Prognostic Factors and Risk Stratification in Malignant Tumor Patients With Sepsis: A Retrospective Cohort Study.

Introduction

Sepsis is a life-threatening complication and a leading cause of mortality among patients with cancer, accounting for up to 40% of sepsis-related intensive care unit (ICU) admissions and carrying a fatality rate exceeding 50% in high-risk subgroups [ 1 – 6 ]. The intersection of malignancy-associated immunosuppression, chemotherapy toxicity, and sepsis-induced organ dysfunction creates a uniquely vulnerable population with disproportionately poor outcomes [ 7 – 10 ]. Sepsis is frequently encountered in patients with cancer, particularly in ICU settings, where it is associated with high morbidity and mortality [ 11 – 14 ]. While cancer-related factors – such as tumor type, stage, and treatment history – have been extensively studied as prognostic determinants, emerging evidence suggests that acute ICU-specific physiological abnormalities may play a more critical role in short-term survival [ 15 – 17 ], underscoring an urgent need for improved risk stratification tools tailored to this population.

Current prognostic models for cancer patients with sepsis predominantly incorporate tumor characteristics, such as malignancy type and metastasis status, and general sepsis severity scores [ 18 ]. However, these models often fail to account for dynamic ICU interventions, such as vasopressor use and mechanical ventilation, or real-time biomarker trends, such as lactate clearance and platelet kinetics, which may better reflect imminent clinical deterioration. This oversight can lead to suboptimal risk stratification, delaying aggressive interventions in high-risk patients.

We hypothesized that combining malignancy-related factors, such as comorbidity burden and treatment history, with dynamic sepsis severity markers would create a better prognostic model for 28-day mortality in cancer-associated sepsis. To test this, we performed a large-scale retrospective study to identify independent predictors of mortality through multivariable modeling, develop and validate a practical nomogram, and measure the model’s usefulness using decision curve and clinical impact analyses.

Material and Methods

Data Source

This retrospective cohort study utilized data from the Medical Information Mart for Intensive Care (MIMIC-IV, version 2.0) [ 19 ], a publicly available, de-identified database containing comprehensive electronic health records for over 250 000 ICU admissions at Beth Israel Deaconess Medical Center (2008–2019). MIMIC-IV integrates clinical data, including vital signs, laboratory results, medications, and procedural records, enabling longitudinal analysis of critical care outcomes.

Study Population

The study cohort was extracted using structured query language (SQL) queries executed in PostgreSQL (v13.0), with the complete code repository, including patient selection, variable extraction, and sepsis and malignancy definitions. Sepsis was defined according to Sepsis-3 criteria (suspected infection + Sequential Organ Failure Assessment (SOFA) score of 2 or higher), in which infection required concurrent antibiotic administration and microbiological culture collection, while SOFA components were extracted from ICU chart events [ 20 ]. Malignancy cases were identified using comprehensive ICD-9 and ICD-10 diagnostic codes, encompassing solid and hematologic malignancies while explicitly excluding benign neoplasms (ICD-9 210-229; ICD-10 D10–D36). Solid tumors were classified by primary anatomical site, including lung (ICD-10 C34.x), breast (C50.x), prostate (C61), colorectal (C18–C20), liver (C22.x; hepatocellular carcinoma C22.0, intrahepatic bile duct C22.1), pancreatic (C25.x), gastric (C16.x), esophageal (C15.x), ovarian (C56), endometrial (C54.x), cervical (C53.x), bladder (C67.x), renal (C64.x), and brain malignancies (C71.x), as well as rarer subtypes, such as head and neck squamous cell carcinoma (C00–C14), thyroid carcinoma (C73), and sarcomas (C49.x). Hematologic malignancies consisted of myeloid neoplasms (acute myeloid leukemia [C92.0]), myelodysplastic syndromes (D46.x), lymphoid malignancies (chronic lymphocytic leukemia [CLL, C91.1], acute lymphoblastic leukemia [C91.0]), lymphomas (Hodgkin lymphoma [C81.x], non-Hodgkin lymphomas [C82–C85], including diffuse large B-cell lymphoma [C83.3]), and plasma cell disorders (multiple myeloma [C90.0] and Waldenström macroglobulinemia [C88.0]). Additionally, rare but clinically significant malignancies were captured, such as neuroendocrine tumors (C7A.x), germ cell tumors (C62–C63), and metastatic cancers of unknown primary (C80). Metastatic disease was rigorously confirmed through either secondary ICD codes (eg, ICD-10 C77–C79 for metastatic sites) or radiology reports documenting distant metastases. This approach ensured systematic capture of malignancy subtypes while minimizing misclassification of benign or premalignant conditions. We included adult patients 18 years of age or older with documented malignancy and sepsis as defined above. Patients with incomplete baseline data (eg, missing SOFA components or lactate within 24 hours of ICU admission) or those discharged or deceased within 24 hours were excluded. The final cohort included 4301 patients, stratified by 28-day survival status for comparative analyses.

