CN115299962A

CN115299962A - Anesthesia depth monitoring method based on bidirectional gate control circulation unit and attention mechanism

Info

Publication number: CN115299962A
Application number: CN202210968028.9A
Authority: CN
Inventors: 吴强; 李浩然; 王浩东
Original assignee: Shandong University
Current assignee: Shandong University
Priority date: 2022-08-12
Filing date: 2022-08-12
Publication date: 2022-11-08

Abstract

The invention discloses an anesthesia depth monitoring method based on a bidirectional gating circulation unit and an attention mechanism. The invention adopts a bidirectional gating circulation unit and a deep learning network of an attention mechanism to analyze the electroencephalogram signals, and is characterized in that: after preprocessing such as denoising and the like is carried out on the original electroencephalogram signals, 46 characteristics of 30-50Hz frequency spectrum characteristics, sequencing entropy, sample entropy, wavelet entropy, 30-47Hz average frequency spectrum, 47-50Hz average frequency spectrum and alpha-ratio of the electroencephalogram signals are extracted, and the characteristic combination comprises linear and nonlinear characteristics, so that the defect that the anesthesia depth cannot be well represented by a single characteristic is overcome; then the 46 characteristics are sent into a BiGRU-Attention network of a bidirectional gating circulation unit and an Attention mechanism to carry out anesthesia depth prediction. According to the method flow, relevant experimental verification is carried out, and the result proves that the method provided by the invention is superior to a single-feature method and a traditional deep learning model method, and has a better anesthesia monitoring effect.

Description

Anesthesia depth monitoring method based on bidirectional gate control circulation unit and attention mechanism

Technical Field

The invention belongs to the field of electroencephalogram signal processing, and particularly relates to an anesthesia depth monitoring method based on a Bidirectional Gated Recurrent Unit (Bi-GRU) and an Attention mechanism (Attention).

Background

Anesthesia is a critical component of modern medicine and is essentially any procedure that is indispensable to surgery. However, excessive or insufficient anesthetic will cause serious physical, mental and psychological trauma to the patient, with the risk of intra-operative recovery being greatest, and therefore it is of great importance to control the anesthetic dosage appropriately.

Although electroencephalographic techniques have emerged in the first half of the last century, only clinicians and neurosciences experts were able to acquire and analyze electroencephalography images using electrodes containing 64 or more channels under strictly controlled laboratory conditions for decades. These high density electrodes are evenly distributed on the human scalp and are used to help researchers discover underlying neural mechanisms associated with motor, cognitive or emotional processing. In recent years, with the development of computer hardware and processor technology, people have gradually deepened understanding of brains, deeper understanding of brain structures and activity rules, and electroencephalogram-based products are also applied to various production and living fields.

In recent years, many methods for monitoring the depth of anesthesia based on electroencephalogram signals have been discovered in the field of anesthesia monitoring. In 1998, rampel et al proposed a Bispectral (BIS) index, which defines an index of depth of anesthesia of 0-100. This algorithm has certain limitations in burst suppression mode (BSP) as indicated by the high BIS index, with the patient still under anesthesia. Electroencephalogram signals exhibit nonlinear or chaotic behavior, and many studies indicate that nonlinear analysis can be used for electroencephalograms in medical applications. The characteristics of the electroencephalogram signals based on the nonlinear dynamics are widely applied to monitoring the depth of anesthesia, such as a hurst index, trend-free fluctuation analysis, entropy and frequency band power ratio.

Many researchers have developed various wavelet entropy algorithms for anesthesia depth monitoring based on wavelet transforms, such as Shannon Wavelet Entropy (SWE), tsalis Wavelet Entropy (TWE), and Renyi Wavelet Entropy (RWE). Wavelet transformation is an effective tool for extracting and analyzing the basic structure of a signal in the time-frequency domain. Wavelet transform, as a method for identifying time spectrum, can automatically adjust the size of a time window, better match the frequency characteristics of signals, and is an ideal tool for signal analysis and processing. In earlier studies, several different rates of electrical activity in each frequency band have been proposed as indicators of the depth of anesthesia. Shah demonstrated that the ratio of alpha and beta frequencies to delta frequency is a useful tool to identify phases of isoflurane anesthesia.

Compared with the traditional signal processing method, the method based on deep learning is widely applied to electroencephalogram signal processing, and a good effect is achieved. In recent years, methods based on deep learning have also achieved good performance in monitoring the depth of anesthesia.

