CN113080864B - Common sleep disease detection method through automatic sleep staging results - Google Patents

Abstract

The invention discloses a common sleep disease detection method through automatic sleep staging results, which comprises the following steps: s1, acquiring sleep data sets of healthy people, patients with nocturnal frontal lobe epilepsy and patients with obstructive sleep apnea; s2, feature extraction; s3, constructing a sleep automatic staging model by combining time attention and a conditional random field; s4, staging the patient data set by adopting transfer learning; s5, constructing a sleep disease detection model: taking the prediction result of the sleep stage as input data, marking a label of a corresponding body state, and making a sleep disease detection data set; and then training through a machine learning Xgboost model to obtain a sleep disease detection model. The method adopts the conditional random field and the time attention model, and can effectively extract the time continuity information of the sleep data; introducing transfer learning to transfer the sleep disease data set to a network of the healthy person data set; the data volume is small, and various sleep disease detections can be completed.

Description

Common sleep disease detection method through automatic sleep staging results

Technical Field

The invention belongs to the technical field of sleep detection, and particularly relates to a common sleep disease detection method based on automatic sleep staging results.

Background

Sleep is one of the most important circadian rhythms in human physiological activity. Sleep quality can affect the performance of many basic activities such as learning, memory and concentration. At present, more and more people suffer from sleep disorders and sleep diseases. Common sleep disorders are nocturnal epilepsy and obstructive apnea syndrome. Symptoms of seizures at night have a significant impact on sleep architecture, manifested as decreased sleep efficiency and the number of REM sleeps, convulsions in the limbs, loss of consciousness, etc. Patients with obstructive apnea have hypomnesis and inability to concentrate, resulting in restlessness, dreaminess, enuresis, impotence, morning headache, etc. Severely persistent patients may be complicated by hypertension, arrhythmia, exhaustion of heart and lung functions, etc. Sleep disorders seriously compromise the health of people. The most important part of the sleep quality assessment is the classification of sleep stages, i.e. the classification of different sleep sessions into periods WAKE, REM, N1, N2 and N3, for five normal classes. Sleep staging aids in the diagnosis of sleep-related disorders. In the traditional sleep staging method, the subject must record a polysomnogram by wearing a polysomnography. Sleep experts classify sleep stages by monitoring signals, which is time consuming, labor intensive and susceptible to subjective influences by sleep experts. Researchers have begun to continually research automated sleep staging methods. The research algorithms of the automatic sleep staging method are roughly divided into two categories, namely traditional machine learning and deep learning by applying an artificial neural network.

Early researchers performed automatic sleep staging by extracting features in conjunction with machine learning. The most common methods of machine learning classification include decision trees, random forests, and support vector machines. Fraiwan is classified by extracting time-frequency characteristics and entropy characteristics and combining a random forest classifier. And (5) extracting time-frequency characteristics by Zhu, and classifying by using a support vector machine. Hassan uses wavelet transforms from tunable Q-factors and random forest classifiers for classification. However, the algorithm combining feature extraction and traditional machine learning generally has the defects of low accuracy, unsuitability for large-scale training samples, low recognition rate in the N1 stage, neglected signal, time continuity of labels and the like, and is low in practicability.

With the development of artificial neural networks, deep learning is gradually popular in the sleep stage field, and is a new research direction in the machine learning field, and forms more abstract high-level representation attribute categories or features by combining low-level features to find distributed feature representation of data. In recent years, deep learning has achieved many results in the field of sleep staging. In 2018, the large-scale multi-channel data set is used by Olesen, a CNN method of 50-layer convolution is adopted, the accuracy rate of 84% is obtained, but the number of CNN layers is too many, the consumed computing resources are too much, and the engineering is not easy to realize. In 2019, by using a 19-layer 1D-CNN model, the accuracy rate of Yildirim and the like is over 90 percent, and the algorithm is suitable for people who normally fall asleep and people who have difficulty falling asleep. However, CNN extracts less time information, has a large number of layers, and is highly jumpy, ignoring temporal continuity. In 2020, Wei Qu et al extract features through CNN, combine attention mechanism and residual neural network to perform sleep staging, obtain an accuracy rate of over 84%, and solve the problem of excessive time consumption of deep learning due to too many network layers. But its accuracy is low and the network structure is complex.

