Sleep stage recognition neural network model establishment method and sleep stage recognition method

Abstract

This invention discloses a sleep stage identification method, comprising: a multimodal data acquisition step; a preprocessing step, which normalizes the acquired physiological signals and segments them according to a set duration to obtain several original signal segments; a Williams RGB compression step, which maps the original signals to an RGB image to obtain a Williams RGB energy map; a feature extraction step, which inputs the Williams RGB energy map into a neural network model for feature extraction; and a comparison and classification step, which compares the feature values ​​with RGB standard thresholds to determine the corresponding sleep stage. This sleep stage identification method avoids directly processing the complex frequency domain features of physiological signals, thus reducing computational complexity. Three physiological signals within a certain time period only need to be represented by a single pixel in the image, achieving physiological signal compression and dimensionality reduction, further reducing computational load. This invention also effectively utilizes medical knowledge graphs to provide interpretable evidence for the sleep stage identification decision-making process.

Description

Technical Field

[0001] This invention relates to the field of biosignal processing and intelligent health device technology, specifically to a method for establishing a neural network model for sleep stage recognition and a method for sleep stage recognition. Background Technology

[0002] Sleep monitoring technology is currently widely used in the medical, health, and research fields. Its core function is to collect physiological signals from the human body through various sensors to assess sleep quality and diagnose sleep disorders. Traditional polysomnography (PSG) systems are the gold standard for diagnosing sleep disorders, and they typically include the collection of the following physiological signals: Electroencephalography (EEG): Used to record the brain's electrical activity, reflecting sleep stages (such as light sleep, deep sleep, and REM sleep).

[0003] Electromyography (EMG): Records the electrical activity of muscles, and is commonly used to monitor the activity of muscles in the jaw or limbs.

[0004] Electrooculography (EOG): Used to help determine sleep stages by recording eye movements.

[0005] Heart rate: Monitors heart activity using electrocardiogram (ECG) or photoplethysmography (PPG) signals.

[0006] Body movement: Recording body movements using accelerometers.

[0007] These signals are acquired using high sampling rates (typically above 128 Hz) and high precision (16-bit) analog-to-digital converters (ADCs) to ensure that minute physiological changes can be captured.

[0008] Existing methods mainly employ traditional machine learning or deep learning for sleep analysis. The process involves: signal preprocessing → feature extraction → feature selection → classifier training → sleep stage identification, etc.

[0009] In existing sleep monitoring technologies, field-based data processing is a common method. Field-based data processing involves dividing the collected raw physiological signals into multiple physically meaningful fields, each representing a specific feature or statistic. The following are common steps and data volume calculations in field-based data processing: (1) Sampling and quantization.

[0010] Assume that the sampling rate for each physiological signal is 128 Hz, that is, 128 data points are collected per second, and each data point is 16-bit data, or 2 bytes.

[0011] Taking five physiological signals (EEG, EMG, EOG, heart rate, and body movement) as an example, the amount of raw data per second is: 5 signals × 128Hz × 2 bytes = 1280 bytes = 1.25kB, and the amount of data per minute is: 1.25kB / s × 60s = 75kB.

[0012] (2) Feature extraction.

[0013] Each signal is typically divided into multiple time windows (e.g., one window every 30 seconds), and time-domain, frequency-domain, and nonlinear characteristics are calculated within each window.

[0014] For example, a 30-second window might contain the following features: Time domain: mean, variance, kurtosis, skewness, RMS, etc. Frequency domain: δ, θ, α, σ, β, γ band power and their ratios. Nonlinearity: sample entropy, LZ complexity, DFA α, etc. Assuming each window has 256 features, and each feature is a 4-byte floating-point number, then the data volume of each window is: 256 features × 4 bytes = 1024 bytes = 1kB. There are two windows within one minute, so the feature data volume is: 1kB / window × 2 windows = 2kB.

[0015] (3) Data summary.

[0016] In addition to feature data, it is also necessary to store protocol information such as timestamps, checksums, and frame headers.

[0017] Assuming this additional information occupies approximately 600 bytes per minute (20 bytes / frame BLE header × 30 frames), the total amount of field-based data per minute is: 75kB (raw waveform) + 2kB (features) + 0.6kB (protocol) ≈ 77.6kB ≈ 78kB.

