Monitoring service robot for old people

Abstract

The invention discloses an old people monitoring service robot which comprises a base, a robot shell is fixedly installed at the top of the base, a main controller is fixedly installed at the top of the robot shell, a containing groove is formed in the front face of the robot shell, and a bracelet is placed in the containing groove. According to the monitoring service robot for the old people, by arranging the cooling fan, hot air in the robot can be exhausted, the situation that the performance and the service life of electronic elements are affected due to the fact that the internal temperature is too high is prevented, the cooling fan blows the hot air from the interior of the robot shell to the cooling groove in the right side through the ventilation opening, and therefore the cooling function is achieved; a dustproof plate can block most of dust under the condition that heat dissipation is not affected, and the interior of the robot is kept clean; and meanwhile, a large amount of dust is accumulated on the surface of the dustproof plate after long-time use, the dustproof plate can be pulled upwards, so that the connecting strips on the two sides are separated from the U-shaped frame, and the dustproof plate is disassembled.

Description

Technical Field

[0001] The present invention belongs to the technical field of robots, and in particular is a robot for elderly care services. Background Art

[0002] With the continuous improvement of people's living standards and the continuous development of robotics technology, home service robots have extremely broad development prospects. The intelligent home will gradually become a reality and will cover all aspects of life in the future. In particular, the application of intelligent robots will lead to smaller and more nuclear-based family structures. This puts enormous pressure on the traditional family model of relying on children for care. This traditional "raising children to provide for old age" model is no longer sustainable. At the same time, accelerated urbanization and changes in the labor market are causing more and more young people to leave their hometowns and move to big cities to seek career opportunities, leading to the increasingly common phenomenon of "empty nesters." These elderly people not only face numerous inconveniences in daily life, such as shopping, medical treatment, and housework, but also suffer from psychological pressures such as loneliness and helplessness.

[0003] For a long time, the primary models of care have been family-based and institutional care. While family care can provide emotional comfort to the elderly, its capacity is diminishing with changes in family structures. While institutional care offers more comprehensive services, high costs, limited beds, and the potential for de-familialization deter many elderly people. Therefore, exploring a new model for elderly care that meets both the daily needs of the elderly and their spiritual well-being has become a pressing issue for society.

[0004] Against this backdrop, the rise of smart technology has brought new solutions to elderly care services. With the continued maturity and widespread adoption of technologies such as the Internet of Things, big data, and artificial intelligence, smart monitoring devices are becoming a crucial aid for home-based elderly care. These devices provide timely and effective safety and health management services by monitoring the elderly's vital signs, living conditions, and environmental changes in real time. Furthermore, the application of emerging technologies such as intelligent robots and telemedicine systems is offering seniors a more convenient and personalized service experience, effectively alleviating the pressures of traditional elderly care models.

[0005] With the advancement of science and technology, robots that accompany the elderly have emerged. Intelligent robots can monitor the vital signs of the elderly, monitor their heart rate, blood pressure, blood sugar, pulse, body temperature and other physiological indicators in real time, detect health abnormalities in time, and provide a basis for medical intervention. At the same time, they can also monitor the elderly's mood, sleep quality, etc., and improve the health management and quality of life of the elderly.

[0006] Therefore, there is an urgent need for an elderly care service robot to meet the elderly's growing needs for health monitoring, emergency rescue, life assistance, etc. in their daily lives. The society's demand for smart monitoring equipment for the elderly at home is increasing. At the same time, most robots are currently equipped with dustproof nets. After long-term use, the dustproof nets are easily clogged, which may affect the air circulation inside the robot, resulting in the electronic components inside the robot being unable to dissipate heat. Long-term high-temperature operation reduces the service life of the electronic components inside the robot. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the prior art and provide a elderly care service robot.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A elderly care service robot comprises a base, characterized in that: a robot housing is fixedly mounted on the top of the base, a main controller is fixedly mounted on the top of the robot housing, a placement slot is defined on the front of the robot housing, a wristband is placed inside the placement slot, a protective component is installed inside the robot housing, a camera is fixedly mounted on the top of the main controller, an alarm is fixedly mounted on the top of the robot housing, and a positioning module, a processing module, a wireless module, and a storage module are fixedly mounted inside the robot housing;

[0010] A vital sign data acquisition module is fixedly installed inside the wristband;

[0011] The protective component includes a heat dissipation groove opened on the right outer surface of the robot shell, two U-shaped frames are fixedly installed on the right outer surface of the robot shell, a dustproof plate is slidably installed between the two U-shaped frames, connecting strips are fixedly installed on the left and right sides of the dustproof plate, and a cooling fan is fixedly installed on the left inner wall of the robot shell.

[0012] Preferably, the two U-shaped frames are located on the left and right sides of the heat dissipation slot, and the two connecting bars are slidably connected to the inside of the U-shaped frame;

[0013] A vent is provided on the left outer surface of the robot shell, and a cooling fan is fixedly installed inside the vent. A wire mesh is fixedly connected inside the vent and on the left side of the cooling fan.

[0014] Preferably, the main controller is electrically connected to the vital sign data acquisition module, the wireless module, the positioning module, the camera, the processing module, the alarm and the storage module respectively;

[0015] The wireless module includes a 4G module, a 5G module, a WIFI module and a Bluetooth module, and the positioning module includes a GPS module;

[0016] The vital sign data acquisition module is electrically connected to the processing module, and the processing module is electrically connected to the alarm.

[0017] Preferably, it also includes:

[0018] The home-based elderly emotion recognition model is configured to classify the elderly's emotions into six types: anger, disgust, fear, happiness, sadness, and surprise. A learning model is built based on the data of the above six types, and the currently collected ECG data is used to determine whether the elderly at home are in one of the above types.

[0019] The pulse recognition and classification model is configured to process the pulse wave signal, filter out noise interference to obtain a high-quality PPG wave signal, and then extract signal features to provide a sufficiently accurate data basis for the data set required by the pulse classification algorithm; use a feature selection algorithm combining mRMR and SVM-RFE to select features from the fused feature set to obtain the optimal feature subset, and use an artificial bee colony algorithm to optimize the parameters of the support vector machine;

[0020] The sub-health status classification and assessment model is configured as follows: using a 116-dimensional pulse feature dataset, a 135-dimensional tongue feature dataset, and a 6-dimensional human body feature dataset, 75% of the data samples are selected as the training set and the remaining 25% as the test set. Four algorithms, SVM, RF, LightGBM, and CatBoost, are used for classification and recognition research. In terms of feature selection, the front LR12 and SVMRE are used for feature selection;

[0021] Preferably, it also includes:

[0022] A multi-dimensional fall state classification and body posture recognition model is configured to: analyze the changing trends of acceleration signals, identify the different data signal changes presented by different postures, obtain different postures, determine the characteristic vectors of walking, running, climbing stairs, and descending stairs, and determine whether they meet the characteristics of falling;

[0023] The classification prediction and identification models for various cardiovascular and cerebrovascular diseases and dangerous conditions are configured as follows: including the logistic classification model, the decision tree classification model, and the random forest algorithm. After preprocessing the data, the relevant classification prediction and identification models for various cardiovascular and cerebrovascular diseases and dangerous conditions based on machine learning methods are established according to the selected features.

[0024] Preferably, it also includes:

[0025] The infection and body inflammation risk identification and classification model is configured as follows: baseline characteristics of elderly patients are obtained, including gender, current smoker, ASA score, comorbidities, and blood measurements; surgical-related data include surgical procedures, anesthesia techniques, laparoscopic surgery, cancer surgery, and surgical severity; surgical checklists are obtained; postoperative infections include urinary tract infection, bloodstream infection, superficial surgical site infection, deep surgical site infection, body cavity infection, and characteristic value 4 inflammation; and infection is assessed according to the definition of infection. All data are guaranteed to be desensitized;

[0026] Based on the obtained data of elderly patients, independent risk factors for postoperative infection were identified by using the inverse probability weighting method, and the correlations derived from these risk factors and the IP weights for postoperative infection were used to construct a traditional logistic regression prediction model;

[0027] The sleep state and effect assessment classification and regression model, as well as the sleep apnea syndrome identification model, are configured as follows: Blood oxygen saturation signals for seven hours of sleep are extracted from sleep monitoring data, combined with vital sign data extracted from medical diagnosis reports as the raw data for training the model. For model training, the training data for each sample represents the subject's vital sign data and blood oxygen saturation, and the label of each sample is its SAS severity. The convolutional neural network is trained based on the raw data.

[0028] A convolutional neural network is used to automatically extract the blood oxygen saturation characteristics and basic vital signs of SAS subjects with different degrees of severity throughout the night. The blood oxygen saturation signal of each record monitored for 7 hours at night is input into the first input of the model. The blood oxygen saturation features are extracted through three convolutional layers. The vital sign data of each record is then input into the second input of the model. The vital sign features are extracted through one convolutional layer. Feature fusion is then performed, and classification prediction is performed based on the subject's blood oxygen saturation characteristics and vital sign characteristics. The predicted SAS severity is finally output.

[0029] Preferably, it also includes:

[0030] An intelligent risk assessment model for heatstroke or heatstroke among elderly people at home is configured as follows: 10% of the data are randomly selected from the original dataset for model testing and validation, the model performance is analyzed using a linear fitting method, the fitness of the model is measured using linear R2, the mean difference between the observed and predicted results is calculated using the Bland-Altman method, the degree of consistency between the prediction and observation is measured using the mean difference and the corresponding 95% confidence limit, the training data is changed to 80% and 70% randomly selected from the original data, and the model performance is evaluated for sensitivity analysis.

[0031] Preferably, it also includes:

[0032] The intelligent evaluation model is configured as follows: using the OpenBCI Cyton development kit as an advanced ECG signal acquisition tool, and using the MATLAB data processing platform to preprocess the collected ECG signals, including noise filtering and signal enhancement, to form an initial sample data set based on different fatigue levels; using the Pan-Tompkins algorithm to perform deep feature extraction on the preprocessed ECG signals, performing Pearson correlation coefficient analysis on the extracted feature parameters, and screening out the most representative preferred indicators through hypothesis testing, and using the SSA algorithm to optimize the parameters of the BP neural network to construct a fatigue recognition model. By inputting the ECG signal characteristics of elderly people at home, their fatigue status can be quickly judged.

[0033] Preferably, it also includes:

[0034] The machine learning exercise risk level classification assessment model is configured as follows: the exercise risk of the elderly is rated as no risk, low risk, medium risk and high risk, the risk status before and after exercise is judged, and the corresponding exercise risk level is calculated in combination with heart rate variability. The HRV of the elderly who exercise outdoors is monitored over a long period of time, the activity level of their autonomic nervous system is tracked, the impact of exercise on the autonomic nervous system is discovered, and the current exercise load and exercise risk of the elderly are assessed. When the exercise load increases significantly, the relevant indicators corresponding to HRV will continue to decrease, reflecting that the elderly's adaptability to exercise is rapidly declining and the possibility of various risk accidents is rapidly increasing.