After the automatic sleep staging result, namely the sleep state, a sleep doctor can judge common sleep diseases such as night epilepsy, obstructive sleep apnea and the like according to the sleep state. However, the current research on automatic sleep disorder detection by computers is mainly directed to single sleep disorders. If Zhao extracts relevant features in EEG, judging whether the tested object has epilepsy through a convolution network. Zarei uses ECG to detect whether the subject suffers from obstructive sleep apnea disease by wavelet transformation and extraction of entropy features. However, there is no study for directly judging various sleep disorders by the result of sleep staging.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a common sleep disease detection method which adopts a conditional random field and a time attention model, can effectively extract time continuity information of sleep data, has small data volume and can finish detection of various sleep diseases and passes through an automatic sleep staging result.

The purpose of the invention is realized by the following technical scheme: a method for detecting common sleep disorders by automatic sleep staging results, comprising the steps of:

s1, data set acquisition: acquiring a sleep dataset of a healthy person, a sleep dataset of a patient with nocturnal frontal lobe epilepsy, a sleep dataset of a patient with obstructive sleep apnea;

s2, feature extraction: the data obtained in the step S1 is segmented, the data is divided into a plurality of data segments with the same length by taking 30 seconds as a step length, and then feature extraction is performed on each data segment;

s3, constructing a sleep automatic staging model by combining time attention and a conditional random field;

s4, staging the patient data set by adopting transfer learning;

s5, constructing a sleep disease detection model: taking the prediction result of the sleep stage as input data, marking a label of a corresponding body state, and making a sleep disease detection data set; and then training through a machine learning Xgboost model to obtain a sleep disease detection model.

Further, the specific implementation method of step S1 is as follows: the data set of the healthy person is a polysomnogram PSG collected by a sleep device, and comprises 5 channels, namely 2 EEG channels, 2 EOG channels and 1 EMG channel;

a Sleep data set for a patient with nocturnal frontal lobe epilepsy was acquired from CAP Sleep Database of Physionet, including 2 EEG channels, 2 EOG channels and 1 EMG channel;

a sleep data set with obstructive sleep apnea was obtained from the sanhengst university hospital/university of dublin school sleep apnea database of Physionet, including 2 EEG channels, 2 EOG channels, 1 EMG channel.

Further, the specific implementation method of step S2 is as follows:

s21, labeling a data set of healthy people, namely a WAKE period, a rapid eye movement period REM, a first stage N1 of a non-rapid eye movement period, a second stage N2 of the non-rapid eye movement period and a third stage N3 of the non-rapid eye movement period;

s22, integrating and labeling the sleep data set of the patient with the nocturnal frontal lobe epilepsy;

s23, integrating and labeling the sleep data set of the patient with obstructive sleep apnea;

s24, dividing the sleep data into a plurality of data sections with the same length by taking 30 seconds as a step length, and corresponding to the labels one by one;

s25, extracting features according to the physiological signals of each channel, and extracting the following features from the sleep data of every 30S: time domain characteristic quantity, frequency domain characteristic quantity and nonlinear dynamics characteristic quantity;

the time domain characteristic quantity comprises statistical characteristic quantity and geometric characteristic quantity; the frequency domain characteristic quantity comprises a power spectral density characteristic quantity and a time frequency characteristic quantity (such as statistical kurtosis, skewness and Hjorth parameters); nonlinear dynamics feature quantities (e.g., LZ complexity) include fractal dimension feature quantities and complexity feature quantities;

part of the feature formula is calculated as follows:

skewness: a measure of the symmetry of a data set with respect to its mean value, the formula:

wherein μ is a mean value and x is a sleep data sequence;

kurtosis: the fourth order standard moment is expressed as:

wherein μ is the mean;

zero crossing rate: refers to the rate at which the sign of a signal changes;

hjorth parameter: including as time-series square mobility Hm and complexity Hc, representing the power spectrum standard deviation ratio and frequency variation:

wherein var is the variance of the sleep data sequence x (t) shifted over time t, Hm is the time-series square mobility;

LZ complexity: the method is a method for representing the rate of new mode appearing in a time sequence, and c (n) is set as LZ complexity of a sequence S and meets the following requirements: when n → ∞ is reached, c (n) tends to be constant n/log_Ln and L are the number of coarse grained sections; the normalized LZ complexity is:

further, the specific implementation method of step S3 is as follows: inputting the extracted feature vector into a Bi-GRU to extract time sequence features, giving high weight to a time sequence with high correlation degree through a Sigmoid function, multiplying the time sequence with the original input feature vector, and inputting an obtained result into an FC to obtain the input of a conditional random field CRF, wherein the formula is Attention (X, X) which is Sigmoid (GRU (X)) X, and X is the input feature vector; inputting a predicted label obtained by a time attention mechanism into a CRF, modeling by using a CRF linear chain method, and decoding an optimal label sequence path by using a Viterbi algorithm; and modifying all data and label sequences to make the data and the label sequences continuously coincide with the label sequences of the expert artificial sleep staging judgment results, and obtaining the sleep automatic staging model when the data and the label sequences are completely coincident.

