Baseline characteristics
Boxplots grouped by ANLR quartiles (Fig. 2) show that NLR levels progressively decreased, while albumin levels gradually increased across quartiles. Baseline characteristics stratified by ANLR quartiles are presented in Table 1. Significant differences were found in age (p < 0.001), gender (p = 0.031), and weight (p < 0.001). All vital signs, including heart rate, respiratory rate, temperature, SBP, and SpO₂, differed significantly across quartiles (all p < 0.001). Among clinical scores, SOFA, APS III, OASIS, and CCI were significantly different (all p < 0.001). For comorbidities, cerebrovascular disease, chronic pulmonary disease, hypertension, septic shock, and AKI showed significant variation (all p < 0.001). All laboratory indicators except potassium showed significant differences across groups (p < 0.05), including RBC, WBC, platelet, lymphocytes, neutrophils, albumin, sodium, ALT, AST, anion gap, INR, PT, BUN, and creatinine (all p < 0.001). Treatment measures such as propofol, midazolam, GC, antibiotics, norepinephrine, dopamine, vasopressin, EN, MV, and CRRT also varied significantly among groups (all p < 0.001). Additionally, all outcome indicators—hospital stay, ICU stay, hospital mortality, ICU mortality, 30-day mortality, and 90-day mortality—were significantly different across ANLR quartiles (all p < 0.001).
Boxplots of NLR (a) and albumin (b) stratified by ANLR quartiles
Primary results
The Kaplan-Meier survival analysis revealed notable differences in survival rates among patients grouped by ANLR quartiles, with those in Quartile 4 demonstrating the most favorable survival outcomes. Statistical comparison using the log-rank test indicated highly significant differences between quartile groups for both 30-day and 90-day mortality (p < 0.001; Fig. 3).
Kaplan-Meier curves for 30-day (a) and 90-day (b) mortality in sepsis patients stratified by ANLR quartiles
The variables included in the multivariable Cox regression analysis (Table S1 Supplementary Material) were selected based on univariable Cox regression results and clinical expertise. The results are presented in Table 2.. In the fully adjusted model (Model 3), ANLR as a continuous variable was independently associated with a reduced risk of both 30-day mortality (HR = 0.68, 95% CI = 0.59–0.79; p < 0.001) and 90-day mortality (HR = 0.85, 95% CI = 0.76–0.94; p = 0.002). When ANLR was analyzed as a categorical variable (quartiles), patients in Quartile 4 had significantly lower risks of mortality compared to those in Quartile 1, for 30-day mortality (HR = 0.44, 95% CI = 0.37–0.53; p < 0.001) and 90-day mortality (HR = 0.45, 95% CI = 0.38–0.53; p < 0.001). A significant trend was observed across ANLR quartiles (p for trend < 0.001) for both time points.
Following adjustment for all covariates, RCS analyses revealed a highly significant overall association between ANLR and both 30-day and 90-day mortality (overall P < 0.001), with evidence of a significant nonlinear pattern (nonlinear P < 0.001) (see Fig. 4).
RCS analysis of ANLR and mortality in patients with sepsis (a) 30-day mortality, (b) 90-day mortality
Subgroup analysis
Following adjustment for relevant covariates, interaction analyses were performed within a set of predetermined subgroups—namely gender, race, cerebrovascular disease, chronic pulmonary disease, diabetes, and use of mechanical ventilation. The findings indicated that the relationship between ANLR and mortality was stable across these categories, as no statistically significant interactions were detected (all interaction P-values exceeded 0.05; see Fig. 5).
Subgroup forest plot of adjusted covariates for mortality in patients with sepsis (a) 30-day all-cause mortality, (b) 90-day all-cause mortality
ROC analysis of ANLR and other predictive markers
The ROC analysis demonstrated that ANLR achieved an AUC of 0.66 for predicting 30-day mortality, with the optimal threshold identified at 0.27. Of note, this predictive capability exceeded that of other individual biomarkers, such as NLR, SOFA score, neutrophil count, albumin, and lymphocyte count. Notably, when SOFA and ANLR were combined as a ratio (SOFA/ANLR), the discriminatory power was further improved, with the AUC for SOFA/ANLR reaching 0.68 and an optimal cutoff value of 12.56. Similar findings were observed for 90-day mortality, where ANLR alone had an AUC of 0.65 (cutoff = 0.30), and SOFA/ANLR demonstrated a higher AUC of 0.67 (cutoff = 10.75), indicating that the SOFA/ANLR ratio substantially enhanced the prognostic performance compared to SOFA alone (Fig. 6).
a ROC curve for 30-day mortality, b ROC curve for 90-day mortality,
Boruta
In accordance with the pre-established protocol, the dataset was partitioned into training and validation cohorts. Comparative analysis revealed no statistically significant differences in baseline characteristics between these two groups (all p-values > 0.05; see Table S2 in the Supplementary Material). Using the Boruta algorithm for feature selection, a total of 24 variables were identified as important predictors for 30-day mortality. Among these, ANLR was ranked as the second most important feature. Other important variables included weight, platelet, RBC, AKI, SBP, heart rate, septic shock, liquid balance, potassium, ALT, WBC, cerebrovascular disease, AST, respiratory rate, SpO₂, INR, PT, creatinine, sodium, age, anion gap, BUN, and temperature. All these variables were identified as “confirmed important” by the Boruta algorithm, and were highlighted in green in the feature importance plot (Fig. 7). These variables were included in the subsequent machine learning modeling.
a Boruta algorithm identifying key variables for predicting 30-day mortality (green indicates important variables), b ROC curve for the validation set, c Decision curves for multiple models in the validation set. d Calibration curves for multiple models in the validation Set. e SHAP summary plot with dependence illustrating variable importance and feature effects in the LightGBM model
Establishment and validation of the prediction model
The predictive model was constructed using 24 key variables highlighted as important features by the Boruta algorithm, all of which were marked in green. The training set comprised 4715 patients, while the validation set included 1573 patients. In the training set, 968 patients (21%) died, while the validation set observed 323 deaths (21%). These event counts ensured compliance with the commonly accepted criterion of a minimum of 10 outcome events per predictor variable.
A range of models was constructed using the selected features, including SVM, Ridge, LightGBM, KNN, ENet, XGBoost, DT, and MLP. LightGBM was identified as the optimal model, demonstrating the lowest Brier score (0.1250), the highest net benefit in decision curve analysis, and the highest AUC (0.821) on the validation set (Fig. 7). The LightGBM model was further interpreted using SHAP values, which identified ANLR as the second most important variable (Fig. 7).
Sensitivity analysis
To assess the robustness of our findings, we performed a multivariable Cox regression analysis using the original (non-imputed) dataset. The results for both 30-day and 90-day mortality were consistent with those obtained from the main analysis (Table S3, Supplementary Material).