A sleep stage staging method based on time series signal convolution and multi-signal fusion

Abstract

The present application relates to a kind of sleep stage staging method based on time series signal convolution and multi-signal fusion, comprising: obtaining the physiological signal data of the sleep state of subject, including: 4 channels of electroencephalogram EEG, 1 channel of electrocardiogram ECG, 2 channels of eye movement signal EOG, 1 channel of electromyogram EMG;The physiological signal data of the sleep state of subject is input to the monitoring model trained, obtains classification result;The monitoring model after training includes: for the one-dimensional convolution of each channel in the input physiological signal data is carried out separately, to extract the time series signal convolution network of the time domain waveform variation feature of the signal of each channel;For extracting the spatial functional connection network between any two channels in physiological signal data;For the classification of physiological signal data according to the time domain waveform variation feature of each channel signal in physiological signal data and the relationship between any two channels to obtain the fully connected network of classification result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sleep stage detection, in particular to a sleep stage staging method based on time series signal convolution and multi-signal fusion. BACKGROUND

[0002] Humans spend about one-third of their time in sleep every day, and sleep is an important physiological process in people's daily life. When sleeping, the whole body is in a low activity state, and all physiological activities including metabolism in the body are at a low level. At the beginning of sleep, blood pressure is basically unchanged, in the middle of the sleep stage, blood pressure is significantly reduced, and at the end of the sleep stage, blood pressure returns to normal level. Insufficient sleep will directly affect people's mental state, easily causing inattention, endocrine disorders and other conditions, affecting normal life, and sleep stage staging is of great significance for understanding human sleep patterns and improving human health.

[0003] Sleep stage staging is an important process in the disease diagnosis process of sleep medical research. The traditional sleep staging method is mainly manually staged by medical workers according to sleep period physiological signal data, which is time-consuming and laborious, and the staging results have subjective differences. The current automatic staging system uses K-means clustering, expert system, hidden Markov model, artificial neural network, random forest, support vector machine, spectral analysis, nonlinear feature analysis and other means. However, these methods first extract frequency domain and time-frequency domain features from each record sample period. In the case of multiple signals, the features from all signals in a period are connected into a feature vector. These features are used to train a classifier to identify the sleep stage staging of a single sample. These methods have the problems of complex detection process, insufficient feature extraction, and low recognition accuracy. SUMMARY

[0004] (I) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present application provides a sleep stage staging method based on time series signal convolution and multi-signal fusion, which solves the technical problems of insufficient use of physiological signals, incomplete manual feature extraction, and low recognition accuracy in the prior art.

[0006] (II) Technical solutions

[0007] In order to achieve the above-mentioned purpose, the main technical scheme adopted by the present application includes:

[0008] The embodiment of the present application provides a sleep stage staging method based on time series signal convolution and multi-signal fusion, comprising:

[0009] S10, acquiring physiological signal data of a subject in a sleep state;

[0010] The physiological signal data comprises: 4 channels of electroencephalogram (EEG), 1 channel of electrocardiogram (ECG), 2 channels of electrooculogram (EOG), and 1 channel of electromyogram (EMG);

[0011] S20, inputting the physiological signal data of the subject in the sleep state into the trained monitoring model to obtain a classification result;

[0012] The trained monitoring model comprises:

[0013] a one-dimensional convolution network for performing one-dimensional convolution on each channel of the input physiological signal data to extract time-domain waveform change features of each channel of the signal;

[0014] a spatial functional connection network for extracting the relationship between any two channels of the physiological signal data;

[0015] a full connection network for classifying the physiological signal data according to the time-domain waveform change features of each channel of the signal and the relationship between any two channels to obtain a classification result.

[0016] Preferably, the monitoring network is trained in advance using training data to obtain the trained monitoring model.

[0017] Preferably, S10 specifically comprises:

[0018] S01, obtaining initial training data based on pre-obtained physiological signal data of multiple subjects in a sleep state for training;

[0019] S02, processing the initial training data using a band-pass filter to obtain training data;

[0020] S03, training the monitoring model using the training data to obtain a trained monitoring model.

