Non-contact non-inductive sleep stage monitoring method based on deep learning
Through deep learning technology, the motion signals are extracted from sleep videos, and the optical flow method and LSTM neural network are used for classification, which solves the problems of contact discomfort and strong subjectivity of existing sleep monitoring devices, and realizes contactless sleep staging Testing improves the reliability of testing and the diagnostic efficiency of sleep quality.
Patent Information
- Application Number
- CN202410834597.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-26
- Publication Date
- 2024-11-12
AI Technical Summary
Existing sleep monitoring equipment has problems such as contact discomfort, strong subjectivity of manual interpretation, and sensitivity to environmental factors, and is costly.
Using a contactless sleep stage monitoring method based on deep learning, the image pyramid is constructed to extract motion signals from sleep videos, and the optical flow method and LSTM neural network are used to classify eye movements and mandibular movements, combining the judgment of different periods of sleep. Design a sleep staging algorithm.
It realizes high comfort, automation and objectivity sleep stage detection, improves the reliability of detection and the clinical diagnosis efficiency of sleep quality, and plays an important role in early intervention of sleep disorders.
Smart Images

Figure CN118942735A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the fields of biomedical engineering and computer vision, and in particular relates to a non-contact sleep staging detection system realized by combining video image processing technology with sleep staging. Background Art
[0002] During sleep, a series of functions of the human brain, muscles, and eyes will change. These changes will vary with the depth of sleep. Monitoring these changes can effectively determine sleep stages. Sleep staging can reflect an individual's health status, sleep quality, and potential sleep disorders. The deep sleep stage is essential for body and brain repair. Lack of deep sleep can affect memory consolidation, immune function, and physical recovery. In addition, the structure of sleep is closely related to the next day's cognitive function, including attention, memory, and problem-solving ability. Abnormal sleep staging patterns may indicate the presence of sleep disorders, such as sleep apnea, insomnia, periodic limb movement disorder, etc.
[0003] Polysomnography is known as the golden method for determining sleep stages. It is a manual visual scoring of electrographic measurements of the brain, eye movements, and jaw muscles. Electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) are used for measurement. This sensor combination, together with sensors for measuring heart and respiratory activity, is collectively referred to as polysomnography (PSG). When using polysomnography for monitoring, it is necessary to install a variety of sensors on the subject, which brings great discomfort to the subject. In addition, even if it monitors many parameters, doctors will still use the subject's medical history and subjective feelings during the monitoring period as one of the basis for evaluation, and the interpretation results are highly subjective. With the development of machine learning, many miniaturized sleep monitoring devices have emerged, such as smart pillows, mattresses, and bracelets. Smart pillows and mattresses use pressure sensors to monitor pressure changes during sleep and analyze the sleep stages of the subject. Smart bracelets record the subject's movements during sleep through built-in sensors such as accelerometers and gyroscopes to determine their sleep state. Although existing sleep monitoring equipment reduces the discomfort of wearing sensors during the monitoring process, its performance may be affected by environmental factors and its cost is relatively high.
[0004] In view of the above problems in sleep monitoring, a non-contact and non-sensing sleep staging monitoring method based on deep learning is designed. For sleep staging, the required motion signals are extracted from the sleep video by constructing an image pyramid and amplified, and then reconstructed to obtain the amplified video. Based on the amplified video, the optical flow method is used to obtain the optical flow curve of eye movement and the optical flow amplitude of mandibular movement. The K-means clustering algorithm and LSTM long short-term memory neural network are used to classify mandibular movement and eye movement, and combined with the judgment of different sleep periods, the sleep staging algorithm is designed to finally obtain the sleep staging results. Summary of the invention
[0005] In view of the discomfort caused by contact monitoring of polysomnography to the subjects and the subjectivity of manual interpretation, as well as the problems existing in the monitoring process of other sleep monitoring equipment, the present invention designs a non-contact and sensorless sleep staging monitoring method based on deep learning to achieve non-contact monitoring of sleep staging.
