CN109833031B - An automatic sleep staging method using multiple physiological signals based on LSTM - Google Patents

Abstract

An automatic sleep staging method based on LSTM and using multiple physiological signals comprises the steps of firstly, acquiring electrocardiosignals, respiration signals and acceleration signals of a tested person; step two, signal processing, step three, extracting features for classification, step four, constructing a model, inputting the manually extracted features into a first-layer long-time memory model, inputting the output probability of the manually extracted features into a second-layer long-time memory model together with the manually extracted features as new features, constructing classifiers for different classification tasks, and step five, using the trained model for classification of sleep stages; the invention adopts various physiological signals including electrocardiosignals, thoracoabdominal respiration signals and head acceleration signals, but does not have electroencephalogram signals, thereby overcoming the defects caused by applying the electroencephalogram to sleep stages; meanwhile, a long-time memory model is adopted, the method is suitable for large-sample and large-data, the time correlation of sleep events is considered, and the accuracy and reliability of sleep staging are improved.

Description

An automatic sleep staging method using multiple physiological signals based on LSTM

technical field

The invention belongs to the technical field of biomedical signal processing, relates to signal processing such as electrocardiogram, respiration and acceleration, in particular to an automatic sleep staging method using multiple physiological signals based on a long-short-term memory model LSTM.

Background technique

Sleep is one of the most important life activities of the human body. Research on sleep patterns and sleep structure can help people improve sleep quality. Since the 1960s, sleep medicine has formed a set of standards after decades of development. Now the most commonly used is the sleep staging standard formulated by the American Academy of Sleep Medicine (AASM) in 2007: Based on polysomnography (PSG), according to the "AASM Sleep and Related Events Interpretation Manual", nocturnal sleep activities are divided into five different periods, namely wakefulness (W), sleep stage I (N1), sleep stage II stage (N2), sleep stage III (N3), rapid eye movement (Rapid Eye Movement, REM). Among them, PSG needs to record multi-channel physiological signals such as EEG, OMG, EMG, and ECG (Electrocardiogram, ECG) at the same time, and then it is interpreted by an experienced professional physician every 30s to obtain clinical classification results. In non-clinical applications, there will also be different classification methods, distinguishing the two categories of wakefulness and sleep (W, Sleep), and distinguishing the three categories of wakefulness, non-rapid eye movement (NREM) and rapid eye movement ( W, NREM, REM), four classifications (W, N1/N2, N3, REM) to distinguish wakefulness, light sleep (LS), deep sleep (Slow Wave Sleep, SWS) and rapid eye movement. Considering the associations and transitions between sleep stages, the four-category approach may be more practical than other classification criteria.

However, the sleep staging method based on PSG technology has many problems in practical application. It is generally carried out in a strict sleep laboratory, but too much signal measurement will affect the comfort of sleep and make the measurement results deviate from the real situation. The high cost also limits the popularity of this method. Therefore, it is of great significance to explore an automatic sleep staging method. At present, the research on automatic sleep staging methods mainly focuses on signal feature extraction and pattern recognition. According to the physiological signals used, it can be divided into three aspects: First, sleep staging research based on EEG. The research in this area is now relatively mature. At present, only using single-lead EEG signals to classify the sleep staging of healthy people can achieve an accuracy rate of more than 90%; The five classifications can achieve an accuracy of 92% and an accuracy of 86% for patients with sleep disorders. Second, sleep staging studies based on cardiopulmonary coupling-related signals. In this aspect, ECG and respiratory signals are mainly used. Huang Wenhan et al. used ECG and respiratory signals to classify patients with sleep-disordered breathing and achieved the highest accuracy rate of 71.9%. Thirdly, the study of sleep staging method based on acceleration signals during sleep, mainly for the classification of wakefulness and sleep, is consistent with the monitoring results of the PSG system up to 91%.

However, the above studies all have certain practical problems. Although the staging method based on EEG signals has a high accuracy rate, EEG is a relatively weak physiological signal and is prone to various interferences, so it has strict requirements on electrodes and acquisition process, and the cost is relatively high. At present, most of the sleep staging methods based on acceleration signals can only be used to distinguish between wakefulness and sleep, and the classification results are poor and have little practical reference significance. Therefore, the research on sleep staging method based on cardiopulmonary coupling has the most practical significance. ECG is a relatively easy-to-obtain physiological signal, with large amplitude and few interference factors, which has great practical application value. At present, there are some sleep staging studies that only use ECG signals. Heart rate variability (HRV) signals are extracted from ECG signals, and sleep staging is performed according to the differences of HRV signals in different sleep stages. The team from the School of Electrical and Electronic Engineering at Hakkari University compared the three-class results of four different classification methods for healthy people. Among them, the highest method achieved an accuracy rate of 87.11%, but the effect of using HRV for sleep staging is currently not very good. Ideal, so the research on sleep staging methods in this area also has great exploratory significance.

