Apprenticeship Learning for a Predictive State Representation of Anesthesia.

Area of Science:

• Anesthesiology
• Artificial Intelligence
• Decision Support Systems

Background:

• Maintaining optimal Depth of Anesthesia (DoA) is critical for patient safety during surgery.
• Current methods rely heavily on anesthesiologist experience, with limited predictive capabilities.
• Anesthesiologist decision-making in drug delivery is complex and influenced by multiple physiological variables.

Purpose of the Study:

• To develop and validate an original decision support algorithm for anesthesiologists.
• To assist in maintaining the optimal Depth of Anesthesia (DoA) through intelligent drug delivery.
• To leverage apprenticeship learning for a data-driven approach to anesthesia management.

Main Methods:

• An algorithm based on Transform Predictive State Representation (TPSR) was developed.
• The model utilized apprenticeship learning, trained on observed anesthesiologist practices.
• Input variables included Heart Rate (HR), Mean Blood Pressure (MBP), Respiratory Rate (RR), and anesthetic agent concentration (AAFi).

Main Results:

• The TPSR-based model demonstrated strong performance in predictions and simulations.
• Low-rank dynamical systems within the model showed the best predictive accuracy.
• An evaluation with six anesthesiologists found 95.7% of the algorithm's actions to be valid.

Conclusions:

• TPSR-based models effectively capture anesthesiologist behavior for predicting DoA.
• The algorithm accurately predicts DoA using commonly monitored physiological variables.
• This AI-driven approach shows significant potential for improving anesthesia management and patient safety.