Data Collection

All clinical variables, including demographics, vitals, laboratory results, and interventions, were extracted from MIMIC-IV tables (chartevents, labevents, procedures_icd) using time-stamped queries, with sepsis-related variables anchored to ICU admission time. We extracted demographic, clinical, and laboratory variables from the MIMIC-IV database, with all admission-related data collected within the first 24 hours of ICU stay. Baseline characteristics included age, sex, height, weight, malignancy type (solid or hematologic), metastatic status, and prior chemotherapy or radiation exposure, infection site, and pathogenic microorganisms Comorbidities, such as hypertension, diabetes, myocardial infarction, congestive heart failure, liver disease (mild or severe), chronic kidney disease, chronic pulmonary disease, cerebrovascular disease, dementia, new-onset atrial fibrillation, hypercoagulable state, immunosuppression, and AIDS, were also recorded. Vital signs recorded at admission included temperature, heart rate, respiratory rate, systolic and diastolic blood pressure, mean arterial pressure, and oxygen saturation (SpO). Laboratory parameters included hematologic markers (white blood cell count, hematocrit, platelet count, hemoglobin), renal and metabolic profiles (creatinine, blood urea nitrogen, anion gap, potassium, sodium, chloride, glucose), coagulation tests (prothrombin time, activated partial thromboplastin time, international normalized ratio), and lactate levels as a key indicator of tissue perfusion. The severity of illness was assessed using validated ICU scoring systems, including the Simplified Acute Physiology Score II (SAPS II), Simplified Acute Physiology Score III (SAPS III), SOFA, Logistic Organ Dysfunction Score, Systemic Inflammatory Response Syndrome score, Glasgow Coma Scale score, Oxford Acute Severity of Illness Score, and Model for End-Stage Liver Disease (MELD) score, while interventions and complications, such as acute kidney injury (Kidney Disease Improving Global Outcomes criteria), invasive mechanical ventilation, cardiopulmonary resuscitation events, and continuous renal replacement therapy, were documented. Hospital course was evaluated through ICU and total hospital length of stay, with 28-day mortality serving as the primary outcome for comparative analysis between the survival and non-survival groups.

Statistical Analysis

To ensure data accuracy prior to analysis, all collected variables underwent rigorous quality control procedures. Missing or implausible values were cross-verified against source documentation when available. Continuous variables were examined for outliers using Tukey’s method (1.5×interquartile range), while categorical variables were checked for consistency with predefined coding schemes.

All statistical analyses were performed using R software (version 4.3.0). Continuous variables are presented as median (interquartile range) or mean (±standard deviation), depending on data distribution, and compared using the Mann-Whitney U test or independent samples t test, as appropriate. Categorical variables are expressed as frequency (percentage), and group comparisons were performed using the chi-square test or Fisher’s exact test.

To assess the association between variables and 28-day mortality, we first conducted univariable logistic regression analysis, with variables showing a significance level of P <0.05 being included in the subsequent multivariable analysis. A backward stepwise selection method (exclusion criterion: P >0.05) was applied to construct the final multivariable logistic regression model, adjusting for potential confounders. Odds ratios (ORs) with 95% CIs were calculated to quantify effect sizes. The goodness-of-fit of the final model was assessed using the Hosmer-Lemeshow test, and multicollinearity was examined using variance inflation factors (VIF <5 considered acceptable). To assess multicollinearity among the candidate predictors, we calculated the VIF, excluding variables with a VIF over 5 from the final multivariable model to ensure stability. Hematocrit (VIF: 6.06), hemoglobin (VIF: 6.03), prothrombin time (VIF: 9.13), and international normalized ratio (VIF: 9.26) exhibited significant multicollinearity and were therefore excluded, while all remaining variables demonstrated acceptable collinearity (VIF ≤5) and were retained for further analysis.

