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

PURPOSE

Forecasting the depth of anesthesia (DOA) is significant for adjusting the dosage of propofol infusion during total intravenous anesthesia. Traditional pharmacokinetic (PK) and pharmacodynamic (PD) models may lack precision in DOA forecasting. Thus, machine learning models optimized to fit the clinical data have been adopted to predict the DOA, showing improved accuracy. These earlier approaches followed the recurrent and/or the feed-forward framework that generates the predictions based on the patient's demographic data. In this work, to explore the potential advantages of using real-time information in DOA forecasting, we proposed using the auto-regressive (AR) frameworks to improve the accuracy of DOA prediction during the induction phase of anesthesia.

METHODS

This study employed two auto-regressive frameworks, intra-loop auto-regressive (ILAR) and real-time auto-regressive (RTAR), respectively, for two different scenarios, with an attention mechanism to predict the DOA during the induction phase. The profiles of 528 patients were considered to forecast the DOA values within the induction phase. 80% (433 patients) of the data were randomly selected for training, while the remaining 105 were used for validation. The model's performance was evaluated using metrics including absolute median performance error (MDAPE), median performance error (MDPE), and root mean square error (RMSE). To assess the performance of our approach, we compared it with a conventional feed-forward framework.

RESULTS

The experimental results indicated that the proposed auto-regressive approach outperformed the framework comprising recurrent and feed-forward networks (implemented using the LSTM, followed by the MLP networks in this paper) in predicting the DOA of the induction phase. The MDAPE values of our ILAR and RTAR frameworks were, respectively, 2.5% and 11.6%; in contrast, the value of the LSTM-MLP framework was 17.7%. With the real-time bispectral Index (BIS) values, the RTAR framework significantly outperforms the LSTM-MLP framework, achieving the least MDAPE and RMSE values and obtaining the closest to zero MDPE values. When the real-time BIS values are unavailable, the proposed ILAR framework has much lower MDAPE values and closer to zero MDPE values than the LSTM-MLP framework.

CONCLUSION

We proposed two auto-regressive frameworks, the ILAR and the RTAR, to predict the DOA values in the induction phase of the anesthetic period. The proposed ILAR and the RTAR can be used to predict the DOA for the two scenarios, depending on whether the sensor can observe the real-time BIS values. The experimental results demonstrated that the BIS values forecast by the proposed approaches were significantly closer to the ground truth than the previous LSTM-MLP framework.