Depth of Anesthesia Analysis and Prediction with Deep Residual Shrinkage Network and Long Short-Term Memory

Precise measurement and control of the depth of anesthesia (DoA) is essential for anesthesiologists to ensure the success of surgery and the safety of patients. Electroencephalogram (EEG) signals, captured by electrodes on the scalp, can indicate the state of consciousness of the brain. However, estimating and predicting DoA via EEG signals still poses a challenge. This paper proposes a refined architecture consisting of a convolution neural network (CNN) based module and a recurrent neural network (RNN) based module, to perform real-time Bispectral Index (BIS), a common commercial indicator of DoA, analysis and BIS prediction. The CNN-based module is a variation of the Deep Residual Shrinkage Network (DRSN), which produces 2 values from the spectrograms of the filtered EEG signals, one representing the real-time analysis of the BIS value, the other representing a crude prediction of the BIS value over a certain period. To simplify the problem, the period is set at 30 seconds in this study. The real-time analysis value, accumulated by time, is then viewed as a time sequence and processed by the RNN-based module, producing one additional value. This newly created value is then concatenated with the crude prediction value to make a refined BIS value prediction. The model is trained on the bio-data of 176 patients who underwent vascular surgery from the public dataset VitalDB. The model achieved satisfying results, with a mean squared error of 42.15 in real-time analysis and 46.92 in prediction. Our model outperformed several existing models in both real-time BIS analysis and BIS prediction and is expected to assist anesthesiologists in preparation for emergencies in clinical practices.