Further, the specific implementation method of step S4 is as follows:

s41, taking the sleep data of the healthy people as a source domain, training a sleep automatic staging model as a basic model, and storing the model and the weight parameters;

s42, taking the sleep data set of the patient suffering from the nocturnal frontal lobe epilepsy as a target domain, sending the sleep data set into a model of a source domain, taking the weight parameter of the source domain as an initial parameter of target domain training for fine adjustment, and storing the fine-adjusted model and parameter with the best result;

s43, finally, taking the sleep data set of the patient with obstructive sleep apnea as a second target domain, taking the model and the parameters stored in the step S42 as the initial model and the parameters of the target domain, fine-tuning again, and storing the fine-tuned model and parameters with the best results;

s44, respectively inputting the sleep data set of healthy people, the sleep data set of the patient suffering from the frontal lobe epilepsy at night and the sleep data set of the patient suffering from the obstructive sleep apnea into the model trained in the step S43 for classification, and obtaining the sleep staging test results of three different types of data sets.

The invention has the beneficial effects that: the invention adopts a conditional random field to extract information between sleep states; by adopting the time attention model, the time continuity information of the sleep data can be effectively extracted; introducing transfer learning to transfer the sleep disease data set to a network of the healthy person data set; the sleep staging results obtained by the model are used as input, the data volume is small, and the detection of various sleep diseases can be completed.

Drawings

FIG. 1 is an overall block diagram of an automatic sleep staging algorithm;

FIG. 2 is a block diagram of a temporal attention model;

FIG. 3 is a schematic diagram of a Bi-GRU network model;

FIG. 4 is a GRU node internal structure;

FIG. 5 is a schematic illustration of transfer learning;

fig. 6 is an overall block diagram of common sleep disorder detection.

Detailed Description

The technical scheme of the invention is further explained by combining the drawings and the specific embodiment.

As shown in fig. 1, the method for detecting common sleep disorders by automatic sleep staging according to the present invention comprises the following steps:

s1, data set acquisition: acquiring a sleep dataset of a healthy person, a sleep dataset of a patient with nocturnal frontal lobe epilepsy, a sleep dataset of a patient with obstructive sleep apnea;

the specific implementation method comprises the following steps: the method for acquiring the water surface data set of the healthy person comprises the following steps: in this embodiment, 17 healthy people are recruited as subjects, 103 sleep polysomnography PSGs at night are collected through sleep devices, and the channels are 7 channels, namely three EEG channels, two EOG channels, one EMG channel and one ECG channel; to maintain consistency with the other two data sets, this embodiment only uses data for 2 EEG channels, 2 EOG channels, 1 EMG channel.

A Sleep data set for a patient with nocturnal frontal lobe epilepsy was obtained from CAP Sleep Database of Physionet. The CPA sleep database is a collection of 108 polysomnography records that have been registered at the Ospedale Maggiore centers of Palma, Italy. Including 3 EEG channels, 2 EOG channels and 1 EMG channel, as well as bilateral tibialis anterior electrograms, respiratory signals and electrocardiograms. Of these, 16 healthy subjects, 40 patients diagnosed with nocturnal frontal lobe epilepsy, 22 REM behavioral disorders, 10 with periodic leg movements, 9 insomnia, 5 lethargy, 4 sleep disordered breathing and 2 with bruxism. Due to the size of the data set, the present example only used sleep data sets of 40 patients diagnosed with nocturnal frontal lobe epilepsy, and the channel selection was consistent with the data set of healthy people, including 2 EEG channels, 2 EOG channels and 1 EMG channel.

A sleep data set with obstructive sleep apnea was obtained from the sanwensen university hospital/university of dublin school sleep apnea database of Physionet, and a total of 25 subjects were diagnosed with obstructive sleep apnea. The recorded signals include: electroencephalograms (C3-a2), electroencephalograms (C4-a1), left eye electrograms, right eye electrograms, EMG, ECG, oronasal airflow (thermistors), chest movements, abdominal motor charts, oxygen saturation, and body position. The data selected in this embodiment includes 2 EEG channels, 2 EOG channels, and 1 EMG channel.