We divided the data into a training set and a validation set according to a 7: 3 ratio. A predictive nomogram was developed based on the independent risk factors identified in the multivariable logistic regression analysis training set. The nomogram was constructed using the ‘rms’ package in R, assigning weighted points to each predictor according to its regression coefficient. To assess its performance, we evaluated discrimination, calibration, and clinical utility in the training set and validation set. The discriminative ability was measured using the area under the receiver operating characteristic (ROC) curve (AUC), while calibration was assessed via a calibration curve with bootstrap resampling (1000 iterations) and the Hosmer-Lemeshow test. Decision curve analysis (DCA) and the clinical impact curve were used to quantify the net benefit of the nomogram across different risk thresholds and demonstrate its clinical applicability. Internal validation was performed using 1000 bootstrap samples to correct for overfitting and estimate optimism-adjusted performance metrics. All statistical analyses were conducted in R (version 4.3.0) with packages including ‘rms’, ‘pROC’, ‘rmda’, and ‘ggplot2’ for visualization. For participants with missing data exceeding 20% in key liver function parameters (alanine aminotransferase, aspartate aminotransferase, or bilirubin), we retained their data and addressed missing values through multiple imputation using chained equations (m=5 imputations) P <0.05 was considered statistically significant.

Outcomes

Baseline Data of Patients With Malignant Tumors and Sepsis

Among 9522 patients diagnosed with metastatic solid tumors, we excluded 537 patients with a hospital stay of less than 24 hours and 1201 patients with more than 20% missing data. Of the remaining cohort, 4301 patients were diagnosed with sepsis and were categorized into survival (n=2949) and non-survival (n=1352) groups. Comparative analysis revealed significant differences between the survival and non-survival groups. Non-survivors had lower weight ( P <0.001), although sex distribution was similar ( P =0.358). Comorbidities such as myocardial infarction ( P =0.008), congestive heart failure ( P =0.010), mild liver disease ( P <0.001), severe liver disease ( P <0.001), chronic kidney disease ( P =0.011), cerebrovascular disease ( P =0.031), and dementia ( P =0.019) were more prevalent in non-survivors.

Vital signs also differed significantly: non-survivors had higher heart rates ( P <0.001) and respiratory rates ( P <0.001) and lower systolic blood pressure ( P <0.001), mean arterial pressure ( P =0.011), and oxygen saturation ( P <0.001). The laboratory test results revealed significant differences between the survival and non-survival groups. The non-survival group had a higher proportion of patients with leukopenia (white blood cell count <4×10 9 /L, P =0.605), lower hematocrit levels ( P =0.004), and more frequent thrombocytopenia (platelet count <150×10 9 /L, P =0.006). Hemoglobin levels were also lower in non-survivors ( P <0.001). Renal function markers were markedly worse in non-survivors, with higher creatinine ( P <0.001) and blood urea nitrogen levels ( P <0.001). Coagulation parameters were significantly more abnormal in non-survivors, with prolonged prothrombin time, elevated partial thromboplastin time, and higher international normalized ratio (all P <0.001). Metabolic disturbances were more pronounced in non-survivors, with wider anion gap, lower chloride levels, and higher potassium levels (all P <0.001). Lactate levels over 4 mmol/L were more frequent in non-survivors ( P <0.001). Additionally, bloodstream infections were more frequent in non-survivors ( P =0.002) ( Table 1 ).

Outcomes of Patients With Malignant Tumors and Sepsis

Significant differences were observed in clinical severity and outcomes between survivors and non-survivors. Non-survivors had shorter hospital stays ( P <0.001) but longer ICU stays ( P =0.001). Disease severity scores were markedly higher in non-survivors, including SAPS III, SAPS II, SOFA, Logistic Organ Dysfunction Score, Charlson Comorbidity Index, MELD, and Oxford Acute Severity of Illness Score (all P <0.001). Non-survivors also had worse organ dysfunction, reflected by lower Glasgow Coma Scale scores ( P< 0.001), higher Systemic Inflammatory Response Syndrome scores ( P <0.001), and higher rates of acute kidney injury (AKI; P <0.001) and septic shock ( P <0.001). Interventions were more frequent in non-survivors, including invasive mechanical ventilation ( P <0.001), cardiopulmonary resuscitation (CPR; P <0.001), and continuous renal replacement therapy ( P <0.001), while vasoactive drug use showed a marginal trend ( P =0.050). These findings highlight the association between higher disease severity, multi-organ dysfunction, aggressive interventions, and mortality in this population ( Table 2 ).