[0021] Preferably, S01 specifically comprises:

[0022] S01-1, setting corresponding sleep stage labels for the pre-obtained physiological signal data of multiple subjects in a sleep state for training;

[0023] The sleep stage labels comprise: a wake period label, a sleep stage I label, a sleep stage II label, a sleep stage III label, and a rapid eye movement period label.

[0024] S01-2, dividing the physiological signal data of the subjects in a sleep state for training into multiple data segments of 30s in length, and each data segment corresponds to a sleep stage label to obtain initial training data.

[0025] Preferably,

[0026] The band-pass filter is a Butterworth band-pass filter.

[0027] The lower limit frequency of the passband of the band-pass filter is 1 Hz, the upper limit frequency of the passband is 45 Hz, and the sampling frequency is 500 Hz.

[0028] Preferably,

[0029] The time series signal convolution network sequentially comprises a first convolution layer, a second convolution layer, a maximum pooling layer, a third convolution layer, a fourth convolution layer, and a fifth convolution layer, and a BatchNorm1d function is arranged after each convolution layer to standardize the data output by the convolution layer, and any two adjacent convolution layers are connected by a ReLU activation function.

[0030] The spatial functional connection network extracts the relationship between each channel in the physiological signal data through formula (1).

[0031] The formula (1) is:

[0032] ;

[0033] Wherein, A represents the adjacency matrix of the relationship between any channel p m and another channel p n in the physiological signal data.

[0034] ReLU is an activation function.

[0035] exp is an exponential function with base e.

[0036] w is a weight vector of the spatial functional connection network.

[0037] Wherein, ;

[0038] T is a vector transpose.

[0039] N is the total number of channels in the physiological signal data.

[0040] Preferably,

[0041] In the time series signal convolution network, the operation of the subsequent layer of any two adjacent convolution layers is:

[0042] ;

[0043] Wherein, M e RB×N×L ;

[0044] parameters representing a convolution kernel;

[0045] represents a standard convolution operation;

[0046] is a convolution kernel;

[0047] is a graph convolution operation;

[0048] M （l） is an operation result of a post-convolution layer of any two adjacent convolution layers;

[0049] is an operation result of a pre-convolution layer of any two adjacent convolution layers;

[0050] R represents a set of rational numbers;

[0051] B represents a batch quantity;

[0052] L represents a length of a time sequence signal.

[0053] Preferably,

[0054] In the time sequence signal convolution network, the first convolution layer has: an input neuron node quantity of 8, an output neuron node quantity of 200, a convolution kernel size of 100, a step length of 10, and a padding of 5;

[0055] The second convolution layer has: an input neuron node quantity of 200, an output neuron node quantity of 200, a convolution kernel size of 100, a step length of 10, and a padding of 5;

[0056] The convolution kernel size of the maximum pooling layer is 5, and the step length is 5;

[0057] The third convolution layer has: an input neuron node quantity of 200, an output neuron node quantity of 300, a convolution kernel size of 10, and a step length of 1;

[0058] The fourth convolution layer has: an input neuron node quantity of 300, an output neuron node quantity of 300, a convolution kernel size of 10, and a step length of 1;

[0059] The fifth convolution layer has: an input neuron node quantity of 300, an output neuron node quantity of 200, a convolution kernel size of 5, and a step length of 1.

[0060] Preferably,

[0061] The full connection network comprises: two linear transformation layers;

[0062] The input neuron node number of the first linear transformation layer of the full connection network is 1500, and the output neuron node number is 200;

[0063] The input neuron node number of the second linear transformation layer of the full connection network is 200, and the output neuron node number is 5.

[0064] Preferably,

[0065] The trained monitoring model is a monitoring model corresponding to a multi-classification cross-entropy loss function of the trained monitoring model when the multi-classification cross-entropy loss function converges;

[0066] The multi-classification cross-entropy loss function is:

[0067] ;

[0068] Wherein, p i is a distribution of probabilities of a sleep stage label output by the monitoring model;

[0069] Wherein, p i = { p 0 , p 1 ,..., p Z－1} ;

[0070] y i is a true probability distribution of a sleep stage label corresponding to the training data;

[0071] Wherein, y i = { y 0 , y 1 ,..., y Z－1} ;

[0072] Z is the number of labels corresponding to the training data.

