A method and device for risk stratification and early warning of post-awakening stroke

Abstract

The application discloses a method and device for risk grading and early warning of post-sleep stroke, and the method comprises the following steps: collecting a video stream during sleep of a monitored person through a non-contact sensor; extracting multi-modal indexes from the video stream, the multi-modal indexes at least including a lateral movement index and a breathing-related index; obtaining historical multi-modal indexes corresponding to historical sleep video streams of the monitored person before the onset of the disease, taking the historical multi-modal indexes as individualized baselines, and setting upper and lower threshold values for each multi-modal index; calculating the multi-modal indexes in the current sleep cycle in real time, and marking the indexes exceeding the upper and lower threshold values as positive indexes; determining a risk grading result according to the number of the positive indexes, and triggering a grading early warning response operation corresponding to the risk grading result. The application realizes long-term, continuous and automatic night monitoring and grading early warning, and a full-automatic closed loop from monitoring to first aid.

Description

Technical Field

[0001] This invention relates to the field of medical early warning technology, and in particular to a method and apparatus for risk classification and early warning of post-awakening stroke. Background Technology

[0002] Post-arousal stroke is characterized by its uncertain onset time and high disability rate, and its prognosis highly depends on early identification and intervention after onset. Currently, nighttime monitoring of stroke risk mainly relies on contact-based physiological signal acquisition devices such as polysomnography, or non-contact monitoring devices based on radar and optical cameras, to continuously collect data during the monitored person's sleep window.

[0003] In existing contactless home monitoring solutions, sensors are typically used to acquire physiological or behavioral data of the monitored individual, and a general classification model pre-trained based on healthy population samples is used to analyze the current data. When the extracted feature data meets preset general abnormality criteria, the system determines that there is a risk of stroke, triggers an alarm, and sends the monitoring data to a remote terminal for manual review by professional medical personnel to ultimately confirm whether emergency medical assistance is needed.

[0004] However, the aforementioned existing technologies have limitations in practical applications. On the one hand, due to the high heterogeneity of physiological and behavioral changes in individuals before and after a stroke, using a general classification model trained based on statistical patterns of healthy individuals for anomaly detection is difficult to adapt to fluctuations in the physiological baseline of different individuals. This results in insufficient generalization ability of the model when faced with individual-specific data, making it highly susceptible to triggering false positive alarms due to normal physiological fluctuations. On the other hand, existing solutions typically rely on single or independent feature thresholds for judgment, lacking a mechanism for constraining the correlation between multidimensional features, making it difficult to effectively eliminate interference from factors such as normal body position changes during sleep over time. Furthermore, existing solutions usually directly transfer to manual review after triggering an alarm, lacking an automated warning grading and step-by-step response mechanism in home settings. This means that all primary abnormal signals require direct screening using professional medical resources, resulting in a serious review bottleneck and waste of medical resources. Therefore, how to improve the accuracy of non-contact sleep monitoring under complex individual differences and construct a reasonable automated warning response mechanism to reduce the burden of manual screening are technical problems that urgently need to be solved by those skilled in the art.

[0005] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for risk stratification and early warning of post-awakening stroke, in order to solve the problems of poor compliance of existing monitoring equipment, weak model generalization ability and lack of multimodal fusion early warning during the sleep window period.

[0007] To achieve the above objectives, the present invention provides a method for risk grading and early warning of post-awakening stroke, comprising the following steps: acquiring video streams during the sleep of a monitored subject using a non-contact sensor; extracting multimodal indicators from the video streams, the multimodal indicators including at least lateral motor indicators and respiratory-related indicators; obtaining historical multimodal indicators corresponding to the monitored subject's historical sleep video streams before the onset of the disease, using the historical multimodal indicators as an individualized baseline, and setting upper and lower thresholds for each multimodal indicator; calculating the multimodal indicators within the current sleep cycle in real time, marking indicators exceeding the upper and lower thresholds as positive indicators; determining the risk grading result based on the number of positive indicators, and triggering a graded early warning response operation corresponding to the risk grading result.

[0008] In one embodiment, the non-contact sensor includes a passive infrared camera, which is fixedly installed on the bedroom wall and facing the monitored person. The passive infrared camera acquires grayscale video streams at a preset frame rate under conditions of no visible light in the environment. Continuous acquisition is started when the monitored person lies down for more than a preset time, and acquisition stops when the preset time is reached or the monitored person gets out of bed for a preset threshold time.

[0009] In one embodiment, the multimodal indicators include the number of times the body rolls over, the overall frequency of body movement, the unilateral limb movement asymmetry index, rolling over preference, limb elevation height, facial orientation percentage, eye opening frequency and duration, respiratory contour frequency and irregularity, whole-body muscle tremors, sleep posture stability, and sweating pattern; wherein, the unilateral limb movement asymmetry index, the rolling over preference, and the facial orientation percentage are lateral movement indicators, the respiratory contour frequency and irregularity are respiratory-related indicators, and the respiratory frequency, the whole-body muscle tremors, the sleep posture stability, and the sweating pattern are symmetrical distribution indicators.

