Intelligent sleep disorder assessment system for convalescents

Abstract

The present application belongs to the field of artificial intelligence, and particularly relates to a kind of convalescent sleep disorder intelligent evaluation system, to solve the problems of strong equipment invasiveness, single data mode and insufficient environmental factor correlation in the prior art.The system includes a non-contact multi-dimensional physiological signal acquisition module, an environmental stress factor monitoring module, an edge side collaborative preprocessing unit, a heterogeneous data fusion server and a deep learning evaluation terminal, which integrates environmental parameters through millimeter wave radar, acoustic and pressure sensing, realizes high-precision, real-time, non-invasive sleep disorder classification and closed-loop intervention, and significantly improves the evaluation accuracy and convalescent effect.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an intelligent assessment system for sleep disorders in convalescent patients. Background Technology

[0002] In the fields of health monitoring and smart elderly care, sleep quality is a key indicator for measuring the physical and mental health of the elderly and caregivers. With the integrated development of biomedical engineering and artificial intelligence technologies, non-invasive health monitoring has become an important component of the elderly care service system. Sleep monitoring technology not only involves the acquisition and processing of physiological signals, but also covers multiple dimensions such as sleep stage identification, disorder diagnosis, and rehabilitation suggestion generation. Effective sleep assessment can provide a scientific basis for medical staff to develop personalized care plans and has important clinical significance for preventing complications such as cardiovascular disease and cognitive impairment.

[0003] Among them, the intelligent sleep disorder assessment system for convalescent settings focuses on integrating multi-source physiological parameters and environmental perception data. This type of system aims to achieve continuous, real-time monitoring of users' sleep behavior throughout the entire process through intelligent terminals and algorithm models. Its core principle lies in using sensing technology to capture weak physiological electrical signals or body movement characteristics, and using deep learning and other methods to automatically extract abnormal sleep patterns and conduct qualitative and quantitative risk assessments, thereby achieving intelligent screening for sleep disorders.

[0004] In current technologies, sleep assessment for caregivers primarily relies on clinical scales and professional polysomnography (PSG) equipment. Traditional scales are limited by the subject's subjective perception and recall bias, often resulting in a lack of objective accuracy. While professional-grade PSG devices offer high accuracy, their complex sensor connections can cause psychological burden and physical discomfort for caregivers, and can even lead to secondary interference due to wire constraints, causing discrepancies between the monitored data and the actual natural sleep state. Furthermore, existing monitoring systems generally suffer from insufficient depth in multimodal data mining, weak ability to identify environmental interference, and a lack of trend analysis across time and space, making it difficult to accurately capture subtle pathological sleep fluctuations. Moreover, they still fall short of meeting the refined needs of professional care services in terms of timely warnings and comprehensive assessment dimensions. Therefore, an optimized intelligent assessment system for sleep disorders in caregivers is desired. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent assessment system for sleep disorders in convalescent patients, in order to solve the problems of existing sleep assessment technologies, such as strong device invasiveness, single data modality, inability to link convalescent environment parameters in real time, and inaccurate classification of sleep disorders for the specific group of convalescent patients.

[0006] The technical solution of the present invention includes: a multidimensional physiological signal non-contact acquisition module, a real-time monitoring module for environmental stress factors, an edge-side data collaborative preprocessing unit, a heterogeneous data fusion and feature engineering server, and a deep learning-based comprehensive assessment terminal for sleep disorders.

[0007] A multidimensional physiological signal non-contact acquisition module is used to acquire micro-dynamic characteristic data of the patient during sleep without physical contact. This module is equipped with an ultra-wideband millimeter-wave radar array, with its operating center frequency set between 77 and 79 GHz. It extracts micrometer-level phase shifts caused by heartbeats and respiratory movements by emitting nanosecond pulse signals and receiving echoes reflected from the chest and body surface. Simultaneously, the module integrates a high-sensitivity acoustic sensor array to collect snoring, coughing, and respiratory noises during sleep. Furthermore, the module includes a pressure-sensing array, placed under the mattress, to acquire data on the patient's body movement frequency, turning movements, and changes in sleeping posture.

[0008] A real-time environmental stress factor monitoring module is used to simultaneously acquire physical and chemical parameters of the sleep environment of the patients. This module is deployed at multiple spatial locations within the treatment area, including a light intensity sensor to monitor the intensity of blue light radiation in specific wavelengths, particularly between 450 and 480 nanometers; a sound pressure level monitoring unit to record the decibel value and frequency distribution of background noise in real time; a temperature and humidity compensation sensor to acquire environmental thermal comfort indicators; and a gas concentration sensor to monitor carbon dioxide concentration, volatile organic compound concentration, and negative oxygen ion level.

[0009] The edge-side data collaborative preprocessing unit receives raw bitstreams from the multidimensional physiological signal non-contact acquisition module and the environmental stress factor real-time monitoring module via physical connection. This unit first performs clock alignment to ensure all modal data remain synchronized on a nanosecond-level time base. Further, the unit removes stationary target clutter caused by environmental reflections and power frequency interference from the power system by executing a digital filtering algorithm. As one embodiment of the invention, for radar echo signals, the unit employs an adaptive cancellation algorithm to extract pure cardiopulmonary motion signals; for acoustic signals, the unit utilizes spectral subtraction and endpoint detection algorithms to extract effective sleep-related acoustic event fragments.

[0010] A heterogeneous data fusion and feature engineering server is used for deep analysis and multi-dimensional feature construction of preprocessed data. The server first maps physiological and environmental signals to a unified spatial coordinate system, constructing a spatiotemporal correlation tensor. For physiological feature extraction, the server extracts frequency domain indices of heart rate variability, including high-frequency, low-frequency, and low-high-frequency ratios, through discrete wavelet transform to characterize the activity state of the autonomic nervous system. Simultaneously, the server automatically identifies obstructive or central sleep apnea events based on the amplitude attenuation ratio and duration of respiratory signals. Regarding environmental feature correlation, the server calculates the conditional probability correlation between environmental noise fluctuations and micro-arousal events in patients, thereby quantifying the degree of interference of environmental factors on sleep structure.

