Early Delirium Prediction Via Serial EEG Trajectories and Machine Learning

Background and Rationale:

Delirium is a critical manifestation of acute brain dysfunction, affecting 10-15% of all hospitalized patients and over 25% of those in intensive care units (ICU). In the trauma ICU, patients are particularly vulnerable due to an inflammatory cascade from repeated surgeries, blood-brain barrier disruption, traumatic brain injury (TBI), and mandatory opioid administration. Despite its clinical significance-including increased mortality and long-term cognitive impairment-early detection remains challenging. Current bedside tools like the CAM-ICU are limited by their periodic nature and dependence on clinician expertise, often missing the rapid neurophysiologic fluctuations that define delirium.

Study Objectives and Methodology:

While previous studies have used electroencephalography (EEG) as a "snapshot" to identify delirium, such cross-sectional approaches often reflect transient sedative depth rather than true neurocognitive vulnerability. This study proposes a longitudinal approach, focusing on the trajectory of change in cortical dynamics over time.

We acquired brief, serial resting-state EEG three times daily for at least three consecutive days from critically ill trauma patients. Using a feasible frontal montage, we quantified a comprehensive set of features, including spectral power (slowing), nonlinear complexity, and phase-based functional connectivity.

Research Hypothesis:

The framework utilizes machine learning (ML) to harness these longitudinal trajectories, aiming to predict delirium vulnerability before formal clinical diagnosis. Furthermore, we hypothesize that eyes-open recordings-by imposing a minimal arousal constraint-will better capture wakeful network integrity and provide superior predictive power compared to traditional eyes-closed recordings, which are often confounded by sedation and drowsiness in the trauma ICU environment.

Clinical Impact:

By identifying the optimal recording condition and establishing an ML-based prediction framework, this study seeks to define a standardized neurophysiologic monitoring strategy. This will ultimately facilitate early intervention and improve the long-term neurological prognosis of severe trauma survivors.

• Inclusion Criteria:

Trauma patients admitted to the Trauma Intensive Care Unit (TICU) who meet the following criteria:

• Patients aged 18 to 65 years.
• Severe trauma patients with an Injury Severity Score (ISS)
• Exclusion Criteria:

Patients with a head Abbreviated Injury Scale (AIS) ≥ 2 Patients with a Richmond Agitation-Sedation Scale (RASS) score ≤ -2 History of neurological disorders (e.g., Parkinson's disease, dementia, cerebrovascular disease) History of major psychiatric disorders (e.g., schizophrenia, bipolar disorder, intellectual disability, autism spectrum disorder) History of illicit drug use disorder or positive results on a urine drug screen for substances other than Benzodiazepines or Tricyclic antidepressants.

Clinical evidence of acute alcohol withdrawal (CIWA-Ar score > 10) History of liver failure or hepatic encephalopathy (Child-Pugh Class B or C) Renal impairment requiring renal replacement therapy (RRT) Inability to perform the Confusion Assessment Method for the ICU (CAM-ICU) due to the following Inability to communicate in Korean Failure to obey commands (unable to follow test instructions) Severe visual or hearing impairment Refusal to undergo CAM-ICU assessment Requirement for isolation due to infectious diseases (e.g., COVID-19, active tuberculosis).