Machine Learning-Based Classification of ICU-Acquired Neuromuscular Weakness: A Comparative Study in Survivors of Critical Illness

Abstract

Classifying the severity of intensive-care-unit-acquired muscle atrophy (ICU-AW) is essential for early prognosis and individualized neurorehabilitation, improving functional outcomes in survivors of critical illness. This study evaluated and compared advanced machine learning (ML) algorithms for classifying neuromuscular atrophy in neurocritical patients. Clinical, biochemical, anthropometric, and morphometric data from 198 neuro-ICU patients were retrospectively analyzed. Six supervised ML models—Support Vector Machine (SVM), Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), TPOT AutoML, AdaBoost, and Multinomial Logistic Regression—were trained using stratified cross-validation, synthetic oversampling, and hyperparameter optimization. Among the most outstanding models, SVM achieved the best performance (accuracy = 93%, ROC-AUC = 0.95), followed by MLP (accuracy = 82.8%, ROC-AUC = 0.93) and XGBoost (accuracy = 80%, ROC-AUC = 0.94). Stability analyses across random seeds confirmed the robustness of SVM and TPOT, with the highest median AUPRC (>0.90). Explainable AI methods (LIME and SHAP) identified BMI, serum albumin, and body surface area as the most influential variables, showing physiologically consistent patterns associated with a classification of muscle loss.