Machine learning-based risk prediction of 28-day mortality for sepsis patients with augmented renal clearance Authors: Yunzhe Wu, Fan Yang, Hongjie Yang, Tong Wu, RuoYu Zhuang, Xiaoli Wang, Yide Lu, Danfeng Dong Published: 2026 ABSTRACT AbstractAugmented renal clearance (ARC) frequently occurs in critically ill septic patients and is known to impact survival outcomes. To address this, we aimed to develop an interpretable machine learning model for early mortality prediction in this high-risk population using the MIMIC-IV database. A total of 518 septic patients with ARC were enrolled, with a 28-day mortality rate of 17.2%. From the first 24 h of ICU admission and the time of ARC onset, we extracted 162 and 88 covariates, respectively. Following data imputation, we applied a three-stage feature selection strategy, consisting of the EPV principle, Bootstrap LASSO with > 88.5% selection frequency, and clinical expert review. This process yielded 9 and 10 key predictors from the two timepoints for subsequent model construction. Among nine machine learning algorithms evaluated, XGBoost achieved the highest discriminative performance (AUC 0.80) using early ICU data. SHAP interpretability analysis revealed that respiratory rate, body temperature, and serum sodium were the most influential predictors. To facilitate further research and external validation, we developed a freely accessible online calculator (https://wuyunzhe.shinyapps.io/arc_prediction/) that enables exploratory bedside risk assessment with automatic stratification. In conclusion, the XGBoost model, leveraging routinely collected early ICU data, provides an interpretable and clinically applicable tool for predicting 28-day mortality in septic patients with ARC, demonstrating potential for future clinical application after prospective validation.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-54630-w.Keywords: Sepsis, Augmented renal clearance, Outcome, Risk prediction, Machine learningSubject terms: Biomarkers, Computational biology and bioinformatics, Diseases, Medical research, Risk factors Introduction Sepsis refers to the body’s inflammatory response and oxidative stress response disorder caused by infection. It leads to multi-system damage to the host, and can eventually progress to septic shock1. Epidemiological data show that the incidence of sepsis in adult patients receiving hospitalization is approximately 189 cases per 100,000 people per year, with a mortality rate of 26.7%. In contrast, the incidence of sepsis in ICU patients is 58 cases per 100,000 people per year, with a mortality rate of 41.9%2. Although sepsis research has advanced continuously in recent years, and clinical practice has been accumulated, there are great differences in age, underlying diseases, and types of infectious pathogens among sepsis patients. It makes the diagnosis and treatment of sepsis challenging and is the main cause of death for patients in intensive care units. Early identification and intervention of factors affecting the adverse prognosis of sepsis can help improve the prognosis of patients3. Augmented renal clearance (ARC) is a state of renal hyperfiltration. Increased renal clearance capacity can accelerate the elimination of drugs through the kidneys. As a result, standard doses may fail to reach effective blood concentrations, thereby increasing the risk of treatment failure4. With an incidence ranging from 30 to 85% in the intensive care unit, ARC represents a prevalent yet frequently overlooked clinical phenomenon5. In septic patients, the development of ARC can significantly increase drug clearance, resulting in subtherapeutic drug levels and potentially contributing to higher mortality6. However, the relationship between ARC and patient outcomes remains controversial. Some studies suggest that ARC may reflect a hyperdynamic state and preserved renal function7, whereas others link it to poor outcomes due to inadequate drug exposure6. Given the pathophysiological complexity of both sepsis and ARC, conventional prognostic tools reliant on single parameters or composite scores may fail to adequately capture the multifaceted risk profile of these patients. Advances in statistical theory and computing power have propelled the adoption of machine learning (ML) in medical research. ML techniques are increasingly applied across a spectrum of clinical tasks, from early diagnosis to outcome prediction, and have demonstrated promising performance in various disease contexts8,9. Notably, several studies indicate that ML-based predictions may outperform those derived from traditional logistic or Cox regression models9,10. Recent machine learning studies have demonstrated impressive performance in predicting sepsis mortality. Gao et al. achieved an AUC of 0.94 using Random Forest on general sepsis populations11, while a novel Teacher-Student architecture attained an AUC of 0.829 on 9476 sepsis patients12. However, these studies focused on general sepsis populations and did not specifically address the unique pathophysiology of patients with augmented renal clearance (ARC). Given that ARC significantly alters drug pharmacokinetics and may influence mortality