[0010] In one embodiment, the step of extracting multimodal indicators from the video stream includes: performing real-time pose estimation on the video stream using an edge computing module or a cloud server, extracting the key point coordinate sequence of the monitored person's body using a lightweight deep learning model, and performing time series analysis based on the key point coordinate sequence to extract the multimodal indicators.

[0011] In one embodiment, a personal profile is established by acquiring multimodal statistical data of the monitored individual during complete sleep cycles without stroke events before the onset of the disease; for symmetrically distributed indicators, the upper and lower limits are the historical mean plus or minus 2 times the historical standard deviation; for skewedly distributed indicators, the upper and lower limits are the historical 5th percentile to 95th percentile; for directional indicators, the lower limit is set at the historical 10th percentile; and for proportional indicators, the upper limit is set at the historical mean plus 1.5 times the standard deviation.

[0012] In one embodiment, after marking indicators that exceed the upper and lower thresholds as positive indicators, a timing constraint is introduced: if the positive indicators do not include at least one lateral movement indicator and at least one respiratory-related indicator, then two additional positive indicators need to be counted to trigger an alert; when the number of positive indicators reaches a second preset number, the first-level alert is skipped and the second-level alert is triggered directly.

[0013] In one embodiment, the step of determining the risk grading result based on the number of positive indicators and triggering a graded early warning response operation corresponding to the risk grading result includes: when the number of positive indicators accumulates to a first preset number, triggering a first-level early warning and performing a local wake-up operation; when the number of positive indicators accumulates to a second preset number, directly skipping the first-level early warning and triggering a second-level early warning; if no release signal is received from the monitored person within a preset time, automatically escalating to a second-level early warning and sending early warning information to the pre-bound associated terminal device; if a feedback signal confirming suspicious discomfort is received from the associated terminal device, triggering a third-level early warning and establishing a video communication connection with a medical emergency terminal.

[0014] This invention also provides a device for risk grading and early warning of post-awakening stroke, which implements the method for risk grading and early warning of post-awakening stroke as described above. The device includes: a non-contact sensor for acquiring video streams during the sleep of a monitored subject; and a processing device communicatively connected to the non-contact sensor, the processing device including: a data processing and feature extraction module for extracting multimodal indicators from the video stream, the multimodal indicators including at least lateral motor indicators and respiratory-related indicators; an individualized baseline establishment module for acquiring historical multimodal indicators corresponding to the monitored subject's historical sleep video streams before the onset of the disease, using the historical multimodal indicators as an individualized baseline, and setting upper and lower thresholds for each multimodal indicator; a risk grading determination module for calculating multimodal indicators in the current sleep cycle in real time, marking indicators exceeding the upper and lower thresholds as positive indicators, and determining the risk grading result based on the number of positive indicators; and a graded early warning response module for triggering a graded early warning response operation corresponding to the risk grading result.

[0015] In some embodiments, the non-contact sensor is a passive infrared camera, which is fixedly installed on the bedroom wall and facing the monitored person. It collects grayscale video streams at a preset frame rate under conditions of no visible light in the environment. When the monitored person lies down for more than a preset time, continuous collection is started, and collection stops when the preset time is reached or when the monitored person gets out of bed for a preset threshold time.

[0016] In one embodiment, the risk grading determination module is configured to: trigger a first-level warning when the number of positive indicators reaches a first preset number; upgrade to a second-level warning if no cancellation signal is received within a preset time; and trigger a third-level warning if a feedback signal confirming suspicious discomfort is received from an associated terminal device. The graded warning response module is configured to: perform local wake-up during a first-level warning; send warning information to a pre-bound associated terminal device during a second-level warning; and establish a video communication connection with a medical emergency terminal during a third-level warning.

[0017] Compared with the prior art, the method and apparatus for risk assessment and early warning of post-awakening stroke provided by the present invention have at least the following beneficial effects: (1) This invention achieves non-contact monitoring by using a passive infrared camera, so that the monitored person does not need to wear any sensors, which improves the long-term compliance of the monitored person (especially the elderly). (2) This invention extracts 11 multimodal indicators such as the number of times the patient turns over, limb symmetry, respiratory irregularity, and sweating pattern at the same time, and establishes upper and lower limit thresholds based on the patient's normal data when the disease is not present. This overcomes the problem of low recognition accuracy of the general population model in stroke patients and effectively reduces the false positive rate. (3) The present invention designs a three-level progressive early warning response link (first level wakes up the patient, second level contacts the family, and third level video connects to the emergency room), and sets a clear threshold for the number of positive indicators and an automatic upgrade mechanism for no response within 10 minutes, forming a fully automatic closed loop from automatic monitoring to medical emergency care, eliminating the defect in the prior art that all abnormal signals need to be reviewed by a physician. (4) The entire device of the present invention uses only a single passive infrared camera, which has low hardware cost and is suitable for home deployment. It realizes risk monitoring and early warning of the sleep window period with uncertain onset time, thereby moving the identification and response node to the sleep period, shortening the delay time of pre-hospital treatment, and effectively improving the patient's prognosis. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method for risk assessment and early warning of post-awakening stroke according to the present invention. Figure 2 This is a schematic diagram of the layout of the device for risk assessment and early warning of post-awakening stroke according to the present invention. Figure 3 This is a schematic diagram of the module of the device for risk assessment and early warning of post-awakening stroke according to the present invention. Detailed Implementation

