358: PREDICTING SEVERITY OF DELIRIUM ON ICU ADMISSION: DEVELOPMENT OF AN AUTOMATED MACHINE LEARNING MODEL

Introduction: Approximately 50% of critically ill patients are affected by delirium. The early detection and mitigation of delirium severity in the intensive care unit (ICU) can significantly reduce the risk of patient morbidity and mortality. Electronic tools that measure and predict delirium severity can help identify such patients upon ICU admission and facilitate an individualized care plan. The primary aim of this study was to develop an automated machine learning model for delirium severity measurement and predict the level of delirium severity upon ICU admission. Methods: This retrospective study (01/01/2018-12/31/2021) extracted demographic and clinical data from the electronic health record (EHR) for adult patients (age>18) at ICU admission. To automate delirium severity measurement, a crosswalk mapped EHR data to CAM-ICU-7 scoring, the gold-standard for delirium severity. The correlation and accuracy of the automated model were evaluated against clinically documented delirium. Five machine learning models (multinomial logistic regression, gradient boosting method, random forest, neural network, and support vector machine) were supplemented with Sequential Optimization (SO) techniques. Their performance was evaluated to develop and finalize the model. Analysis was completed with R, v4.2.2. Results: In total, 38,021 patients were included with a median age of 64 years (52-74), 42.3% female (n=16,093/38,021) and 22% had documented delirium (n=8,222/38,021). The automated delirium severity measurement rule at time of ICU admission significantly correlated (r=0.52, p<.0001) and accurately assigned no/moderate/severe levels in 89% of the documented delirium cases. Multinomial logistic regression with SO performed with 73% ± 3% accuracy compared to other methods. The accuracy of severe delirium prediction was 78% while moderate delirium severity prediction was 33%. The cumulative dose of outpatient benzodiazepines and given upon ICU admission were top features identified to predict severe delirium. Conclusions: This is the first study to automate delirium severity measurement using routine EHR data and to develop a model to predict the level of delirium severity upon ICU admission. Future studies should validate and improve the model in a prospective study.