To develop a prediction model for the prolonged ICU stay in patients with drug poisoning based on machine learning algorithms.

Drug poisoning patients included in the Medical Information Mart for Intensive Care Ⅳ (MIMIC-Ⅳ) and Electronic Intensive Care Unit-Collaborative Research Database (eICU-CRD) were divided into a non-prolonged group (≤ 48 hours) and a prolonged group (> 48 hours) according to the length of ICU stay. A total of 1 342 patients in MIMIC-Ⅳ were divided into a training dataset (939 cases) and a test dataset (403 cases) at a ratio of 7 ∶ 3, with eICU-CRD serving as an external test dataset (2 144 cases). In the training dataset, variables were jointly screened through single-factor analysis and least absolute shrinkage and selection operator (LASSO) regression. Six machine learning algorithms (logistic regression, extreme gradient boosting, light gradient boosting machine, random forest, decision tree, and support vector machine) were used for modeling. Meanwhile, the receiver operating characteristic (ROC) curve, Hosmer-Lemeshow (H-L) test, Brier score, and decision curve analysis (DCA) were used to evaluate the model performance in both internal and external test datasets.

Seven critical predictors were screened out in the training dataset: cerebrovascular disease, liver disease, aspiration pneumonia, sepsis, respiratory rate, sequential organ failure assessment (SOFA) score, and mechanical ventilation. The logistic regression model performed best in the training dataset [area under the curve (AUC) = 0.767, 95% confidence interval (CI) (0.667, 0.868), P < 0.001]. Its AUC was 0.762 [95%CI (0.712, 0.811), P < 0.001] in the internal test dataset, and 0.732 [95%CI (0.708, 0.756), P < 0.001] in the external test dataset. Moreover, the logistic regression model showed good calibration and net returns in both internal and external test datasets.

The logistic regression model constructed in this study consists of seven predictive factors, including cerebrovascular disease, liver disease, aspiration pneumonia, sepsis, respiratory rate, SOFA score, and mechanical ventilation. It can effectively predict the risk of prolonged ICU stay in drug poisoning patients, which is helpful for early clinical identification and intervention.