[0073] Preferably,

[0074] The monitoring model is further provided with an optimizer;

[0075] The optimizer is an SGD optimizer or an Adagrad optimizer.

[0076] (Three) beneficial effects

[0077] The beneficial effects of the present application are: the sleep stage staging method based on time sequence signal convolution and multi-signal fusion of the present application, because the physiological signals adopted are electroencephalogram (EEG), electrocardiogram (ECG), eye movement signal (EOG) and electromyogram (EMG), the limitation of single electroencephalogram is broken through by multi-signal fusion, the characteristics of human body during sleep are more comprehensive, and the accuracy of sleep stage staging is higher. Meanwhile, the sleep stage staging is performed by using the trained monitoring model, so that the problem of insufficient manual feature extraction is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 The flow chart of the sleep stage staging method based on time sequence signal convolution and multi-signal fusion of the present application is shown in the figure.

[0079] Figure 2 The schematic diagram for training the monitoring model in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0080] In order to better explain the present application and facilitate understanding, the present application will be described in detail in combination with the drawings and specific embodiments.

[0081] In order to better understand the above technical solutions, the exemplary embodiments of the present application will be described in more detail with reference to the drawings. Although the exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer, more thorough understanding of the present application and to convey the complete scope of the present application to those skilled in the art.

[0082] Referring to Figure 1 The present embodiment provides a sleep stage staging method based on time sequence signal convolution and multi-signal fusion, comprising:

[0083] S10, acquiring physiological signal data of a subject in a sleep state.

[0084] The physiological signal data includes 4-channel electroencephalogram (EEG), 1-channel electrocardiogram (ECG), 2-channel eye movement signal (EOG) and 1-channel electromyogram (EMG).

[0085] S20, inputting the physiological signal data of the subject in the sleep state into a trained monitoring model to obtain a classification result.

[0086] The trained monitoring model comprises:

[0087] a time sequence signal convolution network for performing one-dimensional convolution on each channel of the input physiological signal data separately to extract time domain waveform change characteristics of the signals of each channel.

[0088] a spatial functional connectivity network for extracting the relationship between any two channels in the physiological signal data.

[0089] a full connection network for classifying the physiological signal data according to the time domain waveform variation characteristics of each channel signal in the physiological signal data and the relationship between any two channels to obtain a classification result.

[0090] wherein the monitoring network is trained in advance using training data to obtain a trained monitoring model.

[0091] Referring to Figure 2 In the actual application of the embodiment, before S10, specifically includes:

[0092] S01, based on the physiological signal data of a plurality of subjects in sleep state obtained in advance for training, obtaining initial training data.

[0093] In the embodiment, the PSG device is used to collect four channel electroencephalogram (EEG) signals, one channel electrocardiogram (ECG) signal, two channel eye movement (EOG) signals, and one channel electromyogram (EMG) signal of the subject, a total of eight channels, as the physiological signal data. Each 30s of the physiological signal data is divided into a data segment, and the label of the corresponding data segment is given. There are 154 subjects, and each subject has 800-900 data segments. The numbers 0-4 are used to represent the sleep stages respectively: "Sleep stage Wake", "Sleep stage N1", "Sleep stage N2", "Sleep stage N3", "Sleep stage R", and the labels are aligned with the data segments.

[0094] In the embodiment, the sleep stage label is converted into a discrete signal sequence , the expression of the label sequence is:

[0095]

[0096] The nth number in the discrete signal sequence is denoted as , and the value is determined by the following formula:

[0097] ;

[0098] wherein, The sleep stage labels are "Sleep stage W", "Sleep stage N1", "Sleep stage N2", "Sleep stage N3" and "Sleep stage R", which represent five sleep stages: wake period, sleep stage I, sleep stage II, sleep stage III and rapid eye movement period, respectively.

[0099] Referring to Figure 2 , S02, processing (i.e., data processing) the initial training data using a band-pass filter to obtain training data.