[0035] Preferably, it also includes:

[0036] The intelligent auxiliary diagnosis model for various TCM syndromes is configured as follows: extracting parameters that can effectively characterize the individual characteristics of TCM syndromes; establishing a TCM pulse syndrome model and training model parameters; calculating the matching distance between the pulse signal to be tested and the TCM syndrome model; judging whether the pulse signal to be tested is the claimed TCM syndrome or which type of TCM syndrome the pulse signal is based on the calculation result of the matching distance. After obtaining the corresponding pulse characteristics, the Gaussian mixture model is used to model the input data. The probability density function of any shape can be approximated by multiple mixed numbers. GMM is used for pulse signals. During recognition, the probability density function of the short-time spectrum feature vector of the pulse signal of different syndrome types is modeled, and then training is performed to obtain the mixed Gaussian probability density function of each syndrome type, which is used as the template of each TCM syndrome type; during identification, the observed feature vector sequence is substituted into the template of each syndrome type, and the calculated maximum posterior probability corresponds to the identified TCM syndrome type. The commonly used parameter estimation of the GMM model is based on the maximum likelihood criterion. The maximum likelihood estimation is to regard the quantity to be estimated as a fixed but unknown quantity, and to find the parameter value that can maximize the probability of the learning sample appearing, and use it as the parameter estimate.

[0037] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0038] 1. This elderly care service robot can significantly improve the comfort and efficiency of home-based care by accurately connecting with the elderly's personalized needs in multiple dimensions such as security, health management, daily life assistance and emotional comfort, creating a safer and healthier home environment, deepening the society's care and respect for the elderly, and jointly building a warm and harmonious social environment for elderly home care.

[0039] 2. The elderly care service robot is equipped with a cooling fan to discharge the hot air inside the robot to prevent the internal temperature from being too high and affecting the performance and life of the electronic components. The cooling fan blows the hot air from the inside of the robot shell through the vents to the heat dissipation slots on the right, thereby achieving the heat dissipation function. The dustproof plate can block most of the dust without affecting the heat dissipation and keep the inside of the robot clean. At the same time, a lot of dust will accumulate on the lower surface of the dustproof plate after long-term use. The dustproof plate can be pulled upward to make the connecting strips on both sides detach from the U-shaped frame, so that the dustproof plate can be disassembled and cleaned. Regular cleaning of the dustproof plate ensures the practical life of the internal electronic components.

[0040] 3. The elderly care service robot, by setting up the coordination between the bracelet, positioning module, processing module, wireless module, vital sign data acquisition module and storage module, wears the bracelet on the wrist of the elderly, uses the vital sign data acquisition module to monitor the elderly's heart rate, heartbeat, blood pressure, blood oxygen saturation, body temperature and weight and other key vital signs, and processes the monitored data through the processing module. When the data is too high, the alarm will sound at the same time to remind family members or medical staff to treat the elderly. The wireless module transmits the processed data to family members or stores it in the storage module. The storage module preserves the elderly's vital sign data, alarm records and other information for a long time to form a complete and detailed personal health file. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a three-dimensional diagram of the structure of the present invention;

[0042] Figure 2 This is a three-dimensional diagram of the dustproof plate structure of the present invention moving upward;

[0043] Figure 3 It is a sectional perspective view of the structure of the present invention;

[0044] Figure 4 It is a left-side stereoscopic view of the structure of the present invention;

[0045] Figure 5 For the present invention Figure 2 A magnified view of the structure at center A;

[0046] Figure 6 This is a structural block diagram of the vital signs monitoring system of the present invention.

[0047] Figure 7 This is a diagram of the architecture of the present invention;

[0048] Figure 8 It is a technical interaction diagram of the present invention;

[0049] Figure 9 Schematic diagram of the data acquisition process of the present invention;

[0050] Figure 10 Schematic diagram of the SAB-SVMKNN structure of the present invention;

[0051] Figure 11 Schematic diagram of PPG signal processing and feature extraction of the present invention;

[0052] Figure 12 Schematic diagram of the time domain characteristics of the PPG signal of the present invention;

[0053] Figure 13 It is the overall flow chart of the Chinese medicine pulse classification algorithm of the present invention;

[0054] Figure 14 Schematic diagram of the key characteristic information of the pulse wave in the time domain of the present invention;

[0055] Figure 15 Schematic diagram of the attention model of the present invention;

[0056] Figure 16 This is a diagram of the SCAttNet model architecture of the present invention;

[0057] Figure 17 This is a schematic diagram of data of a step cycle of the present invention;

[0058] Figure 18 This is a flow chart of the fall recognition process of the present invention;

[0059] Figure 19 This is a flow chart for predicting the risk of postoperative infection in elderly patients in the study of the present invention;

[0060] Figure 20 is a data processing flow chart of the present invention;

[0061] Figure 21 This is a structural diagram of the CNN model of the present invention;

[0062] Figure 22 A flow chart of the construction of a heat stroke model of the present invention;

[0063] Figure 23 The flowchart of the Pan-Tompkins algorithm of the present invention is shown in FIG.

[0064] Figure 24 It is the flow chart of the SSA BP algorithm of the present invention;

[0065] Figure 25 Schematic diagram of the amplitude and time characteristics of the pulse diagram of the present invention;

[0066] Figure 26 It is the principle diagram of the TCM pulse syndrome identification system of the present invention;

[0067] In the figure: 1. Base; 2. Robot shell; 3. Main controller; 4. Placement slot; 5. Bracelet; 501. Vital sign data acquisition module; 6. Protection component; 601. Heat dissipation slot; 602. U-shaped frame; 603. Dustproof plate; 604. Connecting strip; 605. Cooling fan; 7. Camera; 8. Alarm; 9. Positioning module; 10. Processing module; 11. Wireless module; 12. Storage module. DETAILED DESCRIPTION

[0068] The specific embodiments of the present invention are described in detail below.

[0069] The "ranges" disclosed herein are defined in the form of lower and upper limits. A given range is defined by selecting a lower limit and an upper limit, and the selected lower and upper limits define the boundaries of the particular range. Ranges defined in this manner can be inclusive or exclusive and can be combined arbitrarily, i.e., any lower limit can be combined with any upper limit to form a range. For example, if a range of 10 to 50 is listed for a particular parameter, it is understood that ranges of 10 to 40 and 20 to 50 are also contemplated. Furthermore, if the minimum range values ​​listed are 1 and 2, and if the maximum range values ​​listed are 3, 4, and 5, then the following ranges are all contemplated: 1 to 3, 1 to 4, 1 to 5, 2 to 3, 2 to 4, and 2 to 5. In this application, unless otherwise specified, the numerical range "a to b" is an abbreviation for any combination of real numbers between a and b, where a and b are real numbers. For example, the numerical range "0 to 5" means that all real numbers between "0 to 5" are listed herein, and "0 to 5" is simply an abbreviation for these numerical combinations.

[0070] Unless otherwise specified, all embodiments and optional embodiments of the present application can be combined with each other to form a new technical solution.

[0071] Unless otherwise specified, all technical features and optional technical features of this application can be combined with each other to form a new technical solution.

[0072] Unless otherwise specified, all steps of the present application may be performed sequentially or randomly, preferably sequentially. For example, the method includes steps (a) and (b), indicating that the method may include steps (a) and (b) performed sequentially, or may include steps (b) and (a) performed sequentially. For example, the method may further include step (c), indicating that step (c) may be added to the method in any order, for example, the method may include steps (a), (b) and (c), or may include steps (a), (c) and (b), or may include steps (c), (a) and (b), etc.

[0073] Unless otherwise specified, the terms "include" and "comprising" used in this application may be open-ended or closed-ended. For example, "include" and "comprising" may mean that other components not listed may also be included or that only the listed components are included.

[0074] Unless otherwise specified, the reaction is carried out at room temperature and pressure.

[0075] Unless otherwise specified, all parts or percentages are by weight.

[0076] In the present invention, all substances used are known substances and can be purchased or synthesized by known methods.

[0077] In the present invention, the devices or equipment used are all conventional devices or equipment known in the art and are commercially available.

[0078] The following further illustrates the specific implementation of the elderly care service robot of the present invention in conjunction with the embodiments. The elderly care service robot of the present invention is not limited to the description of the following embodiments.

[0079] Example:

[0080] A kind of elderly care service robot, such as Figure 1-6 As shown, it includes a base 1, which is characterized in that: a robot shell 2 is fixedly installed on the top of the base 1, a main controller 3 is fixedly installed on the top of the robot shell 2, a placement slot 4 is opened on the front of the robot shell 2, a wristband 5 is placed inside the placement slot 4, a protection component 6 is installed inside the robot shell 2, a camera 7 is fixedly installed on the top of the main controller 3, an alarm 8 is fixedly installed on the top of the robot shell 2, and a positioning module 9, a processing module 10, a wireless module 11 and a storage module 12 are fixedly installed inside the robot shell 2;

[0081] A vital sign data acquisition module 501 is fixedly installed inside the wristband 5;

[0082] The protective component 6 includes a heat dissipation groove 601 opened on the right outer surface of the robot shell 2. Two U-shaped frames 602 are fixedly installed on the right outer surface of the robot shell 2. A dustproof plate 603 is slidably installed between the two U-shaped frames 602. Connecting strips 604 are fixedly installed on the left and right sides of the dustproof plate 603. A cooling fan 605 is fixedly installed on the left inner wall of the robot shell 2.

[0083] In a possible embodiment, two U-shaped frames 602 are located on the left and right sides of the heat dissipation slot 601, and two connecting bars 604 are both slidably connected to the inside of the U-shaped frame 602;

[0084] A ventilation hole is provided on the left outer surface of the robot housing 2 , and a cooling fan 605 is fixedly installed inside the ventilation hole. A wire mesh is fixedly connected inside the ventilation hole and on the left side of the cooling fan 605 .

[0085] In a possible embodiment, the main controller 3 is electrically connected to the vital sign data acquisition module 501, the wireless module 11, the positioning module 9, the camera 7, the processing module 10, the alarm 8 and the storage module 12 respectively;

[0086] The wireless module 11 includes a 4G module, a 5G module, a WIFI module and a Bluetooth module, and the positioning module 9 includes a GPS module;

[0087] The vital sign data acquisition module 501 is electrically connected to the processing module 10 , and the processing module 10 is electrically connected to the alarm 8 .

[0088] In a possible implementation, the method further includes:

[0089] The home-based elderly emotion recognition model is configured to classify the elderly's emotions into six types: anger, disgust, fear, happiness, sadness, and surprise. A learning model is built based on the data of the above six types, and the currently collected ECG data is used to determine whether the elderly at home are in one of the above types.