risk through distinct mechanisms, a model tailored to this subgroup is warranted. Therefore, this study aimed to develop and validate an interpretable machine learning model to predict 28-day mortality in critically ill sepsis patients with ARC. By providing an accurate and individualized risk assessment, the model seeks to support clinicians in making timely and precise drug dosage adjustments, thereby helping to improve prognosis. Materials and methods Source of data The data for this study were sourced from the Medical Information Mart for Intensive Care (MIMIC-IV) database version 2.2, a publicly available critical care database jointly developed by the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center13. This database contains records from over 70,000 Intensive Care Unit (ICU) admissions between 2008 and 2019. To ensure data utility while protecting patient privacy, temporal data were shifted using a consistent offset for each patient13. MIMIC-IV provides comprehensive clinical information, including demographic characteristics (including age, sex, race), ICU and hospital length of stay, laboratory and hematological test results, imaging reports, and continuous vital signs. Data extraction was performed using SQL by researcher T.W., who holds a credentialed MIMIC-IV license (No. 33461595) and has completed the required human subjects’ protection training. Ethical approval for this study was waived as the database contains fully de-identified data. Definitions and participants In this study, sepsis was defined according to the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). Specifically, patients with a probable infectious cause and an acute increase in total Sequential Organ Failure Assessment (SOFA) scores of ≥ 2 points within the first 24 h of ICU admission were included. For patients with multiple ICU admissions, only data from their first ICU stay were analyzed to avoid potential confounding from repeated admissions. Participants were excluded if they met any of the following criteria: (1) age younger than 18 years; (2) ICU stay duration shorter than 48 h; (3) lack of necessary variables for calculating 24-hour creatinine clearance (24 h-CrCl); (4) receipt of renal replacement therapy (RRT) within the first 7 days after ICU admission; or (5) patients with > 30% missing individual data. Augmented renal clearance (ARC) was defined as a 24 h-CrCl value ≥ 130 mL/min/1.73 m², based on established criteria from prior studies14–18. This threshold was selected to ensure comparability with the classic study by Udy et al.19, in which the patient characteristics (mean age 54.4 years) were similar to those of our cohort (mean age 57.6 years). The 24 h-CrCl was calculated using the following formula: 24 h-CrCl= [UCr × UV]/[SCr × 1440] × [1.73/(0.007184 × height [cm] 0.725 × weight [kg] 0.425 )]). UCr represents the mean urinary creatinine concentration (mg/dL). UV is the total urinary volume (mL) collected over a 24-hour period. SCr is the mean serum creatinine concentration (mg/dL). The result is adjusted for body surface area using the Du Bois formula. All sepsis patients who met our inclusion criteria and had 24 h-CrCl exceeding 130 mL/min/1.73 m² were identified as sepsis patients with ARC and included in the present study. Data extraction and processing Sepsis patients with augmented renal clearance (ARC) were stratified into “survivor” and “non-survivor” groups based on 28-day outcomes. Demographic data including sex, age, race, and Body Mass Index (BMI) were extracted. Underlying conditions were assessed using the Charlson Comorbidity Index (CCI), covering congestive heart failure, myocardial infarction, cerebrovascular disease, peripheral vascular disease, chronic pulmonary disease, dementia, rheumatic disease, renal disease, diabetes, peptic ulcer disease, liver disease, paraplegia, malignancy, and AIDS. Additional clinical variables included infection site, vasopressor use, SOFA score, vancomycin trough levels, and organ support requirements. Vital signs and laboratory parameters were collected at two timepoints: within the first 24 h of ICU admission and at the time of ARC diagnosis. These included heart rate, blood pressure, respiratory rate, body temperature, percutaneous arterial oxygen saturation, blood glucose, partial pressure of oxygen (PaO₂), partial pressure of carbon dioxide (PaCO₂), base excess, hematocrit, hemoglobin, platelet count, white blood cell count, red blood cell count, serum chloride, serum calcium, serum potassium, serum sodium, anion gap, blood urea nitrogen (BUN), alanine aminotransferase (ALT), alkaline phosphatase (ALP), aspartate transaminase (AST), and bilirubin (as detailed in Table S1). Maximum and minimum values for each parameter were recorded to capture physiological variability. As an exploratory epidemiological analysis, this study did not perform a formal sample size calculation. All eligible patients from the MIMIC-IV database were included to ensure statistical power. Missing data imputation Variables with > 30% missingness were excluded from analysis. For variables with ≤ 30% missingness, multiple imputation was performed using the mice (Multivariate