Predictive Application for Early Delirium Detection Subtypes Using GLM's

Delirium is an acute neuropsychiatric dysfunction prevalent in patients admitted to inpatient and intensive care hospital settings. As a multifactorial manifestation, it is typically underdiagnosed and overlooked. Delirium can be categorized, based on motor activity profile, into hypoactive and hyperactive subtypes. Multinomial logistic regression predictive models are often implemented to identify the most influential variables, as they allow for modelling the relationship between predictors and a multinomial dependent variable. In this context, the goal of this paper arises, aiming to develop an application capable of predicting the occurrence of delirium and its subtypes using the GLMs methodology. Subsequently, variable selection was performed using various techniques, with the Elastic Net method having an alpha value of 0.1 showing the best performance. The model achieved combines the multinomial logistic regression algorithm with the elastic net method, which is included in the web application. For the hypoactive subtype, it allowed the selection of 27 variables, resulting in an AUC-ROC of 0.691. The most influential variables include the length of hospitalization in days, alcoholism, analgesics, cardiotonics, as well as the diagnostic group related to toxicity and drugs. Regarding the hyperactive subtype, the model identified 29 relevant variables, with an AUC-ROC of 0.531. The most impactful variables include PCR, age, pO(2), SIRS criteria, and the ER source, specifically UDC1 (Clinical Decision Unit identifies high-priority patients with assigned yellow wristbands). The application is available at https://alexandra-coelho.shinyapps.io/Deliriumdetectionapp/. While further enhancements are possible, this predictive model remains a valuable tool for healthcare professionals diagnosing delirium in emergency rooms.