S2, feature extraction: the data obtained in the step S1 is segmented, the data is divided into a plurality of data segments with the same length by taking 30 seconds as a step length, and then feature extraction is performed on each data segment; the specific implementation method comprises the following steps:

s21, labeling a data set of healthy people, namely a WAKE period, a rapid eye movement period REM, a first stage N1 of a non-rapid eye movement period, a second stage N2 of the non-rapid eye movement period and a third stage N3 of the non-rapid eye movement period;

s22, integrating and labeling the sleep data set of the patient with the nocturnal frontal lobe epilepsy;

s23, integrating and labeling the sleep data set of the patient with obstructive sleep apnea;

s24, dividing the sleep data into a plurality of data sections with the same length by taking 30 seconds as a step length, and corresponding to the labels one by one;

s25, extracting features according to the physiological signals of each channel, and extracting the following features from the sleep data of every 30S: time domain characteristic quantity, frequency domain characteristic quantity and nonlinear dynamics characteristic quantity;

the time domain characteristic quantity comprises statistical characteristic quantity and geometric characteristic quantity; the frequency domain characteristic quantity comprises a power spectral density characteristic quantity and a time frequency characteristic quantity (such as statistical kurtosis, skewness and Hjorth parameters); nonlinear dynamics feature quantities (e.g., LZ complexity) include fractal dimension feature quantities and complexity feature quantities;

part of the feature formula is calculated as follows:

skewness: a measure of the symmetry of a data set with respect to its mean value, the formula:

wherein μ is a mean value and x is a sleep data sequence;

kurtosis: the fourth order standard moment is expressed as:

wherein μ is the mean;

zero crossing rate: refers to the rate at which the sign of a signal changes, e.g., the signal changes from positive to negative, and is expressed as: ZCR ═ count (n | (x)_nx_n-1)<0)；

Hjorth parameter: including as time-series square mobility Hm and complexity Hc, representing the power spectrum standard deviation ratio and frequency variation:

wherein var is the variance of the sleep data sequence x (t) shifted over time t, Hm is the time-series square mobility;

LZ complexity: the method is a method for representing the rate of new modes appearing in a time sequence, and the Lempel-Ziv complexity of the electroencephalogram signals reflects the information content of the electroencephalogram signals and can reveal the relevant rules of brain activities. Let c (n) be the LZ complexity of the sequence S, which satisfies: when n → ∞ is reached, c (n) tends to be constant n/log_Ln and L are the number of coarse grained segments (when the binarization is carried out conventionally, L is 2); the normalized LZ complexity is:

s3, constructing a sleep automatic staging model by combining time attention and a conditional random field; a block diagram of the attention model is shown in fig. 2. Inputting the extracted feature vector into a Bi-GRU shown in fig. 3 and 4 to extract time sequence features, giving high weight to a time sequence with high correlation degree through a Sigmoid function, multiplying the time sequence with the original input feature vector, and inputting an obtained result into an FC to obtain the input of a conditional random field CRF, wherein the formula is Attention (X, X) ═ Sigmoid (GRU (X)) X, and X is the input feature vector; inputting a predicted label obtained by a time attention mechanism into a CRF, modeling by using a CRF linear chain method, and decoding an optimal label sequence path by using a Viterbi algorithm; and modifying all data and label sequences to make the data and the label sequences continuously coincide with the label sequences of the expert artificial sleep staging judgment results, and obtaining the sleep automatic staging model when the data and the label sequences are completely coincident.

S4, staging the patient data set by adopting transfer learning; the transfer learning method shown in fig. 5 is introduced, and the specific implementation method is as follows:

s41, taking the sleep data of the healthy people as a source domain, training a sleep automatic staging model as a basic model, and storing the model and the weight parameters;

s42, taking the sleep data set of the patient suffering from the nocturnal frontal lobe epilepsy as a target domain, sending the sleep data set into a model of a source domain, taking the weight parameter of the source domain as an initial parameter of target domain training for fine adjustment, and storing the fine-adjusted model and parameter with the best result;

s43, finally, taking the sleep data set of the patient with obstructive sleep apnea as a second target domain, taking the model and the parameters stored in the step S42 as the initial model and the parameters of the target domain, fine-tuning again, and storing the fine-tuned model and parameters with the best results;

s44, respectively inputting the sleep data set of healthy people, the sleep data set of the patient suffering from the frontal lobe epilepsy at night and the sleep data set of the patient suffering from the obstructive sleep apnea into the model trained in the step S43 for classification, and obtaining the sleep staging test results of three different types of data sets.

In the above steps S41 to S43, the five-fold cross validation method is adopted in all the model training experiments.