[0019] The method and apparatus for risk grading and early warning of post-awakening stroke proposed in this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of this invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, only for the purpose of conveniently and clearly illustrating the embodiments of this invention. Please refer to the drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0020] Existing post-awareness stroke monitoring and early warning technologies suffer from several limitations: First, insufficient device compliance and excessive monitoring burden. While wearable devices can achieve continuous physiological monitoring, long-term wear presents comfort issues, and the elderly, who are primarily at high risk of stroke, have limited acceptance and operational skills with electronic devices, resulting in low actual usage compliance. While environmental intelligence can avoid discomfort, current non-contact monitoring methods often rely on single-modal data (such as collecting only snoring or facial images), making it difficult to capture the complex behavioral-physiological abnormalities preceding post-awareness stroke. Second, poor model generalization ability and a lack of post-stroke population-specific algorithms. Studies show that sleep staging algorithms trained on healthy individuals show significantly reduced effectiveness when applied to stroke patients, primarily due to the high heterogeneity and individual differences in physiological signals among post-stroke patients, making it difficult for general population models to effectively identify abnormal patterns. Third, the "physician review bottleneck" hinders accurate early warning and clinical implementation. All abnormal signals ultimately require manual review by professional physicians, which not only increases the clinical workload but also makes it difficult for home-based automatic early warning systems to achieve a truly closed-loop response. Fourth, there is a lack of a multimodal fusion framework for the "sleep window" of post-awake stroke. Existing solutions mostly focus on monitoring single risk factors such as electromyography, electroencephalography, atrial fibrillation, and blood pressure fluctuations, failing to perform temporal alignment and joint reasoning between behavioral modalities (turning frequency, limb symmetry, eye-opening duration, etc.) and physiological modalities (breathing rhythm, etc.), and especially lacking targeted algorithm design for the core challenge of "uncertain onset time".

[0021] To address the aforementioned deficiencies, this invention provides a method and apparatus for risk grading and early warning of post-awakening stroke. The apparatus is deployed in the bedroom environment of the monitored individual, and through non-contact sensors, achieves long-term, continuous, and automated nighttime monitoring and graded early warning without contacting the monitored individual's body or interfering with normal sleep.

[0022] Please refer to Figures 1-2 , Figure 1 The method for risk stratification and early warning of post-awake stroke according to an embodiment of the present invention includes the following steps: Step S1: During the sleep of the monitored person 20, a video stream is collected by the non-contact sensor 10 to acquire nighttime image data without physical contact and without disturbing the sleep of the monitored person 20. Step S2: Extract multimodal indicators from the video stream. The multimodal indicators include at least lateral motor indicators and respiratory-related indicators. This is because post-awake stroke, which occurs during sleep, is usually accompanied by unilateral limb motor dysfunction and changes in respiratory rhythm. Therefore, extracting these two types of indicators can capture early physiological and behavioral coordination abnormalities. Step S3: Obtain the historical multimodal indicators corresponding to the historical sleep video stream of the monitored subject 20 before the onset of the disease. Use the historical multimodal indicators as an individualized baseline and set upper and lower thresholds for each multimodal indicator. By using the monitored subject 20's own historical normal data as a baseline, it can adapt to the differences in physiological signals among different individuals and avoid the generalization failure problem caused by the high heterogeneity of the population in the general model trained based on healthy people. Step S4: Calculate the multimodal indicators in the current sleep cycle in real time, and mark the indicators that exceed the upper and lower thresholds as positive indicators, so as to identify characteristics that deviate from the individual's normal physiological state. Step S5: Determine the risk grading result based on the cumulative number of positive indicators during the monitored subject's 20-day sleep period, and trigger the corresponding graded early warning response operation. By grading based on the cumulative number of positive indicators and executing different response operations accordingly, a closed loop from automatic monitoring to emergency medical care is achieved.