Human sleep lasts for a certain period of time, and the physiological signal changes during this period are a time series, and the Long Short Time Memory (LSTM) based on time series has a good ability to deal with such pattern recognition problems. . The LSTM model adds a forgetting gate on the basis of the traditional recurrent neural network, so that the neural network can selectively forget the parameters learned before, thereby avoiding the problem of long-term dependence. Yulita et al. used LSTM for sleep staging research, using three signals of EEG, OMG, and EMG to classify patients with sleep disorders and achieved an accuracy of 86%. Radha et al. Staging studies have also yielded good results.

To sum up, the current research methods on automatic sleep staging all have their own limitations, and there is no simple, reliable and efficient method. The LSTM-based neural network method has good classification ability for time series data, but it has not been well used in the field of sleep staging.

SUMMARY OF THE INVENTION

In order to overcome the shortcomings of the existing automatic sleep staging technology, the purpose of the present invention is to provide an automatic sleep staging method based on LSTM using multiple physiological signals, which is a universal, easy-to-implement and economical sleep staging method. ECG signal, respiration signal and acceleration signal, and then perform feature extraction on them, and use LSTM network to perform sleep staging to realize four different sleep staging tasks, namely two-class, three-class, four-class and five-class respectively, so as to meet the needs of different subsequent occasions. application requirements.

In order to achieve the above object, the technical scheme adopted in the present invention is:

An automatic sleep staging method using multiple physiological signals based on LSTM, including the following steps:

Step 1: Signal Acquisition

The first-lead ECG signal, the first-lead breathing signal and the three-lead acceleration signal of the subject are measured by using an electrocardiogram measuring instrument, a respiratory signal measuring instrument and a three-axis acceleration sensor.

Step 2: Signal processing;

Perform signal processing on the collected one-lead ECG signal, first-lead breathing signal and three-lead acceleration signal, and the three-lead are three mutually perpendicular directions of space x, y, and z; processing the first-lead ECG signal to obtain the final HRV signal and the respiratory wave amplitude signal R _fm ; the processing of the respiratory signal obtains the final RRV signal; the processing of the acceleration signal obtains the integrated acceleration signal acc(n);

Step 3: Feature extraction;

For the four kinds of signals obtained after the original signal processing: HRV, respiratory wave amplitude signal R _fm , RRV, and integrated acceleration signal acc(n), firstly perform segmental processing, with 5 minutes as the window length and 30 seconds as the step The sleep data of the whole night is divided into several data segments of the same length, and then feature extraction is performed on each segment of data, and each feature is normalized. The mean of the series, then divided by the variance of the feature series.

Step 4: Model construction.

The constructed model adopts two layers of LSTM, and each layer has four different classifiers, which are used to realize the staging tasks of two-class, three-class, four-class and five-class respectively.

The input sequence of the first layer of LSTM network is the feature sequence extracted in step 3. The input feature data passes through the first layer of LSTM layer, and there are four parallel classifiers in it. The network structure of each classifier is similar, including 5 layers: input layer, batch normalization layer, Bi-LSTM layer, forgetting layer, fully connected layer.

After the first LSTM layer, each classifier will output data of different dimensions, representing the probability of each class stage, and combining these probabilities can form a new set of features. This new set of features is connected in parallel with the features extracted in step 2 as the input of the second layer Bi-LSTM network. There are also four classifiers in parallel in the second layer LSTM, and the structure of each classifier is the same as that of the previous layer. In the same way, each classifier is trained to obtain the final classification model for prediction.

Step 5: Predict sleep stages

The four classifiers obtained in step 4 are used to predict the sleep period, signal processing and feature extraction are performed on the data to be predicted, and then the extracted features are sent to the classifier in the order of model training, and the sleep stage can be output. result.