[0018] Considering that compressed spectrograms and other information may also be stored in practical applications, the total data volume usually exceeds 100 kB / minute.

[0019] In summary, the processing flow of field-based PSG multimodal data includes raw signal acquisition, feature extraction, and protocol encapsulation, resulting in a large data volume and placing high demands on storage and transmission. The complex feature extraction and model training processes also lead to high computational complexity.

[0020] Secondly, existing methods require complex feature engineering and large amounts of training data. These methods cannot effectively utilize prior knowledge in sleep medicine and lack mechanisms for integration with professional knowledge bases.

[0021] Furthermore, neural network models have strong black-box characteristics, making it difficult to explain the decision-making process, which reduces the practicality and acceptability of the technology and limits its large-scale application. Summary of the Invention

[0022] To address the problem that existing sleep stage identification methods involve large amounts of data, leading to high complexity in feature extraction and model computation, this invention proposes a sleep stage identification method that can solve the aforementioned problems.

[0023] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for sleep stage identification includes: The multimodal data acquisition process includes the simultaneous acquisition of three physiological signals; The preprocessing step involves normalizing the collected physiological signals and segmenting them according to a set duration to obtain several original signal segments. The Williams RGB compression process includes an energy calculation step and a mapping conversion step, wherein the energy calculation step includes: The original signal of each segment is divided according to the range of gray values ​​to obtain 256 sub-segments of the original signal; Calculate the power of the original signal in each segment to obtain the energy value of the original signal in that segment; The mapping transformation step includes: Multiplying the energy value by 255 yields the grayscale value of the original signal of the sub-segment mapped to the color. The gray values ​​of the same sub-segment corresponding to the three physiological signals are respectively assigned to the RGB three channels of the pixel; The pixels of all sub-segments of the original signal are arranged according to the time axis to obtain a Williams RGB energy map; The feature extraction step involves inputting the Williams RGB energy map into a neural network model for feature extraction, and the neural network model outputs the extracted feature values. In the comparison and classification step, the feature value is compared with the RGB standard threshold to obtain the corresponding sleep stage.

[0024] In some embodiments, the mapping conversion step involves arranging the pixels of all sub-segments of the original signal according to the time axis to obtain an original Williams RGB energy map with a size of 1×256. It also includes a step of cropping the number of columns of the original Williams RGB energy map, cropping from both ends and retaining the middle area. The size of the cropped data is 1×224. The Williams RGB energy map with a data size of 1×224 is copied in the column direction to obtain a Williams RGB energy map of 224×224, and then the map is input into the neural network model for feature extraction.

[0025] In some embodiments, in the feature extraction step, the neural network model calculates the RGB values ​​of each pixel in the Williams RGB energy map, and calculates and outputs the mean of the RGB values ​​respectively; The comparison and classification step uses a clustering algorithm for classification, including calculating the Euclidean distance between the mean of the RGB values ​​and each RGB standard threshold, and comparing the RGB standard threshold with the smallest Euclidean distance. The sleep type corresponding to the RGB standard threshold with the smallest Euclidean distance is the sleep type to which the Williams RGB energy map belongs. Each sleep type corresponds to a set of RGB standard thresholds.

[0026] In some embodiments, after the comparison and classification step, the method further includes a step of providing an interpretable output of the identified sleep stages, including: The confidence level of the sleep type corresponding to the Williams RGB energy map and the RGB standard threshold is calculated based on the Euclidean distance, and then output and displayed together with the identified sleep type. The smaller the Euclidean distance, the higher the confidence level.

[0027] In some embodiments, in the feature extraction step, the neural network model calculates the RGB values ​​of each pixel in the Williams RGB energy map, and calculates and outputs the mean of the RGB values ​​respectively; The collected physiological signals include electroencephalogram (EEG) signals, electromyogram (EMG) signals, and electrooculogram (EOG) signals. The comparison and classification step employs a rapid classification method, including: When the mean value of the channel mapped by the electromyography signal is less than the first electromyography threshold M1 and the mean value of the channel mapped by the electrooculography signal is not less than the first electrooculography threshold N1, the sleep stage is judged to be rapid eye movement sleep. When the mean value of the channel mapped by the electromyography signal is less than the second electromyography threshold M2 and the mean value of the channel mapped by the electrooculography signal is less than the second electrooculography threshold N2, the sleep stage is judged to be deep sleep. When the mean value of the channel mapped by the electromyographic signal is not less than the second electromyographic threshold M2 and less than the third electromyographic threshold M3, the sleep stage is judged to be light sleep. When the mean value of the channel mapped by the electromyographic signal is greater than the third electromyographic threshold M3, the sleep stage is judged to be the waking period. Where M1 < M2 < M3; N1 > N2.