[0100] Specifically, in the embodiment, a Butterworth low-pass filter is used to process the initial training data to remove high-frequency noise in the initial training data.

[0101] S03, training the monitoring model using the training data to obtain a trained monitoring model.

[0102] In the practical application of the embodiment, S01 specifically includes:

[0103] S01-1, setting corresponding sleep stage labels for the physiological signal data of the sleep state of the plurality of subjects for training obtained in advance.

[0104] The sleep stage labels include wake period labels, sleep stage I labels, sleep stage II labels, sleep stage III labels and rapid eye movement period labels.

[0105] S01-2, dividing the physiological signal data of the sleep state of the subjects for training into a plurality of data segments each having a length of 30s, and each data segment corresponding to a sleep stage label, to obtain initial training data.

[0106] In the practical application of the embodiment, the band-pass filter is a Butterworth band-pass filter.

[0107] The lower limit frequency of the passband of the band-pass filter is 1Hz, the upper limit frequency of the passband is 45Hz, and the sampling frequency is 500Hz.

[0108] The Butterworth band-pass filter in the embodiment has the characteristic that the frequency curve on the passband is maximally flat without fluctuations, and the gain gradually decreases to 0 on the stop band, which is suitable for processing the physiological signal data collected by the sensor in the application.

[0109] The time series signal convolution network sequentially comprises a first convolution layer, a second convolution layer, a maximum pooling layer, a third convolution layer, a fourth convolution layer and a fifth convolution layer, and a BatchNorm1d function is arranged after each convolution layer to normalize the data output by the convolution layer, and any two adjacent convolution layers are connected by a ReLU activation function.

[0110] Specifically, the time series signal convolution network input data and output data are three-dimensional tensors, and the data structure is (B, N, L), wherein B represents the batch processing quantity; N is the total number of channels in the physiological signal data; and L represents the time series signal length.

[0111] In order to reduce the influence of input values with different distributions on the learning effect of the time series signal convolution network, a BatchNorm1d function is added after each convolution layer to normalize the data, so as to accelerate the training speed of the model and relieve the gradient vanishing problem in training. In this embodiment, the data is normalized in two adjacent convolution layers to make the output data have a mean value of 0 and a variance of 1, so as to accelerate the training, accelerate the convergence speed and stability.

[0112] The spatial functional connection network extracts the relationship between each channel in the physiological signal data through formula (1).

[0113] The formula (1) is:

[0114] ;

[0115] wherein, represents the adjacency matrix of the relationship between any channel p m and another channel p n in the physiological signal data.

[0116] ReLU is an activation function.

[0117] exp is an exponential function with base e.

[0118] w is a weight vector of the spatial functional connection network.

[0119] wherein, .

[0120] T is a vector transpose.

[0121] N is the total number of channels in the physiological signal data.

[0122] In practical application of the embodiment, in the time series signal convolution network, the operation of the next layer of any two adjacent convolution layers is:

[0123] ;

[0124] wherein, M e R B×N×L .

[0125] represents a parameter of a convolution kernel.

[0126] represents a standard convolution operation.

[0127] is a convolution kernel.

[0128] is a graph convolution operation.

[0129] M （l） is an operation result of the next layer of any two adjacent convolution layers.

[0130] is an operation result of the previous convolution layer of any two adjacent convolution layers.

[0131] R represents a set of rational numbers.

[0132] B represents a batch quantity.

[0133] L represents a time series signal length.

[0134] In practical application of the embodiment, in the time series signal convolution network, the first convolution layer has: 8 input channels, 200 output channels, a convolution kernel size of 100, a step of 10, and a padding of 5.

[0135] The second convolution layer has: 200 input channels, 200 output channels, a convolution kernel size of 100, a step of 10, and a padding of 5.

[0136] The convolution kernel size of the maximum pooling layer is 5, and the step is 5.

[0137] The third convolution layer has: 200 input channels, 300 output channels, a convolution kernel size of 10, and a step of 1.

[0138] The fourth convolution layer has: 300 input channels, 300 output channels, a convolution kernel size of 10, and a step of 1.