[0011] This deep learning-based comprehensive sleep disorder assessment terminal automatically generates sleep staging results and disorder assessment reports based on the feature set output by a heterogeneous data fusion and feature engineering server. The terminal is equipped with a deep residual neural network based on a spatiotemporal attention mechanism. This network takes multidimensional physiological feature sequences as input, extracts deep nonlinear semantic information through a multilayer perceptron, and utilizes a long short-term memory network to capture the dynamic evolution of sleep stages. The terminal outputs results including the percentage distribution of wakefulness, light sleep, deep sleep, and REM sleep, as well as quantitative scores for specific disorders such as difficulty falling asleep, sleep maintenance disorders, early awakening, and sleep apnea.

[0012] In one embodiment of the present invention, the ultra-wideband millimeter-wave radar array in the multidimensional physiological signal non-contact acquisition module employs multiple-input multiple-output (MIMO) technology. By configuring four transmitting antennas and four receiving antennas, 16 virtual channels are formed, thereby achieving precise beamforming of the patient's chest cavity position. Through angle estimation and spatial filtering of the 16 channel signals, the system can accurately lock onto the vital signs of the target individual even in environments with interference sources.

[0013] Furthermore, the edge-side data collaborative preprocessing unit employs principal component analysis when processing pressure sensing array signals. This method performs dimensionality reduction processing on pressure data scattered across hundreds or thousands of sensing points, extracting feature vectors representing the movement of the human body's center of mass, thereby accurately determining the patient's turning frequency and bed posture.

[0014] In one embodiment of the present invention, the heterogeneous data fusion and feature engineering server constructs a dynamic physiological rhythm benchmark model. This model acquires historical sleep data of the patients from the past 7 to 14 days and establishes individualized physiological threshold intervals using non-parametric statistical methods. When the real-time collected physiological parameters deviate from this threshold interval by more than 2 standard deviations, the system marks that period as an abnormal sleep fluctuation point.

[0015] Furthermore, the real-time environmental stress factor monitoring module also includes an electromagnetic radiation monitoring unit. This unit is used to sense the intensity of radio frequency electromagnetic fields in the convalescent environment in real time. A heterogeneous data fusion and feature engineering server performs correlation regression analysis between electromagnetic radiation levels and the sleep latency of convalescent patients to assess the potential inhibitory effect of electronic device radiation on melatonin secretion from the pineal gland.

[0016] As one embodiment of the present invention, the deep learning-based comprehensive assessment terminal for sleep disorders also includes a knowledge graph comparison engine. This engine pre-stores expert diagnostic logic from the International Classification of Sleep Disorders, 3rd Edition. The system matches the quantitative indicators output by the deep neural network with the logical rules in the knowledge graph, ensuring that the assessment report not only has statistical accuracy but also conforms to the logical framework of medical diagnosis.

[0017] Furthermore, the system also includes a feedback control closed-loop interface. When the deep learning-based sleep disorder comprehensive assessment terminal detects that environmental stress factors exceed a preset comfort threshold and lead to a decline in sleep quality, the system sends adjustment commands to the sanatorium's environmental control system through this interface, including automatically adjusting the color temperature of dimming lights, reducing the operating noise of the HVAC system, and turning on the negative ion generator of the air purification system.

[0018] In one embodiment of the present invention, the acoustic sensor array in the multidimensional physiological signal non-contact acquisition module employs sound source localization technology. By calculating the time difference of signals arriving at different sensors, the system can distinguish between the snoring of the patient and background noise from outside or adjacent rooms, thereby eliminating the influence of environmental noise on the accuracy of snoring detection.

[0019] Furthermore, the heterogeneous data fusion and feature engineering server incorporates nonlinear dynamics analysis methods when extracting heart rate variability features. By calculating the minor and major axis descriptors of the Poincaré scatter plot and the approximate entropy index, the system can more sensitively reflect the stress load state of patients during sleep.

[0020] As one embodiment of the present invention, the deep learning-based comprehensive sleep disorder assessment terminal employs a federated learning architecture for model iteration. While protecting the privacy of caregivers, the system only uploads gradient information generated from local model training to the central server. By aggregating gradient data from multiple care nodes, the system continuously optimizes the generalization ability of the global sleep disorder identification algorithm.

[0021] Furthermore, the system is configured with a Level 2 security verification protocol. The edge-side data collaboration preprocessing unit performs data anonymization processing before transmitting data to the server. All fields involving identity information are converted into unique hash values ​​to ensure that they cannot be reverse-linked to specific caregivers during the assessment process.

[0022] In one embodiment of the present invention, the temperature and humidity compensation sensor of the real-time environmental stress factor monitoring module adopts a redundant design. Monitoring points are set at the bedside of the sanatorium patient and at the indoor ventilation opening, and the equivalent temperature of the sanatorium patient's perceived environment is calculated by a weighted average algorithm, thereby improving the accuracy of environmental correlation analysis.

[0023] Furthermore, the heterogeneous data fusion and feature engineering server specifically includes a drug metabolism interference correction item during feature construction. This item retrieves the pharmacokinetic parameters of the sedatives or sleep aids taken by the patient based on their electronic medical records, automatically deducting systematic biases caused by drug factors when calculating sleep structure indicators, thereby restoring the patient's true physiological sleep capacity.

[0024] As one embodiment of the present invention, the deep learning-based comprehensive assessment terminal for sleep disorders can generate trend analysis reports at multiple time scales. The report includes not only a detailed staged map of the current night but also a trend map of sleep efficiency evolution over the past 30 days. By fusing long and short-term memory features, the system can identify whether the patient exhibits a long-term evolutionary trajectory of chronic sleep deprivation or circadian rhythm disorder.