The advantages of the invention are: in order to overcome the shortcomings of the existing automatic sleep staging technology, the invention proposes a general, easy-to-implement and economical sleep staging method. First, the only physiological signals used in the present invention are ECG, respiration signal, and acceleration signal. These three physiological signals are easy to obtain and simple to operate, which overcomes the disadvantages of using EEG signals, such as being expensive and complicated. Secondly, the present invention uses the LSTM model, which can make good use of physiological signals during sleep as the correlation before and after the time series, and improve the accuracy of sleep staging. Of course, the present invention has a wide range of applicable scenarios, and can be easily applied to fields such as intensive care wards, sleep departments, and home sleep monitoring, and can also be easily transplanted into portable devices to promote the development of mobile medicine.

Description of drawings

Figure 1 is an overall block diagram of the method.

Figure 2 is the overall block diagram of LSTM model construction.

Figure 3 shows the network structure of each classifier.

specific embodiment

In order to illustrate the operation process of the present invention more clearly, the present invention will be described in detail below with reference to the accompanying drawings and examples.

Referring to Figure 1, an automatic sleep staging method based on LSTM using multiple physiological signals, including the following steps:

Step 1: Signal acquisition:

Use the ECG measuring instrument to collect the ECG signal, select the ECG signal of lead II, the sampling rate is 100Hz, use the respiratory signal measuring instrument to measure the thoracic and abdominal breathing signal as the breathing signal, the sampling rate is 100Hz, and use the triaxial acceleration sensor to measure the patient The head movement information of the prefrontal lobe is used as the acceleration signal, and the sampling rate is 100Hz.

Step 2: Signal processing.

The collected one-lead ECG signal, one-lead breathing signal and three-lead acceleration signal are processed. The three-lead are three mutually perpendicular directions of space x, y, and z. The signal processing steps can be divided into three aspects:

①Processing of ECG signal.

First, filter the original ECG signal;

Secondly, extract the HRV signal from the ECG signal, identify the peak point position of each R wave according to the maximum slope method, and then obtain the time interval between two adjacent peaks by the difference between the adjacent R wave peak positions, but this The time interval sequence obtained is uneven, and then the cubic spline interpolation method is used to interpolate the sequence at equal intervals to become the time series at the target sampling frequency, and finally the final HRV signal can be obtained by taking the reciprocal of the sequence; According to actual needs, downsample the HRV signal and output it;

Then, the respiratory wave amplitude signal R _fm is extracted on the basis of HRV, and the HRV is passed through a third-order Butterworth high-pass filter with a cut-off frequency of 0.15Hz and a third-order Butterworth with a cut-off frequency of 0.5Hz using the frequency modulation method. Voss low-pass filter to obtain the respiratory wave amplitude signal, denoted as R _fm .

②Respiratory signal processing.

Filter the original breathing signal, use a low-pass with a cutoff frequency of 1Hz and a high-pass 3rd-order Butterworth filter with a cutoff frequency of 0.01Hz to filter out the 0.01-1Hz effective components of the respiratory signal, and then extract the effective components from the respiratory signal. Extract the RRV signal from the signal, identify the position of the peak point of each respiratory wave according to the maximum slope method, and then obtain the time interval between two adjacent peaks by the difference between the adjacent peak positions, and then use the cubic spline interpolation method for this. The sequence is interpolated at equal intervals to become a time sequence at the target sampling frequency. Finally, the final RRV signal can be obtained by taking the inverse of the sequence. Finally, the RRV is down-sampled and output according to the actual demand.

③ Processing of acceleration signal.

For the three-axis acceleration signal, calculate the arithmetic square root of the sum of its squares according to formula (1) to obtain the signal acc _temp (n); perform 10-point smoothing filtering on the signal acc _temp (n) to obtain the final comprehensive acceleration signal, record Let acc(n).

Among them, M=3, representing the acceleration signals of the three channels.

Step 3: Feature extraction;

For the four kinds of signals obtained after the original signal processing: HRV, respiratory wave amplitude signal R _fm , RRV, and integrated acceleration signal acc(n), firstly perform segmental processing, with 5 minutes as the window length and 30 seconds as the step The sleep data of the whole night is divided into several data segments of the same length, and then feature extraction is performed on each segment of data, and each feature is normalized. The mean of the series, then divided by the variance of the feature series.