[0139] The fifth convolution layer has: 300 input channels, 200 output channels, a convolution kernel size of 5, and a step of 1.

[0140] In the practical application of the embodiment, the fully connected network comprises two linear transformation layers.

[0141] The first linear transformation layer of the fully connected network has 1500 input neuron nodes and 200 output neuron nodes.

[0142] The second linear transformation layer of the fully connected network has 200 input neuron nodes and 5 output neuron nodes.

[0143] In the practical application of the embodiment, the trained monitoring model is the monitoring model when the multi-classification cross-entropy loss function corresponding to the monitoring model trained by the training data converges.

[0144] The multi-classification cross-entropy loss function is:

[0145] .

[0146] wherein, p i is the distribution of the probability of the sleep stage label output by the monitoring model.

[0147] wherein, p i p 0 , p 1 , p Z－1 .

[0148] y i is the true probability distribution of the sleep stage label corresponding to the training data.

[0149] wherein, y i y 0 , y 1 , y Z－1 .

[0150] Z is the number of labels corresponding to the training data.

[0151] In the practical application of the embodiment, the monitoring model is further provided with an optimizer.

[0152] The optimizer is an SGD optimizer or an Adagrad optimizer. The optimizer in the embodiment is used to update the network parameters after the error backpropagation, so as to accelerate the model training speed and reduce the iteration times.

[0153] ​​The sleep stage staging method based on time sequence signal convolution and multi-signal fusion in the embodiment has the physiological signals of electroencephalogram (EEG), electrocardiogram (ECG), eye movement signal (EOG) and electromyogram (EMG), so that the limitation of single electroencephalogram is broken through, the features of human sleep are more comprehensive, and the accuracy of sleep stage staging is higher. Meanwhile, the sleep stage staging is performed by using the trained monitoring model, so that the problem of insufficient manual feature extraction is avoided.

[0154] Since the system described in the above embodiment of the present application is the system used for implementing the method of the above embodiment of the present application, the specific structure and deformation of the system can be understood by those skilled in the art based on the method described in the above embodiment of the present application, and thus will not be described here. Any system used by the method of the above embodiment of the present application belongs to the scope of the present application.

[0155] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0156] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions.

[0157] It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in the claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the claims, the word "comprising" does not exclude the presence of other elements or steps than those listed in the claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The word "first", "second", "third", etc. does not imply any order. The use of these words is to be construed as being the use of these words in their "ordinal" sense and not to be construed as use of these words as "indicia" to signify different or successive stages or steps in a process.

[0158] Moreover, it is to be understood that the description of the present application set forth herein is illustrative of the present application and is not intended to limit the scope of the present application as defined in the following claims. Various modifications of the preferred embodiment as described herein, will be apparent to those with ordinary skill in the art and can be made without departing from the spirit and scope of the application.

[0159] Although the preferred embodiment of the application has been described, those skilled in the art will be able to make modifications and alterations to this preferred embodiment without departing from the spirit and scope of the application. Accordingly, it is intended that the scope of the application be governed only by the following claims and their equivalents.

[0160] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A sleep stage staging method based on convolution of time series signal and multi-signal fusion, characterized in that, The method comprises: S10, acquiring physiological signal data of a subject in a sleep state; The physiological signal data comprises: 4 channels of electroencephalogram (EEG), 1 channel of electrocardiogram (ECG), 2 channels of eye movement signal (EOG), and 1 channel of electromyogram (EMG); S20, inputting the physiological signal data of the subject in the sleep state into a trained monitoring model to obtain a classification result; The trained monitoring model comprises: a time sequence signal convolution network for performing one-dimensional convolution on each channel of the input physiological signal data separately to extract time domain waveform change features of each channel of the signal; a spatial functional connection network for extracting the relationship between any two channels of the physiological signal data; a full connection network for classifying the physiological signal data according to the time domain waveform change features of each channel of the signal and the relationship between any two channels to obtain a classification result; wherein the monitoring network is trained in advance using training data to obtain the trained monitoring model; The time sequence signal convolution network comprises, in sequence: a first convolution layer, a second convolution layer, a maximum pooling layer, a third convolution layer, a fourth convolution layer, and a fifth convolution layer, and a BatchNorm1d function is arranged after each convolution layer to standardize the data output by the convolution layer, and any two adjacent convolution layers are connected by a ReLU activation function; The spatial functional connection network extracts the relationship between each channel of the physiological signal data by formula (1); The formula (1) is: where K mn = g(p m , p n ) denotes the adjacency matrix of the relationship between any channel p m and another channel p n in the physiological signal data; ReLU is an activation function; exp is an exponential function with e as the base; w is a weight vector of the spatial functional connection network; where w = {w1, w2, w3,..., wn} ; and N} ; T is a vector transpose; N is the total number of channels in the physiological signal data.