Imputation by Chained Equations) package in R (version 4.3.3). The predictive mean matching (PMM) method was selected as the imputation algorithm, which imputes missing values by matching to the observed value with the closest predicted mean, thereby preserving the original distributional characteristics of the data. The number of imputations was set to five (m = 5), based on Rubin’s relative efficiency criterion20. The imputation model included all other variables as predictors, with appropriate generalized linear models applied for categorical variables. Each imputation chain was run for 20 iterations to ensure convergence. All subsequent analyses, including feature selection and model training, were performed independently on each of the five complete imputed datasets. Results were then pooled using the pool() function in R according to Rubin’s rules, yielding combined estimates that account for imputation uncertainty. Feature selection To identify robust predictors, we implemented a three-stage feature selection strategy. First, based on the Events Per Variable (EPV) principle given 89 endpoint events, the maximum number of predictors was constrained to ≤ 8.9 (10 EPV) and ≤ 17.8 (5 EPV) to prevent overfitting21. Second, within this framework, we applied bootstrap-enhanced least absolute shrinkage and selection operator (LASSO) regression with 100 iterations to screen features from two datasets: 166 covariates from the first 24 h of ICU admission and 88 covariates documented at ARC diagnosis. In each iteration, ten-fold cross-validation was performed to select the optimal regularization parameter (λ), retaining features with non-zero coefficients at the lambda.1se threshold22. Variables with selection frequency > 88.5% across iterations were retained, ensuring stability across different data subsamples. Third, all candidate variables were independently reviewed by two senior intensivists to confirm clinical relevance and biological plausibility. Model development and evaluation In the modeling phase, we implemented nine distinct algorithms: Extreme Gradient Boosting (XGBoost), support vector machine (SVM), random forest (RF), decision tree (DT), Gradient Boosting Machine (GBM), Light GBM (LGBM), k-nearest neighbors (KNN), logistic regression (LR), and Firth logistic regression (FL). To address class imbalance, different strategies were applied according to each algorithm’s characteristics: (1) For XGBoost, we set the scale_pos_weight parameter to the ratio of negative to positive samples (429/89 ≈ 4.82), which assigns higher penalties to misclassifications of the minority class; (2) Firth logistic regression was employed specifically for its penalized likelihood approach that inherently handles small sample bias; (3) For other algorithms (RF, SVM, LR, DT, LGBM, GBM), we applied class weight adjustments where available; (4) Algorithms without built-in imbalance handling (KNN) were evaluated with default settings, acknowledging this limitation. Hyperparameter optimization was performed using Bayesian optimization. For XGBoost, we optimized learning rate, maximum tree depth, minimum child weight, and subsample ratio. Other algorithms underwent similar optimization tailored to their respective parameter spaces. The dataset was partitioned using an 80:20 split for training and validation. Model selection prioritized the highest validation set AUC, with comprehensive evaluation including sensitivity, specificity, accuracy, positive (PPV) and negative (NPV) predictive values, F1-score, and Matthews correlation coefficient (MCC). Probability calibration was assessed through calibration curves, while model interpretability was achieved via SHAP value analysis. All analyses were conducted in R version 4.3.3. Web-based prediction tool To facilitate further research and external validation, we developed a freely accessible web-based calculator using the Shiny package in R (version 4.3.3). The final model was exported and integrated into the application, allowing users to input the core predictors and obtain instantaneous mortality risk predictions with automatic risk stratification. All computations are performed locally in the user’s browser to ensure patient data privacy, with no data storage or transmission. The source code is openly available in Supplementary File2. Statistical analysis Continuous and categorical variables were summarized using medians with interquartile ranges and counts with percentages, respectively. Group comparisons employed non-parametric Wilcoxon tests for continuous measures and either χ² or Fisher’s exact tests for proportions as appropriate23 with statistical significance defined as two-tailed p < 0.05. Results Patient characteristics From the MIMIC-IV database, 2673 sepsis patients met the study criteria after screening an initial 32,982 admissions. Among these, 518 (19.37%) were diagnosed with ARC within 7 days of ICU admission (Fig. 1), with a 28-day mortality rate of 17.2% (n = 89). Baseline characteristics are presented in Table 1. Compared to survivors, non-survivors were older (61.15 vs. 56.89 years, p = 0.036), had lower body weight (76.58 vs. 82.85 kg, p = 0.027) and lower