[0025] Furthermore, the multidimensional physiological signal non-contact acquisition module has a real-time early warning function for apnea. When the radar signal detects a respiratory interruption lasting more than 20 seconds, accompanied by severe body movement, the edge-side data collaborative preprocessing unit will trigger an immediate early warning signal. This signal is directly pushed to the mobile terminal of the on-duty medical staff via a high-speed wireless network to prevent the risk of hypoxia caused by sudden respiratory distress.

[0026] As one embodiment of the present invention, the system also integrates a sleep environment optimization suggestion submodule. This submodule customizes a personalized sleep preparation plan for each therapist based on the optimal combination of sleep environment parameters obtained from heterogeneous data fusion and feature engineering server analysis, including suggested sleep time, most suitable indoor temperature, and light intensity gradient suggestions.

[0027] Furthermore, the heterogeneous data fusion and feature engineering server analyzes the energy distribution of rolling-over movements using discrete cosine transform when processing body motion signals. Based on the duration and frequency band of the energy peak, the system distinguishes between normal rolling-over movements and pathological restless legs syndrome movements, thereby improving the accuracy of screening for movement-related sleep disorders.

[0028] In one embodiment of the present invention, all hardware components of the system are designed for low power consumption. The sensor terminal operates via battery power or wireless power, and the communication process uses Bluetooth Low Energy or Narrowband Internet of Things protocol to reduce electromagnetic interference with the overall physical environment of the sanatorium.

[0029] Furthermore, the deep learning-based comprehensive sleep disorder assessment terminal supports concurrent assessments by multiple users. Through cloud-based containerized deployment, the system can simultaneously process real-time data streams from over 100 treatment rooms, ensuring that the generation latency of assessment reports is less than 180 seconds.

[0030] In one embodiment of the present invention, the real-time environmental stress factor monitoring module has a self-calibration function. The system periodically compares the measured values ​​of each sensor, and when it finds that the reading deviation of a certain sensor exceeds a preset 5% deviation threshold, it automatically starts a drift compensation algorithm based on the reference values ​​of neighboring nodes.

[0031] Furthermore, the heterogeneous data fusion and feature engineering server constructed a multidimensional sleep quality index. This index integrates sleep latency, number of awakenings, percentage of deep sleep, respiratory disturbance index, and sensitivity to environmental disturbances, and obtains a scalar score between 0 and 100 through a weighted summation algorithm, providing data support for the quantitative evaluation of recuperation effects.

[0032] As one embodiment of the present invention, the system further includes a data anomaly tracing unit. When a significant anomaly occurs in the evaluation results, this unit can automatically retrieve the original signal slices and environmental parameter records for the corresponding time period for manual review by experts, ensuring the interpretability of the system and the rigor of the evaluation process.

[0033] Furthermore, the support structure of the multidimensional physiological signal non-contact acquisition module adopts an adjustable design, which can automatically adjust the elevation angle and horizontal tilt angle of the millimeter-wave radar according to the height and position of the patient's bed, ensuring that the radio frequency energy beam always covers the patient's torso area.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0035] 1. This solution integrates ultra-wideband millimeter-wave radar, acoustic sensor arrays, and pressure sensing arrays to achieve a completely non-contact multimodal physiological monitoring system. Compared with traditional polysomnography devices, this invention completely eliminates the psychological stress and physical constraint caused to patients by attaching electrodes and wearing lead wires, maximally restoring the sleep performance of patients in a natural state, making the assessment results more realistic and valuable.

[0036] 2. This solution innovatively introduces a real-time monitoring module for environmental stress factors, conducting in-depth spatiotemporal correlation analysis between environmental dimensions such as light, sound, temperature, and air quality and physiological indicators. This allows the system not only to inform about "sleep quality," but also to explain "why the quality is poor" from a causal perspective, accurately identifying specific environmental fluctuations that trigger awakenings or insufficient deep sleep in patients. This comprehensive environmental correlation assessment provides a scientific basis for the precise optimization of the treatment environment.

[0037] 3. This solution employs a deep learning-based assessment terminal, combined with an architecture integrating edge computing and cloud computing. Through residual neural networks and attention mechanisms, the system can capture extremely subtle changes in physiological characteristics, such as rhythmic fluctuations of the autonomic nervous system and weak respiratory disturbance signals. Simultaneously, the application of the federated learning mechanism, while ensuring data privacy, enables continuous model evolution across institutions and regions, significantly improving the sensitivity and specificity of sleep disorder identification for this specific group of convalescent patients, achieving clinical-grade assessment accuracy.

[0038] 4. This solution constructs a closed-loop intelligent intervention architecture. Through the linkage between the feedback control interface and the environmental control system, the system can dynamically adjust the recuperation environment based on real-time sleep assessments. This shift from passive monitoring to proactive intervention transforms the system from an assessment tool into an intelligent management platform for dynamically improving the health status of convalescent patients. The system possesses significant technical advantages in ensuring sleep safety, improving sleep efficiency, and quantifying the effectiveness of convalescent care. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention;

[0040] Figure 2 This is a schematic diagram of the core principle framework of heterogeneous data fusion and comprehensive assessment of sleep disorders in this invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] Example 1

[0043] Please refer to Figure 1 and Figure 2 This invention provides an intelligent assessment system for sleep disorders in convalescent patients, comprising: a multidimensional physiological signal non-contact acquisition module, a real-time monitoring module for environmental stress factors, an edge-side data collaborative preprocessing unit, a heterogeneous data fusion and feature engineering server, and a deep learning-based comprehensive assessment terminal for sleep disorders.

[0044] The multidimensional physiological signal non-contact acquisition module is used to acquire micro-dynamic characteristic data of the patient during sleep without contact with their body. The core of the module is an ultra-wideband millimeter-wave radar array, whose operating center frequency is set between 77 and 79 GHz. It transmits nanosecond-level frequency-modulated continuous wave pulse signals and receives echo signals reflected from the human chest and body surface in real time, extracting micron-level phase shifts caused by heartbeats and respiratory movements. The module also includes a high-sensitivity acoustic sensor array to collect acoustic events such as snoring, coughing, breathing noises, and teeth grinding during sleep. In addition, the module includes a pressure sensing array, which is laid out in a high-density array under the treatment mattress to acquire high-precision data on the dynamic changes in the patient's body movement frequency, turning movements, limb twitching, and sleeping posture.