Multivariate Logistic Analysis of 28-Day Mortality Rate of Patients With Malignant Tumors and Sepsis

Multivariate logistic regression analysis revealed several independent risk factors significantly associated with mortality in critically ill cancer patients with sepsis ( Figure 1 ). Prolonged ICU length of stay was significantly associated with increased risk (OR, 1.065; 95% CI, 1.039–1.092; P <0.001). Other independent predictors were an elevated SAPS III score (OR, 1.013; 95% CI, 1.005–1.021; P =0.002), SOFA score (OR, 1.182; 95% CI, 1.132–1.235; P <0.001), Charlson Comorbidity Index (OR, 1.263; 95% CI, 1.210–1.319, P <0.001), and MELD score (OR, 1.018; 95% CI, 1.002–1.035; P =0.030), the presence of AKI (OR, 1.936; 95% CI, 1.530–2.459; P <0.001), receipt of CPR (OR, 6.171; 95% CI, 2.049–19.627; P =0.002), and the presence of septic shock (OR, 1.382; 95% CI, 1.106–1.725; P =0.004). Additionally, a higher heart rate (OR, 1.011; 95% CI, 1.006–1.016; P <0.001), respiratory rate (OR, 1.046; 95% CI, 1.031–1.061; P <0.001), mean arterial pressure (OR, 1.007; 95% CI, 1.000–1.013; P =0.050), blood urea nitrogen (BUN) level (OR, 1.018; 95% CI, 1.012–1.024; P <0.001), and serum sodium level (OR, 1.042; 95% CI, 1.017–1.067; P <0.001) were significant predictors ( Figure 1 ).

Evaluation of a Nomogram For Predicting 28-Day Mortality With a Training and Validation Set

The nomogram model for predicting mortality risk incorporated the following independent predictors: AKI, CPR, heart rate, respiratory rate, BUN level, septic shock, Charlson Comorbidity Index, serum sodium level, and MELD score. The nomogram model demonstrated robust statistical power, with an overall events-per-variable ratio of 148.56. The events-per-variable ratio was 1377: 1 across all analyzed variables ( Figure 2 ).

The nomogram’s performance was further evaluated in an internal validation cohort, demonstrating strong predictive accuracy. ROC curve analysis demonstrated that the proposed predictive model achieved an AUC of 0.741 (95% CI, 0.722–0.759) in the original cohort, outperforming 2 widely used clinical scoring systems: SAPS III (AUC, 0.726; 95% CI, 0.707–0.745) and SOFA (AUC, 0.696; 95% CI, 0.676–0.716) ( Figure 3A ). The calibration curve of the predictive model in the training set demonstrated close agreement between predicted probabilities and observed outcomes ( Figure 3B ). The bias-corrected curve closely followed the ideal reference line, supported by a calibration slope of 0.868 and a calibration intercept of 0.036, indicating minimal systemic over- or under-prediction. The model’s overall accuracy was further validated by a C-index (concordance statistic) of 0.741, consistent with its discriminative performance in ROC analysis, and a Brier score of 0.185, reflecting good predictive precision for 28-day mortality. DCA confirmed sustained clinical utility, with the model maintaining superior net benefit over default strategies (treat-all or treat-none) across threshold probabilities of 10% to 50% ( Figure 3C ). Finally, the clinical impact curve ( Figure 3D ) demonstrated robust risk stratification, with the number of high-risk patients accurately reflecting true event rates, further supporting the model’s reliability for clinical decision-making. The validation cohort confirmed the model’s robustness, demonstrating strong calibration (slope, 0.94; intercept, 0.014; Brier score, 0.174), superior discrimination (AUC, 0.758; 95% CI, 0.73–0.786; outperforming SAPS III/SOFA), and clinical utility (net benefit in DCA across thresholds), with clinical impact curves aligning predicted risk with observed events ( Figure 4A–4D ).

Discussion

Patients with malignant tumors and sepsis face a poor prognosis, with mortality rates exceeding 30% in high-risk subgroups, underscoring the urgent need for improved risk stratification tools. In this study, we developed and validated a prognostic nomogram to predict 28-day mortality in this vulnerable population. The nomogram demonstrated strong discriminative ability, excellent calibration, and meaningful clinical utility, as evidenced by the DCA and clinical impact curve. These findings suggest that this model outperforms conventional severity scores (eg, SOFA alone and SAPS III) by integrating sepsis-related organ dysfunction and malignancy-specific comorbidities, providing clinicians with a practical tool for individualized risk assessment.