Disclosure of Invention

In order to improve the accuracy of anesthesia monitoring and better realize clinical application, the invention provides an anesthesia depth monitoring method based on a bidirectional gating circulating unit and an attention mechanism. Two common methods for monitoring the depth of anesthesia are available, wherein the depth of anesthesia is judged by observing the physical signs of the patient such as the tension change of the breathing, eyes, skin, digestive tract and skeletal muscle, and the depth of anesthesia is monitored by the commonly used indexes of the anesthesia monitors such as heart rate variability, original electroencephalogram, electroencephalogram Bispectral Index (BIS) and auditory evoked potential index (AEP index). The invention adopts an anesthesia depth monitoring method based on electroencephalogram signals. Because electroencephalogram is a non-stationary time sequence signal, it is usually processed by common RNN such as LSTM and GRU. The method comprises the steps of firstly manually extracting the characteristics of the electroencephalogram signals, then sending the characteristics into a neural network of a Bi-directional gating circulation unit (Bi-GRU) and an Attention mechanism (Attention) to predict the anesthesia depth, and obtaining better anesthesia monitoring accuracy.

The technical scheme adopted by the invention is as follows:

a anesthesia depth monitoring method based on a bidirectional gating circulation unit and an attention mechanism comprises the following steps of firstly manually extracting features of brain telecommunication, then adding a layer of attention mechanism at the tail end of the bidirectional gating circulation unit to form a new network structure, and finally inputting the extracted features into the new network structure for prediction, wherein the specific operation steps are as follows:

data preprocessing: because every 500 points in the data set correspond to a label, every 500 points of the original electroencephalogram signal are divided into a time period; then sending each section of electroencephalogram signal into a 0.5-50Hz Butterworth band-pass filter for filtering, removing a baseline from the filtered signal, and finally removing power frequency interference by adopting a 50Hz wave trap until the pretreatment is finished;

(II) feature extraction: extracting the characteristics of the electroencephalogram signals preprocessed in the step (I), firstly, extracting 30-50Hz frequency spectrum characteristics through short-time Fourier transform, wherein the interval is 0.5Hz, and thus, 40 frequency spectrum characteristics can be obtained; then, sequentially extracting the ordering entropy, the sample entropy, the wavelet entropy, the average frequency spectrum of 30-47Hz, the average frequency spectrum of 47-50Hz and alpha-ratio of the electroencephalogram signals preprocessed in the step (I), wherein the total number of the characteristics is 6; therefore, in the step (II), 46 features including frequency spectrum features and nonlinear features are extracted from each section of electroencephalogram signals;

and (III) sending into a network: firstly, a new network structure of a bidirectional gating circulation unit plus attention mechanism is constructed, and the 46 characteristics extracted in the step (II) are sent to the new network structure for prediction to obtain a prediction result of anesthesia monitoring.

In particular, the new network structure is constructed by building two layers of bidirectional gated cyclic units and then connecting a layer of attention mechanism.

Drawings

FIG. 1 is a network architecture diagram of a bidirectional gated loop unit GRU and Attention mechanism Attention in the present invention;

FIG. 2 is an overall flow chart of the present invention for monitoring the depth of anesthesia;

Detailed Description

The invention is described in detail below with reference to the figures and examples.

Firstly, the invention firstly carries out preprocessing on the original brain electrical signals. The specific process is as follows:

1) Firstly, filtering the original brain electrical signal. Because the frequency range of the brain wave is about 0.5-50Hz, the invention adopts a Butterworth band-pass filter in matlab to filter out the brain wave signal of 0.5-50 Hz;

2) Secondly, baseline removal is carried out on the filtered signal, a detrend function in matlab is adopted, and baseline removal drift is finished;

3) And finally, 50Hz power frequency interference is removed, which is caused by a power system, and the filter trap function in matlab is adopted in the invention to filter the 50Hz power frequency interference.

Secondly, the invention extracts the characteristics of the preprocessed electroencephalogram signals.

1) As shown in fig. 2, first, spectral features of 30-50Hz are proposed by short-time fourier transform (STFT), with an interval of 0.5Hz, which results in 40 spectral features. Short-time fourier transform is the most common time-frequency analysis method, which represents the signal characteristics at a certain time by a segment of signal within a time window. The short-time fourier transform is formulated as:

where s (t) is the brain electrical signal and h (t) is the time window function used to "slice" the signal. In general, the time window is a window having values only over a finite length, and 0 in other regions, thus intercepting the signal over the finite length.