BMI (26.77 vs. 28.81 kg/m2, p = 0.033), and had a higher Charlson Comorbidity Index (4.42 vs. 3.64, p = 0.01). They were also more likely to have underlying cardiovascular disease and metastatic solid tumors, whereas a history of trauma was associated with lower mortality risk. The flowchart of patient selection. Abbreviations: MIMIC, Medical Information Mart for Intensive Care; ICU, intensive care unit; CrCl, creatinine clearance; ARC, augmented renal clearance. Table 1Baseline characteristics of sepsis patients with ARC stratified by 28-day outcome.Variables Overall (518)28-day survival(429)28-day mortality (89)P valueGender, n (%)0.516 Female237 (45.8)193 (45.0)44 (49.4) Male281 (54.2)236 (55.0)45 (50.6)Age, mean (SD)57.62 (17.48)56.89 (17.12)61.15 (18.80)0.036AKI, n (%) a0.429 None182 (35.1)152 (35.4)30 (33.7) I168 (32.4)133 (31.0)35 (39.3) II113 (21.8)96 (22.4)17 (19.1) III55 (10.6)48 (11.2)7 (7.9)Ethnicity, n (%)0.289 Other117 (22.6)90 (21.0)27 (30.3) White329 (63.5)276 (64.3)53 (59.6) Black38 (7.3)34 (7.9)4 (4.5) Latino19 (3.7)17 (4.0)2 (2.2) Asian15 (2.9)12 (2.8)3 (3.4)Weight, mean (SD)81.78 (24.28)82.85 (24.95)76.58 (20.08)0.027Height, mean (SD)169.43 (10.76)169.50(10.65)169.07 (11.33)0.729BMI, mean (SD)28.46 (8.19)28.81 (8.44)26.77 (6.63)0.033Ventilation, n (%)381 (73.6)313 (73.0)68 (76.4)0.59Surgery, n (%)351 (67.8)297 (69.2)54 (60.7)0.148Trauma, n (%)160 (30.9)144 (33.6)16 (18.0)0.006Positive culture, n (%)43 (8.3)35 (8.2)8 (9.0)0.962Time ARC onset, mean (SD)3.38 (2.27)3.43 (2.32)3.15 (1.97)0.281CrCl, mean (SD)234.67 (130.83)234.72(130.43)234.46 (133.47)0.987Infection sites Upper respiratory infection, n (%)67 (12.9)54 (12.6)13 (14.6)0.732 Pneumonia, n (%)127 (24.5)104 (24.2)23 (25.8)0.854 Abdominal infection, n (%)44 (8.5)39 (9.1)5 (5.6)0.389 Urinary infection, n (%)95 (18.3)79 (18.4)16 (18.0)1 Blood infection, n (%)124 (23.9)96 (22.4)28 (31.5)0.091 Soft tissue infection, n (%)10 (1.9)10 (2.3)0 (0.0)0.302 CNS infection, n (%)14 (2.7)14 (3.3)0 (0.0)0.171 Other infection, n (%)242 (46.7)205 (47.8)37 (41.6)0.341Organ Dysfunction SOFA, mean (SD)5.52 (3.13)5.58 (3.12)5.21 (3.16)0.311 Use vasoactive drugs, n (%)209 (40.3)173 (40.3)36 (40.4)1 Dopamine use, n (%)17 (3.3)14 (3.3)3 (3.4)1 Vasopressin use, n (%)44 (8.5)33 (7.7)11 (12.4)0.219 Norepinephrine use, n (%)141 (27.2)118 (27.5)23 (25.8)0.849 Epinephrine use, n (%)14 (2.7)11 (2.6)3 (3.4)0.946 Phenylephrine use, n (%)103 (19.9)88 (20.5)15 (16.9)0.521 Urine-output, mean (SD)2168.48 (1700.80)2195.95 (1738.55)2036.08 (1507.55)0.42Comorbidities CCI, mean (SD)3.78 (2.57)3.64 (2.47)4.42 (2.97)0.01 Myocardial infarct, n (%)55 (10.6)44 (10.3)11 (12.4)0.691 Congestive heart failure, n (%)113 (21.8)88 (20.5)25 (28.1)0.152 Peripheral vascular disease, n (%)26 (5.0)21 (4.9)5 (5.6)0.986 Cerebrovascular disease, n (%)128 (24.7)97 (22.6)31 (34.8)0.022 Dementia, n (%)10 (1.9)6 (1.4)4 (4.5)0.131 Chronic pulmonary disease, n (%)135 (26.1)115 (26.8)20 (22.5)0.475 Rheumatic disease, n (%)16 (3.1)13 (3.0)3 (3.4)1 Peptic ulcer disease, n (%)10 (1.9)10 (2.3)0 (0.0)0.302 Mild liver disease, n (%)84 (16.2)74 (17.2)10 (11.2)0.214 Diabetes without complications, n (%)85 (16.4)74 (17.2)11 (12.4)0.329 Diabetes with complications, n (%)15 (2.9)13 (3.0)2 (2.2)0.957 Diabetes, n (%)93 (18.0)81 (18.9)12 (13.5)0.291 Paraplegia, n (%)45 (8.7)36 (8.4)9 (10.1)0.751 Renal disease, n (%)17 (3.3)16 (3.7)1 (1.1)0.353 Malignant cancer, n (%)75 (14.5)61 (14.2)14 (15.7)0.839 Severe liver disease, n (%)36 (6.9)30 (7.0)6 (6.7)1 Metastatic solid tumor, n (%)24 (4.6)14 (3.3)10 (11.2)0.003 AIDS, n (%)9 (1.7)9 (2.1)0 (0.0)0.351AKI, acute kidney injury; ICU, Intensive care unit; LOS, Lengths of stay; ARC, Augmented renal clearance; min, minimum; max, maximum; BMI, Body mass index; CrCL, Creatinine clearance; CNS, Central nervous system; SOFA, Sepsis-related organ failure assessment; CCI, Charlson comorbidity index; AIDS, Acquired immune deficiency syndrome Baseline characteristics of sepsis patients with ARC stratified by 28-day outcome. 169.50 (10.65) 234.72 (130.43) AKI, acute kidney injury; ICU, Intensive care unit; LOS, Lengths of stay; ARC, Augmented renal clearance; min, minimum; max, maximum; BMI, Body mass index; CrCL, Creatinine clearance; CNS, Central nervous system; SOFA, Sepsis-related organ failure assessment; CCI, Charlson comorbidity index; AIDS, Acquired immune deficiency syndrome LASSO regression analysis of data collected within first 24 h of ICU admission. (A) Tuning parameter (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria; (B) LASSO coefficient profiles of the 162 baseline features; (C) Frequency of variable selection by bootstrap lasso. The y-axis indicates selection frequency, and the x-axis lists the predictor variables. Limited differences in vital signs and laboratory parameters were observed between non-survivors and survivors. However, during the first 24 h of ICU admission, non-survivors had higher maximum diastolic blood pressure (97.53 vs. 91.64 mmHg, p = 0.014), higher minimum and mean respiratory rates (16.41 vs. 14.94 breaths/min, p = 0.019), and lower minimum and mean body temperatures (36.20 vs. 36.40 °C, p = 0.019; 36.86 vs. 37.11 °C, p = 0.03). Laboratory