[0090] The pulse recognition and classification model is configured to process the pulse wave signal, filter out noise interference to obtain a high-quality PPG wave signal, and then extract signal features to provide a sufficiently accurate data basis for the data set required by the pulse classification algorithm; use a feature selection algorithm combining mRMR and SVM-RFE to select features from the fused feature set to obtain the optimal feature subset, and use an artificial bee colony algorithm to optimize the parameters of the support vector machine;

[0091] The sub-health status classification and assessment model is configured as follows: using a 116-dimensional pulse feature dataset, a 135-dimensional tongue feature dataset, and a 6-dimensional human body feature dataset, 75% of the data samples are selected as the training set and the remaining 25% as the test set. Four algorithms, SVM, RF, LightGBM, and CatBoost, are used for classification and recognition research. In terms of feature selection, the front LR12 and SVMRE are used for feature selection;

[0092] In a possible implementation, the method further includes:

[0093] A multi-dimensional fall state classification and body posture recognition model is configured to: analyze the changing trends of acceleration signals, identify the different data signal changes presented by different postures, obtain different postures, determine the characteristic vectors of walking, running, climbing stairs, and descending stairs, and determine whether they meet the characteristics of falling;

[0094] The classification prediction and identification models for various cardiovascular and cerebrovascular diseases and dangerous conditions are configured as follows: including the logistic classification model, the decision tree classification model, and the random forest algorithm. After preprocessing the data, the relevant classification prediction and identification models for various cardiovascular and cerebrovascular diseases and dangerous conditions based on machine learning methods are established according to the selected features.

[0095] In a possible implementation, the method further includes:

[0096] The infection and body inflammation risk identification and classification model is configured as follows: baseline characteristics of elderly patients are obtained, including gender, current smoker, ASA score, comorbidities, and blood measurements; surgical-related data include surgical procedures, anesthesia techniques, laparoscopic surgery, cancer surgery, and surgical severity; surgical checklists are obtained; postoperative infections include urinary tract infection, bloodstream infection, superficial surgical site infection, deep surgical site infection, body cavity infection, and characteristic value 4 inflammation; and infection is assessed according to the definition of infection. All data are guaranteed to be desensitized;

[0097] Based on the obtained data of elderly patients, independent risk factors for postoperative infection were identified by using the inverse probability weighting method, and the correlations derived from these risk factors and the IP weights for postoperative infection were used to construct a traditional logistic regression prediction model;

[0098] The sleep state and effect assessment classification and regression model, as well as the sleep apnea syndrome identification model, are configured as follows: Blood oxygen saturation signals for seven hours of sleep are extracted from sleep monitoring data, combined with vital sign data extracted from medical diagnosis reports as the raw data for training the model. For model training, the training data for each sample represents the subject's vital sign data and blood oxygen saturation, and the label of each sample is its SAS severity. The convolutional neural network is trained based on the raw data.

[0099] A convolutional neural network is used to automatically extract the blood oxygen saturation characteristics and basic vital signs of SAS subjects with different degrees of severity throughout the night. The blood oxygen saturation signal of each record monitored for 7 hours at night is input into the first input of the model. The blood oxygen saturation features are extracted through three convolutional layers. The vital sign data of each record is then input into the second input of the model. The vital sign features are extracted through one convolutional layer. Feature fusion is then performed, and classification prediction is performed based on the subject's blood oxygen saturation characteristics and vital sign characteristics. The predicted SAS severity is finally output.

[0100] In a possible implementation, the method further includes:

[0101] An intelligent risk assessment model for heatstroke or heatstroke among elderly people at home is configured as follows: 10% of the data are randomly selected from the original dataset for model testing and validation, the model performance is analyzed using a linear fitting method, the fitness of the model is measured using linear R2, the mean difference between the observed and predicted results is calculated using the Bland-Altman method, the degree of consistency between the prediction and observation is measured using the mean difference and the corresponding 95% confidence limit, the training data is changed to 80% and 70% randomly selected from the original data, and the model performance is evaluated for sensitivity analysis.

[0102] In a possible implementation, the method further includes:

[0103] The intelligent evaluation model is configured as follows: using the OpenBCI Cyton development kit as an advanced ECG signal acquisition tool, and using the MATLAB data processing platform to preprocess the collected ECG signals, including noise filtering and signal enhancement, to form an initial sample data set based on different fatigue levels; using the Pan-Tompkins algorithm to perform deep feature extraction on the preprocessed ECG signals, performing Pearson correlation coefficient analysis on the extracted feature parameters, and screening out the most representative preferred indicators through hypothesis testing, and using the SSA algorithm to optimize the parameters of the BP neural network to construct a fatigue recognition model. By inputting the ECG signal characteristics of elderly people at home, their fatigue status can be quickly judged.

[0104] In a possible implementation, the method further includes:

[0105] The machine learning exercise risk level classification assessment model is configured as follows: the exercise risk of the elderly is rated as no risk, low risk, medium risk and high risk, the risk status before and after exercise is judged, and the corresponding exercise risk level is calculated in combination with heart rate variability. The HRV of the elderly who exercise outdoors is monitored over a long period of time, the activity level of their autonomic nervous system is tracked, the impact of exercise on the autonomic nervous system is discovered, and the current exercise load and exercise risk of the elderly are assessed. When the exercise load increases significantly, the relevant indicators corresponding to HRV will continue to decrease, reflecting that the elderly's adaptability to exercise is rapidly declining and the possibility of various risk accidents is rapidly increasing.

[0106] In a possible implementation, the method further includes:

[0107] The intelligent auxiliary diagnosis model for various TCM syndromes is configured as follows: extracting parameters that can effectively characterize the individual characteristics of TCM syndromes; establishing a TCM pulse syndrome model and training model parameters; calculating the matching distance between the pulse signal to be tested and the TCM syndrome model; judging whether the pulse signal to be tested is the claimed TCM syndrome or which type of TCM syndrome the pulse signal is based on the calculation result of the matching distance. After obtaining the corresponding pulse characteristics, the Gaussian mixture model is used to model the input data. The probability density function of any shape can be approximated by multiple mixed numbers. GMM is used for pulse signals. During recognition, the probability density function of the short-time spectrum feature vector of the pulse signal of different syndrome types is modeled, and then training is performed to obtain the mixed Gaussian probability density function of each syndrome type, which is used as the template of each TCM syndrome type; during identification, the observed feature vector sequence is substituted into the template of each syndrome type, and the calculated maximum posterior probability corresponds to the identified TCM syndrome type. The commonly used parameter estimation of the GMM model is based on the maximum likelihood criterion. The maximum likelihood estimation is to regard the quantity to be estimated as a fixed but unknown quantity, and to find the parameter value that can maximize the probability of the learning sample appearing, and use it as the parameter estimate.

[0108] By adopting the above technical solutions:

[0109] Technical route description:

[0110] This application takes the home-based elderly care monitoring scenario as the entry point, focusing on the elderly care monitoring needs in different scenarios, and taking data as the main entry point. In the indicator test projects involving basic vital signs, it focuses on developing the functions of various types of equipment and the transmission stability of supporting software, specifically including: continuous optimization of real-time heart rate monitoring and acquisition equipment, optimization of the portability of ECG signal acquisition equipment and real-time data transmission system, optimization of the accuracy of blood oxygen monitoring system and real-time data transmission, optimization of shortening blood pressure measurement time and improving accuracy, and optimization of body temperature and respiratory rate accuracy measurement.

[0111] This application will be based on the data generated by home-based elderly care monitoring, focusing on the development of a set of intelligent evaluation indicator models for basic medical-grade vital signs monitoring indicators. It will use machine learning in the field of artificial intelligence knowledge as the main implementation method, and build a comprehensive intelligent evaluation indicator model and comprehensive indicator evaluation system by integrating basic physiological data, machine learning methods, and existing research content. It will complete the various goals of the project with modular, integrated, and engineering ideas, and in the overall research, it will treat equipment, algorithms, data, software systems, protocols, and standards as a whole to improve the feasibility of the project research. The overall technical solution for home-based elderly care monitoring scenarios is as follows: Figure 7 As shown:

[0112] like Figure 7 As shown in the figure, the overall research plan of the project is divided into the indicator construction layer and the indicator application layer. The indicator construction layer needs to include various basic vital signs collection modules and intelligent assessment application modules. The collected data are systematically constructed to obtain the two modules of the indicator application layer. The key research models and algorithm contents are concentrated in the intelligent assessment application module. Its core is based on artificial intelligence related technologies and designs corresponding artificial intelligence models for the purpose of various current intelligent assessment application goals.

[0113] like Figure 8 As shown, the main challenges currently faced by home vital signs monitoring are data accuracy, real-time performance, portability, assessment accuracy, and monitoring data management. Based on physiological monitoring equipment, this application primarily involves the development of wearable home monitoring devices, medical-grade portable monitoring equipment, and a data service platform for the elderly-healthcare alliance. The project encompasses multiple components, including equipment, software, and algorithms, all integrated and interconnected.

[0114] This application includes the following models:

[0115] (1) Model 1: Emotion recognition model for elderly people at home based on ECG

[0116] 1. Research purpose: Emotions are not only the physiological states of various human feelings, thoughts and behaviors, but also the psychological and physiological reactions produced by various external stimuli. In the home-based elderly care environment, most elderly people face the increased risk of living alone, conflicts with children or neighbors, poor physical and physiological conditions, and being neglected and lacking psychological care. Therefore, various emotions such as anxiety and depression will exist. If the above emotions persist, it will have a greater negative impact on the daily life of the elderly. The emotions of the elderly occupy an important position in daily life and work. Correctly identifying emotions is of great significance in many fields. Starting from the existing home-based elderly care environment, this application uses the electrocardiogram (ECG) of existing wearable smart watches as the data basis to establish an intelligent emotion assessment model for the elderly based on machine learning.

[0117] 2. Emotion recognition system: Emotion models play a vital role in emotion recognition research, and are mainly divided into two categories: discrete models and continuous models. This application achieves the purpose of evaluating the emotions of elderly people at home by constructing a discrete emotion model, and divides the emotions of the elderly into six types: anger, disgust, fear, happiness, sadness, and surprise, and constructs a learning model for the above six types of data. The emotion recognition system can be used to judge whether the elderly at home belong to one of the above types through the currently collected electrocardiogram ECG data. It has two major advantages: First, through the discrete emotion recognition system, the emotions and psychological states of the elderly are divided into 6 categories, including anger, disgust and fear with negative emotions, as well as happiness with positive emotions, which can more comprehensively reflect the psychological state of the elderly at home.

[0118] 3. Data collection: This application is based on the idea of ​​machine learning. Machine learning is a technology that uses algorithms to allow computers to automatically learn data models and patterns to achieve specific tasks. The main goal of machine learning is to allow computers to automatically obtain data models without explicit programming, so as to identify, classify and predict unknown data. Based on the above ideas, this application collects data for model building for elderly people at home: In order to obtain the electrocardiogram signals with significant characteristics corresponding to the above 6 types of discrete emotions, this application obtains training data through data experiments on 100 elderly people at home. The specific process is: prepare external stimulus signals corresponding to the above 6 types of emotions, and let each subject receive these stimuli in a cycle. These stimuli include: film and television clips or sounds that express specific emotions. In the process of receiving these external stimuli, record the electrocardiogram ECG signal of the corresponding subject, and after each time the corresponding stimulus is received, ask the corresponding subject what emotions he just felt, record these emotions, and match them with the corresponding category stimuli. If the data matches, record the current experimental sample of the current subject as valid data, and extract the corresponding data features to build an integrated learning model. The data collection process is as follows: Figure 9 As shown: First, external stimulation is applied to the subjects, and then the subjects are asked to fill out a questionnaire after the stimulation. If the answers are consistent, they are regarded as valid experimental samples for further feature extraction and classification.

