Construction and validation of a machine learning-based risk prediction model for invasive mechanical ventilation in...

This study aimed to create and validate a machine learning (ML) model to predict the likelihood of invasive mechanical ventilation (IMV) in patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD) complicated by respiratory failure.

Data from patients diagnosed with AECOPD and respiratory failure were retrospectively extracted from the Medical Information Mart for Intensive Care-IV (MIMIC-IV). A total of 551 cases were split 7:3 into a training set (385 cases) for model construction and an internal validation set (166 cases). The IMV served as the outcome event. Features were selected with the Boruta algorithm and least absolute shrinkage and selection operator (LASSO). Eight ML algorithms-XGBoost, decision tree (DT), random forest (RF), support-vector machine (SVM), LightGBM, CatBoost, Gaussian naïve Bayes (NB) and K-nearest neighbor (NN)-were trained with 10-fold cross-validation. Model performance was assessed by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, calibration curve, decision curve and clinical impact curve. An external validation cohort of 100 AECOPD-respiratory failure patients admitted to Baoying People's Hospital between January 2020 and August 2025 was collected. The final best model was interpreted with SHapley Additive exPlanations (SHAP) to clarify feature importance and decision logic, and an interactive dynamic nomogram was plotted to increase readability.

Boruta plus LASSO identified total calcium, partial pressure of oxygen (PO₂), oxygen saturation (SpO₂) and sepsis as significant predictors. XGBoost outperformed the other algorithms, achieving an internal validation accuracy of 72.2%, sensitivity of 64.6%, specificity of 79.8 %, F1 score of 69.7% and AUC of 0.813 (95% CI 0.748-0.878). The external validation accuracy reached 76.4%, the sensitivity reached 82.6%, the specificity reached 70.0%, the F1 score reached 78.7%, and the AUC reached 0.840 (95% CI 0.801-0.879). SHAP analysis further indicated that PO₂ and SpO₂ were the primary drivers of model decisions. An interactive dynamic nomogram was successfully constructed.

IMV in AECOPD patients with respiratory failure was associated with total calcium, PO₂, and SpO₂ levels and sepsis. The developed XGBoost model demonstrated good predictive value for IMV in this clinical population.