the low LAR group, the high LAR group exhibited a significantly increased risk of mortality in all three models: Model 1 (HR = 3.18, 95% CI: 2.17–4.65, p < 0.001; p for trend <0.001), Model 2 (HR = 2.80, 95% CI: 1.91–4.12, p < 0.001; p for trend <0.001), and Model 3 (HR = 1.73, 95% CI: 1.10–2.73, p = 0.018; p for trend <0.001).

Likewise, as presented in Table 3, a higher LAR consistently indicated an elevated risk of in-hospital mortality across all models: Model 1 (HR = 1.70, 95% CI: 1.47–1.97, p < 0.001), Model 2 (HR = 1.63, 95% CI: 1.40–1.89, p < 0.001), and Model 3 (HR = 1.50, 95% CI: 1.19–1.90, p = 0.001). The high LAR group also consistently demonstrated higher mortality compared to the low LAR group: Model 1 (HR = 2.98, 95% CI: 2.01–4.40, p < 0.001; p for trend <0.001), Model 2 (HR = 2.74, 95% CI: 1.84–4.07, p < 0.001; p for trend < 0.001), and Model 3 (HR = 1.71, 95% CI: 1.08–2.72, p = 0.022; p for trend <0.001). Restricted cubic spline (RCS) curves further revealed a non-linear positive dose–response relationship between LAR and short-term mortality in critically ill urosepsis patients (Figures 1A,B). Kaplan–Meier survival analysis similarly showed that higher LAR was associated with lower short-term survival rates (Figures 1C,D).

To evaluate the added value of the LAR, we tested its combination with six established severity scores (SOFA, APS III, SAPS II, OASIS, Charlson, APACHE II) for predicting 28-day ICU mortality. The inclusion of LAR universally enhanced the predictive accuracy of all models (Figures 2A–F), with DeLong’s test confirming statistically significant improvements for five of the six scores (SOFA: AUC 0.664 to 0.690, p = 0.004; SAPS II: 0.692 to 0.710, p = 0.017; OASIS: 0.615 to 0.665, p = 0.001; Charlson: 0.618 to 0.675, p < 0.001; APACHE II: 0.653 to 0.695, p < 0.001). A positive, albeit non-significant, increase was also observed for APS III (0.699 to 0.712, p = 0.058).

We conducted subgroup analyses and interaction tests to investigate whether demographic characteristics and comorbidities influenced the association between LAR and 28-day ICU/in-hospital mortality. After multivariable adjustment, the results (Supplementary Figures S1, S2) indicated that the association between LAR and 28-day ICU/in-hospital mortality did not reach statistical significance in certain subgroups. These included patients with comorbidities (e.g., hypertension, chronic kidney disease, diabetes, hyperlipidemia, heart failure, and chronic obstructive pulmonary disease) as well as the subgroup of older male patients. It is important to note that these specific subgroups had relatively smaller sample sizes and lower event rates, which may have resulted in insufficient statistical power. Therefore, the non-significant associations observed in these subgroup analyses should be interpreted with caution, as they may reflect a risk of type II error rather than a true absence of effect. The interaction tests identified a significant interaction effect only between age and LAR for ICU mortality (p for interaction <0.05), which further supports that the non-significant results in other subgroups are likely primarily limited by sample size constraints.

The external validation cohort included 245 patients with urosepsis (28-day mortality = 14.28%). In the Cox regression model adjusted for all potential covariates, a higher LAR was significantly associated with an increased risk of 28-day mortality (HR = 2.12; 95% CI: 1.32–3.41, p = 0.002). Subsequent Kaplan–Meier analysis indicated that patients with LAR scores above the cohort median had significantly higher short-term mortality (log-rank p < 0.001; HR = 7.19, 95% CI: 3.20–16.16; Figure 3A). Furthermore, the RCS curve confirmed a significant non-linear positive dose–response relationship between LAR and ICU short-term mortality in the external cohort (Figure 3B).

In the internal training cohort, the Boruta algorithm identified 15 variables associated with ICU mortality (Figure 4A). Subsequently, random forest (Figure 4B), gradient boosting machine (feature importance score >0.01) (Figure 4C), and LASSO regression (lambda min = −4.947) (Figure 4D) collectively identified six key variables: LAR, SAPS II score, RDW, PTT, TB, and URE (blood urea nitrogen). These were used to construct a risk assessment model for predicting 28-day ICU mortality in critically ill urosepsis patients. The model formula is: Risk Score = (0.458617638 × LAR) + (0.015481692 × SAPSII) + (0.087598812 × RDW) + (0.004594067 × PTT) + (0.011365337 × TB) + (0.004909906 × URE). Compared to traditional severity scoring systems (SOFA, APS III, SAPS II, OASIS, Charlson, APACHE II), this model demonstrated superior sensitivity and specificity in predicting 28-day ICU mortality (Figures 4E–G). The AUC values were 0.749 in the internal training cohort (Figure 4E), 0.742 in the internal validation cohort (Figure 4F), and 0.787 in the external real-world cohort (Figure 4G). These consistent results indicate that the model possesses robust stability and predictive performance across diverse cohorts. The calibration curve and decision curve analysis further demonstrated the robustness and predictive accuracy of the analytical model (Figures 5A,B).