[0028] In some embodiments, the preprocessing step further includes: A Butterworth filter was used to remove 50Hz power frequency interference from the physiological signal; A high-pass filter is used to eliminate baseline drift of the physiological signal.

[0029] In some embodiments, when the physiological signal includes an electroencephalogram (EEG) signal, an alpha wave signal is acquired.

[0030] In some embodiments, when the physiological signal includes electromyography (EMG) signal, a frequency band of 10Hz-200Hz is acquired.

[0031] In some embodiments, when the physiological signal includes electrooculogram (EOG) signals, signals in the 0.1Hz-15Hz frequency band are collected.

[0032] This invention also proposes a sleep stage recognition system, which is based on any of the sleep stage recognition methods described above. The sleep stage recognition system includes: A multimodal data acquisition module, used to simultaneously acquire three physiological signals; The preprocessing module is used to perform preprocessing steps; A compression module, which performs the Williams RGB compression step; A neural network model used to perform the feature extraction step; The classification module is used to perform the comparison classification steps.

[0033] Compared with existing technologies, the advantages and positive effects of this invention are as follows: The sleep stage recognition method of this invention, through the Williams RGB compression step, first divides the three physiological signals into 256 sub-segments according to the grayscale value range, calculates the original signal energy value of each sub-segment, and since the three physiological signals are acquired simultaneously, each sub-segment corresponds to the energy values ​​of the three physiological signals. The energy values ​​of the three physiological signals are multiplied by 255 respectively, that is, the energy value is mapped to the grayscale value of the color. Then, the grayscale values ​​of the same sub-segment corresponding to the three physiological signals are assigned to the RGB three channels of the same pixel, realizing the conversion of the physiological signal into an RGB image representation. This step, on the one hand, avoids directly processing the complex frequency domain features of the physiological signals, reducing computational complexity; on the other hand, the three physiological signals of one sub-segment only need to be represented by one pixel in the image, achieving physiological signal compression and dimensionality reduction, further reducing the amount of computation, and thus reducing the training and computational complexity of the neural network model.

[0034] Secondly, by combining medical expertise, the prior knowledge of sleep medicine is mapped to the characteristics of sleep stages as RGB images. Each sleep stage corresponds to an RGB standard threshold. By comparing the feature values ​​output by the neural network model with the RGB standard threshold in the classification step, the corresponding sleep stage is obtained. On the one hand, this effectively utilizes the medical knowledge graph to provide an interpretable basis for the sleep stage identification decision process. On the other hand, it can further reduce the decision complexity and computational load.

[0035] Other features and advantages of the present invention will become clearer after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. Attached Figure Description

[0036] Figure 1 This is a flowchart of an embodiment of the sleep stage identification method proposed in this invention; Figure 2 This is an embodiment of the sleep stage identification method proposed in this invention, showing the original signal and the Williams RGB energy map obtained by mapping. Figure 3 This is an illustrative mapping of the RGB standard threshold for deep sleep stages in one embodiment of the sleep stage identification method proposed in this invention; Figure 4 This is an illustrative mapping of the RGB standard threshold for the light sleep stage in one embodiment of the sleep stage identification method proposed in this invention; Figure 5 This is an illustrative mapping of the RGB standard threshold for the waking sleep stage in one embodiment of the sleep stage identification method proposed in this invention; Figure 6 This is a graphical mapping of the RGB standard thresholds for REM sleep stages in one embodiment of the sleep stage identification method proposed in this invention. Detailed Implementation

[0037] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1: This example proposes a sleep stage identification method, such as... Figure 1 As shown, it includes: The multimodal data acquisition process includes the simultaneous acquisition of three physiological signals.