Our findings align with and extend prior research on sepsis prognosis in patients with cancer. As shown in Table 2 , non-survivors had significantly higher disease severity scores (SOFA, SAPS III, Logistic Organ Dysfunction Score) and greater comorbidity burdens (Charlson Index, MELD), compared with survivors, reinforcing the critical role of baseline vulnerability in determining outcomes. This is consistent with previous studies reporting that malignancy-associated immunosuppression, pre-existing organ dysfunction, and treatment-related toxicities synergistically worsen sepsis outcomes [ 21 – 23 ]. However, unlike prior models that focus primarily on acute physiological abnormalities, our nomogram incorporates longitudinal prognostic factors, such as the Charlson Index and AKI status, offering a more holistic risk assessment. This approach may help explain why our model achieved superior discrimination, particularly in identifying high-risk patients who might benefit from early, aggressive interventions.

The risk factors identified in our nomogram – the Charlson Comorbidity Index, respiratory rate, and BUN level – have been independently linked to poor outcomes in sepsis; however, their combined prognostic value in patients with cancer has not been previously quantified. Respiratory rate and BUN highlight the prognostic importance of respiratory and renal dysfunction, which are frequently exacerbated by malignancy-related factors, such as chemotherapy-induced nephrotoxicity and pulmonary metastases [ 24 , 25 ]. Notably, the Charlson Comorbidity Index emerged as a key predictor, suggesting that baseline comorbidity burden may be as critical as acute physiological abnormalities in determining outcomes. This has important clinical implications, as it underscores the need for personalized sepsis management strategies that account for both the acute illness severity and pre-existing vulnerabilities.

The clinical applicability of our nomogram is supported by its strong performance in the DCA and clinical impact curve. DCA confirmed that the model provides greater net benefit than “treat all” or “treat none” strategies across a clinically relevant risk threshold (10–50%), making it particularly useful for guiding ICU triage, early resuscitation, and palliative care discussions. The clinical impact curve further demonstrated that, at a 20% risk threshold, the nomogram correctly identifies approximately 250 high-risk patients per 1000, with a manageable number of false positives. This suggests that the model could optimize resource allocation by helping clinicians prioritize patients who may benefit most from intensive monitoring or aggressive interventions [ 26 , 27 ].

These quantitative improvements over existing tools are clinically significant: our model achieved higher AUC values than both the SOFA and SAPS III, with the greatest advantage seen in patients with intermediate risk, in whom conventional scores often misclassify risk. The model’s superior performance stems from its unique integration of malignancy-specific factors, namely the Charlson Comorbidity Index, with dynamic ICU parameters of AKI status and receipt of CPR, variables absent in SOFA and SAPS III but critical for patients with cancer. DCA confirmed these translational benefits, showing our model reduces unnecessary interventions by 15% to 30%, compared with SOFA, while maintaining high sensitivity for true high-risk cases, with a net benefit increase of 0.05 to 0.10 across clinically relevant thresholds.

Despite these strengths, our study has several limitations. First, the retrospective design introduced potential biases in data collection and unmeasured confounding. Our analysis did not capture employment status or socioeconomic factors, which may influence clinical outcomes. Second, the nomogram was derived and validated using a single-center database (MIMIC-IV), which can limit generalizability to other healthcare settings. Most critically, external validation in an independent multicenter cohort was not performed, which restricts the model’s extrapolation to broader populations. Future studies should prospectively validate this nomogram in diverse clinical settings to confirm its robustness. Third, additionally, while our model incorporates key prognostic variables, it does not account for recent advances in sepsis management or treatment status of tumors (eg, chemotherapy, immunotherapy). This information, which can influence outcomes in cancer patients, was not systematically collected in our dataset. Future studies with detailed treatment records could further stratify risks. Finally, some confounding factors affecting the prognosis of cancer patients with sepsis could not be included in this study due to data limitations. These factors – such as infection origin (nosocomial vs community-acquired), nursing home residence, bleeding diatheses, and procalcitonin trends – could not be analyzed. Future studies should explore these variables.