[0119] 4. Feature extraction: This application filters the ECG signal to remove noise, and uses a time-domain and frequency-domain multidimensional feature extraction method to extract corresponding features. For denoising, a variety of commonly used filtering and denoising algorithms used in the field of digital signal processing are used, including low-pass filtering and Butterworth filtering methods. The next step is to perform segmentation on the filtered signal. In order to extract the corresponding time-domain and frequency-domain features, the complete ECG electrocardiogram needs to be segmented and processed to extract several corresponding time-domain features and frequency-domain features. The time-domain features are as follows: P wave: represents the potential change of the depolarization of the atrial muscle of the eigenvalue 1, reflecting the depolarization process of the atrial muscle of the eigenvalue 1. Its morphology and time course abnormalities may be related to the disease of the atrial muscle of the eigenvalue 1. QRS complex: represents the potential change of the depolarization of the ventricular muscle of the eigenvalue 1, and is composed of three closely connected waves: Q wave, R wave and S wave. Its morphology and time course abnormalities often indicate abnormal depolarization of the ventricular muscle of the eigenvalue 1, which may be related to diseases such as ventricular hypertrophy of the eigenvalue 1, myocardial infarction of the eigenvalue 1, and arrhythmia of the eigenvalue 1. The T wave represents the potential changes during ventricular repolarization, particularly late repolarization. Abnormalities in the waveform's shape and direction may be associated with conditions such as ventricular ischemia, ventricular infarction, and electrolyte imbalances. The PR interval is the time interval between the start of the P wave and the start of the QRS complex. A prolonged interval may indicate atrioventricular block, while a shortened interval may be associated with conditions such as preexcitation. The QT interval is the time interval from the start of the QRS complex to the end of the T wave, representing the total time from the onset of ventricular depolarization to the completion of repolarization. A prolonged QT interval may be associated with long QT syndrome, medication effects, or electrolyte imbalances, while a shortened QT interval may be associated with short QT syndrome. Furthermore, the signal is transformed into the frequency domain, extracting the following frequency domain features: Wavelet energy ratio: This is the sum of the last three signals from the three-layer wavelet decomposition to calculate the ratio, yielding the corresponding ratio value. Wavelet Shannon entropy: The Shannon entropy effectively characterizes the uncertainty in the distribution of wavelet coefficients. Fourier transform repeated eigenvalue 1 frequency: The repeated eigenvalue 1 frequency characterizes the primary frequency characteristics of the spectral frequency distribution. Fourier transform average frequency: The average frequency includes not only the primary frequency characteristics of the frequency domain signal, but also a few frequency characteristics. Fourier transform frequency standard deviation: The frequency standard deviation reflects the frequency fluctuation of the frequency domain signal.

[0120] 5. Classification model: This application aims at the emotion recognition model of elderly people at home based on ECG signals. It uses the above-mentioned extracted features to build a multi-level recognition model based on machine learning: According to the guidance of ensemble learning ideas and methods, this paper builds a machine learning model SAB-SVMKNN for emotion recognition. Figure 10The outermost layer of the model uses a Stacking ensemble model, where the first-level learners use a Bagging algorithm and an Adaboost algorithm, respectively. Support vector machine classification models are used for serial and parallel training in both algorithms. Since the output of the first-level learner is a two-dimensional classification result label, the second-level learner uses the KNN classification algorithm to further learn and classify the first-level learner's results. The raw data processed by the feature correlation algorithm is divided into a training set and a test set and fed into the bone model. The current model contains two base learners: one using the Adaboost ensemble algorithm and the other using the Bagging ensemble algorithm. Therefore, the training data is also divided into two parts, each used for training the base learners. The two models are then used to make predictions on the test set data. The predictions are then averaged and used as prediction sample data for the second-level learners. The Stacking structure indicates that the greater the differences between the individual learners that make up the Stacking first-level learner, the lower the risk of overfitting the model for more complex problems. Therefore, using two ensemble learning models with completely different ensemble methods can achieve balanced improvements in accuracy and generalization performance.

[0121] 6. Data acquisition scheme: The device used in this application is a wearable smart watch. As a personal device worn by the elderly on a daily basis, the wearable smart watch is intended to integrate a variety of health monitoring functions. It can monitor key physiological indicators such as heart rate, blood pressure, and blood oxygen saturation in real time, record life data such as number of steps and sleep quality, and synchronize the data to the cloud server through the APP for remote viewing by family members and doctors. Elderly people at home can extract the corresponding ECG electrocardiogram signals by using wearable smart watches or wearable smart rings. In the model construction process of this application, the main user objects selected are elderly people at home. Elderly people at home are more affected by emotions and are more likely to have negative emotions. The number of elderly people at home used in the early model construction of this application is 100. Because it corresponds to 6 categories of quantitative discrete emotion modeling, this application sets up control groups with different emotion levels during data collection. The emotions contained in these control groups include: anger, disgust, fear, happiness, sadness and surprise. Since the number of samples of elderly people at home who are tested for building the model is limited, in order to further improve the accuracy of the emotion monitoring model for elderly people at home, the data collection process will conduct a cyclic test of 6 emotion signals for these 100 samples. In this way, it is equivalent to that every 100 samples are used to build 6 emotion prediction models, and the number of samples is expanded from 100 to 600. The model is more accurate, and the applied stimuli can include various stimulus signals that affect emotions, audio, video, pictures, etc. The ECG signals after stimulation are recorded, and actual emotions and feelings are recorded using questionnaires. This processing method can achieve two benefits: first, the amount of data is expanded through cyclic experiments, from 100 subjects to 600 subjects. The second benefit is that the cyclic emotional stimulation can provide research data for the subsequent emotional continuity of the project, because human emotions will influence each other and are continuous. These data are conducive to optimizing the emotion recognition model that may be developed in the future for elderly people living at home, thereby achieving more accurate emotion prediction for elderly people living at home.

[0122] (2) Model 2: Pulse recognition and judgment classification model based on traditional Chinese medicine theory

[0123] 1. Research purpose: With the changes in society and the times, the proportion of elderly people living at home has gradually increased. Factors such as living alone for a long time, lack of social communication, poor living environment, and eating habits have increased the incidence rate of vascular diseases of elderly people at home. Hypertension, high blood sugar, high blood lipids and excessive obesity are the main reasons why the number of elderly people suffering from vascular diseases of characteristic value 1 at home remains high and has increased instead of decreased. This application uses the wearable smart watch corresponding to the project as a carrier to collect photoelectric volume pulse wave (PPG) signals, and solves the problems existing in the above-mentioned elderly people at home from the perspective of traditional Chinese medicine. In the diagnosis of traditional Chinese medicine pulse, taking the pulse is one of the most commonly used methods of Chinese medicine doctors, which is the "cut" of the four examinations of traditional Chinese medicine "look, listen, ask and feel". This diagnostic method has disadvantages to some extent. The pulse diagnosis of taking the pulse is mostly based on the experience of doctors. In this case, it may be affected by subjective factors, which leads to a large error between the results and the facts. This application automates, standardizes and visualizes pulse diagnosis based on the collected modern medical physiological signals, and uniformly and automatically standardizes the photoelectric volume pulse wave signals collected in modern laboratories and displays them on the application interface.

[0124] 2. Feature extraction: Due to some equipment limitations, the photoelectric volumetric pulse wave signal collected by the front-end device sensor has a lot of interference noise, which has a significant impact on the accuracy of signal feature extraction and the accuracy of TCM pulse classification. Therefore, before conducting the pulse classification algorithm research experiment, it is necessary to process the pulse wave signal and filter out the noise interference to obtain the PPG high-quality wave signal, and then perform signal feature extraction to provide a sufficient and accurate data basis for the data set required by the pulse classification algorithm. There are two main types of noise in the pulse wave signal: one is the low-frequency baseline drift noise caused by normal physiological activities such as human breathing. This type of noise will be eliminated by the Hilbert-Huang transform method; the other is the motion artifact interference noise caused by human movement, such as panting, coughing, talking, shouting, etc. This type of noise will be eliminated by calculating and analyzing the divergence value. The reliability and practicality of the algorithm in the state of human movement are verified through experimental analysis. The signal filtering processing and feature extraction process are as follows: Figure 11 As shown:

[0125] The time domain signal feature is the most intuitive and easy to extract signal feature in the pulse wave signal. Its calculation method is simple and closely related to the waveform of the pulse wave. The morphological features of the PPG signals of different TCM pulses of different people are different. Therefore, the extraction and application of the signal features of the PPG signal in the time domain can help improve the classification accuracy of TCM pulses. After obtaining the pre-processed signal, the time domain and frequency domain features of the PPG signal can be extracted. The pulse wave diagram is shown in the figure below. Figure 12As shown in the figure: the extracted time signal features are t1, t2, t3, and t4; the amplitude signal features are h1, h2, h3, and h4; the area signal features are S1 and S2, corresponding to the area of ​​the rising branch and the area of ​​the falling branch respectively; the slope signal features are s1=h1 / t1 and s2=h2 / t4, corresponding to the slope of the rising branch and the slope of the falling branch respectively.

[0126] The frequency domain signal is the amplitude energy value describing the energy size in each frequency band corresponding to the Hilbert marginal spectrum analysis. Each marginal spectrum energy value is normalized, that is, the ratio of the amplitude energy value on each frequency band to the total of all amplitude energy values ​​is obtained.

[0127] 3. Sample acquisition: This study selected 20 elderly people living at home, carried out the above-mentioned feature extraction in a home-based elderly care environment, and obtained the corresponding pulse characteristics. In the model of this application, six different types of pulse signals are taken into account: flat pulse, thin pulse, slippery pulse, string pulse, astringent pulse, and wiry pulse. Flat pulse: The waveform of the flat pulse has three maximum points, that is, there are three peaks, namely the main wave peak, the tidal wave peak, and the dicrotic wave peak. Ping pulse is the normal pulse of healthy people. Traditional Chinese medicine describes it as "calm and powerful, neither fast nor slow, with a gentle rhythm." Xi pulse: The waveform of Xi pulse has an inconspicuous peak and two obvious peaks: the main wave peak and the dicrotic wave peak. In traditional Chinese medicine theory, Xi pulse is an abnormal pulse. "Thinness means small energy." People with abnormal pulse have thin and small pulses. Compared with Ping pulse, the waveform has small fluctuations. The external manifestations are small and blocked blood vessels, slow blood flow, poor blood flow, poor complexion, insufficient blood and energy, and a weak and cold body. Slippery pulse: The waveform of slippery pulse has two very obvious peaks. The main wave peak is steep and prominent, and the dicrotic wave is wide and long-lasting. In traditional Chinese medicine theory, the slippery pulse is smooth and smooth, "as smooth as a bead", which is a favorable pulse. String pulse: The waveform of string pulse has two peaks with close distances, that is, the main wave peak and the dicrotic wave peak appear at similar times. The amplitude difference is small, and the dicrotic wave is high and wide. It feels "slender, straight, and taut like a violin string," with poor elasticity. This is a precursor to vascular sclerosis, indicating a certain degree of arterial aging. This is often seen in vascular diseases such as vascular sclerosis and coronary artery disease. A wiry pulse: The waveform of a wiry pulse is single-peaked, with a single maximum point at the main wave peak. This peak is wide, and the ascending and descending branches rise and fall gently. This indicates resistance to blood flow in the blood vessels, slowing blood flow and increasing blood viscosity. This is often seen in conditions such as polycythemia and vascular sclerosis. A wiry pulse: The waveform of a wiry pulse has a single main wave peak, but it is narrow. In Traditional Chinese Medicine, a pulse sensation of "empty in the center and full on both sides" indicates a low blood volume, poor vascular saturation, and poor blood flow. This is often seen in conditions such as excessive blood loss, severe vomiting and diarrhea that result in a lack of local blood supply, and severe damage to Yang Qi.