[0023] Among them, such as Figure 2As shown, in step S1 above, the non-contact sensor 10 includes a passive infrared camera, which is fixedly installed high on the bedroom wall or in a corner, covering the entire bed area and facing the monitored person 20. In some embodiments, the wavelength range of the passive infrared camera can be set to 850nm-940nm, and it has night vision capabilities, enabling the passive infrared camera to capture grayscale video streams at a preset frame rate (e.g., 30 frames / second) in environments with no visible light. It is understood that the passive infrared camera does not emit visible light when working in complete darkness, thereby avoiding interference with the normal sleep of the monitored person 20. When the monitored person 20 enters the bed area and remains in a lying position for more than a preset time (e.g., 2 minutes), a continuous acquisition mode is activated; acquisition continues until a preset time (e.g., 8 or 9 am the next day) or until the monitored person 20 gets out of bed and stands for a preset threshold time (e.g., 30 seconds), at which point acquisition stops. If the monitored person returns after leaving bed (to use the toilet) at night, the data collection will automatically resume, ensuring that the passive infrared camera's data collection time covers the entire sleep cycle.

[0024] In some embodiments, step S2, the step of extracting multimodal indicators from the video stream, includes: performing real-time pose estimation on the video stream using an edge computing module or a cloud server; extracting the key point coordinate sequence of the monitored subject's body using a lightweight deep learning model (such as an embedded version of MediaPipe or OpenPose); and performing temporal analysis based on the key point coordinate sequence to extract the multimodal indicators. The specific analysis and extraction process can be based on general knowledge, and this invention does not limit this. The key point coordinate sequence includes 15 key points: head, shoulders, elbows, wrists, hips, knees, ankles, overall facial view, eyes, forehead wrinkles, corners of the mouth, nasolabial folds, eyebrows, overall lower limb view, and overall upper limb view.

[0025] In some embodiments, the multimodal indicators in step S2 specifically include 11: number of times the body rolls over, overall body movement frequency, unilateral limb movement asymmetry index, rolling over preference, limb elevation height, facial orientation percentage, eye opening frequency and duration, respiratory contour frequency and irregularity, whole-body muscle tremors, sleep posture stability, and sweating pattern. Among these, the unilateral limb movement asymmetry index, the rolling over preference, and the facial orientation percentage are lateral movement indicators; the respiratory contour frequency and irregularity are respiratory-related indicators; and the respiratory frequency, whole-body muscle tremors, sleep posture stability, and sweating pattern are symmetrical distribution indicators.

[0026] Furthermore, the specific extraction and calculation logic for each of the aforementioned multimodal indicators is as follows: (1) Number of times the body turns over: The event of the body turning over is detected when the overall orientation of the body changes between lying on the left or right side and lying on the back. Each change is counted as 1 time.

[0027] (2) Overall body movement frequency: Calculate the displacement velocity of the center point of the trunk (i.e. the midpoint of the two hips) on the bed plane and count the frequency of events exceeding 5cm / s.

[0028] (3) Unilateral limb movement asymmetry index: The movement amplitude of the left and right wrists and ankles is counted respectively (displacement amplitude > 3cm is counted as 1 effective movement), wherein the asymmetry index = (left movement number - right movement number) / (left movement number + right movement number).

[0029] (4) Turning over preference: Calculate the percentage of time spent in left lateral position, right lateral position and supine position throughout the night, and take the one with the highest percentage as the preferred side.

[0030] (5) Limb elevation height: The maximum vertical distance between the left and right wrists and the bed plane was measured, and the average and maximum values ​​for the whole night were taken.

[0031] (6) Facial orientation percentage: Calculate the deflection angle of the facial normal vector relative to the bed plane and the center line of the pillow based on the key points of the head, and count the percentage of time spent in the left, right and supine orientations.

[0032] (7) Eye opening frequency and duration: Based on the pixel intensity change in the eye region or a pre-trained eye opening and closing classifier, detect the number of eye opening events at night and the duration of each event.

[0033] (8) Frequency and irregularity of breathing profile: The breathing waveform is extracted by the periodic longitudinal displacement of the key points of the thoracic cavity (i.e., from the midpoint of the shoulders to the midpoint of the hip), and the breathing frequency (breaths / min) and the coefficient of variation of the breathing interval are calculated as indicators of irregularity.

[0034] (9) Whole-body muscle tremors: Fourier analysis was performed on the high-frequency micro-vibrations of key limb points during non-rolling and non-voluntary movement to extract the tremor energy spectral density; wherein the definition of high-frequency micro-vibrations can be a frequency of 4-15Hz and an amplitude of <5mm. It should be noted that in this embodiment, the whole-body muscle tremors are acquired by a passive infrared camera; in other embodiments, millimeter-wave radar can also be used to acquire them, and the present invention does not specifically limit the acquisition of this indicator.

[0035] (10) Sleeping posture stability: Calculate the drift speed (cm / h) of the body's center of mass (i.e. the center of gravity of the key points of the whole body) within a 15-minute sliding window, and the total length of the center of mass trajectory throughout the night.