[0040] Currently, polysomnography (PSG) is the gold standard for diagnosing sleep disorders, and it typically involves the acquisition of the following physiological signals: Electroencephalography (EEG): Used to record the brain's electrical activity, reflecting sleep stages (such as light sleep, deep sleep, and REM sleep).

[0041] Electromyography (EMG): Records the electrical activity of muscles, and is commonly used to monitor the activity of muscles in the jaw or limbs.

[0042] Electrooculography (EOG): Used to help determine sleep stages by recording eye movements.

[0043] Heart rate: Monitors heart activity using electrocardiogram (ECG) or photoplethysmography (PPG) signals.

[0044] Body movement: Recording body movements using accelerometers.

[0045] These signals are acquired using high sampling rates (typically above 128 Hz) and high precision (16-bit) analog-to-digital converters (ADCs) to ensure that minute physiological changes can be captured.

[0046] In the multimodal data acquisition step of this embodiment, any three of the above physiological signals can be acquired simultaneously.

[0047] The preprocessing step involves normalizing the collected physiological signals and segmenting them according to a set duration to obtain several original signal segments.

[0048] To ensure signal quality, the sampling frequency is uniformly set to 200Hz.

[0049] In some embodiments, sleep is segmented according to a 30-second standard sleep segmentation window.

[0050] Normalization is used to unify signals of different dimensions into the 0-1 range, preparing for the subsequent steps of mapping physiological signals to RGB images.

[0051] The Williams RGB compression process includes an energy calculation step and a mapping conversion step, wherein the energy calculation step includes: The original signal of each segment is divided according to the range of gray values ​​to obtain 256 sub-segments of the original signal.

[0052] Calculate the power of the original signal in each segment to obtain the energy value of the original signal in that segment.

[0053] The mapping transformation steps include: Multiplying the energy value by 255 gives the grayscale value of the original signal in that segment mapped to the color.

[0054] The gray values ​​of the same segment corresponding to the three physiological signals are assigned to the RGB three channels of the pixel respectively.

[0055] The pixels of all sub-segments of the original signal are arranged according to the time axis to obtain a Williams RGB energy map.

[0056] This step achieves an innovative mapping of multimodal physiological signals to the RGB image space, transforming complex physiological signal features into intuitive color image representations, such as... Figure 2 As shown, the left side is a segment of the original signal, which includes three physiological signal characteristics. The right side is the Williams RGB energy map obtained by mapping the original signal on the left side.

[0057] In the feature extraction step, the Williams RGB energy map is input into the neural network model for feature extraction, and the neural network model outputs the extracted feature values.

[0058] The comparison and classification steps involve comparing the feature values ​​with the RGB standard threshold to determine the corresponding sleep stage.

[0059] The neural network model is pre-trained. Its training method involves mapping historical prior physiological signals to obtain Williams RGB energy maps in the manner described above. The Williams RGB energy maps are used as inputs to the neural network model, and the RGB values ​​of the physiological signals are used as outputs for training.

[0060] The sleep stage recognition method in this embodiment, through the Williams RGB compression step, first divides the three physiological signals into 256 sub-segments according to their grayscale value ranges. The energy value of the original signal in each sub-segment is calculated. Since the three physiological signals are acquired simultaneously, each sub-segment corresponds to the energy values ​​of the three physiological signals. These energy values ​​are multiplied by 255, thus mapping the energy values ​​to grayscale values. Then, the grayscale values ​​of the same sub-segment corresponding to the three physiological signals are assigned to the RGB three channels of the same pixel, converting the physiological signals into an RGB image representation. This step reduces computational complexity by avoiding direct processing of the complex frequency domain features of the physiological signals. Furthermore, since the three physiological signals in one sub-segment only need to be represented by a single pixel in the image, physiological signal compression and dimensionality reduction are achieved, further reducing computational load and consequently decreasing the training and computational complexity of the neural network model.

[0061] Secondly, by combining medical expertise, the prior knowledge of sleep medicine is mapped to the characteristics of sleep stages as RGB images. Each sleep stage corresponds to an RGB standard threshold. By comparing the feature values ​​output by the neural network model with the RGB standard threshold in the classification step, the corresponding sleep stage is obtained. On the one hand, this effectively utilizes the medical knowledge graph to provide an interpretable basis for the sleep stage identification decision process. On the other hand, it can further reduce the decision complexity and computational load.

