Development and exploratory external assessment of a machine learning model for early delirium prediction in non-mechanically ventilated patients with sepsis.

Delirium is a common and severe complication of sepsis in critically ill patients and is associated with increased mortality and prolonged hospitalization. Early identification of high-risk individuals remains challenging, particularly among non-mechanically ventilated patients who are less frequently represented in prediction studies. Using routinely collected electronic health record data, we performed a retrospective multicenter study to establish and validate an early prediction model for delirium in non-mechanically ventilated adults with sepsis. Patients from the MIMIC-IV database were used for model development, and an independent cohort from the eICU Collaborative Research Database was used for exploratory external assessment. A total of 3,015 patients in the development cohort and 1,119 patients in the exploratory test cohort were included. Among the evaluated algorithms, XGBoost achieved the best discrimination, with an area under the receiver operating characteristic curve of 0.807 in internal validation. The eICU cohort contained only 10 delirium cases (incidence < 1%), forming a sparse-event dataset; although the XGBoost model yielded an AUC of 0.833 in this cohort, the calibration and decision curve analysis results should be interpreted with extreme caution due to limited positive events. Lower Glasgow Coma Scale scores, higher illness-severity indices, exposure to sedative or analgesic agents, and metabolic disturbances such as elevated anion gap and abnormalities in hematocrit and calcium were the strongest contributors to prediction. These predictors align with recognized features of sepsis-associated brain dysfunction. This externally validated model may support early risk stratification and targeted preventive strategies in this vulnerable population.