S5, constructing a sleep disease detection model: as shown in fig. 6, the prediction result of sleep stages is used as input data, labels of corresponding body states are marked, and a sleep disease detection data set is made, namely, a healthy person is 0, a person with frontal lobe epilepsy at night is 1, and a person with obstructive sleep apnea is 2; and then training through a machine learning Xgboost model to obtain a sleep disease detection model, wherein the Xgboost searches for optimal parameters by adopting a grid search method.

It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.

Claims (4)

1. A method for generating a common sleep disease detection model through an automatic sleep staging result is characterized by comprising the following steps:

s1, data set acquisition: acquiring a sleep dataset of a healthy person, a sleep dataset of a patient with nocturnal frontal lobe epilepsy, a sleep dataset of a patient with obstructive sleep apnea;

s2, feature extraction: the data obtained in the step S1 is segmented, the data is divided into a plurality of data segments with the same length by taking 30 seconds as a step length, and then feature extraction is performed on each data segment;

s3, constructing a sleep automatic staging model by combining time attention and a conditional random field; the specific implementation method comprises the following steps: inputting the extracted feature vector into a Bi-GRU to extract time sequence features, giving high weight to a time sequence with high correlation degree through a Sigmoid function, multiplying the time sequence with the original input feature vector, and inputting an obtained result into an FC to obtain the input of a conditional random field CRF, wherein the formula is Attention (X, X) which is Sigmoid (GRU (X)) X, and X is the input feature vector; inputting a predicted label obtained by a time attention mechanism into a CRF, modeling by using a CRF linear chain method, and decoding an optimal label sequence path by using a Viterbi algorithm; modifying all data and label sequences to make the data and the label sequences continuously coincide with the label sequences of the expert artificial sleep staging judgment results, and obtaining a sleep automatic staging model when the data and the label sequences are completely coincident;

s4, staging the patient data set by adopting transfer learning;

s5, constructing a sleep disease detection model: taking the prediction result of the sleep stage as input data, marking a label of a corresponding body state, and making a sleep disease detection data set; and then training through a machine learning Xgboost model to obtain a sleep disease detection model.

2. The method for generating a common sleep disorder detection model through an automatic sleep staging result as claimed in claim 1, wherein the step S1 is implemented by: the data set of the healthy person is a polysomnogram PSG collected by a sleep device, and comprises 5 channels, namely 2 EEG channels, 2 EOG channels and 1 EMG channel;

a Sleep data set for a patient with nocturnal frontal lobe epilepsy was acquired from CAP Sleep Database of Physionet, including 2 EEG channels, 2 EOG channels and 1 EMG channel;

a sleep data set with obstructive sleep apnea was obtained from the sanhengst university hospital/university of dublin school sleep apnea database of Physionet, including 2 EEG channels, 2 EOG channels, 1 EMG channel.

3. The method for generating a common sleep disorder detection model through an automatic sleep staging result as claimed in claim 1, wherein the step S2 is implemented by:

s21, labeling a data set of healthy people, namely a WAKE period, a rapid eye movement period REM, a first stage N1 of a non-rapid eye movement period, a second stage N2 of the non-rapid eye movement period and a third stage N3 of the non-rapid eye movement period;

s22, integrating and labeling the sleep data set of the patient with the nocturnal frontal lobe epilepsy;

s23, integrating and labeling the sleep data set of the patient with obstructive sleep apnea;

s24, dividing the sleep data into a plurality of data sections with the same length by taking 30 seconds as a step length, and corresponding to the labels one by one;

s25, extracting features according to the physiological signals of each channel, and extracting the following features from the sleep data of every 30S: time domain feature quantity, frequency domain feature quantity and nonlinear dynamics feature quantity.

4. The method for generating a common sleep disorder detection model through an automatic sleep staging result as claimed in claim 1, wherein the step S4 is implemented by:

s41, taking the sleep data of the healthy people as a source domain, training a sleep automatic staging model as a basic model, and storing the model and the weight parameters;

s42, taking the sleep data set of the patient suffering from the nocturnal frontal lobe epilepsy as a target domain, sending the sleep data set into a model of a source domain, taking the weight parameter of the source domain as an initial parameter of target domain training for fine adjustment, and storing the fine-adjusted model and parameter with the best result;

s43, finally, taking the sleep data set of the patient with obstructive sleep apnea as a second target domain, taking the model and the parameters stored in the step S42 as the initial model and the parameters of the target domain, fine-tuning again, and storing the fine-tuned model and parameters with the best results;

s44, respectively inputting the sleep data set of healthy people, the sleep data set of the patient suffering from the frontal lobe epilepsy at night and the sleep data set of the patient suffering from the obstructive sleep apnea into the model trained in the step S43 for classification, and obtaining the sleep staging test results of three different types of data sets.

Images