[0128] 4. Model structure: In order to obtain better TCM pulse classification results, the three different types of signal features, namely time domain, time-frequency domain and nonlinear system signal features, are fused to obtain a high-dimensional, multi-category signal feature set to fully express the effective information in the classification target. The feature selection algorithm combining mRMR and SVM-RFE is used to select the fused feature set to obtain the optimal feature subset. The artificial bee colony (ABC) algorithm is used to optimize the parameters of the support vector machine (SVM) to further improve the classification model's recognition performance for pulse classification targets. The overall process of the classification algorithm is as follows: Figure 13 As shown:

[0129] 5. Data Acquisition Solution: This application utilizes wearable smart watches, smart bracelets, and wearable smart rings to collect volumetric pulse wave (PPG) data. As the heart beats, the pulse wave propagates through the peripheral blood vessels. The pulse signal is the most intuitive indicator of pulse condition and vascular health. Therefore, by analyzing the relationship between the pulse wave and the pulse condition, the pulse wave can be used to reflect the health of human blood vessels. Pulse waves can be divided into two types based on their propagation mode: pressure pulse waves and volume pulse waves. Due to differences in blood vessel resistance and diameter, blood circulation velocities vary. When the pulse wave propagates through arteries, it generates fluctuations, which in turn generate pressure. This is called a pressure pulse wave. In contrast, during blood circulation, as blood flows through microvessels, the microvascular blood volume pulse wave exhibits periodic pulse variations under the heart pulse. This process is called a volume pulse wave. The data collected in this study is a volume pulse wave. The main subjects selected as the data source for this application are elderly people living at home. Taking into account the complexity of the modeling experiment and the required sample requirements, the number of samples of elderly people living at home selected in this application is 20. According to the theory of traditional Chinese medicine, pulse is divided into 6 types of pulse, namely flat pulse, fine pulse, slippery pulse, stringy pulse, astringent pulse, and wiry pulse. In modern medicine, a normal pulse indicates that the subject's characteristic value is 1, the internal organs function normally, the blood supply is unobstructed, the vascular endothelium wall is smooth and elastic, and the blood flow is smooth. It is the standard pulse of a normal person. Generally speaking, the blood vessels of the subject with this pulse are very healthy, the qi and blood are sufficient, the veins run smoothly, and the pulse is strong. When an abnormal pulse occurs, it will be reflected in different waveforms of the pulse wave. The pulse wave characteristics mainly involved in the data extraction of this application are the main wave peak, the dicrotic wave peak and the tidal wave peak. According to experiments and corresponding traditional Chinese medicine research, different pulse wave parameters correspond to different pulse characteristics. This application corresponds to 20 research experimental samples for research, of which at least 3 people with 6 different pulse patterns are required. By cyclically collecting the volume pulse waves of the corresponding samples to expand the data set, a more accurate model construction can be achieved.

[0130] (3) Model 3: Classification and evaluation model of sub-health status based on the set theory of traditional Chinese and Western medicine

[0131] 1. Research purpose: With the development of the times, people are paying more and more attention to health issues. According to surveys, among the young and middle-aged people who are in good health, 70% are in sub-health status. This application is aimed at the elderly in the context of home-based care, and their sub-health problems will be more prominent. Traditional Chinese medicine is an important "preventive treatment" and non-invasive diagnosis model, which is very consistent with the definition of sub-health based on personal feelings without obvious symptoms. With the continuous deepening of the objective research on traditional Chinese medicine diagnosis in recent years, it has provided unlimited possibilities for traditional Chinese medicine to identify sub-health status. Based on the problem of home-based care, this application uses a wearable smart watch that can collect volumetric pulse waves PPG, combined with home camera equipment that can take high-definition pictures, to realize a traditional Chinese medicine sub-health filling classification evaluation model based on pulse and tongue.

[0132] 2. Feature extraction of pulse signal: Perform wavelet denoising on the extracted pulse wave signal: Select one-dimensional multi-layer wavelet decomposition technology, use bior4.4 wavelet basis, perform 8-layer wavelet decomposition on the pulse signal, remove 50Hz power frequency and some other high-frequency interference, and then reconstruct it to obtain a more ideal denoised pulse signal. The pulse wave signal is smoothed to remove the influence of burrs, with a smoothing coefficient of 0.9; the periodic average value is obtained through Fourier transform, and a window of length Ls (Ls is the largest odd number less than 0.85*Tx) is set, moving from the starting point of the signal to the end point of the signal. The pulse wave signal segment intercepted by this window is ps = [1, 2, ...,], and when (+1) / 2 is the maximum point and the maximum extreme point of the ps signal segment, it is determined to be the main wave peak point B; then, within the signal segment, on the left side of the horizontal coordinate of B, a small window of 50 points is taken, and the minimum point and the minimum extreme point in the small window is the main wave starting point A; when a pair of points A and B are found, the window continues to slide to the right to form a new window to find new points A and B. After noise removal, the pulse wave signal is further normalized according to the min-max normalization method. This application extracts 32 types of pulse time domain features of the left and right hands (a total of 64 types). The key information of the pulse signal is as follows Figure 14 As shown: the starting point is a, the end point of a single cycle is g, the main wave peak is b, the pre-dicrotic wave trough c and peak d, the descending gorge e, the dicrotic wave peak f, the main wave's fastest rising and falling points a1 and b1, and the dicrotic wave's fastest falling point f1. The peak amplitudes are hb, hd, and hf. The two trough amplitudes are hc and he. The five time information points are the abscissas tb, tc, td, te, and tf corresponding to the three peaks and two troughs, and the cycle length is T.

[0133] 3. Feature extraction of tongue image signals: The tongue image data used in this section comes from 308 tongue images collected by this application in the scenario of elderly people at home, with an image size of 1080*1440. The deep learning technology model used in the experiment requires manual standard data for training. Use Labelme software to label the edges of the tongue. Use the SCAttNet model to segment the tongue. The attention model (AM) was initially used for machine translation, and has gradually developed into an important concept in the field of neural networks. It has a wide range of applications in natural language processing, image recognition and other fields. AM draws on the selective attention mechanism of human vision, that is, the human visual system tends to focus on important information in the image that helps judgment and ignores other useless information. The core idea of ​​AM is to select information that is more critical to the task goal from a lot of information. Figure 15 This is a common AM configuration, and both its encoder and decoder use recurrent neural networks (RNNs). RNNs mainly consist of input layers, hidden layers, and output layers. Its characteristic is that it stores information from previous network layers and applies it to the calculation of the current output layer. In this way, the hidden layers are connected, and the outputs of the hidden and output layers at the previous moment are included in the input. Figure 15 As shown in Figure 3, the attention allocation mechanism is reflected in that the intermediate semantics C is continuously adjusted according to the current output during the progressive process, and each C corresponds to a possibly different attention allocation probability distribution.

[0134] 4. This application uses a semantic segmentation network based on spatial and channel attention (SCAttNet). The network integrates lightweight spatial and channel attention modules, which can adaptively refine features. It consists of two parts: a backbone network for feature extraction and an attention module. The attention module consists of a cascade of a channel attention module and a spatial attention module. After the image is input, the backbone network first extracts features, and then inputs the feature map into the channel attention module for refinement. The refined channel feature map is then input into the spatial attention module for refinement on the spatial axis. Finally, the semantic segmentation result is obtained through convolution and SoftMax operations. The network structure is as follows Figure 16 shown.

[0135] Sub-health identification model: After data processing, this application uses a 116-dimensional pulse feature data set, a 135-dimensional tongue feature data set, and a 6-dimensional human body feature data set. 75% of the data samples are selected as training sets, and the remaining 25% are used as test sets (keeping the ratio of healthy samples and sub-healthy samples unchanged). Four algorithms, SVM, RF, LightGBM, and CatBoost, are used for classification and recognition research. Support vector machines (SVM) are a two-class classification model with strong generalization ability and are suitable for small sample learning. The purpose of SVM is to find a hyperplane to segment samples. LightGBM (Light Gradient Boosting Machine) is an improved GBDT algorithm developed by Microsoft Research Asia. It can be used for tasks such as sorting, classification, and regression. It has the advantages of high accuracy, fast speed, and support for category features. CatBoost is an improved GBDT algorithm released by Yandex, a Russian search giant. It uses symmetric decision trees (oblivious trees) as the base learner. It does not require too many parameters, has high training accuracy, can efficiently process categorical variables, and effectively prevent overfitting. In terms of feature selection, three methods including front LR12 and SVMREF are used for feature selection.

[0136] 6. Data acquisition scheme: When performing data collection in this application, wearable smart watches, smart bracelets and wearable smart rings can be used to realize the data collection of volume pulse wave PPG. As the eigenvalue 1 organ beats, the pulse wave propagates in the peripheral blood vessels, and the complete pulse signal can be collected. This application can use the camera function of the intelligent nursing robot to collect and analyze tongue data. The data collection and analysis process of pulse and tongue can be divided into a modeling stage and an application stage. The application stage is based on the data collection requirements of the modeling stage, including collection content, collection carrier, collection frequency and time, etc. The sample data is input and the results are output. In the data modeling stage, the data acquisition scheme is as follows: 100 middle-aged and elderly people at home who have been diagnosed with various sub-health syndromes by traditional Chinese medicine are selected. Because the requirements for data indicators are too high, the restrictions on sample age are lowered in the data collection modeling stage of this application. The 100 samples were required to contain an equal or relatively even distribution of syndrome types. Data extraction was performed on these 100 samples. The specific method was to record basic information for each sample and extract the corresponding 6-dimensional human feature dataset. Furthermore, 116-dimensional pulse characteristics and 135-dimensional tongue image data were extracted according to the model requirements. Using a cross-validation method, 75 individuals from each sample that met the above criteria were selected as the basis for model establishment. The remaining 25 individuals were used as the corresponding test optimization data for targeted model optimization.