[0036] (11) Sweating pattern: Based on the gray-scale gradient distribution changes of the face or neck in infrared images, combined with ambient temperature compensation, regional sweating events and sweating intensity are inferred. It is understood that the sampling frequency, sliding window, and outlier removal method of the above multimodal indicators can be adjusted according to the specific sensor characteristics, and this invention does not limit them.

[0037] Further, in step S3, multimodal statistical data of the monitored subject 20 during multiple consecutive stroke-free sleep cycles (e.g., the first 7 nights or 10 consecutive complete sleep cycles) before the onset of the disease are obtained to establish a personal profile. The mean, median, standard deviation, and percentile of each multimodal indicator are recorded to form the individualized baseline. Different threshold setting strategies are adopted based on the data distribution characteristics of each multimodal indicator. For symmetrically distributed indicators (including respiratory rate, whole-body muscle tremors, sleep posture stability, and sweating patterns), the upper and lower limits (i.e., normal range) are taken as the historical mean plus or minus 2 times the historical standard deviation. For skewed distribution indicators or indicators with clear upper and lower boundaries (including the number of times a person rolls over, the frequency of overall body movement, the index of unilateral limb movement asymmetry, and the frequency and duration of eye opening), the upper and lower limit thresholds (i.e. the normal range) are taken from the 5th percentile to the 95th percentile in history. For indicators with clear directionality (such as limb elevation height), only the lower limit threshold is set at the 10th percentile of history. If the value is below this lower limit threshold, it is judged as abnormal. For proportional indicators (such as turning-over preference and facial orientation percentage), only an upper limit threshold is set at the historical mean plus 1.5 standard deviations. Values ​​exceeding this upper limit threshold are considered abnormal. For example, if the percentage of time spent on one side exceeds the historical mean plus 1.5 standard deviations, it is considered abnormal.

[0038] Furthermore, the individualized baseline can be set to update automatically weekly or monthly, or be manually reset under the guidance of a doctor. The individualized baseline setting of this invention establishes a baseline using the patient's normal sleep data before the onset of illness, effectively overcoming the generalization failure problem caused by the high heterogeneity of physiological signals in general models trained on healthy individuals, thus improving the individual applicability and accuracy of the early warning model. Simultaneously, a differentiated threshold setting strategy is adopted for different data distribution characteristics, making the definition of the normal range more consistent with statistical regularities and clinical pathological manifestations, effectively reducing the overall false positive rate of the system. In addition, the mechanism of periodic automatic updating or manual resetting of the baseline allows it to adapt to the slow drift of the patient's physiological signals, preventing an increase in false alarms due to baseline rigidity and ensuring the reliability of long-term home monitoring.

[0039] Based on the individualized baseline, the statistical values ​​of the aforementioned 11 multimodal indicators within the current sleep cycle are calculated in real time during each nighttime monitoring period and compared with the upper and lower thresholds of the individualized baseline. When a multimodal indicator exceeds its upper threshold or falls below its lower threshold, the multimodal indicator is marked as a positive indicator. Furthermore, to reduce the false positive warning rate and achieve a closed-loop response, a temporal constraint is introduced: if the positive indicators do not include at least one lateral movement indicator (such as unilateral limb movement asymmetry index, turning preference, and facial orientation percentage) and at least one respiratory-related indicator (such as the frequency and irregularity of breathing contours), then two additional positive indicator counts are required to trigger an warning.

[0040] In some embodiments, the tiered early warning response operation specifically includes the following: When the number of positive indicators accumulates to a first preset number (e.g., 4), a first-level warning is triggered, and a local wake-up operation is performed. A wake-up signal is emitted through the bedside audio-visual alarm (such as a gradually increasing buzzer sound and a breathing light), and a notification is pushed to the monitored person, requiring them to confirm that they are conscious within a preset time (e.g., 10 minutes) by pressing a button, raising their arm, using a preset gesture, or responding with a voice. If the alarm is successfully deactivated, the event is recorded but no external alarm is issued.

[0041] When the number of positive indicators accumulates to a second preset number (e.g., 6), the first-level warning is skipped, and a second-level warning is triggered. Alternatively, if no release signal is received from the monitored individual within a preset time (e.g., 10 minutes), the warning is automatically upgraded to the second-level warning. Simultaneously, a warning message is sent via a wireless communication network (4G / 5G / Wi-Fi) to a pre-bound associated terminal device (e.g., a family member's or caregiver's smartphone). The warning message includes a summary of the current positive indicators, a risk score, and encrypted video clips from the past 15 minutes. The family member or caregiver can view the monitored individual's real-time video and historical trends through the associated terminal device and make a preliminary remote assessment. If it is confirmed that the monitored individual has no obvious discomfort and can communicate normally, the alarm can be remotely lifted. If the family member confirms a suspicious discomfort signal through the associated terminal device, or if the family member cannot be contacted, a third-level warning is triggered.