[0006] The technical solution of the present invention is a non-contact and non-sensing sleep staging monitoring method based on deep learning, which comprises the following steps:
[0007] Step 1: Build a sleep monitoring platform;
[0008] Three cameras are placed on the upper, left and right sides of the test subject's body to obtain video images of the test subject during sleep;
[0009] Step 2: Segment the video image obtained by the camera located above the tester's body in step 1 to obtain a video image of the tester's jaw area and a video image of the eye area;
[0010] Step 3: Build an image pyramid, perform multi-scale analysis on the video frames, amplify specific motion signals, and finally reconstruct the video to visualize these amplified changes;
[0011] Step 4: Input the jaw region video and eye region video obtained in step 2 into the algorithm framework constructed in step 3, respectively, extract and amplify the motion signals of the jaw region and the eye region, and output the jaw region video image after amplifying the motion signal and the eye region video image after amplifying the eye movement frequency;
[0012] Step 5: The dense optical flow method is used to obtain the cumulative optical flow values of the jaw region and the eye region after the enlarged movement of the jaw region video and the eye region video obtained in step 4;
[0013] Step 6: Smooth and detrend the accumulated optical flow values obtained in step 5;
[0014] Step 7: Accumulate the amplitude of the optical flow of the video image of the mandibular area obtained in step 6 as the monitoring result of the mandibular area;
[0015] Step 8: Accumulate the optical flow curve of the eye area video image obtained in step 6 as the monitoring result of eye movement;
[0016] Step 9: Classify the eye movement optical flow curves according to the corresponding time based on the electrooculogram data collected synchronously by multi-lead sleep breathing monitoring when shooting the video, and divide them into rapid eye movement and non-rapid eye movement;
[0017] Step 10: Build an eye movement classification model based on LSTM, use the data in step 9 to train the model, divide the training set and the validation set to train the network until the accuracy of the validation set reaches the set threshold, and the network training is completed;
[0018] Step 11: Build a deep learning-based sleep staging monitoring algorithm framework, and implement sleep staging based on medical sleep staging standards by combining eye movement and jaw movement.
[0019] Wherein, the step 2 is specifically as follows:
[0020] Step 2.1: Segment the face area of the video image acquired by the camera located above the tester's body and extract the tester's face video image;
[0021] Step 2.2: Detect facial key points on the facial video image obtained in step 2.1 to obtain the locations of 68 key points on the tester's face;
[0022] Step 2.3: Segment the video image obtained in step 2.1 again using the facial key point positions obtained in step 2.2;
[0023] Step 2.4: Find the lower boundary of the face, the position of the mouth, and the left and right boundaries of the mandible through the position of the key points of the face, and segment the mandible area video image into a rectangular area;
[0024] Step 2.4: By identifying the key points of the left and right eyes of the face, find the key points representing the left eye corner, the right eye corner, the uppermost part of the eye socket and the lowermost part of the eye socket, and segment the eye area video image into a rectangular area through the above four key points.
[0025] Wherein, the step 3 is specifically as follows:
[0026] Step 3.1: Construct a 4-layer Laplacian pyramid for each frame of the video. Each layer of the pyramid represents the representation of the original image at a different scale, and the resolution of each layer is half of the previous layer.
[0027] Step 3.2: Apply bandpass filtering to each layer of the pyramid in order to amplify only changes within a specific frequency range (0.4-3Hz);
[0028] Step 3.3: Multiply the filtered signal by the set amplification factor (alpha) to increase the visibility of these small movements;
[0029] Step 3.4: Add the amplified motion back to the original frames, reconstruct each frame using the modified pyramid, and combine the reconstructed frames into a new video file containing the amplified small motion changes.
[0030] Wherein, the step 6 is specifically as follows:
[0031] Step 6.1: Initialize an empty list to store the smoothed data, and traverse each data point of the accumulated optical flow;
[0032] Step 6.2: Calculate the new smoothed value by multiplying the previous smoothed data point by the smoothing factor factor and adding the result obtained by multiplying the current data point by (1-factor). Add the smoothed value to the list initialized in step 6.1 and traverse to obtain the weighted moving average data.
[0033] Step 6.3: Use Savitzky-Golay filter to further smooth the weighted moving average data obtained in step 6.2;
[0034] Step 6.4: For the data obtained in step 6.3, use a quadratic polynomial to fit the data and time series to obtain a quadratic equation that best fits the data;
[0035] Step 6.5: Calculate the value of the trend line using the coefficients of the polynomial and the time series. Subtract the value of the trend line from the data obtained in step 6.3 to obtain the detrended data.
[0036] Wherein, the step 10 is specifically as follows:
[0037] Step 10.1: Build a two-layer LSTM network, each layer outputs 24-dimensional features;
[0038] Step 10.2: Add two fully connected layers after the LSTM layer. The first fully connected layer contains 4800 nodes and the second fully connected layer contains 32 nodes.