[0045] A real-time environmental stress factor monitoring module is used to simultaneously acquire physical and chemical parameters of the sleep environment of the patients. The module is deployed at multiple spatial locations in the treatment area to form a three-dimensional spatial monitoring grid. A light intensity sensor is used to monitor specific wavelengths, especially the blue light component radiation intensity between 450 and 480 nanometers. Its sampling frequency is set to 10 times per second to capture any subtle light fluctuations. A sound pressure level monitoring unit is used to record the decibel value and frequency distribution of background noise in real time, and decomposes the sound signal into low-frequency, mid-frequency, and high-frequency components through fast Fourier transform. Temperature and humidity compensation sensors are distributed at the bedside and indoor ventilation openings to acquire environmental thermal comfort indicators. A gas concentration sensor integrates the monitoring of carbon dioxide concentration, volatile organic compound concentration, and negative oxygen ion level.

[0046] The edge-side data collaborative preprocessing unit receives the raw bitstream sent by the multidimensional physiological signal non-contact acquisition module and the environmental stress factor real-time monitoring module via a physical bus or high-speed wireless communication interface. The unit first performs clock alignment, using a precise time protocol to ensure all modal data remain synchronized on a nanosecond-level time base, preventing feature misalignment caused by data drift. The unit further executes a digital filtering algorithm to remove clutter interference caused by static reflections from environmental objects and 50Hz power frequency interference generated by the power system. For the millimeter-wave radar echo signal, the unit uses an adaptive cancellation algorithm to separate pure heartbeat and respiratory signals. For acoustic signals, the unit uses spectral subtraction for background noise reduction and combines it with an endpoint detection algorithm to extract effective sleep-related acoustic event fragments.

[0047] A heterogeneous data fusion and feature engineering server is used for deep analysis and multidimensional feature construction of preprocessed multidimensional data. The server first maps physiological and environmental signals acquired by various sensors to a unified spatial coordinate system, constructing a spatiotemporal correlation tensor with time and spatial axes. For physiological feature extraction, the server performs multi-scale decomposition of heart rate variability using discrete wavelet transform to extract frequency domain indicators reflecting the balance of the autonomic nervous system. Simultaneously, the server automatically identifies obstructive, central, or mixed sleep apnea events based on the amplitude attenuation ratio and duration of respiratory signals. Regarding environmental feature correlation, the server quantifies the physical interference intensity of environmental factors on sleep structure by calculating the conditional probability correlation between environmental noise pulse fluctuations and micro-arousal events in patients.

[0048] This deep learning-based comprehensive sleep disorder assessment terminal automatically generates sleep staging results and disorder assessment reports based on the structured feature set output by the heterogeneous data fusion and feature engineering server. The terminal deploys a deep residual neural network based on a spatiotemporal attention mechanism. This network takes heart rate, respiration, body movement, and acoustic feature sequences as joint inputs, extracts deep nonlinear semantic information through a multilayer perceptron, and uses a long short-term memory network to capture the dynamic evolution of sleep stages throughout the night. The terminal outputs results including the percentage distribution of wakefulness, light sleep, deep sleep, and REM sleep, as well as quantitative scores for specific disorders such as difficulty falling asleep, sleep maintenance disorders, early awakening, and sleep apnea.

[0049] A smart assessment system for sleep disorders in sanatorium patients begins by acquiring raw data through a multi-dimensional physiological signal non-contact acquisition module. During operation, a millimeter-wave radar array generates a high-linearity ramp voltage via a signal generator, controlling a voltage-controlled oscillator to produce a frequency-modulated continuous wave signal. After passing through a power amplifier, the signal is radiated towards the sanatorium patient's area via a transmitting antenna array. When the electromagnetic waves encounter the human chest cavity, they scatter and reflect, with the reflected waves carrying vital signs being captured by the receiving antenna array. The system is equipped with four transmitting antennas and four receiving antennas, forming 16 virtual channels through multiple-input multiple-output (MIMO) technology. This antenna arrangement significantly increases the virtual aperture, thereby improving angular resolution.

[0050] For each signal received by a virtual channel, the system first performs down-conversion processing to convert it into an intermediate frequency signal. The frequency offset of the intermediate frequency signal is proportional to the target distance, while the phase of the signal is closely related to the small displacement of the target. The edge-side data collaborative preprocessing unit performs analog-to-digital conversion on the intermediate frequency signal with a sampling depth of 16 bits, and uses a digital signal processing chip to perform a distance-dimensional fast Fourier transform, thereby determining the precise spatial coordinates of the patient on the bed.

[0051] In order to extract the weak heartbeat signal from the complex echo, the edge-side data collaborative preprocessing unit executed an adaptive cancellation algorithm; the algorithm subtracts the large-amplitude chest movement component caused by respiration from the mixed signal by establishing a reference noise model, thereby exposing the sub-millimeter micro-vibration signal caused by cardiac ejection.

[0052] Meanwhile, the acoustic sensor array acquires sound signals through a microelectromechanical system (MEMS) microphone. To eliminate ambient background noise, the system employs spatial filtering technology based on sound source localization. By calculating the time difference between the arrival of the sound signal at different microphone units, the system can construct the spatial pointing vector of the sound source in real time. Only when the sound source is located within a preset spherical envelope of the patient's mouth and nose area is the acoustic event marked as a valid signal. For example, when the system detects an instantaneous sound pressure peak, if the sound source localization result indicates that the sound originates from outside the window, the system will automatically identify it as external interference and eliminate it. If the localization result indicates that it originates from the patient, the system will further call a feature recognition algorithm to determine whether it is simple snoring or breathing noise accompanied by airway obstruction.