[0062] Taking electroencephalogram (EEG) signals as an example: The raw sampled physiological signal (e.g., EEG = -45.2 μV) cannot be directly used as an RGB value because the RGB value range is 0-255, while the EEG value range is approximately ±100 μV. It needs to undergo Williams RGB compression, which specifically includes energy calculation and mapping conversion.

[0063] Each Williams RGB energy map corresponds to a 30-second segment of the original signal data. The sampling rate of the 30-second signal is 200Hz, meaning that each signal has 6000 sampling points.

[0064] Compressing 6000 time points into 256 pixels results in a compression ratio of 6000 ÷ 256 = 23.44, which corresponds to the signal energy within a time period of (30 ÷ 256 = 0.117 seconds) per pixel.

[0065] Then, the energy of the EEG alpha wave (8-13Hz) was calculated. alpha_energy = power of the filter (eeg_segment, 8-13Hz), the result is an energy value, such as 0.85. This energy value is calculated based on the normalized original signal, is a relative value, and has no unit.

[0066] In the mapping transformation step, the grayscale value of the original signal of this segment mapped to the color is: alpha_energy × 255 = 0.85 × 255 = 217.

[0067] Similarly, the gray values ​​corresponding to the electromyographic (EMG) signal and the electrooculographic (EOG) signal can be obtained.

[0068] The grayscale values ​​of the same segment corresponding to the three physiological signals are assigned to the RGB channels of the pixel, respectively. For example, the grayscale value corresponding to the electroencephalogram (EEG) signal is assigned to the R channel, the grayscale value corresponding to the electromyography (EMG) signal is assigned to the G channel, and the grayscale value corresponding to the electrooculogram (EOG) signal is assigned to the B channel. The combination of this R, G, and B values ​​yields the color of the first pixel. The pixels of all segments of the original signal are arranged along the time axis to obtain a Williams RGB energy map.

[0069] In industry standard practice, the actual size of image convolution is 224×224. During the mapping transformation step, the pixels of all sub-segments of the original signal are arranged along the time axis to obtain an original Williams RGB energy map, which is 1×256 in size. To directly adapt to the input requirements of standard CNN models (such as VGG and ResNet), some embodiments also include a step of cropping the number of columns in the original Williams RGB energy map, cropping from both ends and retaining the middle region. The size of the cropped data is 1×224. This preserves details in the central region, maximizing information retention.

[0070] The Williams RGB energy map with a data size of 1×224 is copied in the column direction to obtain a Williams RGB energy map of 224×224, and then the map is input into the neural network model for feature extraction.

[0071] In some embodiments, during the feature extraction step, the neural network model calculates the RGB values ​​of each pixel in the Williams RGB energy map, and calculates and outputs the mean of the RGB values ​​respectively.

[0072] The collected physiological signals include electroencephalogram (EEG), electromyogram (EMG), and electrooculogram (EOG).

[0073] In some embodiments, the preprocessing step further includes: A Butterworth filter was used to remove 50Hz power frequency interference from physiological signals in order to improve the accuracy of classification.

[0074] A high-pass filter is used to eliminate baseline drift of physiological signals.

[0075] In some embodiments, when the physiological signals include electroencephalogram (EEG) signals, alpha wave signals are collected to reflect the electrical activity of the cerebral cortex.

[0076] In some embodiments, when the physiological signals include electromyographic signals, signals in the 10Hz-200Hz frequency band are collected to monitor muscle tension.

[0077] In some embodiments, when the physiological signals include electrooculogram (EOG) signals, signals in the 0.1Hz-15Hz frequency band are collected to record eye movements.

[0078] In the comparative classification step, based on medical standards, the following four RGB standard thresholds for sleep stages can be given: Table 1. RGB standard thresholds for four sleep stages Sleep stage R-mean (EEG) G-mean (EMG) B-mean (EOG) Total Energy Core features Deep sleep 0.333 0.111 0.056 0.500 Low muscle tone + low eye movement Light sleep 0.333 0.333 0.167 0.833 moderate muscle tone lucidity 0.333 0.778 0.222 1.333 High muscle tone REM sleep 0.333 0.056 0.444 0.833 Lack of muscle tone + hyperocular movement like Figures 3-6 The image shows a graphical mapping of the RGB standard thresholds for the four sleep stages.

