Abstract
This study aimed to develop models for predicting the 30-day mortality of sepsis-associated delirium (SAD) by multiple machine learning (ML) algorithms. On the whole, a cohort of 3,197 SAD patients were collected from the Medical Information Mart for Intensive Care (MIMIC)-IV database. Among them, a total of 659 (20.61%) patients died following SAD. The patients who died were about 73.00 (62.00, 82.00) years old and mostly male (56.75%). Recursive feature elimination (RFE) was used to distinguish risk factors. Subsequently, six ML algorithms including artificial neural network (NNET), gradient boosting machine (GBM), adaptive boosting (Ada), random forest (RF), eXtreme Gradient Boosting (XGB) and logistic regression (LR) were employed to establish models to predict the 30-day mortality of SAD. The performance of models was assessed via both discrimination and calibration by cross-validation with 100 resamples. Overall, 10 independent predictors, including Glasgow Coma Scale (GCS), Sequential Organ Failure Assessment (SOFA), anion gap (AG), continuous renal replacement therapy (CRRT), temperature, mean corpuscular hemoglobin concentration (MCHC), vasopressor, blood urea nitrogen (BUN), base excess (BE), and bicarbonate were identified as independent predictors for the 30-day mortality of SAD. The validation cohort demonstrated that all these six models had relatively favorable differentiation, while among them, the GBM model had the highest area under the curve (AUC) of 0.845 (95% Confidence Interval (CI): 0.816, 0.874). Furthermore, the calibration curve of these six models was close to the diagonal line in the validation sets. As for decision curve analysis, the predictive models were clinically useful as well. Based on real-world research, we developed ML models to provide personalized predictions of delirium-related mortality in sepsis patients, potentially enabling clinicians to identify high-risk SAD patients more promptly.