[0042] Upon triggering a Level 3 alert, with family members already at the bedside, an automatic video communication connection is established with the medical emergency terminal (such as the nearest emergency center or stroke center). This connection transmits at least 30 seconds of continuous real-time video, the current risk score, a trend chart of monitoring indicators from the past 24 hours, and an individualized baseline reference range. Emergency physicians use the medical emergency terminal to conduct a rapid video neurological function assessment (such as facial paralysis, limb movement, and speech testing). If a suspected stroke is confirmed, an ambulance is immediately dispatched, and patient information is sent to the emergency department, achieving a fully automated closed loop from monitoring to emergency care.

[0043] Accordingly, embodiments of the present invention also provide a device for risk grading and early warning of post-awakening stroke, such as... Figure 3 As shown, the device includes a non-contact sensor 10 and a processing device 30 communicatively connected to the non-contact sensor 10. In this embodiment, the processing device 30 is a processor (e.g., an edge computing chip) integrated inside the non-contact sensor 10. The non-contact sensor 10 and the processing device 30 interact via an internal bus to achieve localized data processing and rapid response. The non-contact sensor 10 is used to collect video streams during the monitored subject's sleep. The processing device 30 includes a data processing and feature extraction module 31, an individualized baseline establishment module 32, a risk grading determination module 33, and a graded early warning response module 34, to implement the risk grading and early warning method for post-awakening stroke described in the above embodiment.

[0044] Specifically, the data processing and feature extraction module 31 receives the video stream transmitted from the non-contact sensor 10 and extracts multimodal indicators from the video stream. These multimodal indicators include at least lateral movement indicators and respiratory-related indicators. The individualized baseline establishment module 32 acquires historical multimodal indicators corresponding to the subject's historical sleep video stream before the onset of illness, uses these historical multimodal indicators as an individualized baseline, and sets upper and lower thresholds for each multimodal indicator. The risk grading determination module 33 calculates the multimodal indicators within the current sleep cycle in real time, marks indicators exceeding the upper and lower thresholds as positive indicators, thereby identifying characteristics deviating from the individual's normal physiological state, and determines the risk grading result based on the number of positive indicators. The graded early warning response module 34 triggers a graded early warning response operation corresponding to the risk grading result output by the risk grading determination module 33.

[0045] Furthermore, to optimize the accuracy and continuity of non-contact monitoring, the non-contact sensor 10 is a passive infrared camera. The passive infrared camera is fixedly installed high on the bedroom wall or in a corner, covering the entire bed area and facing the monitored person. The wavelength range of the passive infrared camera can be set to 850nm-940nm, and it has night vision capabilities, acquiring grayscale video streams at a preset frame rate (e.g., 30 frames / second) in environments with no visible light. The processing device 30, as the processor built into the passive infrared camera, can directly acquire the grayscale video stream for local analysis, eliminating the need to upload massive amounts of video data to the cloud and reducing network transmission latency. When the monitored person enters the bed area and remains in a lying position for more than a preset time (e.g., 2 minutes), the passive infrared camera starts a continuous acquisition mode; acquisition continues until a preset time (e.g., 8 or 9 am) or until the monitored person gets out of bed and stands for a preset threshold time (e.g., 30 seconds), at which point acquisition stops. If the monitored person leaves bed at night (to use the toilet) and returns, acquisition automatically resumes.

[0046] To reduce the false positive warning rate and achieve a closed-loop response, the risk grading judgment module 33 within the processing device 30 is configured to: trigger a first-level warning when the number of positive indicators reaches a first preset number (e.g., 4); upgrade to a second-level warning if no release signal is received from the monitored person within a preset time (e.g., 10 minutes); and trigger a third-level warning if a feedback signal confirming suspicious discomfort is received from the associated terminal device (30). When the number of positive indicators reaches a second preset number (e.g., 6), a second-level warning is directly triggered. The graded warning response module 34 within the processing device 30 is configured to: perform local wake-up during a first-level warning; send warning information to the pre-bound associated terminal device 40 during a second-level warning; and establish a video communication connection with the medical emergency terminal 50 during a third-level warning.