2. The method of claim 1, wherein, Before S10, specifically comprising: S01, acquiring initial training data based on pre-acquired physiological signal data of multiple subjects in a sleep state for training; S02, processing the initial training data using a band-pass filter to obtain training data; S03, training the monitoring model using the training data to obtain a trained monitoring model.

3. The method of claim 2, wherein, The S01 specifically comprises: S01-1, setting corresponding sleep stage labels for the pre-acquired physiological signal data of multiple subjects in a sleep state for training; The sleep stage labels comprise: a wake period label, a sleep stage I label, a sleep stage II label, a sleep stage III label, and a rapid eye movement period label; S01-2, dividing the physiological signal data of the subjects in the sleep state for training into multiple data segments with a length of 30s, and each data segment corresponds to a sleep stage label to obtain initial training data.

4. The method of claim 3, wherein: The band-pass filter is a Butterworth band-pass filter; The lower limit frequency of the passband of the band-pass filter is 1 Hz, the upper limit frequency of the passband is 45 Hz, and the sampling frequency is 500 Hz.

5. The method of claim 4, wherein: In the time sequence signal convolution network, the operation of the latter layer of any two adjacent convolution layers is: where M e R B×N×L ; Φ represents the parameters of the convolution kernel. * denotes a standard convolution operation; g θ is a convolution kernel; * G denotes a graph convolution operation; M (l) is the operation result of the latter convolutional layer of any two adjacent convolutional layers; is the operation result of the previous convolution layer of any two adjacent convolution layers; R denotes a set of rational numbers; B denotes a batch size; L denotes a length of a time series signal.

6. The method of claim 5, wherein, in the time series convolution network, a first convolution layer has: an input neuron node number of 8, an output neuron node number of 200, a convolution kernel size of 100, a stride of 10, and a padding of 5; a second convolution layer has: an input neuron node number of 200, an output neuron node number of 200, a convolution kernel size of 100, a stride of 10, and a padding of 5; a max pooling layer has a convolution kernel size of 5 and a stride of 5; a third convolution layer has: an input neuron node number of 200, an output neuron node number of 300, a convolution kernel size of 10, a stride of 1; a fourth convolution layer has: an input neuron node number of 300, an output neuron node number of 300, a convolution kernel size of 10, a stride of 1; a fifth convolution layer has: an input neuron node number of 300, an output neuron node number of 200, a convolution kernel size of 5, a stride of 1.

7. The method of claim 6, wherein, the fully connected network comprises: two linear transformation layers; a first linear transformation layer of the fully connected network has an input neuron node number of 1500 and an output neuron node number of 200; a second linear transformation layer of the fully connected network has an input neuron node number of 200 and an output neuron node number of 5.

8. The method of claim 7, wherein, the trained monitoring model is a monitoring model corresponding to a multi-classification cross-entropy loss function of the trained monitoring model converging when the trained monitoring model is trained by using the training data; the multi-classification cross-entropy loss function is: where p i is a distribution of probabilities of the corresponding sleep stage labels output by the monitoring model; where p i = {p0, p1,..., p Z－1} ; y i a true probability distribution of sleep stage labels corresponding to the training data; wherein y i = {y0, y1,..., y Z－1}; Z denotes a number of labels corresponding to the training data.

9. The method of claim 8, wherein, the monitoring model is further provided with an optimizer; the optimizer is an SGD optimizer or an Adagrad optimizer.

Images