[0137] (4) Model 4: Multi-dimensional fall state classification and body posture recognition based on image recognition, radar recognition, or sensor recognition

[0138] 1. Research purpose: Human posture refers to the various movement modes and states that people present in their daily lives, including regular movement modes such as walking and running, as well as irregular movement states such as sitting, lying, and accidental falls. In addition, for home-based elderly care scenarios, population aging is a social problem that cannot be ignored. According to the sixth national census report, there are more than 170 million elderly people over the age of 60, and empty-nest and single elderly people account for a large number. The elderly's own physical functions decline, and their self-protection ability becomes weaker. Once an accident occurs, if there are no relatives around, it is easy to cause unnecessary consequences. Their health deserves more attention. This application strives to use the wearable portable devices equipped by the project and effective posture recognition methods, so that relatives of elderly people at home can not only understand the elderly's daily exercise status, but also detect accidents (falls) in time to avoid the situation from developing in a bad direction. This application is for wearable portable devices. Using a smartphone with an embedded three-axis acceleration sensor as a data acquisition facility, a wearable portable device is simulated to conduct experimental activities on the daily movement of the human body, collect three-dimensional acceleration data during movement, and propose relevant methods to identify four basic daily movement postures (walking, running, going upstairs and downstairs). The transitions between various postures are further analyzed, and special fall conditions are identified. This application starts with the data in the existing elderly care scenario, selects multiple people to specify relevant actions, and conducts multiple experiments. First, relevant data is collected, the data is analyzed, and finally the data features of the posture are extracted, a mathematical model is established, and a feasible posture recognition method is formed. This method is used to realize the recognition of human body movement postures and the recognition of transitions between different postures.

[0139] 2. Gait feature analysis based on acceleration sensors: When the human body is at rest (such as standing or sitting), the force on the acquisition device is balanced, and the corresponding acceleration data remains almost unchanged, basically its offset value, which is 9.8 in the vertical direction and 0 in the horizontal direction. Regular movements (such as walking and running) are all generated by periodic steps of the lower limbs. The force conditions during stepping will eventually be reflected in the changes in the acceleration data. By analyzing the changing trends of the acceleration signal, we can scientifically identify the different data signal changes presented by different postures, which in turn can be applied to different posture recognition. Figure 17This is a graph of acceleration data from a single stride in walking posture. The data was collected at a frequency of 100 Hz. During a stride, the vertical downward direction is the positive y-axis direction, the horizontal forward direction is the positive x-axis direction, and the horizontal left direction is the positive z-axis direction. Marked points ① and ② in the figure represent the primary force points during a stride. The data in the figure show that before point ① (approximately the first 40 seconds), the three-axis acceleration data remain almost constant, with the y-axis (vertical) at 9.8 and the x- and z-axis (horizontal) at 0. Significant fluctuations occur between points ① and ② (approximately 50-80 seconds). The y-axis exhibits two distinct peaks (approximately 20 seconds) and a trough (approximately 3 seconds). The x-axis exhibits a second distinct peak (approximately 20 seconds), while the z-axis exhibits relatively little fluctuation. Furthermore, the locations of the x-axis peak, the second y-axis peak, and the z-axis maximum are very close (no more than 10 data points, approximately 0.1 seconds apart), indicating they occur at nearly the same time. After relatively obvious fluctuations, the three-axis data simultaneously entered the final relatively static interval.

[0140] The selected data time window should be moderate. Using a shorter rectangular window will not cover the entire stride cycle and will increase computational complexity. Excessively long rectangular windows can lead to significant latency in real-time systems. It's important to note that stride cycle segmentation, discussed later in this chapter, is calculated within the time window. Since the data required for cycle segmentation may fall within the boundaries of the time window, the boundary points of the time window must be excluded during the calculation. Only cycle segmentation points that fall completely within the current time window must be calculated and stored. The resulting data is filtered and denoised to produce preprocessed data. Further movement posture selection is performed to determine the stride feature vectors for walking, running, and stair climbing and descending.

[0141] 3. Analysis of elderly people’s fall identification based on posture recognition: The elderly’s fall process can be regarded as a posture transition, which is different from the natural, purposeful and selective slow sitting and lying of the human body. It is caused by a relatively rapid and unexpected external factor, but the final result is relatively similar. After the human body falls, the heavy eigenvalue 1 no longer undergoes displacement changes, which is a special process of transition from motion to static state. Falling is a transition from motion to static state. It is necessary to first determine the continuous posture before and after. The front posture is a walking posture, and the back posture is a relatively static posture. The fall recognition flow chart is as follows: Figure 18As shown: Its main meaning is: to test the data segment d = {1, 2, ...,} step cycle segmentation, extract the characteristics of each step cycle; according to the level-by-level feature region matching method, identify the step posture of each step, the recognition result v = {1, 2, ...,}; according to the continuous posture judgment, the v obtained in the above is subjected to continuous posture judgment, and the continuous posture segment s = {1, 2, ...,} in this data segment is obtained; the conversion feature of each conversion interval is calculated = v(um,,

[0142] +1); fall discrimination in transition interval;

[0143] The above-mentioned motion posture recognition model can more accurately identify the occurrence of falls among elderly people living at home.

[0144] 4. Data acquisition plan: The acceleration recognition involved in this application can be completed through the project's corresponding smart watch with an acceleration sensor. In the data model establishment stage and application stage, the above-mentioned smart wearable smart watch can be used to complete the collection of relevant data. The specific collection plan involved is: select 100 middle-aged and elderly people. Use millimeter wave radar and wearable smart watches with acceleration sensors to recognize different postures, simulate the posture characteristics of walking, running, going up and down stairs, and use the cross-validation method to extract the acceleration and millimeter wave radar monitoring data of 75 people for data model training. The remaining 25 people are used as the judgment results of the current model verification, and 4 cross-validation cycles are performed to achieve cross-validation and parameter optimization for all samples. For the recognition of fall status, due to safety considerations, 100 young people can set up soft cushions to simulate various fall situations to collect data. Similarly, the data collection and modeling logic should refer to the above-mentioned cross-validation method: 75 young people are used to simulate falls and record relevant data for modeling. The remaining 25 young people were divided into other gait groups and fall combinations according to the proportion for data verification, and four cycles were performed with targeted parameter tuning.

[0145] (5) Model 5: Classification, prediction and identification model of various cardiovascular and cerebrovascular diseases and dangerous conditions based on machine learning methods

[0146] 1. Research purpose: Cardiovascular and cerebrovascular diseases are one of the major diseases that threaten the elderly living at home. They not only bring health and economic burdens to the elderly, but also further increase the difficulties in the development of the country's public health cause. It is necessary to construct a classification model for the dangerous state of cardiovascular and cerebrovascular diseases in the elderly based on various risk factors. The shortcomings of traditional cardiovascular and cerebrovascular management are mainly reflected in: lack of disease prevention. For cardiovascular and cerebrovascular diseases, early prevention is the most effective intervention measure. The elderly in my country lack knowledge about cardiovascular and cerebrovascular diseases. Generally, they choose to go to the hospital for treatment after they have obvious abnormal symptoms. In addition, the hospital mainly treats the elderly with cardiovascular and cerebrovascular diseases. There is a lack of monitoring of the elderly before the disease occurs at home and early warning of various physiological indicators. Another important reason is the low level of informatization. According to the report of the Health Information Center, 40% of the cities and counties in my country have established independent information centers. The staffing of the information centers is relatively insufficient, which has brought great obstacles to the management of cardiovascular and cerebrovascular diseases. The current main informatization means can only record the medical conditions of the elderly who have already developed the disease, and do not cover the collection of the elderly's diet, exercise, and various physiological indicators of the body. The current situation is difficult to meet the personalized needs of the elderly. Hospitals currently adopt a hierarchical management system for sick elderly people and cannot provide specific treatment plans for each patient. This treatment method is not only not conducive to the patient's recovery, but also wastes medical resources. In addition, the elderly have a low level of self-management. Due to the limitations of their living, social, and living environments, they lack relevant knowledge of health management, which seriously affects their later recovery. Doctors cannot monitor patients' various indicators at all times except for periodic review to check for further recurrence. Similarly, community medical resources are not well developed. 80% of rehabilitation and health care measures for cardiovascular and cerebrovascular patients are completed in the community and at home. At the community level, my country lacks excellent general practitioners and medical resources, which reduces the prevention effect of cardiovascular and cerebrovascular diseases. From the perspective of my country's health organizations, community hospitals and other grassroots health organizations are the best institutions for cardiovascular and cerebrovascular prevention and rehabilitation. However, due to the long-term shortage and uneven distribution of medical resources in my country, the community has not played a corresponding role in the management of cardiovascular and cerebrovascular diseases.

[0147] 2. Construction of a cardiovascular classification prediction model: Based on the relevant data collected by the elderly people at home using portable monitoring devices based on electrode patches, wearable smart watches and other devices involved in this application, combined with the elderly people's own medical test and monitoring data, a classification prediction model for cardiovascular and cerebrovascular diseases is constructed. Currently, there are many indicators for scoring cardiovascular and cerebrovascular diseases. The most widely used is the Framingham scoring table established by the US FHS. The probability of cardiovascular and cerebrovascular diseases in the next ten years is predicted based on the test cholesterol level and non-cholesterol level. The specific indicators are shown in Table 11.

[0148] Table 35 Framingham Score (Cardiovascular and Cerebrovascular Risk Score)

[0149]

[0150] Table 29 Cardiovascular and cerebrovascular risk score table

[0151] In addition to the Framingham score, the Essen score can also be used as a scoring system with better results and capabilities, as shown in Table 12:

[0152] Table 30 Essen score table (cardiovascular and cerebrovascular risk score table)

[0153]

[0154]

[0155] Combined with other relevant research and proven medical treatment and diagnosis schemes, this application uses the above indicators to perform subsequent classification and prediction of cardiovascular and cerebrovascular diseases.

[0156] Table 31 Eigenvalue 1 Vascular disease prediction index table adopted by this application

[0157]

[0158] After obtaining the above indicators, we combined them with the corresponding cardiovascular and cerebrovascular disease risk signatures from the 335 elderly individuals identified in this application to establish a multi-level machine learning cardiovascular and cerebrovascular disease prediction model. This model utilized the following machine learning sub-models: logistic classification, decision tree classification, and random forest algorithm. After preprocessing the data, we then established relevant machine learning-based classification and prediction models for various cardiovascular and cerebrovascular disease and risk status based on the selected features.

[0159] 3. Data acquisition scheme: The cardiovascular classification prediction model involved in this application is completed by data collection through portable monitoring equipment of electrode patches and wearable smart watches, and is combined with statistical survey data of information on gender, age, smoking status and biochemical monitoring data such as total cholesterol and high-density lipoprotein, combined with resting heart rate and blood pressure of electrode patches and wearable smart watches, bracelets and rings to model. The specific data acquisition method is: extract 300 samples of elderly people at home with the above-mentioned biochemical indicators, of which 300 samples must also meet the following conditions: According to the cardiovascular score table, the above 300 samples can be divided into three categories of people with low risk, medium risk and high risk of cardiovascular and cerebrovascular diseases, of which 100 samples of elderly people at home with each level of risk are all included, and these samples can be registered to obtain the above-mentioned basic information survey. And it is completed using the wearable smart watch, wearable bracelet, wearable ring or electrode patch portable monitoring device designed in this application. According to the risk prediction model formulated in this application, the variable and immutable factors of the above 300 samples were collected, and matched with the cardiovascular risk labels of the corresponding samples, and the modeling was completed using a multi-class machine learning model. The modeling process was carried out using cross-validation, that is, according to the principle of split sampling, 90% of the samples at different levels of low-risk, medium-risk and high-risk were used each time to build and train the machine learning model, and the remaining 10% of the samples were used for machine learning model parameter evaluation and model optimization. That is, among the above 300 samples, 30 samples were used each time to predict the model parameters and adjust them, and 270 samples were used to train the overall cardiovascular risk prediction machine learning model. This process needs to be repeated 10 times for each unit.

