Machine Learning-Based Prediction of Subsyndromal Delirium in Intensive Care Patients.

BACKGROUND

Subsyndromal delirium is common in intensive care unit (ICU) patients and can prolong hospital stay, increase costs, and worsen prognosis. Advance intervention to prevent subsyndromal delirium would be valuable.

OBJECTIVE

To construct a machine learning-based model to predict the risk of subsyndromal delirium in ICU patients.

METHODS

This prospective cohort study included data from 447 patients hospitalized in the ICU between September 2023 and August 2024. Eight independent predictors of subsyndromal delirium were identified by least absolute shrinkage and selection operator and multivariate logistic regression analyses. Four machine learning models and a logistic regression model were constructed and validated to obtain the optimal algorithmic model.

RESULTS

Subsyndromal delirium occurred in 90 ICU patients (20.1%). Richards-Campbell Sleep Questionnaire score, sedative use, restraint tape use, receipt of continuous renal replacement therapy or extracorporeal membrane oxygenation, intra-abdominal pressure, prealbumin level, history of alcohol consumption, and stroke were independent predictors of subsyndromal delirium. Of the 4 machine learning models constructed, the random forest model had the best comprehensive performance (area under the receiver operating characteristic curve, 0.885; F1 score, 0.629).

CONCLUSIONS

Four machine learning-based risk prediction models and 1 traditional logistic regression model were developed to predict risk of subsyndromal delirium in ICU patients. Choosing a suitable model for early screening of ICU patients at high risk of subsyndromal delirium would allow medical staff to formulate individualized early intervention plans, which could improve patients' prognosis and save medical costs.