Construction and validation of a machine learning-based prediction model for 48-hour reintubation risk in...

In the ICU, reintubation after extubation in mechanically ventilated patients is often followed by adverse clinical events and is associated with longer ICU and hospital stays as well as increased mortality. Therefore, timely and accurate assessment of reintubation risk is clinically important for supporting extubation decisions and early management. Although several scoring systems and prediction models have been proposed, machine learning approaches may offer additional value by integrating multidimensional clinical information and potentially improving predictive performance.

This retrospective observational study included mechanically ventilated patients admitted to the intensive care unit (ICU) of the First Affiliated Hospital of Jinzhou Medical University between January 2022 and October 2025. Patients were randomly allocated at a 7:3 ratio to a training set (n = 496) and a test set (n = 211). In the training set, random forest-recursive feature elimination (RF-RFE) and the least absolute shrinkage and selection operator (LASSO) were used for feature selection, and prediction models were developed using seven machine learning algorithms. Model performance in the test set was evaluated by discrimination [area under the receiver operating characteristic curve (AUROC)], calibration (calibration curve), and clinical utility [decision curve analysis (DCA)] to identify the best-performing model. The final prediction tool was presented as both a static nomogram and a web-based dynamic nomogram.

A total of 707 mechanically ventilated patients were included, and the 48-h reintubation rate was 17.39% (123/707). Compared with the other six models, the LASSO-logistic regression (LASSO-LR) model achieved superior discrimination in the test set (AUROC = 0.879, 95% CI 0.814-0.935) and showed the best calibration (Brier score = 0.090, 95% CI 0.063-0.119). DCA indicated that this model provided a measurable net clinical benefit for predicting reintubation within 48 h after extubation among mechanically ventilated patients. Accordingly, LASSO-LR was selected as the optimal model and further implemented as a general static nomogram and a web-based dynamic nomogram (https://predict-for-reintubation-within-48-hours.shinyapps.io/dynnomapp/).

We developed and compared seven models for predicting reintubation risk after extubation in mechanically ventilated patients, among which the LASSO-LR model demonstrated the best overall performance. Visualizing the model as static and dynamic nomograms that integrate key predictors may facilitate early identification of patients at high risk of reintubation and support targeted preventive and management strategies in clinical practice.