Prediction of delirium occurrence using machine learning in acute stroke patients in intensive care unit

Abstract

Introduction: Delirium, frequently experienced by ischemic stroke patients, is one of the most common neuropsychiatric syndromes reported in the Intensive Care Unit (ICU). Stroke patients with delirium have a high mortality rate and lengthy hospitalization. For these reasons, early diagnosis of delirium in the ICU is critical for better patient prognosis. Therefore, we developed and validated prediction models to classify the real-time delirium status in patients admitted to the ICU or Stroke Unit (SU) with ischemic stroke.

Methods: A total of 84 delirium patients and 336 non-delirium patients in the ICU of Ajou University Hospital were included. The 8 fixed features [Age, Sex, Alcohol Intake, National Institute of Health Stroke Scale (NIHSS), HbA1c, Prothrombin time, D-dimer, and Hemoglobin] identified at admission and 12 dynamic features [Mean or Variability indexes calculated from Body Temperature (BT), Heart Rate (HR), Respiratory Rate (RR), Oxygen saturation (SpO2), Systolic Blood Pressure (SBP), and Diastolic Blood Pressure (DBP)] based on vital signs were used for developing prediction models using the ensemble method.

Results: The Area Under the Receiver Operating Characteristic curve (AUROC) for delirium-state classification was 0.80. In simulation-based evaluation, AUROC was 0.71, and the predicted probability increased closer to the time of delirium occurrence. We observed that the patterns of dynamic features, including BT, SpO2, RR, and Heart Rate Variability (HRV) kept changing as the time points were getting closer to the delirium occurrence time. Therefore, the model that employed these patterns showed increasing prediction performance.

Conclusion: Our model can predict the real-time possibility of delirium in patients with ischemic stroke and will be helpful to monitor high-risk patients.