Dynamic Multimodal Delirium Warning After Cardiac Surgery

Overview

After heart surgery, up to half of all patients may develop a state of sudden confusion called postoperative delirium. This condition can lead to longer time on a breathing machine, extended stays in the intensive care unit (ICU), and a slower overall recovery. Currently, doctors have no reliable way to predict delirium early enough to take preventive action. This study aims to build a computer-based early warning system. The system will combine continuous, real-time measurements of brain waves (EEG), the oxygen level in the brain, and heart and blood pressure function. It will also include information about each patient's health status. By analyzing all of these signals together, the model is designed to give an alert 1 to 6 hours before delirium might start, giving the care team a window of time to intervene. The study will take place in the ICU at Zhongda Hospital, Southeast University. Adults between 18 and 80 years old who are admitted to the ICU after heart surgery will be invited to participate. All patients will receive the usual standard of care; the study does not test any new treatment. Participation means the investigators will continuously record the brain, oxygen, and heart signals that are already being monitored, and a researcher will regularly assess the patient's thinking and alertness with a simple bedside check.

Postoperative delirium (POD) following cardiac surgery has an incidence of 20-50% and is associated with prolonged mechanical ventilation, extended ICU and hospital stay, and increased mortality. Its pathophysiology involves a complex interplay of cerebral hypoperfusion, neuroinflammation, blood-brain barrier disruption, and neurotransmitter imbalances-a "multiple-hit" model. Current clinical assessment relies on static risk stratification or post-hoc diagnostic tools , which lack the dynamic, pre-symptomatic warning capability needed for timely intervention. The increasing availability of multimodal ICU monitoring (EEG, near-infrared spectroscopy, invasive hemodynamics) and advanced time-series deep learning provides an unprecedented opportunity to capture the evolution of physiological uncoupling before the clinical manifestation of delirium. Objective: This study aims to develop and validate a multimodal deep learning model that fuses continuous central electrophysiology (frontal EEG), regional cerebral oxygen saturation (rScO₂), macro-hemodynamic parameters, and clinical static/dynamic risk factors to provide a dynamic early warning of POD 1-6 hours in advance. Study Design: This is a single-center, prospective, observational cohort study. Setting and Population: The study will enroll consecutive adult patients (18-80 years) admitted to the Department of Critical Care Medicine at Zhongda Hospital, Southeast University, after cardiac surgery (CABG, valve repair/replacement, major aortic surgery, or combined procedures) between May 1, 2026 and December 30, 2027. All eligible patients must have multimodal monitoring including continuous EEG, bilateral frontal rScO₂, and invasive arterial blood pressure, and must provide informed consent. Key Exclusion Criteria: Pre-existing dementia, psychiatric illness or long-term antipsychotic use precluding accurate delirium assessment; severe hepatic (Child-Pugh C) or renal insufficiency (eGFR \<30 mL/min/1.73 m²); significant brain injury or seizure history; inability to obtain adequate signal quality; expected death within 24 hours. Data Collection and Monitoring: Data will be captured at multiple time windows: preoperative baseline, intraoperative period, and postoperative time points (immediately upon ICU arrival, 6h, 24h, 48h, and at the moment delirium is first detected). Preoperative phenotyping includes demographics, MoCA, Clinical Frailty Scale, and EuroSCORE II. Continuous EEG features (power spectral density, burst suppression ratio, complexity indices) and rScO₂ (baseline, desaturation events \>20%, autoregulation index COx) will be recorded. Hemodynamic variables include heart rate, beat-to-beat blood pressure variability, and, where available, derived cardiac output metrics. Comprehensive clinical data (laboratory values, sedation/analgesic dosing, vasoactive-inotropic score, mechanical ventilation parameters, and SOFA/APACHE II scores) will be collected concurrently. Delirium Assessment: Trained research staff will assess delirium using the Confusion Assessment Method for the ICU (CAM-ICU) in conjunction with the Richmond Agitation-Sedation Scale (RASS). Assessments occur at baseline, postoperatively when the patient is awake, and at scheduled intervals, with documentation of first onset, duration, and subtype (hyperactive, hypoactive, mixed). Model Development and Analysis: All signals will be time-aligned to construct a high-resolution multimodal time-series dataset. The primary predictive model will be based on a Transformer or Long Short-Term Memory (LSTM) architecture employing a sliding window approach (e.g., input: preceding 2 hours; output: predicted delirium risk in the next 1-6 hours). The dataset will be split into training, validation, and test sets (7:1.5:1.5). Primary performance metrics are area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive/negative predictive values, and the achievable warning lead time. Model interpretability will be explored using SHAP or attention weight analysis. The added value of multimodal fusion will be quantified by comparing the full model against unimodal baselines (clinical data only, EEG only, or rScO₂ only). Outcomes: The primary outcome is the occurrence of POD. Secondary outcomes include duration of mechanical ventilation, ICU length of stay, and hospital length of stay. Significance: By delineating the temporal trajectories of EEG, cerebral oxygenation, and systemic hemodynamics preceding delirium, this study will provide new pathophysiological insights into the "cerebral perfusion-metabolism-electrical activity" uncoupling hypothesis and deliver a clinically implementable early warning framework to enable proactive brain-directed interventions.