findings during this period showed higher levels of glucose, hematocrit, hemoglobin, and serum sodium in non-survivors (Supplementary Table S1). At the time of ARC diagnosis, non-survivors exhibited higher heart rates (100.1 vs. 94.1 bpm, p = 0.003) and higher respiratory rates (23.47 vs. 21.98 breaths/min, p = 0.004). Laboratory parameters showed significant elevations in electrolytes, including calcium, chloride, and sodium among non-survivors (Supplementary Table S2). Taken together, these findings suggest that higher diastolic blood pressure, respiratory rate, and specific laboratory markers (glucose, hematocrit, hemoglobin, sodium, calcium, and chloride) may be associated with higher mortality in sepsis patients with ARC. Limited predictive utility of individual indicators and feature selection Initially, we assessed the predictive performance of individual laboratory parameters, vital signs, and the SOFA score for 28-day mortality. As a standalone predictor, the SOFA score achieved an AUC of only 0.58. All other single indicators also showed limited predictive ability, with AUC values below 0.6 (Supplementary Data Figure S1). These results indicate that neither the SOFA score nor any single clinical parameter provides adequate discrimination for mortality risk in ARC patients, underscoring the need for a dedicated prediction tool. Using the three-stage feature selection process described in the Methods section, we identified 9 parameters from the early ICU admission dataset and 10 parameters from the ARC diagnosis dataset (Supplementary Tables S3 and S4). All selected variables were confirmed by clinical experts, forming a concise yet clinically meaningful predictor set for subsequent model development. Model performance and calibration To predict 28-day mortality in sepsis patients with ARC, we employed nine machine learning models using parameters selected by Bootstrap LASSO regression. When trained on features from the first 24 h of ICU admission (Supplementary Table S3), XGBoost demonstrated the best predictive performance among all algorithms tested in our study, achieving an AUC of 0.80, outperforming RF (AUC = 0.70), SVM (AUC = 0.78), DT (AUC = 0.64), KNN (AUC = 0.46), LR (AUC = 0.74), GBM (AUC = 0.77), LGBM (AUC = 0.62) and FL (AUC = 0.76) (Fig. 3A). The optimal hyperparameters for each algorithm, including learning rate (eta), the number of estimators, and the maximum tree depth, are detailed in the supplementary files (Supplementary File 1). When using parameters selected at the time of ARC diagnosis (See Supplementary Data Table S4), all nine models yielded AUC values below 0.80 (Fig. 3B). The receiver operating characteristic (ROC) curves for the nine models in the validation set for data within 24 h of ICU admission (A) and at ARC onset (B), respectively. Abbreviation: SVM, support vector machine; XGBoost, Extreme Gradient Boosting; KNN, k-nearest neighbors; GBM, Gradient Boosting Machine; LGBM, Light GBM; FL, Firth logistic regression; AUC, area under the curve. Beyond AUC, we assessed comprehensive performance metrics using the 24-hour ICU admission data (Supplementary Table S5). XGBoost achieved the highest AUC (0.80), with an accuracy of 0.74, a sensitivity of 0.65, a specificity of 0.75, a PPV of 0.34, and a NPV of 0.91. Firth logistic regression showed complementary characteristics, achieving a high sensitivity (0.778) and NPV (0.933), along with an F1 of 0.459 and a MCC of 0.333. Similarly, GBM also showed high sensitivity (0.83) and the highest NPV (0.94), but with lower specificity (0.588). In contrast, several models exhibited notable limitations: SVM and KNN failed to identify any deaths (sensitivity 0.00), while random forest missed most deaths (sensitivity 0.18). Calibration analysis revealed a trade-off between discrimination and calibration (Fig. 4). While XGBoost showed systematic deviation from ideal calibration, Firth logistic regression demonstrated well-calibrated probability estimates. This trade-off reflects the complementary nature of the two approaches: XGBoost prioritizes ranking accuracy for risk stratification, whereas Firth regression provides more reliable absolute risk estimates. Given our primary clinical goal of identifying high-risk ARC patients for targeted monitoring, XGBoost was selected as the primary model. Calibration curves for the nine ML models in the validation set. Abbreviations: SVM, support vector machine; XGBOOST, Extreme Gradient Boosting; DT TREE, Decision Tree; RF, Random Forest; LOGISTIC, logistic regression; KNN, k-nearest neighbors; GBM, Gradient Boosting Machine; LGBM, Light GBM; FL, Firth logistic regression. The data source is the test results within 24 h of entering the ICU. Model interpretation To gain deeper insights into how the XGBoost model predicts outcomes in sepsis patients with ARC, we employed SHAP values for model interpretation. The SHAP summary plot identified respiratory rate, body temperature and serum sodium levels as the top three features with the highest SHAP values, indicating their significant contributions to the model’s predictions. These parameters were deemed the most critical in determining patient outcomes (Fig. 5A,B). To further validate the model’s interpretability, we randomly selected two patients from the validation set and visualized their predictions using SHAP force plots. For the surviving patient, the model’s output value f(x) was 1.51, reflecting a high probability of survival (Fig. 5C). In contrast, for the non-surviving patient, the model’s output value f(x) was − 2.12, indicating a high probability of mortality (Fig. 5D). These findings demonstrate that the XGBoost model not only achieves accurate predictions but also provides transparent and interpretable insights into the contribution of individual features. SHAP summary plot for the top nine clinical features and SHAP force analysis for explaining individual’s prediction results. (A) Ranking of the feature importance by SHAP. The matrix plot shows the importance of each feature in the development of the final prediction model; (B) the attributes of the features in the black box model. Each line represents features, and the abscissa is the SHAP value. yellow dots represent higher feature values, and dark purple dots represent lower values; (C and D) positive (C) and negative (D) force plot. The yellow bars and purple bars represent the parameters predicting death and survival in the model, respectively. The length of the bar for each feature indicates the importance of that feature in making the prediction. Abbreviation: SHAP, Shapley additive explanation; na, serum sodium; t,temperature; rr, respiratory rate. To enable further research and external validation, we developed a freely accessible online calculator, available at: https://wuyunzhe.shinyapps.io/arc_prediction/. This tool allows clinicians to input the nine core predictors, generating real-time mortality risk predictions with automatic risk stratification (low < 50% and high ≥ 50%). However, the calculator is intended for research, educational, and hypothesis‑generation purposes only, and is not yet validated for independent clinical decision-making. Discussion Sepsis patients are at high risk for developing augmented renal clearance (ARC). This condition was reported in 13–51% of cases during ICU stays14. As a life-threatening complication, ARC significantly alters pharmacokinetic profiles and diminishes the efficacy of essential medications such as antibiotics. As a result, it exacerbates the clinical challenge of managing this vulnerable population14. Therefore, accurate prediction of mortality risk is crucial for optimizing treatment strategies. To address this need, our study leveraged the MIMIC-IV database to develop and validate multiple machine learning models for predicting 28-day mortality in sepsis patients with ARC. We analyzed baseline characteristics between survivors and non-survivors at 28 days among sepsis patients with ARC. The mortality group exhibited distinct demographic and clinical features, including older age, higher CCI scores, and a greater prevalence of cerebrovascular disease and metastatic solid tumors. These observations are consistent with established literature that elderly patients with pre-existing chronic conditions tend to have more severe consequences in response to the systemic inflammation in sepsis24–26. Our findings align with the well-documented “obesity paradox” in critical care27. Patients with elevated BMI may have better physiological reserves and metabolic adaptability. This may help them face the metabolic demands of sepsis and ARC. The potential mechanisms underlying this paradox in septic patients with ARC were unclarified, probably related to the enhanced energy reserves, altered inflammatory responses, or improved stress tolerance during acute illness28. Trauma patients also exhibited better survival outcomes. Previous studies suggest that trauma patients are generally younger and have fewer pre-existing comorbidities29. This may partially explain their improved survival. Additionally, the pathophysiological mechanisms underlying trauma-associated sepsis may differ from those of other sepsis etiologies, potentially contributing to variations in both the occurrence of ARC and clinical outcomes among these patients30. However, whether these differences in baseline characteristics are determinants of survival in sepsis patients with ARC remains unclear. This uncertainty underscores the need for more comprehensive clinical epidemiological analyses in future studies to elucidate these potential relationships. Compared to single-parameter prediction, machine learning models demonstrated superior predictive accuracy in this high-risk ARC population. Among these models, XGBoost achieved the highest discriminative ability (AUC 0.80). Consistent with our results, the SOFA score showed limited discriminative ability in this ARC cohort and did not differ between survivors and non-survivors. This finding suggests that generic severity scores may not adequately capture mortality risk in ARC patients, further supporting the need for an ARC‑specific prediction tool. However, the performance comparison revealed an important trade-off between discrimination and calibration. XGBoost excelled in risk ranking. In contrast, Firth logistic regression specifically