Deep learning guided propofol ketamine dosing and inflammation trajectories in elderly burns

Area of Science:

• Artificial Intelligence in Medicine
• Anesthesiology
• Critical Care Medicine

Background:

• Elderly burn patients experience severe hyperinflammation and reduced hemodynamic reserve.
• Optimizing combined propofol-ketamine anesthesia for this group is challenging due to limited data on anti-inflammatory mechanisms and safety parameters.

Purpose of the Study:

• To develop and validate a machine learning model for predicting inflammatory marker trajectories.
• To integrate predictions with a reinforcement learning agent for individualized propofol-ketamine dosing.

Main Methods:

• Retrospective multi-cohort analysis of ICU databases (MIMIC-IV, eICU).
• Utilized an Event-Transformer model with Neural ODE dynamics for inflammatory marker prediction.
• Employed a safety-constrained offline reinforcement learning agent (CPQL) for anesthetic dosing.

Main Results:

• The Event-Transformer model significantly improved IL-6 prediction accuracy (70.1% reduction in MAE) and inflammatory spike detection (AUROC 0.814).
• The CPQL agent improved MAP target range adherence (62.3% to 71.8%), reduced vasopressor use (27.0% to 18.4%), and decreased peak CRP by 21.3%.
• CPQL achieved 0.0% safety constraint violations, unlike unconstrained RL (4.2%), while reducing propofol exposure by 12.1% with adjunct ketamine.

Conclusions:

• An integrated AI pipeline can personalize propofol-ketamine anesthesia for elderly burn patients.
• The system shows potential for improving hemodynamic stability and reducing inflammatory burden.
• This framework can serve as a decision-support tool for future clinical trials.