Answers are grounded in this paper and cite their sources.
This retrospective study develops a day-1 prediction model for sepsis-associated encephalopathy in ICU patients with AKI, suggesting clinical utility for real-time risk stratification.
Sepsis-associated encephalopathy (SAE) is common in the intensive care unit (ICU) and portends worse short- and long-term outcomes. To enable real-time bedside use and multicenter deployment, we aimed to develop a parsimonious, transparent day-1 prediction model using routinely available variables while preserving discrimination, calibration, and clinical utility. Using MIMIC-IV (2008–2022), we conducted a single-center retrospective study of adult sepsis patients with KDIGO-defined AKI. Predictors were restricted to the first 24 hours after ICU admission; the endpoint was any in-ICU SAE (“ever” vs “never”). After multiple imputation (m = 5), 44 baseline variables were standardized and entered into LASSO with 20-fold cross-validation. A 3-rule clinical screen (24 hours availability; non-treatment; low collinearity) distilled LASSO-selected features to a four-predictor logistic model; performance was internally validated (bootstrap) and compared with an XGBoost benchmark. SHAP analyses supported interpretability. Among 6780 ICU stays (training n = 4746; validation n = 2034), SAE occurred in 69.8%. The final 4 predictors were age, SAPS II, serum sodium, and mean arterial pressure (MAP). Discrimination was stable (AUC 0.734 training; 0.739 validation) with excellent calibration (validation CITL = −0.045; slope = 0.996; Brier = 0.182). Decision-curve analysis showed greater net benefit than XGBoost across thresholds 0.15 to 0.55; although AUCs were similar, XGBoost calibrated worse (CITL = −0.289; slope = 0.729). SHAP ranked contributions as SAPS II, sodium, age, and MAP, indicating a near-linear sodium-risk rise within 138 to 144 mmol/L, age-related risk above ~70 years, and a U-shaped MAP effect with protection around 55 to 75 mm Hg. We developed and validated a four-factor nomogram that uses only routine day-1 data to stratify SAE risk rapidly and transparently, outperforming a complex learner in calibration and net benefit. This parsimonious, interpretable tool highlights modifiable targets (sodium, individualized MAP) and provides a pragmatic foundation for multicenter validation and EMR-embedded early warning and intervention strategies.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: