Journal Article

Development and validation of a machine learning model for early prediction of intensive care unit acquired weakness

Van den Berghe, Greet, Grandas, Fabian Güiza, Coppens, Grégoire, Nakano, Felipe Kenji, Vens, Celine, Vanhorebeek, Ilse, Van Aerde, Nathalie

2025

Overview

Background Early identification of potential high cost and high need patients on the ICU may assist in the development of targeted protocols, which allows proper resource utilization and initialization of preventive care. Weakness acquired in the ICU developed within the first week is an independent predictor of both short and long-term adverse outcomes, nonetheless early prediction is challenging. We aimed to develop and validate a machine learning model for ICU acquired-weakness (ICU-AW), using data readily available within the first 24 h of ICU admission. Methods Patients from the EPaNIC trial (NCT00512122, N  = 4640) who were assessed for muscle weakness at day 9 (IQR 8–13), after ICU-admission, using the Medical Research Council (MRC) sum. Patients are diagnosed with ICU-AW if their MRC is lower than 48. The final subset contains N  = 600. Our models were internally validated using 100 repetitions of fivefold cross validation. We compared three predictive models: (i) a random forest and (ii) a logistic regression model built using descriptors available at day 1, (iii) a random forest using only APACHE II as a descriptor. Both random forests contain 150 trees. Results The training set comprised 600 patients where the incidence of ICU-AW was 38.6% (232/600). The AUROC of the random forest with all descriptors and the logistic regression were 76% and 74%, respectively. The random forest (RF) achieved a specificity of 62% and a sensitivity 79%, whereas the logistic regression yielded 69% and 68%, respectively. The RF identified APACHE II, creatinine, SOFA PaO2/FiO2, bilirubin, BMI, age, glycemia upon admission, morning glycemia and sepsis as the most relevant descriptors. Lastly, the RF also presented very good calibration and clinical usefulness for a wide range of risk thresholds. Conclusions Machine learning models, especially random forests, can be used to predict if patients are at risk of developing ICU-AW, using data available within 24 h of admission. This tool allows prognostication early in an adult general critically ill patient population, with the potential to detect high cost and high need patients who benefit from different levels of care.

• Publisher: Springer International Publishing,Springer Nature B.V,Springer,SpringerOpen
• Subject: C-reactive protein / Calibration / Clinical outcomes / Creatinine / Critical Care Medicine / Critical illness / Datasets / Diabetes / Infections / Intensive / Intensive care / Intensive Care Unit / Machine learning / Medical research / Medicine / Medicine & Public Health / Methodologies / Muscle weakness / Patients / Random forest / Regression analysis / Sepsis / Supervised Machine Learning