Conclusions

In this large cohort study, we developed and internally validated a prognostic nomogram to predict 28-day mortality in patients with malignant tumors and sepsis. The model demonstrated excellent discrimination, calibration, and clinical utility, outperforming conventional severity scores by incorporating acute physiological abnormalities and malignancy-specific risk factors. While further external validation is needed, this nomogram provides a practical tool for risk stratification, potentially guiding early interventions, ICU triage, and personalized treatment strategies in this population with high mortality.

Acknowledgments

We sincerely thank Dr. Lina Zhao for her expert review and validation of the statistical analyses in this study.

Abbreviations

acute kidney injury

area under the ROC curve

blood urea nitrogen

cardiopulmonary resuscitation

decision curve analysis

intensive care unit

Model For End-Stage Liver Disease

Medical Information Mart for Intensive Care

odds ratio

receiver operating characteristic

Simplified Acute Physiology Score

Sequential Organ Failure Assessment

variance inflation factor

Footnotes

Financial support: This work was supported by the grant from the Construction of China-Europe International Academic Exchange Platform for Critical Care Medicine (25YFKFYS00890)

Conflict of interest: None declared

Ethics Approval and Consent to Participate: This study utilized the MIMIC-IV database (version 2.0), which was approved by the Institutional Review Boards of the Beth Israel Deaconess Medical Center (2001-P001699/14) and the Massachusetts Institute of Technology (approval No. 33690380). The database is de-identified and compliant with the Health Insurance Portability and Accountability Act (HIPAA), waiving the need for individual patient consent.

Patient Permission/Consent Statement: Patient-related information was de-identified and no privacy has been exposed; therefore, informed consent was waived.

Declaration of Figures’ Authenticity: All figures submitted have been created by the authors who confirm that the images are original with no duplication and have not been previously published in whole or in part.

Availability of Data and Materials

The dataset used in this study (MIMIC-IV v2.0) is publicly available under a data use agreement after completing required training ( https://physionet.org/content/mimiciv/2.0/ ). The analysis code for this study has been deposited in GitHub.

References

1. Kochanek M, Schalk E, von Bergwelt-Baildon M, et al. Management of sepsis in neutropenic cancer patients: 2018 guidelines from the Infectious Diseases Working Party (AGIHO) and Intensive Care Working Party (iCHOP) of the German Society of Hematology and Medical Oncology (DGHO) Ann Hematol. 2019;98(5):1051–69. doi: 10.1007/s00277-019-03622-0.

2. Nates JL, Pène F, Darmon M, et al. Nine-I Investigators. Septic shock in the immunocompromised cancer patient: A narrative review. Crit Care. 2024;28(1):285. doi: 10.1186/s13054-024-05073-0.

3. Yang YS, Liu CY, Pei MQ, et al. Quercetin protects against sepsis-associated encephalopathy by inhibiting microglia-neuron crosstalk via the CXCL2/CXCR2 signaling pathway. Phytomedicine. 2024;134:155987. doi: 10.1016/j.phymed.2024.155987.

4. Xie L, Zhang G, Wu Y, et al. Protective effects of Wenqingyin on sepsis-induced acute lung injury through regulation of the receptor for advanced glycation end products pathway. Phytomedicine. 2024;129:155654. doi: 10.1016/j.phymed.2024.155654.

5. Bajgai B, Suri M, Singh H, et al. Naringin: A flavanone with a multifaceted target against sepsis-associated organ injuries. Phytomedicine. 2024;130:155707. doi: 10.1016/j.phymed.2024.155707.

6. Li Y, Li H, Hou S, et al. MPEP ameliorates sepsis-associated delirium by modulating hippocampal mGluR5 signaling and pyramidal neuron excitability in mice. Phytomedicine. 2025;147:157166. doi: 10.1016/j.phymed.2025.157166.

7. Lin SP, Zhu L, Shi H, et al. Puerarin prevents sepsis-associated encephalopathy by regulating the AKT1 pathway in microglia. Phytomedicine. 2023;121:155119. doi: 10.1016/j.phymed.2023.155119.

8. Ma J, Wang J, Wang J, et al. Araloside A alleviates sepsis-induced acute lung injury via PHD2/HIF-1α in macrophages. Phytomedicine. 2024;135:156089. doi: 10.1016/j.phymed.2024.156089.

9. Rampes S, Ma D. The potential impact of COVID-19 disease caused multi-organ injuries on patients’ surgical outcomes. Anesthesiol Perioper Sci. 2023;1(1):4. doi: 10.1007/s44254-023-00004-8.