designed for rare events through penalized likelihood estimation demonstrated superior calibration, indicating its strength in identifying high-risk patients and ruling out mortality. This complementary performance reflects the distinct strengths of the two approaches. As a gradient-boosted decision tree algorithm, XGBoost excels at capturing complex nonlinear relationships and interactions among the nine selected predictors. This contributes to its superior discriminative ability. Its capacity to handle heterogeneous data types and resist overfitting through built-in regularization makes it suited for real-world clinical datasets31. XGBoost has been successfully applied across diverse clinical contexts, including predicting acute kidney injury in sepsis patients32, evaluating mortality in COVID-19 patients33, and diagnosing secondary hypertension34. These applications confirme its versatility and reliability in clinical research35. Conversely, Firth logistic regression’s superior calibration aligns with its theoretical properties. The penalty term shrinks extreme predictions toward the mean, yielding more conservative but better-calibrated probability estimates36. Such estimates are particularly valuable when accurate absolute risk estimates are needed for individual patient counseling. Additionally, its higher sensitivity (0.78 vs. XGBoost’s 0.65) suggests that Firth regression may be preferable when the clinical priority is to minimize missed deaths. Given our primary clinical goal of risk stratification to guide monitoring intensity and resource allocation in ICU settings, XGBoost’s superior discriminative ability makes it the preferred model for identifying high-risk ARC patients. However, we acknowledge that its calibration limitations warrant attention. Recent methodological advances suggest that calibration can be improved through techniques such as Platt scaling or isotonic regression37. We plan to explore in future iterations. Furthermore, sophisticated ensemble approaches such as stacking and blending may further enhance predictive performance by strategically combining the complementary strengths of XGBoost and Firth regression38. The rapidly evolving clinical course of sepsis in ICU patients poses significant challenges for timely prognosis and intervention39. Our study reveals that ARC in septic patients typically develops at a median of 3.38 days after ICU admission. This finding aligns with the work of Mikami et al.40, who reported that ARC onset occurs within 3 days of admission in approximately half of the cases, and within 1 week in most of the remaining cases. These observations highlight a critical window for early risk stratification in septic patients. Machine learning models using data from the first 24 h of ICU admission performed better than those using data collected at ARC diagnosis. This provides important pathophysiological insights. It suggests that physiological parameters measured during the initial ICU period may more sensitively reflect the underlying pathological state of sepsis patients. ARC may represent a manifestation of the hypermetabolic state in sepsis. Its development is likely driven by early inflammatory and hemodynamic disturbances41. Since patients undergo various medical interventions upon ICU admission, these treatments may mask critical warning signs. Therefore, initial ICU parameters may more directly reflect these fundamental pathological processes. The limited predictive utility of isolated laboratory parameters underscores the complex pathophysiology of sepsis with ARC. Through SHAP analysis, we identified three key abnormalities in non-survivors: elevated respiratory rate, lower body temperature, and higher serum sodium. These are common prognostic markers in general sepsis populations. However, their prominence in our ARC-specific model reflects additional pathophysiological significance in the context of enhanced renal function. Tachypnea in general sepsis primarily reflects metabolic compensation. In ARC patients, however, elevated respiratory rate represents a compensatory response to inflammatory-mediated mitochondrial dysfunction and glycolytic activation. This response increases CO₂ production42. Similarly, hypothermia takes on added meaning in the ARC context. In general sepsis, it typically indicates immune paralysis. In our ARC cohort, hypothermia may be attributed to increased cutaneous blood flow in hyper filtering states. This is compounded by vasodilation from sedative-analgesic medications, leading to accelerated core heat loss43. Notably, previous studies indicate that patients with core temperature < 36 °C within 24 h of ICU admission face a 28-day mortality rate of 54.3%. This rate is approximately three times higher than that of patients with fever43. Regarding hypernatremia in ARC patients, it provides a direct window into the “hyperfiltration-hypo reabsorption” renal phenotype. Increased sodium filtration and reduced reabsorption time lead to osmotic diuresis44. The convergence of these three abnormalities signals a critical transition to decompensated sepsis-ARC. Together, they point to a common underlying mechanism: enhanced