[0047] The tiered early warning response module 34 includes a first-level early warning unit, a second-level early warning unit, and a third-level early warning unit. When a first-level early warning is triggered, the first-level early warning unit performs a local wake-up operation, sending a wake-up signal through a bedside audible and visual alarm and pushing a notification to the monitored person, requiring them to confirm consciousness within a preset time (e.g., 10 minutes) by pressing a button, raising their arm, or speaking. If successfully deactivated, the system records the information but does not issue an external alarm. When a second-level early warning is triggered, the second-level early warning unit sends early warning information via a wireless communication network to a pre-bound associated terminal device 40 (e.g., a family member's or caregiver's smartphone). The information includes a summary of positive indicators, a risk score, and encrypted video clips from the past 15 minutes. The family member remotely assesses the situation via the associated terminal device 40; if no discomfort is found, the alarm is deactivated. When a third-level early warning is triggered, the third-level early warning unit establishes a video communication connection with a medical emergency terminal 50 (e.g., the nearest emergency center or stroke center), transmitting continuous real-time video for more than 30 seconds, a risk score, a trend chart of monitoring indicators from the past 24 hours, and an individualized baseline reference range. Emergency physicians conduct video neurological function assessments via medical emergency terminal (40). If a suspected stroke is confirmed, an ambulance is dispatched and the information is sent to the emergency department.

[0048] In one specific embodiment, during the monitored subject's (20's) nighttime sleep, a passive infrared camera captures a grayscale video stream at 30 frames per second and transmits it directly to its built-in processing device (30). The data processing and feature extraction module (31) within the processing device (30) extracts 15 body key point coordinate sequences using a lightweight deep learning model and calculates 11 multimodal indicators frame-by-frame or second-by-second. The risk grading and determination module (33) acquires the indicator statistics for the current night in real time and compares them with the individualized upper and lower limit thresholds stored in the individualized baseline establishment module (32). Indicators exceeding the thresholds are marked as positive.

[0049] As sleep progresses, the number of positive indicators accumulates dynamically locally. When the cumulative number reaches 4, the Level 1 warning unit activates an audio-visual wake-up call. If the monitored person does not respond within 10 minutes, or if the number of positive indicators directly reaches 6, the Level 2 warning unit sends encrypted video and data to the associated terminal device (30). The family member of the associated terminal device 40 reviews the video and provides a confirmation signal. If the feedback confirms a suspected discomfort, the Level 3 warning unit automatically connects to the medical emergency terminal 50, and the emergency physician conducts a neurological function assessment and dispatches emergency resources via video. This process does not require manual alarm triggering, and the acquisition of video streams, feature extraction, and risk assessment are all completed in a closed loop within the sensor's local processing device 30. Only when an alarm is triggered is summary data sent or video communication established, which significantly reduces data transmission latency and protects the patient's home privacy.

[0050] In summary, the method and device for risk stratification and early warning of post-awareness stroke provided by this invention employs a passive infrared camera as a non-contact sensor 10 for collecting multimodal indicator behaviors, and integrates the processing device 30 within it as an edge computing node. This avoids the burden of the monitored person wearing the device, solves the problem of poor compliance of existing wearable devices, and at the same time, localized processing avoids the high bandwidth consumption and privacy leakage risks caused by continuous video uploading, making it suitable for long-term home deployment. Secondly, by extracting 11 multimodal indicators and combining them with the self-calibration mechanism of the individualized baseline establishment module 32, the method utilizes the patient's own asymptomatic... The data setting threshold overcomes the problem of weak generalization ability of general models trained on healthy populations in stroke patients, reducing the false alarm rate caused by individual differences. Furthermore, by setting a multi-indicator joint counting rule with time-series constraints through the risk grading judgment module 33, and combining it with the three-level progressive early warning response link of the graded early warning response module 34, false positive events are intercepted at the first and second level response stages, and only highly suspicious events are pushed to the medical emergency terminal 50, reducing the review burden of emergency physicians and realizing a fully automatic closed loop from monitoring to emergency treatment. This solves the problems of manual review bottleneck and inability to achieve home-based closed loop in the existing technology.

[0051] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0052] In the description of this invention, it should be understood that the terms "center," "height," "thickness," "upper," "lower," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0053] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0054] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0055] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for risk stratification and early warning of post-awake stroke, characterized in that, Includes the following steps: Video streams are collected during the subject's sleep using non-contact sensors; Multimodal metrics are extracted from the video stream, including at least lateral motion metrics and respiratory-related metrics; The historical multimodal indicators corresponding to the historical sleep video stream of the monitored person before the onset of the disease are obtained. The historical multimodal indicators are used as individualized baselines, and upper and lower thresholds are set for each multimodal indicator. The system calculates multimodal indicators in real time during the current sleep cycle and marks indicators that exceed the upper and lower thresholds as positive indicators. The risk classification result is determined based on the number of positive indicators, and a graded early warning response operation corresponding to the risk classification result is triggered.

2. The method for risk stratification and early warning of post-awake stroke as described in claim 1, characterized in that, The non-contact sensor includes a passive infrared camera, which is fixedly installed on the bedroom wall and facing the monitored person. The passive infrared camera acquires grayscale video streams at a preset frame rate under conditions of no visible light in the environment. Continuous data collection begins when the monitored person lies in bed for more than a preset time and stops when the preset time is reached or when the monitored person leaves the bed for a preset threshold period.