10. Li H, Dai R. Perspective from anesthesiologists on the therapy of critically ill patients with COVID-19. Anesthesiol Perioper Sci. 2023;1(1):5. doi: 10.1007/s44254-023-00009-3.

11. Li Y, Zhao L, Yu Y, et al. Conservative oxygen therapy in critically ill and perioperative period of patients with sepsis-associated encephalopathy. Front Immunol. 2022;13:1035298. doi: 10.3389/fimmu.2022.1035298.

12. Zhao L, Zhang Z, Wang P, et al. NHH promotes aepsis-associated encephalopathy with the expression of AQP4 in astrocytes through the gut-brain Axis. J Neuroinflammation. 2024;21(1):138. doi: 10.1186/s12974-024-03135-2.

13. Wang P, Yan J, Shi Q, et al. Relationship between nonhepatic serum ammonia levels and sepsis-associated encephalopathy: A retrospective cohort study. Emerg Med Int. 2023;2023:6676033. doi: 10.1155/2023/6676033.

14. Zhao L, Fan Y, Wang Z, et al. The blood pressure targets in sepsis patients with acute kidney injury: An observational cohort study of multiple ICUs. Front Immunol. 2022;13:1060612. doi: 10.3389/fimmu.2022.1060612.

15. Mirouse A, Vigneron C, Llitjos JF, et al. Sepsis and cancer: An interplay of friends and foes. Am J Respir Crit Care Med. 2020;202(12):1625–35. doi: 10.1164/rccm.202004-1116TR.

16. Liu Z, Mahale P, Engels EA. Sepsis and risk of cancer among elderly adults in the United States. Clin Infect Dis. 2019;68(5):717–24. doi: 10.1093/cid/ciy530.

17. Nathan N, Sculier JP, Ameye L, et al. Sepsis and septic shock definitions in patients with cancer admitted in ICU. J Intensive Care Med. 2021;36(3):255–61. doi: 10.1177/0885066619894933.

18. Poirion OB, Jing Z, Chaudhary K, et al. DeepProg: An ensemble of deep-learning and machine-learning models for prognosis prediction using multi-omics data. Genome Med. 2021;13(1):112. doi: 10.1186/s13073-021-00930-x.

19. Johnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1. doi: 10.1038/s41597-022-01899-x.

20. Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3) JAMA. 2016;315(8):801–10. doi: 10.1001/jama.2016.0287.

21. Nazer L, Lopez-Olivo MA, Cuenca JA, et al. All-cause mortality in cancer patients treated for sepsis in intensive care units: A systematic review and meta-analysis. Support Care Cancer. 2022;30(12):10099–109. doi: 10.1007/s00520-022-07392-w.

22. Hensley MK, Donnelly JP, Carlton EF, et al. Epidemiology and outcomes of cancer-related versus non-cancer-related sepsis hospitalizations. Crit Care Med. 2019;47(10):1310–16. doi: 10.1097/CCM.0000000000003896.

23. El Haddad H, Chaftari AM, Hachem R, et al. Biomarkers of sepsis and bloodstream infections: The role of procalcitonin and proadrenomedullin with emphasis in patients with cancer. Clin Infect Dis. 2018;67(6):971–77. doi: 10.1093/cid/ciy331.

24. Cuenca JA, Nates JL, Laserna A, et al. Long-term survival of patients with cancer, sepsis, and vasopressor requirements based on lactate levels. Crit Care Explor. 2024;6(4):e1070. doi: 10.1097/CCE.0000000000001070.

25. López R, Pérez-Araos R, Baus F, et al. Outcomes of sepsis and septic shock in cancer patients: Focus on lactate. Front Med (Lausanne) 2021;8:603275. doi: 10.3389/fmed.2021.603275.

26. Zhao L, Wang Y, Ge Z, et al. Mechanical learning for prediction of sepsis-associated encephalopathy. Front Comput Neurosci. 2021;15:739265. doi: 10.3389/fncom.2021.739265.

27. Zhao L, Li Y, Wang Y, et al. Development and validation of a nomogram for the prediction of hospital mortality of patients with encephalopathy caused by microbial infection: A retrospective cohort study. Front Microbiol. 2021;12:737066. doi: 10.3389/fmicb.2021.737066.

Associated Data

Data Availability Statement

The dataset used in this study (MIMIC-IV v2.0) is publicly available under a data use agreement after completing required training ( https://physionet.org/content/mimiciv/2.0/ ). The analysis code for this study has been deposited in GitHub.