[0079] Next, we define a range for each RGB channel to determine which specific sleep stage the input image belongs to: Table 2 RGB Channel Threshold Range

[0080] Based on the above, in some embodiments, in the feature extraction step, the neural network model calculates the RGB values ​​of each pixel in the Williams RGB energy map, and calculates and outputs the mean of the RGB values ​​respectively.

[0081] The comparison and classification steps employ a clustering algorithm for classification, including calculating the Euclidean distance between the mean of RGB values ​​and each RGB standard threshold, and comparing the RGB standard threshold with the smallest Euclidean distance. The sleep type corresponding to the RGB standard threshold with the smallest Euclidean distance is the sleep type to which the Williams RGB energy map belongs.

[0082] Each sleep type corresponds to a set of RGB standard thresholds.

[0083] In some embodiments, after the comparison and classification step, the method further includes an interpretable output step for the identified sleep stages, including: The confidence level of the sleep type corresponding to the Williams RGB energy map and the RGB standard threshold is calculated based on the Euclidean distance and displayed together with the identified sleep type. The smaller the Euclidean distance, the higher the confidence level.

[0084] In some embodiments, the method for Euclidean distance transpose confidence is as follows: confidence = max(0, 1 - distance / 0.5) * 100, where 0.5 is the maximum tolerance distance. Since the smaller the Euclidean distance, the higher the confidence, the confidence of the identified Williams RGB energy map to which the sleep type belongs is the highest, thus realizing the interpretability of the sleep stage identification decision process.

[0085] In some embodiments, the comparison and classification step employs a rapid classification method and further includes: When the mean value of the channel mapped by the electromyography (EMG) signal is less than the first EMG threshold M1 (e.g., 0.15) and the mean value of the channel mapped by the electrooculography (EOG) signal is not less than the first EOG threshold N1 (e.g., 0.35), the sleep stage is determined to be REM sleep. This is defined as G < 0.15 and B ≥ 0.35.

[0086] When the mean value of the channel mapped by the electromyography (EMG) signal is less than the second EMG threshold M2 (e.g., a value of 0.25) and the mean value of the channel mapped by the electrooculography (EOG) signal is less than the second EOG threshold N2 (e.g., a value of 0.10), the sleep stage is determined to be deep sleep. This is equivalent to G < 0.25 and B < 0.10.

[0087] When the mean value of the channel mapped by the electromyographic signal is not less than the second electromyographic threshold M2 and less than the third electromyographic threshold M3 (e.g., a value of 0.55 can be taken), the corresponding sleep stage is judged to be light sleep. According to the above values, that is: 0.25 ≤ G < 0.55.

[0088] When the mean value of the channel mapped by the electromyographic signal is greater than the third electromyographic threshold M3, the corresponding sleep stage is judged as the waking period. This is calculated according to the above value, i.e., G ≥ 0.55.

[0089] Where M1 < M2 < M3, N1 > N2.

[0090] Example 2: This example proposes a sleep stage recognition system, which is based on the sleep stage recognition method described in Example 1. The sleep stage recognition system includes: A multimodal data acquisition module, used to simultaneously acquire three physiological signals; The preprocessing module is used to perform preprocessing steps; A compression module, which performs the Williams RGB compression step; A neural network model used to perform the feature extraction step; The classification module is used to perform the comparison classification steps.

[0091] For details on the sleep stage identification method, please refer to Example 1, which will not be repeated here.

[0092] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for identifying sleep stages, characterized in that, include: The multimodal data acquisition process includes the simultaneous acquisition of three physiological signals; The preprocessing step involves normalizing the collected physiological signals and segmenting them according to a set duration to obtain several original signal segments. The Williams RGB compression process includes an energy calculation step and a mapping conversion step, wherein the energy calculation step includes: The original signal of each segment is divided according to the range of gray values ​​to obtain 256 sub-segments of the original signal; Calculate the power of the original signal in each segment to obtain the energy value of the original signal in that segment; The mapping transformation step includes: Multiplying the energy value by 255 yields the grayscale value of the original signal of the sub-segment mapped to the color. The gray values ​​of the same sub-segment corresponding to the three physiological signals are respectively assigned to the RGB three channels of the pixel; The pixels of all sub-segments of the original signal are arranged according to the time axis to obtain a Williams RGB energy map; The feature extraction step involves inputting the Williams RGB energy map into a neural network model for feature extraction, and the neural network model outputs the extracted feature values. In the comparison and classification step, the feature value is compared with the RGB standard threshold to obtain the corresponding sleep stage.