[0039] Step 10.3: Use the Softmax output layer to predict the class label of the sample, i.e., REM vs. NREM.
[0040] Wherein, the step 11 is specifically as follows:
[0041] Step 11.1: Use K-means clustering algorithm to classify the mandibular optical flow amplitude into three categories: wakefulness, sleep, and rapid eye movement;
[0042] Step 11.2: Use K-means clustering algorithm to classify the eye movement optical flow amplitude into two categories: wakefulness and sleep;
[0043] Step 11.3: Add the awakening part in step 11.1 to the classified data in step 11.2 according to the corresponding time to obtain two-category sleep data: awakening and sleeping;
[0044] Step 11.4: Further classify the sleep data in step 11.3, and check the sleep data from the beginning in chronological order. If the "sleep" state is detected for the first time to last for more than 5 minutes, the first "sleep" point of the continuous "sleep" state is regarded as the critical point of sleep, called the "sleep point";
[0045] Step 11.5: Reversely check the sleep data in step 11.3. If the first detected "sleep" state lasts for more than 5 minutes, it will be regarded as the last period of "sleep", and the last "sleep" point of the continuous "sleep" state is the "awakening point";
[0046] Step 11.6: Modify the data before the "sleep point" and after the "wake-up point" in step 11.3 to the "wake-up" state;
[0047] Step 11.7: In the data obtained in step 11.6, the eye movement classification model trained in step 10 is used to further classify the "sleep" state part to obtain data with three categories: wakefulness, rapid eye movement, and non-rapid eye movement;
[0048] Step 11.8: Since the duration of the REM period can generally last from 5 to 40 minutes, when the interval between two "REM" states is less than 5 minutes, it is determined that the states during this period are all "REM" states;
[0049] Step 11.9: After step 11.8, three categories of sleep state data are obtained: wakefulness, non-rapid eye movement, and rapid eye movement, and all categories are arranged in time series to obtain the final sleep stage.
[0050] The present invention is a non-contact and non-sensing sleep staging system based on deep learning. The system first segments the acquired video images into video images of the mandibular area and video images of the eye area, then constructs an image pyramid, performs Euler video amplification on specific movements of the mandibular area and the eye area, calculates the mandibular movement optical flow amplitude and the cumulative optical flow of eye movements using the optical flow method, builds an eye movement classification model based on the LSTM long short-term memory neural network to achieve two classifications of eye movements: rapid eye movement and rapid no eye movement, and then builds a sleep staging algorithm based on the optical flow method. First, sleep is classified into three categories based on mandibular optical flow, and then classified into two categories based on eye movement optical flow. The "awakening" part of the three categories is merged into the two categories. Then, the "sleep point" and "awakening point" are determined based on the two categories, and the sleep periods before the "sleep point" and after the "awakening point" are divided into the "awakening" state. The trained eye movement classification model is then used to classify the eye movements in the "sleep" state. The results are further analyzed based on the characteristics of different sleep periods. Finally, all classifications are arranged in time series to obtain the sleep stage situation for the whole night. The present invention provides a highly comfortable and highly automated sleep monitoring system for testers, which realizes the detection of sleep stages in a non-contact and imperceptible manner, improves the reliability and objectivity of the detection, and plays a key role in the clinical diagnosis of sleep quality and the clinical treatment and early intervention of patients with sleep disorders. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 This is the EVM video amplification flow chart
[0052] Figure 2 This is the eye movement optical flow data flow chart
[0053] Figure 3 Mandibular movement monitoring flow chart
[0054] Figure 4 Eye movement classification monitoring flow chart
[0055] Figure 5 Sleep stage flow chart DETAILED DESCRIPTION
[0056] In conjunction with the accompanying drawings, a non-contact, sensorless sleep staging monitoring method based on deep learning of the present invention is described in detail below:
[0057] Step 1: Build a sleep monitoring platform. Place three cameras on the upper, left, and right sides of the tester's body to obtain video images of the tester during sleep;
[0058] Step 2: Segment the video image obtained by the camera located above the tester's body in step 1 to obtain a video image of the tester's jaw area and a video image of the eye area;
[0059] Step 2.1: Segment the face area of the video image acquired by the camera located above the tester's body and extract the tester's face video image;
[0060] Step 2.2: Detect facial key points on the facial video image obtained in step 2.1 to obtain the locations of 68 key points on the tester's face;
[0061] Step 2.3: The video image obtained in step 2.1 is segmented again using the facial key point positions obtained in step 2.2.