Respiratory wave amplitude signal R _fm , the extractable features include: the median value of the entire data segment divided by the sequence range; the median value of the quartile and quartile interval sequences divided by the sequence range; integer The mean of the segment data divided by the variance; the median value of the respiratory wave peak-peak time interval series; the median value of the respiratory wave peak-trough time interval series; the median value of the respiratory wave trough-peak time interval series; the respiratory wave peak-peak time interval The median value of the interval series is divided by the sequence range; the median value of the respiratory trough-peak time interval series is divided by the sequence range; the median value of the respiratory wave trough-peak time interval series is divided by the sequence range.

For HRV and RRV, the features that can be extracted include: the mean of the entire data; the variance of the entire data; the mean divided by the variance of the entire data; the sum of the power spectrum of the low frequency band (0.01Hz-0.04Hz) of the entire data; The sum of the power spectrum of the middle frequency band (0.04Hz-0.15Hz) of the entire data; the sum of the power spectrum of the high frequency (0.15Hz-0.4Hz) of the entire data; the sum of the total power spectrum of the entire data; the high frequency of the entire data The ratio of the sum of the power spectrum of (0.15Hz-0.4Hz) to the sum of the power spectrum of the mid-band (0.04Hz-0.15Hz).

For the integrated acceleration signal acc(n), the features that can be extracted include: the mean of the entire data; the variance of the entire data; the mean of the entire data divided by the variance; the median of the entire data divided by the sequence range; quarter The median of the series of 1 quantile and quartile intervals divided by the series range.

A total of 30 features are listed above. On this basis, the data segment is still selected with a window length of 5 minutes and a step size of 30 seconds, and the extracted feature sequences are arranged in descending order. The value at three-tenths of the position is used as a feature of the segment, and finally, a total of 90 feature parameters are obtained. Due to differences between individuals, each feature needs to be normalized before the model is constructed. The normalization method is to subtract the mean of the feature sequence from each person's features over the night, and then divide by the variance of the feature sequence.

Step 4: Model construction.

The model constructed by the present invention adopts two-layer LSTM, and each layer has four different classifiers to realize the staging tasks of two-class, three-class, four-class and five-class, referring to FIG. 2 .

The input sequence of the first layer LSTM network is the 90 feature sequences extracted in step 2. The input feature data passes through the first layer of LSTM layer, which contains four parallel classifiers. The network structure of each classifier is similar. Referring to Figure 3, it includes five network layers: input layer, batch normalization layer, Bi-LSTM layer , forgetting layer, and fully connected layer.

After the first LSTM layer, each classifier will output data of different dimensions, representing the probability of each class stage, and these probabilities are combined to form 14 new columns of features. Connect these 14 new features in parallel with the features extracted in step 2 as the input of the second layer of LSTM network. There are also four classifiers in parallel in the second layer of LSTM, and the structure of each classifier is the same as the previous layer. , after each classifier is trained with a large number of samples, the final classification model for prediction is obtained;

Step 5: Predict sleep stages

The four classifiers obtained in step 4 are used to predict sleep stages, and corresponding instruments are used to measure the ECG signal, respiratory signal and acceleration signal of the remaining subjects, perform signal processing and feature extraction on the data, and then extract the extracted features according to The order in which the model is trained is fed into the classifier, and the result of sleep staging can be output.

Claims (4)