2) And secondly, extracting four nonlinear characteristics of ordering entropy, sample entropy, wavelet entropy and alpha-ratio. The Permutation Entropy (PE) is an average Entropy parameter that measures the complexity of a one-dimensional time series, and is similar to the LyaPunov index in terms of performance reflecting the complexity of the one-dimensional time series. Sample Entropy (SampEn) measures the complexity of a time series by measuring the magnitude of the probability of generating a new pattern in a signal, the greater the probability of generating a new pattern, the greater the complexity of the sequence. Wavelet Entropy (WE) is a quantitative index for measuring the ordered and disordered levels of multi-scale dynamic behaviors of a nonlinear signal, and can provide information of the complexity of a nonlinear dynamic process of the signal. Alpha-ratio is a spectrum analysis method, the ratio of the spectrum of an alpha frequency band and the spectrum of 30-50Hz is subjected to log removal, and the complexity of the electroencephalogram signal is measured by utilizing the parameter. The four nonlinear characteristics utilized by the invention are all calculated by adopting the corresponding functions in matlab. Finally, the invention also adopts average frequency spectrums of 30-47Hz and 47-50Hz, and because the frequency spectrums of 30-50Hz are calculated, the two average frequency spectrum characteristics can directly average the frequency spectrums of 30-47Hz and the frequency spectrums of 47-50 Hz. Therefore, a total of 46 features are obtained, and the feature extraction step is ended.

Thirdly, constructing a BiGRU-Attention network of a bidirectional GRU filling expectation mechanism. At present, a plurality of people use recurrent neural networks such as LSTM, GRU and the like to process the anesthesia monitoring problem, as shown in figure 1, the invention provides a bidirectional GRU plus Attention network structure, which has two main advantages compared with a single LSTM or a single GRU, one of which is that the bidirectional GRU structure enables the network to not only remember useful historical information, but also remember the future state as a reference in the training process, which is the bidirectional meaning, because in some problems, the output at the current moment is not only related to the previous state, but also possibly related to the future state, and the anesthesia monitoring problem is just the same; and secondly, adding an Attention layer after the output of the Bi-GRU layer, wherein the Attention layer calculates the weight of each time sequence, then weights and takes the vectors of all time sequences as characteristic vectors, and then classifies the vectors.

Fourthly, inputting 46 characteristics into a BiGRU-Attention network to obtain an accuracy result of anesthesia monitoring, and comparing the accuracy result with characteristics such as wavelet entropy and the like or an LSTM network structure which are independently adopted. The experimental result shows that the accuracy of the method is superior to that of other network models. The results are shown in Table 1.

TABLE 1

Anesthesia monitoring method	Rate of accuracy
		Entropy of wavelets	0.8201±0.1006
Sample entropy	0.7182±0.1363
		Alpha ratio	0.8354±0.0867
30-47Hz	0.8285±0.0842
		47-50Hz	0.8282±0.09
Ordering entropy	0.8373±0.0739
		LSTM	0.8479±0.0748
The invention of BiGRU-Attention	0.8523±0.0827

Obviously, the experimental result 0.8523 of the BiGRU-Attention structure is superior to other experimental results. The overall process of the present invention is shown in FIG. 2.

Claims

1. A anesthesia depth monitoring method based on a bidirectional gating circulation unit and an attention mechanism comprises the following steps of firstly manually extracting features of brain telecommunication, then adding a layer of attention mechanism at the tail end of the bidirectional gating circulation unit to form a new network structure, and finally inputting the extracted features into the new network structure for prediction, wherein the specific operation steps are as follows:

data preprocessing: because each 500 points in the data set corresponds to a labelLabel (Sticks)Therefore, every 500 points of the original brain electrical signal are divided into a time period; then sending each section of electroencephalogram signal into a 0.5-50Hz Butterworth band-pass filter for filtering, removing a baseline from the filtered signal, and finally removing power frequency interference by adopting a 50Hz wave trap until the pretreatment is finished;

(II) feature extraction: extracting the characteristics of the brain electrical signals preprocessed in the step (I), firstly extracting 30-50Hz spectrum characteristics through short-time Fourier transform, wherein the interval is 0.5Hz, and thus obtaining 40 spectrum characteristics; then, sequentially extracting the ordering entropy, the sample entropy, the wavelet entropy, the average frequency spectrum of 30-47Hz, the average frequency spectrum of 47-50Hz and alpha-ratio of the electroencephalogram signals preprocessed in the step (I), wherein the total number of the characteristics is 6; therefore, in the step (two), 46 features including the frequency spectrum feature and the nonlinear feature are extracted from each segment of electroencephalogram signal;

2. The method for monitoring the depth of anesthesia based on a bidirectional gated cyclic unit and an attention mechanism as claimed in claim 1, wherein the new network structure is constructed by building two layers of bidirectional gated cyclic units and then connecting one layer of attention mechanism.