Machine learning model predicts 28-day mortality in sepsis patients with diabetic retinopathy

Patients with diabetic retinopathy who develop sepsis face a notoriously unpredictable course in the intensive care unit, and clinicians have lacked a prognostic tool designed specifically for this comorbid population. A new retrospective study published in BMC Endocrine Disorders addresses that gap by building and validating a machine learning model that estimates a patient’s risk of death within 28 days using only clinical data gathered during the first 24 hours of ICU admission.

The research team, led by Heling Wang, Taiyu Chen and colleagues, drew on two independent electronic health record databases: MIMIC-IV, which supplied 525 patients with diabetic retinopathy who met the Sepsis-3 criteria between 2008 and 2022, and eICU-CRD, which contributed 86 patients for external testing. The 28-day mortality rate was 12.95 percent in the MIMIC-IV cohort and substantially higher, at 33.72 percent, in the eICU cohort, underscoring how variable outcomes are across centers and why a calibrated risk estimate could matter at the bedside.

To identify the variables that carry prognostic signal, the authors combined two feature-selection strategies: least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. The convergence of these approaches distilled the final model down to just four predictors, a parsimony that simplifies deployment in busy clinical settings.

Seven machine learning algorithms were then trained on data collected within the first 24 hours of ICU admission, with the prediction landmark fixed at 24 hours after admission. The MIMIC-IV data were divided into training and internal validation sets by 7:3 stratified random sampling, while the eICU data served as a fully external validation set. Algorithm choice was settled through nested cross-validation within the training set.

Performance was assessed with three complementary metrics. The area under the curve, which measures discrimination between patients who die and those who survive, reached 0.870 in the training set, 0.805 in the MIMIC-IV internal validation set and 0.833 in the eICU external validation set. Values in this range indicate useful, though not perfect, discrimination, and the fact that the external figure held up despite the eICU cohort’s much higher mortality rate suggests the model captures risk signals that generalize beyond a single database.

Discrimination alone does not guarantee clinical usefulness, so the team also examined calibration curves, which test whether predicted probabilities match observed death rates, and decision curve analysis, which quantifies net benefit across a range of risk thresholds. Both validation cohorts showed favorable calibration and positive clinical net benefit across threshold probabilities from 0.0 to 1.0, meaning the model added value compared with treating all patients or none in the evaluated ranges.

To open the black box, the researchers applied SHapley Additive exPlanations, or SHAP, to interpret the optimal model. This interpretability layer reveals which of the four predictors pushed a given patient’s estimated risk up or down rather than issuing an unexplained number.

The authors frame the tool as a quantitative reference for coordinated intervention between ophthalmology and intensive care, reflecting the hypothesis that the systemic microvascular dysfunction underlying diabetic retinopathy may accelerate infection progression and organ failure in sepsis. A web-based application built on the model is available, allowing the risk calculation to be explored outside the research environment.

Several limitations deserve emphasis. The study is retrospective and observational, so the model estimates risk in the populations represented by the two databases rather than demonstrating that using it improves outcomes; the eICU cohort is small at 86 patients; and the four-predictor structure means the model depends on the specific variables available in these records. Even so, the dual-database validation, the modest predictor count and the transparent interpretation together mark a concrete step toward risk stratification for this comorbid patient group.

Subject of Research: Machine learning prediction of 28-day mortality in sepsis patients with diabetic retinopathy using MIMIC-IV and eICU-CRD data

Article Title: Machine learning-driven predictive model for 28-day mortality risk in patients with diabetic retinopathy complicated with sepsis: dual-database validation in MIMIC-IV and eICU-CRD

Article References: Wang, H., Chen, T., Yang, Y., Yang, Q., & Gu, Y. (2026). Machine learning-driven predictive model for 28-day mortality risk in patients with diabetic retinopathy complicated with sepsis: dual-database validation in MIMIC-IV and eICU-CRD. BMC Endocrine Disorders. https://doi.org/10.1186/s12902-026-02651-x

Image Credits: AI Generated

DOI: 10.1186/s12902-026-02651-x

Keywords: diabetic retinopathy, sepsis, machine learning, mortality prediction, MIMIC-IV, eICU-CRD, LASSO regression, Boruta algorithm, SHAP, ICU, risk prediction, BMC Endocrine Disorders