[0053] During data acquisition, the pressure sensing array scans all sensing points 20 times per second; the pressure value generated by each point is converted into an 8-bit digital signal; the edge-side data collaborative preprocessing unit uses principal component analysis to reduce the dimensionality of data from thousands of points, extracting the first and second principal component feature vectors representing the movement of the human body's center of mass; by analyzing the energy distribution of the feature vectors in the time domain, the system can accurately distinguish whether the patient is in a quiet sleep state or in a state of frequent adjustment such as turning over, limb shaking, or falling asleep.

[0054] The real-time environmental stress factor monitoring module acts as a recorder of external variables during system operation; the light intensity sensor uses a wide dynamic range photosensitive element, capable of sensing light changes from 0.01 lux to 1000 lux; in particular, the system performs weighted integration on blue light irradiance in the 450 nm to 480 nm band, because the photon energy in this band can inhibit neural activity in the paraventricular nucleus of the hypothalamus by stimulating melanopsin cells in the retina, thereby blocking the synthesis of melatonin; the system uploads this irradiance data in real time to a heterogeneous data fusion and feature engineering server for analyzing the causal relationship between light pollution and prolonged sleep latency;

[0055] The temperature and humidity compensation sensor adopts a redundant design, with monitoring points set at the bedside and air conditioning vents. Due to the complex airflow organization in the sanatorium room, the temperature at a single point often cannot represent the equivalent temperature felt by the patient. The system uses a weighted average algorithm, combined with data from the real-time wind speed sensor, to calculate an equivalent comfort index that reflects the body's thermal balance. The gas concentration sensor monitors carbon dioxide concentration using infrared nondispersive technology. If the indoor carbon dioxide concentration rises rapidly in the early stages of sleep, it often indicates a stuffy feeling during sleep due to poor ventilation, which will serve as an important dimension for assessing a decline in sleep quality.

[0056] After receiving all the preprocessed data, the heterogeneous data fusion and feature engineering server begins to construct a multidimensional feature vector space. The server performs discrete wavelet transform to decompose the heart rate waveform. By decomposing the signal into detail coefficients and approximation coefficients of different frequency subbands, the system can sensitively capture the high-frequency and low-frequency components in heart rate variability. The high-frequency components are usually related to vagal nerve activity, while the low-frequency components are jointly regulated by the sympathetic and vagal nerves. The server calculates the ratio of high-frequency to low-frequency components as a key feature characterizing the sleep depth and autonomic nerve tone of the patients.

[0057] As one embodiment of the present invention, a heterogeneous data fusion and feature engineering server constructs a dynamic physiological rhythm benchmark model; the model does not use a fixed universal threshold, but learns by acquiring historical sleep data of a specific patient over the past 7 to 14 days; the system uses non-parametric statistical methods, such as kernel density estimation, to establish an individualized physiological threshold range for each patient; for example, for a patient with a low baseline heart rate, the normal sleep heart rate fluctuation range will be narrower than that of an ordinary person;

[0058] Based on this benchmark model, the system monitors the degree of deviation of physiological parameters in real time; the judgment logic follows the following mathematical model:

[0059]

[0060] in, Represents the current real-time collected physiological parameter values, including but not limited to instantaneous heart rate, respiratory rate, or kinetic energy values; This represents the average physiological parameters of the patient during the same sleep stage over the past 14 days; This represents the historical standard deviation of the parameter. When the absolute value of the difference between the real-time parameter and the historical mean exceeds twice the standard deviation, the heterogeneous data fusion and feature engineering server will automatically record this moment as an abnormal sleep fluctuation point and give it higher weight in subsequent obstacle assessments.

[0061] The heterogeneous data fusion and feature engineering server also includes a dedicated drug metabolism interference correction item during feature construction. The system retrieves patients' electronic medical records through a secure gateway to identify whether they have taken sedative drugs such as diazepam or zopiclone. The server has a built-in pharmacokinetic parameter library for common sleep aids, dynamically calculating residual blood drug concentrations during sleep based on administration time and drug half-life. When calculating sleep stage indicators, the system automatically deducts forcibly prolonged deep sleep or artificially shortened sleep latency caused by medication, thus restoring the patient's true physiological sleep capacity without medication dependence. This has extremely high clinical value for evaluating the effectiveness of rehabilitation.

[0062] After receiving the complete feature vector, the deep learning-based comprehensive assessment terminal for sleep disorders initiates multi-layer neural network inference. The deep residual neural network solves the gradient vanishing problem in deep networks through skip connections, enabling the model to learn extremely subtle feature associations. The spatiotemporal attention mechanism layer weights the input sequence. For example, if a sudden peak in environmental noise and a momentary increase in heart rate occur simultaneously in a sequence, the attention mechanism will automatically increase the weights of these two features, thereby accurately determining that this is a micro-arousal event caused by external noise, rather than a spontaneous sleep stage transition.

[0063] The deep learning-based comprehensive assessment terminal for sleep disorders also integrates a knowledge graph comparison engine. The engine pre-stores thousands of diagnostic logic rules from the 3rd edition of the International Classification of Sleep Disorders. The probability distribution results output by the neural network are entered into the engine for secondary verification. For example, if the neural network initially determines that the patient has severe sleep apnea, the knowledge graph comparison engine will further check whether the apnea index reaches the medical diagnostic standard of more than 15 times per hour, and combine it with the oxygen saturation prediction model to form a logical loop to ensure that the final assessment report conforms to the diagnostic thinking of medical experts.

[0064] The system's feedback control closed-loop interface comes into play after evaluating the terminal output results. If the system detects that the carbon dioxide concentration in the current environment is too high or the temperature and humidity deviate from the comfortable range, which directly leads to the sanatorium staff frequently turning over or an increase in the proportion of light sleep, the system will send instructions to the sanatorium's environmental control system through the local area network. The dimming lights will automatically adjust to a low color temperature mode to reduce blue light interference, and the HVAC system will switch to a silent operation mode and automatically adjust the set value according to the equivalent temperature calculated by the system, thereby achieving dynamic and coordinated optimization of the environment and physiological state.