[0062] Step 2.4: Find the lower boundary of the face, the position of the mouth, and the left and right boundaries of the mandible through the position of the key points of the face, and segment the mandible area video image into a rectangular area;
[0063] Step 2.4: By identifying the key points of the left and right eyes of the face, find the key points representing the left eye corner, the right eye corner, the uppermost part of the eye socket and the lowermost part of the eye socket, and segment the eye area video image into a rectangular area through the above four key points.
[0064] Step 3: Build an image pyramid, perform multi-scale analysis on the video frames, amplify specific motion signals, and finally reconstruct the video to visualize these amplified changes;
[0065] Step 3.1: Construct a 4-layer Laplacian pyramid for each frame of the video. Each layer of the pyramid represents the representation of the original image at a different scale, and the resolution of each layer is half of the previous layer.
[0066] Step 3.2: Apply bandpass filtering to each layer of the pyramid in order to amplify only changes within a specific frequency range (0.4-3Hz);
[0067] Step 3.3: Multiply the filtered signal by an amplification factor (alpha=1) to increase the visibility of these small movements;
[0068] Step 3.4: Add the amplified motion back to the original frames, reconstruct each frame using the modified pyramid, and combine the reconstructed frames into a new video file containing the amplified small motion changes;
[0069] Step 4: Input the jaw region video and eye region video obtained in step 2 into the algorithm framework constructed in step 3, respectively, extract and amplify the motion signals of the jaw region and the eye region, and output the jaw region video image after amplifying the motion signal and the eye region video image after amplifying the eye movement frequency;
[0070] Step 5: The dense optical flow method is used to obtain the cumulative optical flow values of the jaw region and the eye region after the enlarged movement of the jaw region video and the eye region video obtained in step 4;
[0071] Step 6: Smooth and detrend the accumulated optical flow values obtained in step 5;
[0072] Step 6.1: Initialize an empty list to store the smoothed data, and traverse each data point of the accumulated optical flow;
[0073] Step 6.2: Calculate the new smoothed value by multiplying the previous smoothed data point by the smoothing factor (factor=0.8) and adding the result of multiplying the current data point by (1-factor). Add the smoothed value to the list in 6.1 and obtain the weighted moving average data.
[0074] Step 6.3: Use Savitzky-Golay filter to further smooth the weighted moving average data obtained in step 6.2;
[0075] Step 6.4: For the data obtained in step 6.3, use a quadratic polynomial to fit the data and time series to obtain a quadratic equation that best fits the data;
[0076] Step 6.5: Calculate the value of the trend line using the coefficients of the polynomial and the time series. Subtract the value of the trend line from the data obtained in step 6.3 to obtain the detrended data.
[0077] Step 7: Accumulate the amplitude of the optical flow of the video image of the mandibular area obtained in step 6 as the monitoring result of the mandibular area;
[0078] Step 8: Accumulate the optical flow curve of the eye area video image obtained in step 6 as the monitoring result of eye movement;
[0079] Step 9: Classify the eye movement optical flow curves according to the corresponding time based on the electrooculogram data collected synchronously by multi-lead sleep breathing monitoring when shooting the video, and divide them into rapid eye movement and non-rapid eye movement;
[0080] Step 9.1: Separately read the electrooculographic data of the left and right eyes collected synchronously using multi-lead sleep breathing monitoring;
[0081] Step 9.2: Based on the definition of rapid eye movement, videos with conjugate, irregular, sharp-peaked eye movements and initial deflection peak time < 500ms were classified as rapid eye movement;
[0082] Step 9.3: The videos that are still unclassified after the classification in step 9.2 are further classified, and the videos without rapid eye movement and during sleep are classified as non-rapid eye movement.
[0083] Step 10: Build an eye movement classification model based on LSTM, use the data in step 9 to train the model, divide the training set and the validation set in a ratio of 8:2 to train the network until the accuracy of the validation set reaches more than 95%, and the network training is completed;
[0084] Step 10.1: Build a two-layer LSTM network, each layer outputs 24-dimensional features;
[0085] Step 10.2: Add two fully connected layers after the LSTM layer. The first fully connected layer contains 4800 nodes and the second fully connected layer contains 32 nodes.
[0086] Step 10.3: Use the Softmax output layer to predict the class label of the sample, i.e., REM vs. NREM.