[0160] (6) Model 6: Infection and body inflammation risk identification and classification model using physiological indicators and artificial intelligence related technologies

[0161] 1. Study Objective: Due to their diminished physical condition and immune system, elderly individuals living at home face a higher risk of infection, especially when they develop an illness requiring home care or rehabilitation. Using deep learning models to analyze factors influencing infection risk in these individuals, and to identify perioperative factors, such as home environmental factors, lifestyle habits, and health management practices, that predict and / or promote infection, will facilitate the development of more targeted and effective preventive measures, thereby improving the health of these individuals and reducing their risk of infection and its potential serious consequences. This is an observational cohort study enrolling elderly residents from multiple community-based or home-based care systems in China between 2014 and 2023. Deep learning training was performed using a training to validation data ratio of 8:2. The training data was used to identify factors associated with postoperative infection risk in elderly individuals living at home, while the validation data was used to evaluate the model's effectiveness and accuracy in real-world applications. This application aims to develop and validate a deep learning-based model for predicting the risk of postoperative infection in elderly individuals living at home.

[0162] 2. Infection Indicators and Model Methods: Baseline characteristics of elderly patients were obtained, including sex, current smoker status, American Society of Surgeons (ASA) score, comorbidities (coronary artery disease, heart failure, diabetes, metastatic cancer, cirrhosis, stroke, COPD / asthma, and other), and blood measurements (hemoglobin, serum creatinine, sodium, and white blood cells). Surgical data included surgical procedure (orthopedic, gynecologic, urologic and renal, upper gastrointestinal, lower gastrointestinal, hepatobiliary, vascular, breast, head and neck, plastic and skin, visceral, thoracic, and other), anesthetic technique (general, spinal, epidural, and sedation / local), laparoscopic, cancer surgery, and surgical severity (mild, moderate, and severe). Surgical checklists were also obtained. Postoperative infections included urinary tract infection, bloodstream infection, superficial surgical site infection, deep surgical site infection, body cavity infection, and inflammation. Infections were assessed according to the Centers for Disease Control (CDC) infection definition, and all data were de-identified.

[0163] Based on the obtained data of elderly patients, independent risk factors for postoperative infection were identified by using the inverse probability (IP) weighting method. The correlation derived from these risk factors and the IP weights of postoperative infection was used to construct a traditional logistic regression prediction model. This application used sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV) and AUC to evaluate the predictive ability of the established conventional infection model. In addition, this application also split the original data set into a training data set and a validation data set in a ratio of 3:1 to develop and verify the deep learning model. Among them, various neural network-based postoperative infection prediction models were established based on the training set patient data. The sensitivity, specificity, NPV, PPV and accuracy of various neural network-based models were evaluated based on the test set patient data. This application allocated more patients to the training data set to ensure that the neural network was well trained.

[0164] In this study, deep learning is one of the machine learning models using a multi-layer neural network, whose hierarchical computational design is partly inspired by the structure of biological neurons and is used to generate outputs of the probability of postoperative infectious complications. The output layer consists of response variables. For each neuron, a weight is attached, indicating the effect of the corresponding neuron and all the data passing through the neural network signal. The signal is first processed by integrating all input signals and an activation function to transform the output of the neuron. For a specific neural network, the observed data is used to train the neural network, where the neural network learns an approximation of the relationship by iteratively adjusting its parameters. This application uses a training data set to fit the deep learning model and a validation set to evaluate the predictive performance. Hyperparameter (e.g., weight) tuning is performed iteratively based on the training data set using a backpropagation algorithm to identify the optimal values ​​of the parameters learned during the training process, which are called candidate neural network prediction models. Then, this application selects the best neural network prediction model with the maximum AUC value among these candidate neural networks based on the validation data set. This application uses the Youden index to select the optimal threshold (i.e., sensitivity + specificity - 1). The corresponding sensitivity, specificity, NPV, and PPV are further calculated based on the optimal threshold. Finally, the accuracy of the classification of the probability of postoperative infection in patients was compared using the area under the curve (AUC), with higher AUC values ​​indicating better prediction model accuracy. Statistical analysis data were expressed as the number (percentage) of elderly patients or odds ratios (OR) with 95% confidence intervals (95% CI), as appropriate. All analyses were performed using R software 3.6.2, with the neural network package used to train and establish neural network prediction models and the pROC package used to calculate sensitivity, specificity, NPV, PPV, AUC, and accuracy. Confidence intervals were calculated using the bootstrap method, with 2000 replicates. Statistical significance was set at P < 0.05.

[0165] The corresponding model flow chart is as follows Figure 19As shown, a multi-layer perceptron-based artificial neural network (ANN) reflects the complex functional relationship between risk factors and the response variable using a backpropagation algorithm, logit activation function, and error function, taking into account possible nonlinear associations. Deep learning uses backpropagation to instruct the machine on how to change its internal parameters to predict the optimal desired output of the response variable. This makes ANNs a valuable predictive toolbox. Indeed, deep learning models have been successfully applied in healthcare for predicting clinical events, disease classification, and electronic health record data enhancement. However, the application of deep learning in detecting diseases and complications in elderly patients is limited. In this study, a deep learning approach was applied to assess postoperative infection in elderly patients. Deep learning was trained and validated using a database of multiple elderly patients. Recent studies have also demonstrated the promising performance of deep learning in predicting disease progression. The modeling approach designed in this application provides clinicians with direct and rapid computational guidance for predicting the likelihood of postoperative infection. The deep learning model proposed in this application can be used to identify elderly homebound patients at high risk for postoperative infection. These findings suggest the potential use of this model to help physicians justify interventions that could significantly impact the perioperative management of elderly patients themselves, providing effective health protection for homebound elderly patients.

[0166] 3. Data acquisition plan: The infection and body inflammation risk identification and classification model involved in this application has the following specific data acquisition requirements: 600 samples of various types of elderly people at home who meet the conditions are selected, and infection and body risk prediction models are constructed for these samples. The specific conditions are: chronic disease data, which serve as important influencing features of the body infection and inflammation risk prediction model. The comorbidities of the above 600 samples are collected, including coronary artery disease, heart failure, diabetes, metastatic cancer, cirrhosis, stroke, COPD / asthma and other types of acute and chronic comorbidities. Surgery-related data includes surgery types, such as orthopedic surgery, gynecological surgery, urology and renal surgery, upper gastrointestinal surgery, lower gastrointestinal surgery, hepatobiliary surgery, vascular surgery, breast surgery, head and neck surgery, plastic and skin surgery, visceral surgery, chest surgery, and other unrecorded surgery types. The type and location of surgery are highly correlated with the risk of postoperative infection and inflammation. It is expected that this will contribute significantly to the machine learning model for postoperative risk prediction for elderly people living at home. Anesthesia technology is also a type of surgery-related data and is another important factor affecting surgical success. Anesthesia types can include general anesthesia, spinal anesthesia, epidural anesthesia, and sedation / local anesthesia. In addition, surgeries can be divided into mild, moderate, and severe surgeries according to their severity. According to the actual situation and the relevant research and analysis conducted with the hospital in the early stage of this application, the model prediction label classification for the risk of postoperative infection and inflammation of elderly people at home can be divided into urinary tract infection, bloodstream infection, superficial surgical site infection, deep surgical site infection, body cavity infection and inflammation with a characteristic value of 4. The above 6 types of infection and inflammation basically cover all risk types of infection and inflammation risks related to elderly people at home. Therefore, at the data acquisition and preparation level, it is necessary to find 100 samples of the above 6 different types of infection or combined infection for the calculation, performance evaluation and parameter optimization of the machine learning model. In the actual modeling process, the 10-fold cross-validation method is used to select 10 out of 100 cases of each type of infection for current model performance evaluation and further optimization of model parameters. 90 cases are selected for the construction of various machine learning models.

[0167] (7) Model 7: Sleep state and effect evaluation classification regression model and sleep apnea syndrome identification model based on various physiological data and human sleep characteristics

[0168] 1. Study Objective: Sleep apnea syndrome (SAS) is a sleep disorder characterized by recurring interruptions and cessations of airflow from the mouth and nose for periods exceeding 10 seconds during sleep. The primary effects of SAS are chronic intermittent hypoxia and recurrent arousals during sleep, severely impacting sleep quality and daytime mental well-being. Untreated SAS can lead to complications such as hypertension, coronary disease, and stroke. It is a common health problem among the elderly in my country, but its diagnosis rate is significantly limited by limited hospital resources, currently reaching less than 10%. The apnea-hypopnea index (AHI) is a key indicator of SAS in subjects. Existing methods typically acquire physiological signal data from sleep monitoring, perform noise reduction filtering on the signals, segment the signals, manually extract signal features, and then train a neural network to predict whether a segment is normal or a sleep respiratory event. The AHI is then calculated based on the occurrence of sleep respiratory events. Finally, the severity of SAS in the subject is determined based on the AHI. These methods require signal noise reduction and filtering, resulting in weak signal anti-interference capabilities. Furthermore, since the signal must be segmented, event-labeled data is also required. However, in reality, event-labeled data is not readily available, and this segmentation approach cannot directly diagnose the severity of sleep apnea syndrome. In light of these current situations and challenges, this paper innovatively proposes a solution that aims to automatically extract features of blood oxygen saturation in patients of varying severity through a convolutional neural network, integrating the entire blood oxygen saturation signal with basic physiological parameters. The features extracted by the convolutional neural network are then combined with the subject's vital signs and classified using a fully connected layer, thereby enabling direct prediction of SAS severity and effectively improving the efficiency of SAS identification and diagnosis. This method relies on detailed sleep monitoring data and professional medical diagnosis reports provided by a medical institution, and constructs a deep convolutional neural network model with a precise feature value of 1. In particular, considering the special needs of the elderly in a home environment, the design of this model takes into account both convenience and accuracy, so that even in a home environment, a preliminary assessment of the severity of SAS can be made through data obtained by simple monitoring equipment, providing the elderly with more timely and effective health management and intervention measures.