3. The method for risk assessment and early warning of post-awake stroke as described in claim 1, characterized in that, The multimodal indicators include the number of times the body rolls over, the overall frequency of body movement, the unilateral limb movement asymmetry index, rolling over preference, limb elevation height, facial orientation percentage, eye opening frequency and duration, respiratory contour frequency and irregularity, whole-body muscle tremors, sleep posture stability, and sweating pattern; wherein, the unilateral limb movement asymmetry index, the rolling over preference, and the facial orientation percentage are lateral movement indicators, the respiratory contour frequency and irregularity are respiratory-related indicators, and the respiratory frequency, the whole-body muscle tremors, sleep posture stability, and sweating pattern are symmetrical distribution indicators.

4. The method for risk assessment and early warning of post-awake stroke as described in claim 1, characterized in that, The steps for extracting multimodal metrics from the video stream include: performing real-time pose estimation on the video stream using an edge computing module or a cloud server, extracting the key point coordinate sequence of the monitored person's body using a lightweight deep learning model, and performing time-series analysis based on the key point coordinate sequence to extract the multimodal metrics.

5. The method for risk assessment and early warning of post-awake stroke as described in claim 1 or 3, characterized in that, To establish a personal profile, multimodal statistical data of the monitored individuals were obtained from complete sleep cycles without stroke events before the onset of the disease. For symmetrically distributed indicators, the upper and lower limits were set as the historical mean plus or minus 2 times the historical standard deviation. For skewedly distributed indicators, the upper and lower limits were set as the historical 5th percentile to the 95th percentile. For directional indicators, the lower limit was set as the historical 10th percentile. For proportional indicators, the upper limit was set as the historical mean plus 1.5 times the standard deviation.

6. The method for risk assessment and early warning of post-awake stroke as described in claim 1, characterized in that, After marking indicators that exceed the upper and lower thresholds as positive indicators, a time constraint is introduced: if the positive indicators do not include at least one lateral movement indicator and at least one respiratory-related indicator, then two additional positive indicators need to be counted to trigger an alert. When the number of positive indicators reaches a second preset number, the first-level warning is skipped and the second-level warning is triggered directly.

7. The method for risk assessment and early warning of post-awake stroke as described in claim 1, characterized in that, The step of determining the risk grading result based on the number of positive indicators and triggering the graded early warning response operation corresponding to the risk grading result includes: When the number of positive indicators reaches a first preset number, a first-level warning is triggered and a local wake-up operation is performed. When the number of positive indicators reaches a second preset number, the first level warning is skipped and the second level warning is triggered. If no release signal is received from the monitored party within a preset time, the alert will automatically escalate to the second level and send an alert message to the pre-bound associated terminal devices. If a feedback signal confirming a suspected discomfort is received from the associated terminal device, a third-level warning is triggered and a video communication connection is established with the medical emergency terminal.

8. An apparatus for risk assessment and early warning of post-awake stroke, used to implement the method for risk assessment and early warning of post-awake stroke as described in any one of claims 1 to 7, characterized in that, include: Non-contact sensors are used to capture video streams during the sleep of the monitored person; A processing device, communicatively connected to the non-contact sensor, the processing device comprising: The data processing and feature extraction module is used to extract multimodal indicators from the video stream, wherein the multimodal indicators include at least lateral motion indicators and respiratory-related indicators; The individualized baseline establishment module is used to obtain the historical multimodal indicators corresponding to the historical sleep video stream of the monitored person before the onset of the disease, use the historical multimodal indicators as the individualized baseline, and set upper and lower limit thresholds for each multimodal indicator. The risk grading and determination module is used to calculate multimodal indicators in the current sleep cycle in real time, mark indicators that exceed the upper and lower thresholds as positive indicators, and determine the risk grading result based on the number of positive indicators. The graded early warning response module is used to trigger graded early warning response operations corresponding to the risk grading results.

9. The device for risk assessment and early warning of post-arousal stroke as described in claim 8, characterized in that, The non-contact sensor is a passive infrared camera, which is fixedly installed on the bedroom wall and facing the monitored person. It captures grayscale video streams at a preset frame rate under conditions of no visible light in the environment. Continuous data collection begins when the monitored person lies in bed for more than a preset time and stops when the preset time is reached or when the monitored person leaves the bed for a preset threshold period.

10. The device for risk assessment and early warning of post-arousal stroke as described in claim 8, characterized in that, The risk grading and judgment module is configured to: trigger a first-level warning when the number of positive indicators reaches a first preset number; upgrade to a second-level warning if no cancellation signal is received within a preset time; and trigger a third-level warning if a feedback signal confirming suspicious discomfort is received from an associated terminal device. The graded warning response module is configured to: perform local wake-up during a first-level warning; send warning information to a pre-bound associated terminal device during a second-level warning; and establish a video communication connection with a medical emergency terminal during a third-level warning.

Images