[0087] Step 11: Build a deep learning-based sleep staging monitoring algorithm framework, and implement sleep staging based on medical sleep staging standards, combined with eye movement and jaw movement;
[0088] Step 11.1: During the REM period, the EMG tension is lower than that in any other period, so the K-means clustering algorithm is used to classify the mandibular optic flow amplitude into three categories: wakefulness, sleep, and REM;
[0089] Step 11.2: Based on the sharp peak of the eye movement wave during the rapid eye movement period, the K-means clustering algorithm is used to classify the eye movement optical flow amplitude into two categories: wakefulness and sleep;
[0090] Step 11.3: Add the awakening part in step 11.1 to the classified data in step 11.2 according to the corresponding time to obtain new binary sleep data (awake and sleep);
[0091] Step 11.4: Further classify the sleep data in step 11.3, and check the sleep data from the beginning in chronological order. If the "sleep" state is detected for the first time to last for more than 5 minutes, the first "sleep" point of the continuous "sleep" state is regarded as the critical point of sleep, called the "sleep point";
[0092] Step 11.5: Reversely check the sleep data in step 11.3. If the first detected "sleep" state lasts for more than 5 minutes, it will be regarded as the last period of "sleep", and the last "sleep" point of the continuous "sleep" state is the "awakening point";
[0093] Step 11.6: modify the data before the "sleep point" and after the "wake-up point" in the data obtained in step 11.3 to the "wake-up" state;
[0094] Step 11.7: In the data obtained in step 11.6, the "sleep" part in the binary data is further classified. The eye movement classification model trained in step 10 is used, and the eye movement monitoring results in step 8 corresponding to the "sleep" state part are used as input. Then, the "sleep" part can be further classified into "rapid eye movement" and "non-rapid eye movement" to obtain data with three classifications;
[0095] Step 11.8: Since the duration of the REM period can generally last from 5 to 40 minutes, when the interval between two "REM" states in the data of step 11.7 is less than 5 minutes, it is determined that the states during this period are all "REM" states;
[0096] Step 11.9: After step 11.8, three categories of sleep state data are obtained: wakefulness, non-rapid eye movement, and rapid eye movement. All categories are arranged in time series to obtain the final sleep stage.
Claims
1. A non-contact, sensorless sleep staging monitoring method based on deep learning, the method comprising the following steps: Step 1: Build a sleep monitoring platform; Three cameras are placed on the upper, left and right sides of the test subject's body to obtain video images of the test subject during sleep; Step 2: Segment the video image obtained by the camera located above the tester's body in step 1 to obtain a video image of the tester's jaw area and a video image of the eye area; Step 3: Build an image pyramid, perform multi-scale analysis on the video frames, amplify specific motion signals, and finally reconstruct the video to visualize these amplified changes; Step 4: Input the jaw region video and eye region video obtained in step 2 into the algorithm framework constructed in step 3, respectively, extract and amplify the motion signals of the jaw region and the eye region, and output the jaw region video image after amplifying the motion signal and the eye region video image after amplifying the eye movement frequency; Step 5: The dense optical flow method is used to obtain the cumulative optical flow values of the jaw region and the eye region after the enlarged movement of the jaw region video and the eye region video obtained in step 4; Step 6: Smooth and detrend the accumulated optical flow values obtained in step 5; Step 7: Accumulate the amplitude of the optical flow of the video image of the mandibular area obtained in step 6 as the monitoring result of the mandibular area; Step 8: Accumulate the optical flow curve of the eye area video image obtained in step 6 as the monitoring result of eye movement; Step 9: Classify the eye movement optical flow curves according to the corresponding time based on the electrooculogram data collected synchronously by multi-lead sleep breathing monitoring when shooting the video, and divide them into rapid eye movement and non-rapid eye movement; Step 10: Build an eye movement classification model based on LSTM, use the data in step 9 to train the model, divide the training set and the validation set to train the network until the accuracy of the validation set reaches the set threshold, and the network training is completed; Step 11: Build a deep learning-based sleep staging monitoring algorithm framework, and implement sleep staging based on medical sleep staging standards by combining eye movement and jaw movement.