2. The sleep stage identification method according to claim 1, characterized in that, In the mapping and conversion step, the pixels of all sub-segments of the original signal are arranged according to the time axis to obtain an original Williams RGB energy map with a size of 1×256. It also includes a step of cropping the number of columns of the original Williams RGB energy map, cropping from both ends and retaining the middle area. The size of the cropped data is 1×224. The Williams RGB energy map with a data size of 1×224 is copied in the column direction to obtain a Williams RGB energy map of 224×224, and then the map is input into the neural network model for feature extraction.

3. The sleep stage recognition method according to claim 1, characterized in that, In the feature extraction step, the neural network model calculates the RGB values ​​of each pixel in the Williams RGB energy map, and calculates and outputs the mean of the RGB values ​​respectively. The comparison and classification step uses a clustering algorithm for classification, including calculating the Euclidean distance between the mean of the RGB values ​​and each RGB standard threshold, and comparing the RGB standard threshold with the smallest Euclidean distance. The sleep type corresponding to the RGB standard threshold with the smallest Euclidean distance is the sleep type to which the Williams RGB energy map belongs. Each sleep type corresponds to a set of RGB standard thresholds.

4. The sleep stage identification method according to claim 3, characterized in that, Following the comparison and classification step, the method further includes an interpretable output step for the identified sleep stages, including: The confidence level of the sleep type corresponding to the Williams RGB energy map and the RGB standard threshold is calculated based on the Euclidean distance, and then output and displayed together with the identified sleep type. The smaller the Euclidean distance, the higher the confidence level.

5. The sleep stage identification method according to claim 1, characterized in that, In the feature extraction step, the neural network model calculates the RGB values ​​of each pixel in the Williams RGB energy map, and calculates and outputs the mean of the RGB values ​​respectively. The collected physiological signals include electroencephalogram (EEG) signals, electromyogram (EMG) signals, and electrooculogram (EOG) signals. The comparison and classification step employs a rapid classification method, including: When the mean value of the channel mapped by the electromyography signal is less than the first electromyography threshold M1 and the mean value of the channel mapped by the electrooculography signal is not less than the first electrooculography threshold N1, the sleep stage is judged to be rapid eye movement sleep. When the mean value of the channel mapped by the electromyography signal is less than the second electromyography threshold M2 and the mean value of the channel mapped by the electrooculography signal is less than the second electrooculography threshold N2, the sleep stage is judged to be deep sleep. When the mean value of the channel mapped by the electromyographic signal is not less than the second electromyographic threshold M2 and less than the third electromyographic threshold M3, the sleep stage is judged to be light sleep. When the mean value of the channel mapped by the electromyographic signal is greater than the third electromyographic threshold M3, the sleep stage is judged to be the waking period. Where M1 < M2 < M3; N1 > N2.

6. The sleep stage identification method according to claim 1, characterized in that, The preprocessing step also includes: A Butterworth filter was used to remove 50Hz power frequency interference from the physiological signal; A high-pass filter is used to eliminate baseline drift of the physiological signal.

7. The sleep stage identification method according to any one of claims 1-6, characterized in that, When the physiological signals include electroencephalogram (EEG) signals, alpha wave signals are acquired.

8. The sleep stage identification method according to any one of claims 1-6, characterized in that, When the physiological signal includes electromyography (EMG) signal, signals in the 10Hz-200Hz frequency band are collected.

9. The sleep stage identification method according to any one of claims 1-6, characterized in that, When the physiological signal includes electrooculography (EOG) signal, signals in the 0.1Hz-15Hz frequency band are collected.

10. A sleep stage recognition system, characterized in that, It is based on the sleep stage recognition method according to any one of claims 1-9, wherein the sleep stage recognition system comprises: A multimodal data acquisition module, used to simultaneously acquire three physiological signals; The preprocessing module is used to perform preprocessing steps; A compression module, which performs the Williams RGB compression step; A neural network model used to perform the feature extraction step; The classification module is used to perform the comparison classification steps.

Images