Development and Validation of a Machine Learning‑based Prediction Model for ICU‑Acquired Weakness in Critically Ill Patients: A Prospective Multicenter Cohort Study

ICU-acquired weakness (ICU-AW) is a common and devastating complication of critical illness associated with prolonged mechanical ventilation and increased mortality, whereas externally validated tools for early risk stratification remain insufficient. This study aimed to construct and validate machine learning models to predict ICU-AW risk in critically ill patients.

A total of 535 patients from three medical centers were divided into a training set (n = 276), an internal validation set (n = 119) and an external validation cohort (n = 140). Candidate predictors were restricted to variables available within the first 48 hours of ICU admission. Five algorithms including logistic regression (LR), random forest (RF), support vector machine (SVM), classification and regression tree (CART) and XGBoost were trained with 10-fold cross-validated grid search and compared using DeLong tests. Six predictors were retained: age, Glasgow Coma Scale score, SOFA score, sepsis, renal failure and hypoalbuminemia.

All five models displayed favorable predictive capacity in the internal validation set (AUROC 0.906-0.939). In external validation, SVM (0.853, 95% CI 0.779-0.915) and LR (0.841, 0.765-0.906) outperformed RF (0.802), XGBoost (0.790) and CART (0.685), and the difference between SVM and LR was not statistically significant (DeLong P = 0.085). Tree-based algorithms presented near-perfect apparent performance that attenuated markedly in the external cohort, indicating overfitting. The LR model, which offered statistically comparable discrimination with full transparency, was converted into a nomogram (external AUROC 0.841) with positive net clinical benefit across threshold probabilities of 1-81%.

A six-predictor logistic model based on variables available within 48 hours of ICU admission predicted ICU-AW comparably to tuned machine learning classifiers and showed stable cross-center performance. The nomogram may assist early risk stratification, and prospective studies are warranted to confirm whether model-guided care improves patient outcomes.