[0065] The ultra-wideband millimeter-wave radar array in the multidimensional physiological signal non-contact acquisition module features an adjustable support structure during deployment. The system's built-in tilt sensor can detect the radar's elevation and horizontal tilt angles. If the patient's bed height is adjusted, the system will automatically control the micro-motor to rotate and adjust the antenna array's direction, ensuring that the radio frequency energy beam always covers the patient's chest and abdominal areas, thereby guaranteeing that the received vital signs signal intensity is always within the optimal signal-to-noise ratio range.

[0066] Before data is transmitted to the server, the edge-side data collaboration preprocessing unit strictly performs data anonymization processing. The system is configured with a Level 2 security verification protocol, and all sensitive fields involving names, ID numbers, and medical record numbers are converted into unique hexadecimal character strings through a salted hash algorithm. In the entire heterogeneous data fusion and deep learning inference process, the system only processes the hashed anonymized data, ensuring that even if the data is illegally intercepted in the transmission link, it cannot be reverse-linked to the personal privacy of specific caregivers.

[0067] The deep learning-based comprehensive assessment terminal for sleep disorders employs a federated learning architecture for model iteration. In a multi-center deployment environment, local data from each rehabilitation node is not stored in the database; model training and gradient update information are generated only on the local terminal. The central server periodically collects gradient data from each node, uses a weighted average algorithm to aggregate model parameters, generates a higher-performance global model, and redistributes it to each assessment terminal. This approach effectively protects privacy while leveraging the distribution characteristics of massive heterogeneous data, significantly improving the system's generalization ability and accuracy in identifying sleep disorders in rehabilitation patients of different physical conditions and ages.

[0068] When extracting heart rate variability features, the heterogeneous data fusion and feature engineering server introduces nonlinear dynamic analysis methods in addition to conventional frequency domain analysis. By calculating the short and long axis descriptors of the Poincaré scatter plot, the system can gain insight into the regulatory flexibility of the autonomic nervous system from the nonlinear fluctuations of the heartbeat interval.

[0069] The approximate entropy index, as a key parameter for measuring the complexity of physiological signals, is defined by the following formula:

[0070]

[0071] In the above formula, This represents the length of the pattern match, which is typically set to 2 in this system. This represents the similarity tolerance, set to 0.2 times the standard deviation of the original signal; It describes the signal in The logarithmic average probability value of the distribution pattern in the dimensional space; the heterogeneous data fusion and feature engineering server can measure the stress load state of the convalescent patients during sleep by calculating the approximate entropy of the heart rate sequence in real time; if the approximate entropy value is significantly reduced, it indicates that the cardiovascular regulatory system of the convalescent patient is in a rigid, high-stress pathological state, and the system will use this as an important basis for long-term health risk warning in the sleep quality assessment report.

[0072] The deep learning-based comprehensive assessment terminal for sleep disorders can generate trend analysis reports at multiple time scales. The reports not only show the sleep stage map of a single night, but also map the evolution trend of sleep efficiency over the past 30 days, 90 days, and even longer periods by aggregating historical data. By utilizing the hidden layer features of the long short-term memory network, the system can identify whether the patient has a cumulative process of chronic sleep deprivation, or whether there is a long-term evolution trajectory of rhythm disorder caused by seasonal changes or changes in treatment plans, providing scientific quantitative support for physicians to adjust rehabilitation plans.

[0073] The multidimensional physiological signal non-contact acquisition module also has a real-time early warning function for apnea; when the edge-side data collaborative preprocessing unit monitors radar waveforms, if it detects that the low-frequency sine wave amplitude representing respiratory motion disappears for more than 20 seconds and is accompanied by violent struggling body movements captured by the pressure sensing array, the system will immediately determine it as an emergency apnea event; the early warning signal is directly pushed to the mobile terminal in the sanatorium duty room with the highest priority through a high-speed wireless network, and the real-time physiological parameter slices of the corresponding room automatically pop up on the duty screen, thus providing medical staff with a timely intervention window before the risk of hypoxia occurs;

[0074] When processing body movement signals, the heterogeneous data fusion and feature engineering server uses discrete cosine transform to perform a refined analysis of the energy distribution of rolling over movements. Based on the distribution characteristics of energy peaks in the frequency domain, the system can effectively distinguish between normal posture adjustment rolling over and pathological restless legs syndrome movements. The energy of normal rolling over is usually concentrated in the low frequency band and is short in duration, while restless legs syndrome is characterized by high-frequency, regular lower limb twitching. Through this precise frequency domain identification, the system significantly improves the initial screening accuracy of sleep disorders related to movement disorders and reduces the frequency of false alarms.

[0075] The real-time environmental stress factor monitoring module has a complete self-calibration function; the system regularly compares the measurement values ​​of multiple sensors deployed in the same area and refers to a high-precision standard reference node; if the reading deviation of a certain sensor exceeds the preset 5% deviation threshold, the system will automatically start the drift compensation algorithm based on spatial interpolation of neighboring nodes, and adjust the software correction coefficient of the sensor to eliminate the measurement error caused by hardware aging or contamination, thereby ensuring data consistency during long-term operation.

[0076] Heterogeneous data fusion and feature engineering servers construct a multidimensional sleep quality index, which is a comprehensive scalar score calculated using a weighted summation algorithm.

[0077]

[0078] in, Represents a multidimensional sleep quality index; Represents the total number of feature terms involved in the calculation; The normalized characteristic index values ​​include sleep latency score, number of awakenings score, deep sleep duration percentage score, respiratory disorder index score, and environmental disturbance sensitivity score. The weight coefficients representing each feature item are determined by the system's deep learning model through regression iteration based on expert scores from a massive number of clinical samples. The final scores range from 0 to 100, providing intuitive and standardized data support for the quantitative assessment of convalescent effects and the comparison of rehabilitation progress among convalescent patients.

