Eight Routine ICU Variables Predict Mortality in Sepsis-Heart Failure Patients

Intensive care physicians now have a new statistical tool for one of the most dangerous patient combinations they face: people admitted to the intensive care unit who are simultaneously coded for sepsis and heart failure. In a retrospective cohort study published in Cardiovascular Diabetology – Endocrinology Reports, researchers built and internally validated a nomogram that predicts ICU mortality using eight variables already collected in routine care. The work draws on the eICU Collaborative Research Database, a multicentre dataset covering more than 200,000 hospital admissions across 335 ICUs in 208 US hospitals between 2014 and 2015.

The clinical stakes are high. Sepsis and heart failure are each among the most challenging conditions in modern critical care, and when they co-occur, mortality rates reportedly far exceed those of either condition alone. The authors describe this combined phenotype as one in which dysregulated inflammatory responses and impaired cardiac pump function can escalate rapidly into refractory shock and multiple organ dysfunction. Yet existing prognostic assessment in ICUs relies mainly on generalised scoring systems such as APACHE and SOFA, which were validated across broad critically ill populations but lack specificity for particular condition combinations. The study set out to fill that gap with a tool built from readily accessible parameters.

To assemble the study population, the team searched the eICU-CRD for patients carrying both sepsis and heart failure diagnostic codes, using a list of ICD-9 and ICD-10 codes, and additionally classified patients with suspected infection and a Sequential Organ Failure Assessment score of two or higher as having sepsis under the Sepsis 3.0 criteria. Ultimately, 1,394 patients met the combined diagnostic coding criteria. Of these, 183 patients, or 13.3 percent, died during their ICU stay. The cohort was randomly split in a 7:3 ratio into a training set of 975 patients and a validation set of 419 patients, with no significant differences in baseline characteristics between the two groups.

Variable selection began with Least Absolute Shrinkage and Selection Operator, or LASSO, regression to reduce the risk of overfitting. After cross-validation determined the optimal penalty parameter, 13 clinical indicators advanced to multivariable logistic regression, which identified eight independent predictors: mechanical ventilation, lactate, respiratory rate, white blood cell count, age, platelet count, systolic blood pressure and oxygen saturation. Mechanical ventilation carried the strongest association with death, with an odds ratio of 4.34. Lactate, respiratory rate, white blood cell count and age each raised risk, while platelet count, systolic blood pressure and arterial oxygen saturation emerged as protective factors in this cohort, likely marking less severe physiological disturbance rather than true biological protection.

The final nomogram converts each of the eight variables into points, sums them, and maps the total to a probability of ICU mortality. Discriminatory performance, measured by the area under the receiver operating characteristic curve, was 0.826 in the training cohort and 0.798 in the validation cohort, placing the model in the good range on the training side and at the threshold between acceptable and good on the validation side. The authors note that the confidence intervals, particularly the wide 0.732 to 0.864 interval in the validation set, suggest the true performance may vary.

Calibration, meaning the agreement between predicted probabilities and observed outcomes, was assessed with calibration curves adjusted by bootstrap correction, the Brier score and the Hosmer-Lemeshow test. Calibration curves approached the ideal diagonal in both datasets, Brier scores were 0.090 for training and 0.091 for validation, and both Hosmer-Lemeshow tests returned p-values above 0.05. Decision curve analysis and clinical impact curves further indicated net benefit across a range of threshold probabilities in both datasets.

The model also outperformed traditional severity scores on reclassification metrics in this population. Compared with APACHE, the nomogram showed a net reclassification improvement of 0.868 in the training set and 0.833 in the validation set, with integrated discrimination improvement values of 0.172 and 0.139 respectively. Against SOFA, the corresponding NRI values were 0.914 and 0.896, with IDI values of 0.173 and 0.191. These figures suggest the eight routine variables capture prognostic information that universal critical illness scores do not fully address in this specific subgroup.

Sensitivity analyses tested the model’s robustness. When in-hospital mortality replaced ICU mortality as the endpoint, with 189 deaths across the full cohort, the AUC was 0.784 in the training set and 0.769 in the validation set, closely tracking the primary results. Excluding mechanical ventilation from the model lowered the training AUC to 0.761 and the validation AUC to 0.730, indicating that ventilation contributes meaningful incremental value but that the model retains reasonable performance without it. Missing data in the predictor variables ranged from zero to about 4.5 percent and were handled with multiple imputation by chained equations, with complete-case sensitivity analyses yielding consistent results.

The authors are unusually explicit about the limits of their tool. The case definition rested entirely on ICD codes entered by attending physicians, and the eICU-CRD does not consistently provide echocardiography, natriuretic peptide or invasive haemodynamic data, so the heart failure diagnosis could not be verified against objective measures of cardiac structure or function. The coding-based definition also cannot distinguish between sepsis-induced cardiomyopathy and sepsis arising in pre-existing chronic heart failure, nor between heart failure subtypes such as reduced or preserved ejection fraction. Moreover, most predictors, including lactate, respiratory rate and white blood cell count, are non-specific markers of acute illness severity rather than heart failure-specific indicators; the model omits echocardiographic parameters, natriuretic peptides, troponin, vasopressor use, volume status and formal heart failure classification.

For these reasons, the study frames the nomogram as an internally validated ICU mortality prediction model for a high-risk, code-defined population rather than a cardiology-specific or clinically actionable instrument. Internal validation through randomised splitting supports internal validity, but no external validation from independent cohorts in other health systems has yet been performed, and the authors state that external validation is an absolute prerequisite for any clinical application. Until then, they position the tool as a hypothesis-generating risk stratification aid that could complement, but never replace, clinical judgement, direct diagnostic evaluation and evidence-based treatment decisions at the bedside.

Subject of Research: Development and internal validation of a nomogram predicting ICU mortality in patients coded for both sepsis and heart failure using the eICU-CRD database.

Article Title: Development and internal validation of an ICU mortality prediction model for patients with concurrent sepsis and heart failure: a retrospective cohort study based on eICU-CRD

Article References: Ye, M., Huang, H., Lv, S., Lv, D., & Wang, X. (2026). Development and internal validation of an ICU mortality prediction model for patients with concurrent sepsis and heart failure: a retrospective cohort study based on eICU-CRD. Cardiovascular Diabetology – Endocrinology Reports, 12(1), Article 53. https://doi.org/10.1186/s40842-026-00323-2

Image Credits: AI Generated

DOI: 10.1186/s40842-026-00323-2

Keywords: sepsis, heart failure, ICU mortality, nomogram, eICU-CRD, prediction model, retrospective cohort, LASSO regression, internal validation, risk stratification, APACHE, SOFA