Highlights
• A multicenter cohort (n=1,449) produced a validated, multivariable nomogram that predicts sepsis in patients presenting with acute gastrointestinal bleeding (AGIB).
• Seven routinely available variables—chronic kidney disease (CKD), respiratory rate (RR), neutrophil-to-lymphocyte ratio (NLR), C‑reactive protein (CRP), creatinine (Cr), activated partial thromboplastin time (APTT), and fibrinogen (FIB)—drove the model.
• Discrimination was strong across datasets (AUC 0.827 training; 0.836 internal validation; 0.884 external validation), with close calibration (MSEs 0.00094, 0.00791, 0.00045) and demonstration of clinical utility by decision curve analysis (DCA).
• The model was delivered via an online, real‑time warning system intended to enable dynamic monitoring and early, preemptive management.
Background — Clinical context and unmet need
Acute gastrointestinal bleeding (AGIB) is a common emergency with substantial morbidity and mortality that often requires rapid triage and critical care. In the context of AGIB, secondary infection and progression to sepsis magnify risk and are associated with a large increase in in‑hospital mortality. Early identification of patients at high risk of developing sepsis is therefore essential to direct monitoring intensity, expedite diagnostics, and guide empiric interventions while avoiding unnecessary resource use in low‑risk patients.
Existing general critical‑illness risk scores (for example, APACHE II and SOFA) and AGIB-specific scores (for example, Glasgow‑Blatchford Score, GBS) were not designed specifically to predict subsequent sepsis in the AGIB population and may lack sensitivity or specificity for this outcome. A practical, validated, and dynamic risk tool that uses routinely available clinical and laboratory data to provide real‑time sepsis risk estimates in patients with AGIB addresses an important clinical gap.
Study design and methods
The study by Jiang et al. (EClinicalMedicine, 2025) is a multicenter cohort conducted in China between January 2020 and July 2024 that aimed to develop and validate a real‑time predictive model for sepsis in adult patients admitted with AGIB. The overall cohort included 1,449 patients (median age 65 years; 68.7% male). Patients with hospital stay <24 hours, or with pre‑existing infection or sepsis at presentation, were excluded.
Design and cohorts:
- Training cohort: retrospective, n = 878 (lead center)
- Prospective internal validation: n = 187 (prospectively enrolled at lead center)
- External validation: n = 384 (three independent tertiary hospitals)
Outcome definition: Sepsis was defined according to Sepsis‑3 criteria, i.e., life‑threatening organ dysfunction caused by a dysregulated host response to infection.
Model development: Multivariable logistic regression was used to develop a nomogram. Candidate predictors were chosen from routinely collected clinical signs and laboratory parameters. Model performance was assessed using area under the receiver operating characteristic curve (AUC) for discrimination, calibration curves and mean squared error (MSE) for calibration, decision curve analysis (DCA) for clinical utility, and SHapley Additive exPlanations (SHAP) for model interpretability. The model’s discrimination was compared with GBS, APACHE II, and SOFA scores. An online platform implementing the nomogram provided real‑time risk monitoring and alerting.
Key findings and results
Incidence and outcomes: Among the 1,449 patients with AGIB, 223 (15.4%) developed sepsis. Sepsis was associated with markedly increased in‑hospital mortality: 23.7% in the sepsis group versus 6.8% in the non‑sepsis group (p < 0.001).
Final model predictors: Seven variables emerged as the key predictors included in the nomogram:
- Chronic kidney disease (CKD)
- Respiratory rate (RR)
- Neutrophil‑to‑lymphocyte ratio (NLR)
- C‑reactive protein (CRP)
- Creatinine (Cr)
- Activated partial thromboplastin time (APTT)
- Fibrinogen (FIB)
Model discrimination: The nomogram demonstrated strong discrimination across datasets:
- Training set AUC 0.827 (95% CI 0.759–0.888)
- Internal validation AUC 0.836 (95% CI 0.776–0.896)
- External validation AUC 0.884 (95% CI 0.816–0.952)
Comparative performance: The authors report that the new model outperformed some established scores (GBS, APACHE II) in discriminating patients who subsequently developed sepsis. While SOFA is designed to quantify acute organ dysfunction and remains an important comparator, the nomogram’s specific combination of AGIB‑relevant and inflammatory/coagulation variables improved sepsis prediction in this population.
Calibration and clinical utility: Calibration curves showed excellent agreement between predicted probabilities and observed sepsis rates across risk strata, with mean squared errors of 0.00094 (training), 0.00791 (internal validation), and 0.00045 (external validation). Decision curve analysis indicated net benefit across a clinically relevant range of threshold probabilities, supporting potential clinical utility for risk‑based decision making.
Interpretability and model insights: SHAP analysis ranked the relative importance of predictors and illustrated how changes in specific variables shifted individual patient risk. For example, higher CRP and NLR, prolonged APTT, higher creatinine, and the presence of CKD increased predicted sepsis risk; higher fibrinogen had a complex association reflecting coagulation‑inflammation interplay.
