Stratified Causal Inference for Intensive Care Unit Risk Prediction: Informatics-Based Modeling of Anesthetic Drug...

Area of Science:

• Anesthesiology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Postoperative intensive care unit (ICU) admission impacts 15-20% of surgical patients, increasing morbidity and healthcare costs.
• Current anesthetic dosing relies on empirical guidelines, lacking individualized risk assessment.
• Fentanyl-propofol combinations are commonly used, but their dose-response relationship with ICU admission risk is complex.

Purpose of the Study:

• To develop and evaluate a stratified, causal machine learning framework for identifying optimal fentanyl-propofol combinations.
• To predict postoperative ICU admission risk using electronic health record (EHR) data.
• To enable precision anesthesia and individualized clinical decision support.

Main Methods:

• Analysis of 67,134 perioperative EHRs from UC Irvine Medical Center (2017-2022).
• Utilized a hierarchical learning framework to estimate causal effects and control for confounding variables.
• Identified 6 dose-sensitive patient subgroups through stratified analysis, with postoperative ICU admission as the primary endpoint.

Main Results:

• High-risk fentanyl (>5 mcg/kg) and propofol (<1 mg/kg) combinations increased ICU admission risk by 36%.
• Six patient subgroups exhibited distinct dose-response patterns.
• Vulnerable populations (high glucose, elevated creatinine) showed increased risk even at standard doses.
• Optimal dose ranges determined: propofol 1.25-4.25 mg/kg and fentanyl 3.5-4.0 mcg/kg.

Conclusions:

• Fentanyl-propofol combinations have complex, nonlinear dose-response relationships with ICU admission risk.
• High-dose combinations increase risk synergistically; specific subgroups need enhanced monitoring.
• Findings support individualized dosing algorithms and risk assessment tools to reduce postoperative ICU utilization, pending external validation.