renal function driving altered drug pharmacokinetics. From a hypothesis-generating perspective, this triad may serve as potential clinical signals in ARC patients. Even subtle deviations could prompt consideration of therapeutic drug monitoring, dose escalation of renally cleared antibiotics, and heightened vigilance for treatment failure. Importantly, the model is currently exploratory and requires external validation before it can be considered for routine clinical use. Despite these promising findings, several limitations should be acknowledged. First, the absence of external validation is the primary limitation of this study. Our model was developed and validated solely on the MIMIC-IV database. Therefore, its performance may not generalize to other populations. This represents a significant gap, as the generalizability of our findings requires verification through additional clinical cohorts. Future studies should prioritize collecting and analyzing external datasets to validate the model’s performance across diverse populations and clinical settings. Second, the relatively small sample size limits statistical power and may lead to overfitted models. Although we employed Bootstrap LASSO and cross-validation to mitigate this risk, validation in larger cohorts remains necessary. Third, while we compared our model with the SOFA score, other established ICU prognostic tools (e.g., SAPS II, APACHE II) were not included due to missing component variables in the MIMIC-IV database. Future studies using datasets with complete information should conduct such comparisons to further evaluate the added value of our model. Fourth, while we identified several differential indicators between survivors and non-survivors, the precise relationships between these parameters and patient outcomes remain to be fully elucidated. Specifically, the clinical relevance and optimal cutoff values of these predictive markers need further investigation to provide clinicians with more definitive guidance. Fifth, the ARC definition used in this study (CrCl ≥ 130 mL/min/1.73 m²), while consistent with key prior studies, is not universally standardized. Definitions vary across the literature, with thresholds ranging from 120 to 150 mL/min and methods including both measured and estimated CrCl. This variability represents a limitation, as our findings may not fully generalize to populations defined by alternative criteria. To ensure long‑term clinical utility after deployment, we have developed a maintenance protocol for the XGBoost model based on dual‑dimension monitoring. Specifically, the Population Stability Index (PSI) will be used to detect data drift, with thresholds of 0.1 for mild drift, and 0.25 for severe drift. Mild drift triggers incremental learning. Severe drift triggers full retraining. Performance will be monitored monthly against initial baselines, with alerts triggered by AUC decline more than 5% or an F1 decline more than 10%45. Incremental learning will be implemented using the xgb.update function in the R xgboost package, while full retraining will be required when AUC declines more than 10%, F1 declines more than 15%, or PSI exceeds 0.25. The complete R code for this protocol is provided in Supplementary File 3. In conclusion, this study demonstrates that the XGBoost model achieves optimal performance in predicting 28-day mortality for sepsis patients with ARC using the MIMIC database. Our findings suggested that utilizing data from the first day of ICU admission yields superior predictive accuracy. Furthermore, through the application of SHAP values and force plots, we have enhanced the model’s interpretability. This provides an interpretable tool with potential future utility for predicting 28-day mortality in septic patients with ARC. Currently, the model should be viewed as a proof‑of‑concept for ARC‑specific risk stratification. Therefore, prospective validation and impact analysis are required before any clinical application can be considered. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Acknowledgements None. Author contributions Conceptualization: Dong D, Lu Y, Wang X; Methodology: Wu Y, Wu T; Investigation: Wu Y, Yang H, Yang F, Zhuang R; Visualization: Wu Y, Yang F; Funding acquisition: Dong D, Wang X; Project administration: Wu T; Supervision: Lu Yi; Writing – original draft: Wu Yun, Yang F; Writing – review & editing: Dong D. Funding This work was supported by the “Shanghai “Rising Stars of Medical Talents” Youth Development Program-Clinical Laboratory Practitioner Program” and “Shanghai Municipal Health Commission’s Seed Program for Medical New Technology Research and Translation [Grant No. 2024ZZ2045]”. Data availability Supporting data is available upon reasonable request from corresponding authors. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yunzhe Wu, Fan Yang,and Hongjie Yang contributed equally to this work. Contributor Information Yide Lu, Email: yidelu@sina.com. Danfeng Dong, Email: ddf40688@rjh.com.cn. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Supplementary Material 5 Data Availability Statement Supporting data is available upon reasonable request from corresponding authors.