Implementation: An online platform integrated the nomogram and allowed dynamic, real‑time risk updates as new vital signs and lab values were entered. The system generated alerts to flag rising sepsis risk, enabling clinicians to escalate monitoring, obtain cultures, or consider early empiric therapy per local protocols.
Interpretation and clinical implications
This study provides a pragmatic and validated tool tailored to the AGIB population to estimate the risk of subsequent sepsis using variables that are commonly available at presentation or during early hospitalization. Key implications include:
- Risk stratification: The nomogram can help triage patients for higher‑intensity monitoring (for example, earlier ICU transfer or step‑up care) and prioritize diagnostic evaluation for infection (blood cultures, imaging).
- Targeted interventions: High‑risk patients could be evaluated for early, guideline‑concordant interventions, such as rapid source control and timely antimicrobial therapy when warranted by clinical assessment and microbiology.
- Resource allocation: In settings with constrained ICU capacity, a validated prediction tool can assist in allocating monitoring and treatment resources to patients most likely to benefit.
- Dynamic monitoring: The online real‑time system aligns with modern clinical workflows and supports repeated risk re‑assessment rather than a single static prediction.
Strengths
Major strengths include a relatively large overall sample size, multicenter design with prospective internal validation and independent external validation, use of routinely available predictors, formal calibration and clinical utility assessment (DCA), and model interpretability via SHAP. The delivered online platform demonstrates feasibility for real‑world integration and dynamic use.
Limitations and cautions
Despite encouraging performance, several limitations should guide clinical adoption and future work:
- Geographic and population generalizability: The study was conducted in tertiary hospitals in China. External validation in other healthcare systems, ethnic groups, and community hospital settings is warranted.
- Selection criteria: Patients with hospitalization <24 hours and those with pre‑existing infection or sepsis were excluded, which limits applicability to very short admissions and to patients already infected on arrival.
- Potential confounding and residual bias: As with any observational model development, unmeasured confounding may influence predictor associations. Prospective impact studies are needed to quantify clinical benefit.
- Alert burden and decision thresholds: Implementation of real‑time alerts risks generating false positives and clinician alert fatigue. Local calibration of decision thresholds and integration with stewardship/triage pathways are essential.
- Outcomes beyond in‑hospital sepsis: The model predicts sepsis occurrence; effects on downstream outcomes (e.g., time to antibiotics, ICU length of stay, long‑term mortality) require prospective evaluation in implementation trials.
Expert commentary and guideline context
Sepsis‑3 emphasizes early recognition of organ dysfunction from infection and prompt intervention. Tools that specifically identify patients at high risk of developing sepsis after a sentinel event such as AGIB complement clinician judgement and standard sepsis screening. However, prediction tools should not replace bedside assessment or clinical judgment; rather, they should act as adjuncts to focus attention and expedite guideline‑based care.
Prior to broad deployment, institutions should undertake local validation, assess how alerts will change workflow, and measure patient‑centered outcomes. A measured implementation strategy that includes clinician education, appropriate threshold selection, and monitoring of alert performance and clinical outcomes is recommended.
Conclusions and next steps
The real‑time nomogram developed by Jiang et al. represents a promising, validated approach to identify AGIB patients at elevated risk for sepsis, using readily available predictors and an interpretable model framework. The tool demonstrated strong discrimination, tight calibration, and potential clinical utility in multicenter validation, and it was operationalized through an online monitoring platform.
Future priorities include external validation in diverse healthcare settings, prospective implementation studies that assess impact on clinical decision making and patient outcomes, and refinement of thresholds to balance sensitivity and specificity while minimizing alert fatigue. Integration with electronic health record systems and stewardship programs will be central to realizing the potential benefits of dynamic sepsis risk monitoring in AGIB.
Funding
The study was funded by the National Key R&D Program of China, the National Natural Science Foundation Project of China, the Knowledge Innovation Program of Wuhan, and the Cross Innovation Talent Project of Renmin Hospital of Wuhan University.
ClinicalTrials.gov
No clinicaltrials.gov identifier was reported for this observational model development and validation study.
References
1. Jiang G, Sun S, Wang Q, Liu Z, Huang C, Quan F, Zuo X, Peng T, Xu J, Duan H, Barajas‑Martínez H, Zhang D, Hu D, Zhan L. Real‑time risk prediction model for sepsis in patients with acute gastrointestinal bleeding: development and multicenter validation of a dynamic monitoring tool. EClinicalMedicine. 2025 Oct 24;90:103574. doi: 10.1016/j.eclinm.2025.103574. PMID: 41209656; PMCID: PMC12595275.
2. Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis‑3). JAMA. 2016;315(8):801–810. doi:10.1001/jama.2016.0287.
3. Evans L, Rhodes A, Alhazzani W, et al. Surviving Sepsis Campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med. 2021;47(11):1181–1247. doi:10.1007/s00134-021-06506-y.
