Predicting respiratory decompensation in mechanically ventilated adult ICU patients

• DOI: 10.3389/fphys.2023.1125991
• Publication Date: 2023-04-14T05:40:17Z
• Authors: Yvette Tan; Michael Young; Akanksha Girish; Beini Hu; Zina Kurian; Joseph L. Greenstein; Han Kim; Raimond L Winslow; James Fackler; Jules Bergmann
• Keywords: Hyperparameter; Decompensation

Introduction: Mechanical ventilation is a life-saving treatment in the Intensive Care Unit (ICU), but often causes patients to be at risk of further respiratory complication. We created statistical model utilizing electronic health record and physiologic vitals data predict Center for Disease Control Prevention (CDC) defined Ventilator Associated Complications (VACs). Further, we evaluated effect temporal resolution feature generation method choice on accuracy such constructed model. Methods: random forest occurrence VACs using records chart events from adult Medical Information Mart III (MIMIC-III) database. trained machine learning models two patient populations 1921 464 based low high frequency availability. Model features were generated both basic summaries tsfresh, python library that generates large number derived time-series features. Classification determine whether will experience VAC one hour after 35 h was performed classifier. Two different sample spaces conditioned five varying extraction techniques identify most optimal selection resulting best discrimination. Each dataset assessed K-folds cross-validation (k = 10), giving average area under receiver operating characteristic curves (AUROCs) accuracies. Results: After selection, hyperparameter tuning, extraction, performing used automatically achieved an AUROC 0.83 ± 0.11 0.69 0.10. Discussion: Results show potential viability predicting learning, indicate higher-resolution larger set by tsfresh yield better AUROCs compared lower-resolution manual