2. The non-contact, sensorless sleep staging monitoring method based on deep learning as claimed in claim 1, characterized in that: The step 2 is specifically as follows: Step 2.1: Segment the face area of the video image acquired by the camera located above the tester's body and extract the tester's face video image; Step 2.2: Detect facial key points on the facial video image obtained in step 2.1 to obtain the locations of 68 key points on the tester's face; Step 2.3: Segment the video image obtained in step 2.1 again using the facial key point positions obtained in step 2.2; Step 2.4: Find the lower boundary of the face, the position of the mouth, and the left and right boundaries of the mandible through the position of the key points of the face, and segment the mandible area video image into a rectangular area; Step 2.4: By identifying the key points of the left and right eyes of the face, find the key points representing the left eye corner, the right eye corner, the uppermost part of the eye socket and the lowermost part of the eye socket, and segment the eye area video image into a rectangular area through the above four key points.
3. The non-contact, sensorless sleep staging monitoring method based on deep learning as claimed in claim 1, characterized in that: The step 3 is specifically as follows: Step 3.1: Construct a 4-layer Laplacian pyramid for each frame of the video. Each layer of the pyramid represents the representation of the original image at a different scale, and the resolution of each layer is half of the previous layer. Step 3.2: Apply bandpass filtering to each layer of the pyramid to amplify only changes in the frequency range of 0.4-3Hz; Step 3.3: Multiply the filtered signal by a set amplification factor to increase the visibility of these small movements; Step 3.4: Add the amplified motion back to the original frames, reconstruct each frame using the modified pyramid, and combine the reconstructed frames into a new video file containing the amplified small motion changes.
4. The non-contact, sensorless sleep staging monitoring method based on deep learning as claimed in claim 1, characterized in that: The step 6 is specifically as follows: Step 6.1: Initialize an empty list to store the smoothed data, and traverse each data point of the accumulated optical flow; Step 6.2: Calculate the new smoothed value by multiplying the previous smoothed data point by the smoothing factor factor and adding the result obtained by multiplying the current data point by (1-factor). Add the smoothed value to the list initialized in step 6.1 and traverse to obtain the weighted moving average data. Step 6.3: Use Savitzky-Golay filter to further smooth the weighted moving average data obtained in step 6.2; Step 6.4: For the data obtained in step 6.3, use a quadratic polynomial to fit the data and time series to obtain a quadratic equation that best fits the data; Step 6.5: Calculate the value of the trend line using the coefficients of the polynomial and the time series. Subtract the value of the trend line from the data obtained in step 6.3 to obtain the detrended data.
5. The non-contact, sensorless sleep staging monitoring method based on deep learning as claimed in claim 1, characterized in that: The step 10 is specifically as follows: Step 10.1: Build a two-layer LSTM network, each layer outputs 24-dimensional features; Step 10.2: Add two fully connected layers after the LSTM layer. The first fully connected layer contains 4800 nodes and the second fully connected layer contains 32 nodes. Step 10.3: Use the Softmax output layer to predict the class label of the sample, i.e., REM vs. NREM. Wherein, the step 11 is specifically as follows: Step 11.1: Use K-means clustering algorithm to classify the mandibular optical flow amplitude into three categories: wakefulness, sleep, and rapid eye movement; Step 11.2: Use K-means clustering algorithm to classify the eye movement optical flow amplitude into two categories: wakefulness and sleep; Step 11.3: Add the awakening part in step 11.1 to the classified data in step 11.2 according to the corresponding time to obtain two-category sleep data: awakening and sleeping; Step 11.4: Further classify the sleep data in step 11.3, and check the sleep data from the beginning in chronological order. If the "sleep" state is detected for the first time to last for more than 5 minutes, the first "sleep" point of the continuous "sleep" state is regarded as the critical point of sleep, called the "sleep point"; Step 11.5: Reversely check the sleep data in step 11.
3. If the first detected "sleep" state lasts for more than 5 minutes, it will be regarded as the last period of "sleep", and the last "sleep" point of the continuous "sleep" state is the "awakening point"; Step 11.6: Change the data before the "sleep point" and after the "wake-up point" in step 11.3 to the "wake-up" state; Step 11.7: In the data obtained in step 11.6, the eye movement classification model trained in step 10 is used to further classify the "sleep" state part to obtain data with three categories: wakefulness, rapid eye movement, and non-rapid eye movement; Step 11.8: Since the duration of REM period can generally last from 5 to 40 minutes, when the interval between two "REM" states is less than 5 minutes, it is determined that the state during this period is the "REM" state; Step 11.9: After step 11.8, three categories of sleep state data are obtained: wakefulness, non-rapid eye movement, and rapid eye movement, and all categories are arranged in time series to obtain the final sleep stage.
Citation Information
Cited By
Analysis system and therapeutic planning for sleep-related breathing disorder by adding imaging interpretation
WO2026210474A1