[0079] The system also includes a data anomaly tracing unit to improve the credibility and transparency of the assessment results. When a major anomaly occurs in the assessment results output by the deep learning-based sleep disorder comprehensive assessment terminal, such as a sudden and significant drop in sleep quality score, the tracing unit will automatically retrieve the original radar echo slices, acoustic event recordings, and original waveform records of environmental parameters for the corresponding time period. These original evidence chains are encapsulated in a protected data package for manual verification by medical experts, thereby ensuring that the system not only provides results but also provides underlying evidence for expert review, enhancing the rigor of the system in medical and convalescent scenarios.

[0080] The system's hardware components are all designed for extremely low power consumption. The sensor terminals are powered by long-life lithium batteries or by environmental wireless power supply, eliminating the need for large-scale high-voltage wiring in the sanatorium rooms. The communication process uses low-power Bluetooth or narrowband IoT protocols to minimize radio frequency transmission power, thereby minimizing the disturbance of the electromagnetic environment to the overall physical environment of the sanatorium area and creating an ideal green, low-electromagnetic-radiation sleep space for sanatorium patients.

[0081] Example 2

[0082] Based on the above embodiments, this embodiment further details the electromagnetic radiation monitoring unit in the real-time environmental stress factor monitoring module and its interaction mechanism with the heterogeneous data fusion and feature engineering server.

[0083] The electromagnetic radiation monitoring unit uses a wideband isotropic antenna to sense the intensity of radio frequency electromagnetic fields in the frequency range of 100 kHz to 6 GHz in real time. In the convalescent environment, the main sources of electromagnetic radiation include mobile phone wireless signals, wireless local area network signals and stray electromagnetic fields generated by various electronic devices. The monitoring unit converts the sensed electromagnetic field power density into digital signals in real time and sends them to the heterogeneous data fusion and feature engineering server.

[0084] The heterogeneous data fusion and feature engineering server performed a correlation regression analysis between electromagnetic radiation levels and the sleep latency of patients. The server established a mathematical model to assess the potential inhibitory effect of electronic device radiation on the secretion of melatonin from the pineal gland in patients. In the regression analysis, the system excluded the interference of known variables such as light and noise, extracted electromagnetic radiation intensity as the independent variable, and the time from turning off the lighting equipment to entering a light sleep stage as the dependent variable. By analyzing several weeks of data, the system was able to identify individual patients who were sensitive to electromagnetic radiation and provide reasonable suggestions for the use of electronic devices.

[0085] The temperature and humidity compensation sensor in the real-time environmental stress factor monitoring module adopts a redundant design, with monitoring points not only set at the bedside, but also added monitoring points in the area where the convalescent's lower limbs are located and at the ceiling vents; the heterogeneous data fusion and feature engineering server calculates the equivalent temperature of the convalescent's perceived environment through a weighted average algorithm;

[0086] The calculation of equivalent temperature not only considers dry-bulb temperature, but also incorporates relative humidity and local air velocity. Because different parts of the patient's body have varying sensitivities to temperature during sleep, the system assigns a 45% weight to the readings from the bedside sensor, a 35% weight to the lower limb sensor, and a 20% weight to the indoor background environment sensor. Through this multi-point fusion calculation method, the system can more realistically recreate the microenvironment of the patient's body surface, thereby more accurately analyzing the correlation between thermal discomfort and sleep-wake cycles.

[0087] The deep learning-based comprehensive sleep disorder assessment terminal supports multi-user concurrent assessments. The system adopts a cloud-based containerized deployment approach, using container orchestration technology to dynamically allocate computing resources. When the number of monitoring rooms in the sanatorium increases, the system automatically starts new computing container instances to handle concurrent data streams. Under this architecture, even when processing real-time physiological and environmental data streams from more than 100 sanatorium rooms simultaneously, the latency from data collection to assessment report generation can be stably controlled within 180 seconds, ensuring the timeliness of the assessment.

[0088] The acoustic sensor array in the multidimensional physiological signal non-contact acquisition module achieves beamforming through microelectromechanical systems array technology. When processing acoustic signals, the system not only focuses on the loudness of the sound, but also finely characterizes the envelope features of snoring by calculating the zero-crossing rate and short-time energy of the signal. The heterogeneous data fusion and feature engineering server utilizes these acoustic envelope features, combined with the apnea features extracted by radar, to more accurately classify sleep apnea syndrome, distinguishing between loud snoring caused by obstructive apnea and the disappearance of breathing caused by central apnea, thus providing a basis for subsequent precision medical intervention.

[0089] When constructing features, the heterogeneous data fusion and feature engineering server also extracted the extremely low-frequency components of heart rate variability through discrete wavelet transform. These components typically reflect long-term fluctuations in body temperature regulation and the endocrine system. The system couples the changing trends of the extremely low-frequency components with the long-term temperature fluctuations obtained by the real-time monitoring module of environmental stress factors to explore the potential regulatory effect of environmental temperature stability on the endocrine rhythm of convalescent patients.

[0090] Example 3

[0091] This embodiment further illustrates the implementation details of the feedback control closed-loop interface and personalized sleep preparation scheme in an intelligent assessment system for sleep disorders in convalescent patients;

[0092] The feedback control closed-loop interface connects to the building automation system of the nursing home through a standardized industrial communication protocol. When the sleep disorder comprehensive assessment terminal based on deep learning determines that the resident has entered a deep sleep period, the system will automatically issue a noise reduction command to the interface. At this time, the building automation system will reduce the fan speed of the HVAC system and shut down all unnecessary mechanical ventilation equipment to control the background noise below 30 decibels, thereby prolonging the duration of deep sleep. When the system predicts that the resident is about to enter the awakening stage, it will slowly increase the color temperature and brightness of the indoor dimming lights through the interface to simulate the sunrise process, achieving a natural and gentle awakening and avoiding psychological stress caused by loud alarm sounds.

