Development and Validation of a Predictive Model for ICU-acquired Weakness in Sepsis Patients: An Interpretable Machine-learning Approach.

Abstract

Objective: Intensive-care-unit-acquired weakness (ICU-AW), including critical illness polyneuropathy (CIP), critical illness myopathy (CIM), and critical illness neuromyopathy, is a common neuromuscular complication of sepsis. An interpretable machine-learning model for the early prediction of ICU-AW in patients with sepsis was developed and validated using the Medical Information Market for Intensive Care (MIMIC)-IV v3.1 database and local hospital data.

Methods: A total of 3,842 adult patients who met the Sepsis-3 criteria were enrolled to create the MIMIC-IV database. ICU-AW was defined as per International Classification of Diseases codes in the MIMIC cohort and with a Medical Research Council score of ≤ 48 in the external cohort. Baseline demographics, vital signs, severity scores, and laboratory data within the first 48 h of intensive care unit (ICU) admission were recorded. Features were selected using least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. The dataset was split into training and validation sets in a 7:3 ratio. Seven machine-learning models were constructed: LightGBM, XGBoost, logistic regression, Naïve Bayes, random forest, CatBoost, and a support vector machine. Model performance was assessed in terms of the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, calibration curves, and decision curve analysis. SHapley Additive explanations (SHAP) analysis was used to interpret the optimal model.

Results: Among 3,842 patients, 203 (5.28%) were diagnosed with CIM/CIP. Seven key features were selected using the LASSO and Boruta methods. The random forest model performed the best, with an AUC of 0.772 in the validation set and 0.753 in the external cohort. It exhibited good calibration and the highest net benefit. The SHAP analysis revealed that early antibiotic use, early mechanical ventilation, sequential organ failure assessment scores, and age were the main predictors of ICU-AW.

Conclusion: A random forest model using early ICU data could effectively predict the risk of ICU-AW in patients with sepsis and offer interpretation via SHAP. Thus, it may serve as a clinical decision-making tool for early risk identification and optimized prevention.