1. an automatic sleep staging method utilizing multiple physiological signals based on LSTM, is characterized in that, comprises the following steps: Step 1: Signal Acquisition Using ECG measuring instrument, respiratory signal measuring instrument and triaxial acceleration sensor to measure the subject's one-lead ECG signal, one-lead breathing signal and three-lead acceleration signal; Step 2: Signal processing; Signal processing is performed on the collected one-lead ECG signal, one-lead breathing signal and three-axis acceleration signal, and the three-lead are three mutually perpendicular directions of space x, y, and z; the final HRV is obtained by processing the one-lead ECG signal signal and respiratory wave amplitude signal R _fm ; the processing of the respiratory signal obtains the final RRV signal; the processing of the three-axis acceleration signal obtains the integrated acceleration signal acc(n); Step 3: Feature extraction; For the four kinds of signals obtained after the original signal processing: HRV, respiratory wave amplitude signal R _fm , RRV, and integrated acceleration signal acc(n), firstly perform segmental processing, with 5 minutes as the window length and 30 seconds as the step The sleep data of the whole night is divided into several data segments of the same length, and then feature extraction is performed on each segment of data, and each feature is normalized. the mean of the series, then divided by the variance of the feature series; Step 4: Model construction; The constructed model adopts two-layer LSTM, and each layer has four different classifiers to realize the staging tasks of two-class, three-class, four-class and five-class; The input sequence of the first layer of LSTM network is the feature sequence extracted in step 3. The input feature data passes through the first layer of LSTM layer, and there are four parallel classifiers in it. The network structure of each classifier is similar, including 5 network layers. : input layer, batch normalization layer, Bi-LSTM layer, forgetting layer, fully connected layer; After the first LSTM layer, each classifier will output data of different dimensions, representing the probability of each category, and combining these probabilities can form a new set of features; this new set of features, with The features extracted in step 2 are connected in parallel as the input of the second layer of LSTM network. There are also four classifiers in parallel in the second layer of LSTM. The structure of each classifier is the same as that of the previous layer. Classification model for prediction; Step 5: Predict sleep stages Use the four classifiers obtained in step 4 to predict sleep stages, perform signal processing and feature extraction on the data to be predicted, and then send the extracted features to the classifier in the order in which the model was trained to output the sleep stages. result. 2. a kind of automatic sleep staging method utilizing multiple physiological signals based on LSTM according to claim 1, is characterized in that, the processing of the respiratory signal described in step 1 is specifically: Filter the original respiratory signal, use a low-pass with a cutoff frequency of 1Hz and a high-pass 3rd-order Butterworth filter with a cutoff frequency of 0.01Hz to filter out the effective components of the respiratory signal 0.01-1Hz, and then extract the effective components from the effective respiratory signal. Extract the RRV signal from , identify the peak position of each respiratory wave according to the maximum slope method, and then obtain the time interval between two adjacent peaks by the difference between the adjacent peak positions, and then use the cubic spline interpolation method to this sequence Interpolate at equal intervals to become the time series at the target sampling frequency. Finally, the RRV signal can be obtained by taking the reciprocal of the sequence. Finally, the RRV is down-sampled and output according to the actual demand. 3. a kind of automatic sleep staging method utilizing multiple physiological signals based on LSTM according to claim 1, is characterized in that, the processing of the acceleration signal described in step 1 is specifically: For the three-axis acceleration signal, calculate the arithmetic square root of the sum of its squares according to formula (1) to obtain the signal acc _temp (n); perform 10-point smoothing filtering on the signal acc _temp (n) to obtain the comprehensive acceleration signal, denoted as acc (n)

Among them, M=3, representing the acceleration signals of the three channels. 4. a kind of automatic sleep staging method utilizing multiple physiological signals based on LSTM according to claim 1, is characterized in that, The features that can be extracted for the respiratory wave amplitude signal R _fm described in step 3 include: the median value of the entire piece of data divided by the sequence range; the median value divided by the quarter quantile and the third quarter quantile interval sequence In the range of the series; the mean value of the entire data divided by the variance; the median value of the respiratory wave peak-peak time interval series; the median value of the respiratory wave peak-trough time interval series; the median value of the respiratory wave trough-peak time interval series; respiration The median value of the peak-peak time interval series divided by the sequence range; the median value of the respiratory wave trough-peak time interval series divided by the sequence range; the median value of the respiratory wave trough-peak time interval series divided by the sequence range; The features that can be extracted for the HRV and RRV described in step 3 include: the mean value of the entire piece of data; the variance of the entire piece of data; the mean value of the entire piece of data divided by the variance; sum; the sum of the power spectrum of the middle frequency band 0.04Hz-0.15Hz of the whole piece of data; the sum of the power spectrum of the whole piece of data high frequency 0.15Hz-0.4Hz; the sum of the total power spectrum of the whole piece of data; the whole piece of data high frequency 0.15Hz -The ratio of the sum of the power spectrum of 0.4Hz to the sum of the power spectrum of the mid-band 0.04Hz-0.15Hz; For the comprehensive acceleration signal acc(n) described in step 3, the extractable features include: the mean value of the entire segment of data; the variance of the entire segment of data; the mean value of the entire segment of data divided by the variance; the median value of the entire segment of data divided by the sequence range ; the median of the series of quartile and quartile intervals divided by the series range; A total of 30 features are listed above. On this basis, the data segment is still selected with a window length of 5 minutes and a step size of 30 seconds, and the extracted feature sequences are arranged in descending order. The value at three-tenths of the position is used as a feature of the segment, and finally, a total of 90 feature parameters are obtained. Due to differences between individuals, each feature needs to be normalized before the model is constructed. The normalization method is to subtract the mean of the feature sequence from each person's features over the night, and then divide by the variance of the feature sequence.

Images