[0093] The system also integrates a sleep environment optimization suggestion submodule, which runs on a heterogeneous data fusion and feature engineering server. Based on the specific patient's optimal sleep data over a period of time, the submodule inversely derives the corresponding combination of environmental parameters. For example, analysis reveals that a patient experiences the shortest sleep latency and the highest percentage of deep sleep when the indoor temperature is 22 degrees Celsius, the carbon dioxide concentration is below 600 parts per million, and the background noise exhibits pink noise characteristics. Based on this, the submodule customizes a personalized sleep preparation plan for the patient, including suggested bedtime, suggested indoor light intensity gradients before bedtime, and ventilation suggestions before sleep. These suggestions are pushed to the patient and caregivers via a mobile application, achieving a complete closed loop from environmental monitoring to proactive environmental optimization.

[0094] In the federated learning architecture of the deep learning-based comprehensive assessment terminal for sleep disorders, a differential privacy algorithm is introduced in the gradient uploading stage to further enhance security. Before sending the local model gradient to the central server, the system injects a small amount of Laplace noise into the gradient data. This process ensures that even if the central server is compromised, attackers cannot infer the original physiological characteristic data of a single sanatorium room through reverse engineering, thus building a strong defense for data privacy at the algorithm level.

[0095] The radar array in the multidimensional physiological signal non-contact acquisition module has human posture recognition capabilities; the system can identify whether the patient is lying on their side, supine, or prone by performing point cloud imaging processing on the echo signals; the heterogeneous data fusion and feature engineering server statistically correlates posture data with the frequency of apnea to identify position-dependent sleep apnea; if it is found that the patient's apnea is most severe when lying supine, the report generated by the assessment terminal will clearly recommend that the patient sleep in a side-lying position, and the nursing system will assist the patient in adjusting their sleeping posture through physical means;

[0096] When performing data desensitization, the edge-side data collaborative preprocessing unit also automatically destroys the original recordings after feature extraction of acoustic signals; after the system extracts the frequency, intensity and respiratory noise features of snoring, it immediately erases the original audio bitstream from the local buffer and only retains the feature vectors to upload to the server; this mechanism eliminates the possibility of leakage of the private sounds of the sanatorium from the physical source.

[0097] When evaluating the sleep quality index, the heterogeneous data fusion and feature engineering server also incorporates a meteorological data correction term. The system retrieves external meteorological data such as outdoor atmospheric pressure and geomagnetic activity index in real time through the Internet interface. Studies have shown that some sensitive groups may experience fluctuations in sleep quality when there are drastic changes in atmospheric pressure or geomagnetic storms. By incorporating these external macroeconomic variables into the model, the server can more fairly assess the actual contribution of the sanatorium environment and avoid misjudging physiological fluctuations caused by natural climate as defects in the sanatorium environment.

[0098] Finally, the deep learning-based comprehensive sleep disorder assessment terminal can generate a visual, interactive sleep report. The report displays the evolution of sleep structure through 3D charts and allows caregivers to retrieve environmental parameter comparison charts at specific time points by clicking on those points. This clearly shows the possible triggers behind each micro-awakening event, such as a sudden corridor noise or an abnormal fluctuation in room temperature. This intuitive analysis method greatly improves the targeting and efficiency of nursing interventions.

[0099] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart assessment system for sleep disorders in convalescent patients, characterized in that, include: The multidimensional physiological signal non-contact acquisition module is used to acquire microdynamic characteristic data of the patient during sleep without contact with the patient's body; The real-time monitoring module for environmental stress factors is used to synchronously collect physical and chemical parameters in the sleep environment of the patients. The edge-side data collaborative preprocessing unit is used to perform time alignment, noise filtering, and effective event extraction on the multidimensional physiological signals and environmental stress factor data. Heterogeneous data fusion and feature engineering server, used to perform spatiotemporal mapping, feature construction and individualized benchmark modeling on preprocessed multi-source data; A deep learning-based comprehensive sleep disorder assessment terminal is used to automatically generate sleep staging results and quantitative assessment reports of sleep disorders based on fusion features.

2. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The multidimensional physiological signal non-contact acquisition module includes an ultra-wideband millimeter-wave radar array, a high-sensitivity acoustic sensor array, and a pressure sensing array laid under the mattress.

3. The intelligent assessment system for sleep disorders in convalescent patients according to claim 2, characterized in that, The acoustic sensor array employs sound source localization technology to distinguish between the snoring emitted by the patient and ambient background noise.

4. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The real-time environmental stress factor monitoring module includes a light intensity sensor, a sound pressure level monitoring unit, a temperature and humidity compensation sensor, a gas concentration sensor, and an electromagnetic radiation monitoring unit.

5. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, When processing the pressure sensing array signal, the edge-side data collaborative preprocessing unit uses principal component analysis to extract the human body's center of mass movement feature vector in order to identify the frequency of turning over and the posture in bed.

6. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The heterogeneous data fusion and feature engineering server constructs a dynamic physiological rhythm benchmark model and identifies abnormal sleep fluctuation points based on individual historical sleep data.

7. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The heterogeneous data fusion and feature engineering server is equipped with a drug metabolism interference correction item, which is used to deduct the systematic deviation of sleep structure indicators caused by sedative or sleep aid drugs based on the electronic medical record information of the patients.

8. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The deep learning-based comprehensive assessment terminal for sleep disorders integrates a knowledge graph comparison engine to verify the consistency between the quantitative assessment results output by the neural network and the diagnostic logic rules in the international classification standards for sleep disorders.

9. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The system also includes a feedback control closed-loop interface, which sends adjustment commands to the sanatorium's environmental control system to optimize lighting, noise, or air quality when environmental stressors are detected to exceed comfort thresholds and cause a decline in sleep quality.

10. The intelligent assessment system for sleep disorders in convalescent patients according to claim 1, characterized in that, The system is equipped with a data anomaly tracing unit, which is used to automatically retrieve the original physiological signal slices and environmental parameter records of the corresponding time period when a major anomaly occurs in the evaluation results, for medical experts to manually review.

Images