Forecasting anesthetic depth using an auto-regressive transformer in propofol infusion during the induction phase

Area of Science:

• Anesthesiology
• Machine Learning in Medicine
• Pharmacometrics

Background:

• Accurate depth of anesthesia (DOA) forecasting is crucial for safe propofol infusion during total intravenous anesthesia.
• Traditional pharmacokinetic (PK) and pharmacodynamic (PD) models often lack the precision required for real-time DOA prediction.
• Machine learning (ML) models have shown promise in improving DOA prediction accuracy by fitting clinical data.

Purpose of the Study:

• To explore the advantages of real-time information for DOA forecasting.
• To propose and evaluate novel auto-regressive (AR) frameworks for enhanced DOA prediction accuracy during anesthesia induction.
• To improve the precision of propofol dosage adjustments through more accurate DOA forecasting.

Main Methods:

• Two auto-regressive frameworks, intra-loop auto-regressive (ILAR) and real-time auto-regressive (RTAR), were developed and implemented with an attention mechanism.
• The models were trained and validated using data from 528 patients undergoing anesthesia induction, with 80% for training and 20% for validation.
• Performance was assessed using absolute median performance error (MDAPE), median performance error (MDPE), and root mean square error (RMSE), compared against a conventional feed-forward (LSTM-MLP) framework.

Main Results:

• The proposed ILAR and RTAR frameworks significantly outperformed the LSTM-MLP framework in predicting DOA during anesthesia induction.
• MDAPE values for ILAR and RTAR were 2.5% and 11.6%, respectively, compared to 17.7% for LSTM-MLP.
• The RTAR framework, utilizing real-time bispectral index (BIS) values, achieved the lowest MDAPE and RMSE, demonstrating superior performance.

Conclusions:

• The ILAR and RTAR auto-regressive frameworks provide accurate DOA predictions during the anesthetic induction phase.
• These models offer flexibility, applicable in scenarios with or without real-time BIS monitoring.
• The proposed AR approaches significantly improve the accuracy of BIS value forecasting compared to previous LSTM-MLP methods.