A 24-hour landmark machine learning model for predicting new-onset sepsis-associated encephalopathy

Sepsis-associated encephalopathy (SAE) is a common neurological complication of sepsis and is associated with prolonged intensive care unit (ICU) stay, extended mechanical ventilation, and poor outcomes. This study aimed to develop and internally validate a 24-hour landmark machine learning-based model for predicting new-onset SAE after the first ICU day in patients with sepsis.

This retrospective observational study included 537 ICU patients meeting the Sepsis-3 criteria. This retrospective study developed a 24-hour landmark prediction model. Clinical variables collected during the first 24 h after ICU admission were used to predict new-onset SAE occurring after the 24-hour landmark time point among patients who survived, remained in the ICU, and were SAE-free during the first ICU day. Demographic characteristics, comorbidities, vital signs, laboratory parameters, organ function scores, therapeutic interventions, and prognosis-related variables were collected. Missing data were handled by multiple imputation. The cohort was randomly divided into a training set and an internal test set at a 7:3 ratio. Model performance was evaluated using internal validation only. Six models were developed, including logistic regression, random forest, XGBoost, support vector machine, k-nearest neighbors, and naive Bayes. Model performance was assessed using discrimination, calibration, and decision curve analysis.

SAE occurred in 215 patients, corresponding to an incidence of 40.0%. Compared with non-SAE patients, patients with SAE had greater disease severity, higher SOFA and APACHE II scores, increased respiratory rate, worse renal and inflammatory profiles, and higher rates of vasoactive drug use, mechanical ventilation, and continuous renal replacement therapy. In the internal test set, logistic regression showed the highest AUC among the candidate models, with an AUC of 0.767, accuracy of 0.698, and sensitivity of 0.754. Calibration analysis showed acceptable agreement between predicted and observed risks, and decision curve analysis indicated potential net benefit. A nomogram was constructed to visualize the final model.

The logistic regression-based model showed moderate predictive performance and may support preliminary 24-hour landmark risk stratification for new-onset SAE in ICU patients with sepsis. Further multicenter prospective validation is required.