Hierarchical Poincaré analysis for anaesthesia monitoring

Although the degree of dispersion in Poincaré plots of electroencephalograms (EEG), termed the Poincaré-index, detects the depth of anaesthesia, the Poincaré-index becomes estranged from the bispectral index (BIS) at lighter anaesthesia levels. The present study introduces Poincaré-index_20–30 Hz, targeting the 20- to 30-Hz frequency, as the frequency range reported to contain large electromyogram (EMG) portions in frontal EEG. We combined Poincaré-index_20–30 Hz with the conventional Poincaré-index_0.5–47 Hz using a deep learning technique to adjust to BIS values, and examined whether this layered Poincaré analysis can provide an index of anaesthesia level like BIS. A total of 83,867 datasets of these two Poincaré-indices and BIS-monitor-derived parameters were continuously obtained every 3 s from 30 patients throughout general anaesthesia, and were randomly divided into 75% for a training dataset and 25% for a test dataset. Two Poincaré-indices and two supplemental EEG parameters (EMG_70–110 Hz, suppression ratio) in the training dataset were trained in a multi-layer perceptron neural network (MLPNN), with reference to BIS as supervisor. We then evaluated the trained MLPNN model using the test dataset, by comparing the measured BIS (mBIS) with BIS predicted from the model (PredBIS). The relationship between mBIS and PredBIS using the two Poincaré-indices showed a tight linear regression equation: mBIS = 1.00 × PredBIS + 0.15, R = 0.87, p < 0.0001, root mean square error (RMSE) = 7.09, while the relationship between mBIS and PredBIS simply using the original Poincaré-index_0.5–47 Hz was weaker (R = 0.82, p < 0.0001, RMSE = 7.32). This suggests the 20- to 30-Hz hierarchical Poincaré analysis has potential to improve on anaesthesia depth monitoring constructed by simple Poincaré analysis.