Rationale: Intraoperative hypotension, commonly defined as mean arterial pressure (MAP) below 65 mmHg, occurs in up to 60% of patients undergoing general anesthesia and is independently associated with acute kidney injury, myocardial injury, stroke, and 30-day mortality. Current approaches to hemodynamic management are reactive and rely on individual clinician judgment. Machine learning-based prediction systems (e.g., the Hypotension Prediction Index) have shown potential in reducing hypotension burden, but do not provide actionable dosing recommendations. Reinforcement learning (RL) offers a fundamentally different approach: learning optimal sequential decision-making policies from clinical data. The RL-PRAIS system (Reinforcement Learning-based Perioperative Real-time Anesthesia Intelligent System) consists of a Transformer-based patient state encoder and a model-based RL framework (patient model plus policy model) that generates real-time vasoactive agent dosing recommendations.
Hypothesis: The RL-PRAIS system can learn intraoperative hemodynamic management decisions that approximate those of experienced anesthesiologists and, when deployed as a decision support tool, can reduce the burden of intraoperative hypotension compared to standard care.
Study Design: This is a prospective, dual-center, parallel-group, randomized, controlled, proof-of-concept superiority trial. The study follows a four-phase stepwise validation framework: (1) model development using retrospective electronic health records (n = 7,216), (2) retrospective validation (n = 75), (3) prospective deployment study (n = 40), and (4) proof-of-concept randomized controlled trial (n = 40, 20 per arm). This registration covers Phase 4 (the RCT).
Objectives: Primary Objective: To evaluate the concordance between RL-recommended vasoactive agent dosing and attending anesthesiologist-executed dosing, quantified by mean absolute error (MAE) in norepinephrine-equivalent units. Secondary Objectives: To compare intraoperative hemodynamic outcomes between the RL-guided and standard care groups, including time-weighted average of MAP below 65 mmHg (TWA-MAP \< 65), percentage of time with MAP in the target range of 65-100 mmHg (MAP TIR), incidence of hypotensive events, cumulative vasoactive agent consumption, MAP variability, and clinician acceptance rate of RL recommendations. Exploratory Objectives: To collect pilot data on 30-day major adverse cardiac or cerebrovascular events (MACCE), acute kidney injury, perioperative myocardial injury, PACU length of stay, hospital length of stay, and 30-day all-cause mortality.
Interventions: Intervention Group (RL-Guided): The RL-PRAIS system provides real-time dosing recommendations for vasoactive agents (norepinephrine, phenylephrine, and ephedrine) based on continuous invasive arterial blood pressure monitoring and patient state features. Recommendations are displayed on a bedside screen at 1-minute intervals. The attending anesthesiologist retains full clinical autonomy and is responsible for all final dosing decisions. Control Group (Standard Care): The anesthesiologist manages intraoperative hemodynamics according to institutional standard practice and clinical judgment, with continuous invasive arterial blood pressure monitoring but without input from the RL-PRAIS system.
Study Population: Adult patients aged 18 to 85 years, ASA physical status I-IV, scheduled for elective non-cardiac surgery under general anesthesia with continuous invasive arterial blood pressure monitoring. Surgery is expected to last at least 2 hours. Key exclusion criteria include emergency surgery, severe obesity (BMI \>= 35), and allergy to study-related agents.
Statistical Considerations: The sample size of 40 participants (20 per arm) is designed for a proof-of-concept evaluation, consistent with the benchmarking study (Wang et al., Nature Medicine 2023, RL-DITR, n = 16 in feasibility RCT). The primary analysis will compare MAE between groups using a two-sided paired or independent t-test (or non-parametric equivalent). Secondary hemodynamic outcomes will be compared using the Wilcoxon rank-sum test. Exploratory outcomes will be reported descriptively without formal hypothesis testing.
