Early prediction of 28-day mortality in ICU patients with pneumonia-induced sepsis using plasma microbial cell-free DNA and interpretable machine learning: a retrospective cohort study

Pneumonia-induced sepsis is a leading cause of ICU mortality, yet early risk stratification at ICU admission remains challenging. Plasma microbial cell-free DNA (mcfDNA) is an infection-related circulating biomarker that may offer prognostic information beyond routine parameters. We developed and internally validated an interpretable machine-learning model incorporating plasma mcfDNA to predict 28-day mortality in ICU patients with pneumonia-induced sepsis.

We conducted a retrospective, single-center cohort study including adult ICU patients with pneumonia-induced sepsis admitted between July 2023 and August 2025. Baseline demographic, clinical, and laboratory variables-including plasma mcfDNA copy number-were collected within 24 h of ICU admission. Features were selected using least absolute shrinkage and selection operator (LASSO) regression. Six machine-learning algorithms were trained and evaluated. Model performance was assessed using multiple metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, positive predictive value, negative predictive value, sensitivity, specificity, and F1 score. Repeated fivefold cross-validation was additionally performed for internal assessment of the final model, and formal calibration assessment included the Brier score, calibration intercept, calibration slope, and calibration plot. Model interpretability was examined using Shapley Additive Explanations (SHAP).

A total of 187 patients were included; 62 (33.2%) died within 28 days. LASSO identified six predictors: arterial blood pH, white blood cell count, prothrombin time ratio, international normalized ratio, lactate dehydrogenase, and plasma mcfDNA copy number. Among the evaluated models, the random forest model showed the most balanced overall performance. In repeated fivefold cross-validation, the final random forest model demonstrated favorable internal discriminative performance. Direct comparison between a reduced model based on routine clinical variables alone and an extended model additionally including plasma mcfDNA showed improved discrimination, calibration, and threshold-dependent classification performance for the extended model. SHAP analysis identified plasma mcfDNA as a major contributor to model prediction.

An interpretable machine-learning model integrating plasma mcfDNA with routinely available variables showed potential for early risk stratification of 28-day mortality in ICU patients with pneumonia-induced sepsis. These findings suggest that plasma mcfDNA may provide added prognostic information, but they should be considered exploratory and require confirmation in larger prospective multicenter studies with external validation.