[0169] 2. Model and Method Introduction: This study introduces an innovative deep convolutional neural network (CNN) approach for health monitoring of elderly individuals living at home. This approach can automatically and efficiently predict the severity of sleep apnea syndrome (SAS) using blood oxygen saturation signals throughout the night and basic vital signs (such as age, gender, height, weight, and body mass index (BMI)). This innovation eliminates the need for complex expert analysis and judgment, nor does it require tedious segmentation of blood oxygen saturation signals. Instead, it directly utilizes up to 7 hours of full-night monitoring data to comprehensively capture respiratory changes during sleep, offering greater versatility. A literature review revealed that many alternative approaches to traditional SAS detection have been developed in recent years. These methods are based on physiological signals such as respiration, blood oxygen saturation signals, snoring, and electrocardiograms. However, all of these methods involve data preprocessing, feature extraction, and feature selection of physiological signals. While feature engineering is crucial, this process requires considerable domain expertise, especially with high-dimensional data. Compared to traditional methods, this approach not only simplifies the operational process but also significantly improves prediction accuracy and efficiency. This technology is particularly suitable for hospitals in second- and third-tier cities with relatively limited medical resources, as it requires only basic blood oxygen monitoring equipment and basic patient vital signs to conduct preliminary screening and assessment of SAS. Furthermore, this method is more versatile and can cover a wider range of elderly people living at home, providing them with more convenient and timely health monitoring services. This effectively promotes the early detection and intervention of SAS, significantly improving the diagnosis rate and safeguarding the health of the elderly.

[0170] 3. Introduction to research content: The data used in this application are actual clinical data provided by the hospital. These data are collected from PSG or portable PSG devices. Each data record includes the subject's blood oxygen saturation signal from the night sleep monitoring and a medical diagnosis report. The diagnosis report contains the doctor's diagnosis of the severity of sleep apnea syndrome. During the training process, the data is divided into training set, validation set and test set in a ratio of 3:1:1. The hospital provides two sets of data, namely sleep monitoring data and medical diagnosis report. The sleep monitoring data contains the physiological signal monitoring data of the subject's night sleep. Depending on the PSG equipment used for collection, the number of channels and sampling frequency of the data are also different. This study uses a single blood oxygen signal as the judgment signal, so the blood oxygen saturation data is first extracted from the sleep monitoring data, and the downsampling frequency is 1Hz.

[0171] Basic vital signs data include the subject's age, gender, height, weight and body mass index (BMI). The diagnostic report contains the subject's AHI value. According to medical standards, patients are divided into four severity levels: normal, mild, moderate and severe. The subject's SAS severity is used as a training label. The subject's blood oxygen saturation signal for 7 hours at night is extracted from the sleep monitoring data, and combined with the vital signs data extracted from the medical diagnostic report as the original data for training the model. For model training, the training data of each sample represents the subject's vital signs data and blood oxygen saturation, and the label of each sample is its SAS severity. Finally, the convolutional neural network is trained based on the original data. See the complete data processing flow for details. Figure 20 .

[0172] A convolutional neural network is used to automatically extract blood oxygen saturation characteristics and basic vital signs for subjects with SAS of varying severity throughout the night. Automatic feature extraction via CNN replaces traditional expert feature extraction. Each recorded blood oxygen saturation signal from a 7-hour nighttime monitoring session is first fed into the model's first input, where it is passed through three convolutional layers to extract the blood oxygen saturation features. Each recorded vital sign data is then fed into the model's second input, where it is passed through a single convolutional layer to extract the vital sign features. Feature fusion is then performed, combining the subject's blood oxygen saturation and vital sign features for classification and prediction. The model's final output is the predicted severity of SAS. This complete model framework.

[0173] In this application, for the elderly population living at home, given the lack of segmentation labels, an innovative method is proposed to use convolutional neural network (CNN) technology to extract blood oxygen saturation and basic signs features to predict the severity of sleep apnea syndrome (SAS). Specifically, the blood oxygen saturation signal is continuously collected for 7 hours from the original night sleep monitoring data as the primary input of the model; at the same time, the basic signs information of the subjects recorded in the doctor's diagnosis report is combined as an auxiliary input (see Figure 21 ). This application designs a structure containing three convolutional layers to deeply mine the key features in the blood oxygen saturation signal. Specifically, an original CNN model is used to automatically extract features from each recorded blood oxygen saturation signal, and the Relu nonlinear activation function is selected as the activation function of the CNN model. At the same time, the KaiMing initialization method is used to initialize the parameters of the convolution layer. This method has a better effect on the model of the nonlinear activation function. For vital sign data, a layer of convolutional layer is used for effective feature extraction. Subsequently, through feature fusion technology, these two types of feature information are organically combined, and finally sent to the fully connected layer and Softmax layer for intelligent classification, thereby outputting the severity assessment of SAS for each elderly person at home.

[0174] This application can provide strong technical support for hospitals and community medical service institutions, and bring good news to the vast number of SAS patients among the elderly living at home. It can help to timely detect and intervene in this potential health risk, effectively alleviate the medical burden caused by delayed diagnosis, and improve the quality of life and health management level of the elderly.

[0175] 3. Data acquisition plan: The sleep state and effect evaluation classification and regression model and the sleep apnea syndrome identification model involved in this application require sleep monitoring data and medical diagnosis reports. The physiological related indicator feature is blood oxygen saturation, which can be achieved through the smart wearable physiological data collection watch of this application, as well as some basic physical information indicator data such as age, gender, height, weight and body mass index (BMI). The above data are used to classify and predict the sleep apnea situation of the elderly. For each sample, the corresponding blood oxygen saturation monitoring throughout the night is used. Because the subjects are always in a sleeping or lying state, the collected blood oxygen data is more accurate and stable. The specific data acquisition quantity requirement for the model construction process corresponding to this application is: obtain 400 samples of middle-aged and elderly people with different degrees of sleep apnea, as well as the corresponding basic information indicators and blood oxygen saturation monitoring indicators corresponding to the whole night. According to the corresponding research requirements of this application, the 400 different samples should include those without sleep apnea problems, as well as those with mild, moderate and severe sleep apnea problems. After the above data is fused with the multimodal features proposed in this application, it is trained through a machine learning model to obtain a corresponding prediction model.

[0176] (8) Model 8: Intelligent Heatstroke or Heatstroke Risk Assessment Model for Elderly People at Home

[0177] 1. Study Objective: In recent years, with the significant intensification of global climate change, the frequency of extreme heat events has shown an alarming upward trend. This phenomenon poses an increasingly severe health threat and challenge to elderly people living at home. High temperatures not only exacerbate fluid and electrolyte imbalances in the elderly but can also lead to serious health problems such as heatstroke and heat stroke, which can be life-threatening. Furthermore, the elderly are often less adaptable to extreme heat due to decreased physical function, weakened thermoregulation, and potential chronic illnesses, making them particularly vulnerable to this climatic phenomenon. Therefore, with the increasing frequency of extreme heat events, predicting the risk of heat stroke has become increasingly important and urgent. This is to ensure a safe living environment for elderly people at home, provide timely and effective heatstroke prevention and cooling measures, and enhance their awareness and self-protection capabilities in hot weather. However, existing predictions of heat stroke-related health impacts have limited representativeness and validation. This study used three years of data from a typical high-temperature city in my country to predict the incidence of heat stroke and tested the importance ranking of model parameters, including meteorological and socioeconomic status (SES) factors.

[0178] 2. Data source: This application is based on the heat stroke risk assessment model for elderly people living at home, and on the high temperature heat stroke reporting system of the Chinese Center for Disease Control and Prevention. It collects heat stroke data from the summer of 2012-2014 in cities such as Shanghai and Chongqing, and combines it with information from the Provincial and Municipal Statistical Yearbook. The model treats cities as categorical variables and incorporates the Baidu search index as a predictor because it is highly correlated with meteorological parameters. Since there is no direct urban Internet penetration rate, provincial data is used instead. At the same time, the NDVI difference and urban impervious area ratio of the Chinese Academy of Sciences' geospatial data cloud are used as geographical variables, combined with meteorological data and socioeconomic statistics (SES) factors to construct a prediction model. All variables are considered to be at least one day behind, in order to achieve the goal of predicting heat stroke risk one day in advance.

[0179] 3. Data Processing: For data preprocessing, missing data were removed from the original dataset. Furthermore, to assess the model's accuracy under high-temperature conditions, the "Regulations on the Administration of Heatstroke Prevention and Cooling" define days with a maximum temperature exceeding 35°C as high temperatures. Therefore, this study only included data on days with a maximum temperature above 35°C. The Boruta algorithm was used to screen variables to minimize computational time for the subsequent random forest model development. The key to introducing the Boruta algorithm is to ensure that the variables in our model are significantly correlated with the outcome variable. By incorporating randomness into the system and collecting results from a collection of random samples, the misleading effects of random fluctuations and correlations can be reduced. The Boruta algorithm has been used in some studies, but not in the field of heatstroke prediction. A random forest method was used to develop a correlation model between the number of heatstroke cases and the selected variables. In this study, 90% of the data was randomly selected as training data. The remaining 10% was used as testing data. To select model parameters, the number of variables (mtry) and number of decision trees (ntree) corresponding to the maximum model R2 were chosen to achieve a better model fit. The correctness of the model was verified through 10-fold cross validation, and the trained model was applied to the experimental data to evaluate the model.

[0180] 4. Model Structure: This application randomly selected 10% of the data from the original dataset for model testing and validation. Model performance was analyzed using a linear fitting method, and the linear R² was used to measure the model's fitness. The Bland-Altman method was also used to calculate the mean difference between the observed and predicted results. The mean difference and the corresponding 95% confidence limits were used to measure the degree of agreement between the predictions and observations. The Bland-Altman method, applied to two sets of data from a single observation, is suitable for evaluating the fitness of the random forest model's predictions. The training data was changed to 80% and 70% randomly selected from the original data, and model performance was evaluated for sensitivity analysis.

[0181] The structure of the model is as follows Figure 22 As shown:

[0182] 5. Data Acquisition Solution: The intelligent heatstroke or heatstroke risk assessment model for elderly people living at home involved in this application requires the collection of physiological health data of the random forest model used in this application, including pulse wave data and blood oxygen saturation data, body temperature and city temperature, and predicts the rate of heatstroke among elderly people living at home. The body temperature data and pulse wave data corresponding to this application can be obtained through the corresponding equipment of this application, that is, wearable intelligent physiological indicator collection watches, bracelets and rings. The data indicators of 400 elderly people living in different urban ambient temperatures are obtained, among which the degree of urban heat is predicted by querying According to the statistical yearbook and corresponding statistical analysis, it is necessary to ensure that the number of samples of the above 400 cases of heat stroke in different cities is the same, and the degree of heat stroke is divided into no heat stroke, mild heat stroke, moderate heat stroke and severe heat stroke. There should be 100 data for each indicator. The multimodal data feature extraction method is used, combined with the time domain and frequency domain analysis feature extraction method, to perform multi-dimensional feature extraction on the pulse wave data corresponding to the heat stroke or heat stroke at the same time of the above samples. Combined with the corresponding city heat degree and blood oxygen saturation value, the degree and situation of heat stroke are used as prediction labels to establish a random forest prediction classification model.

[0183] (9) Model 9: Intelligent Assessment Model for Physiological and Psychological Fatigue

[0184] 1. Purpose of the study: Fatigue is a natural physiological reaction of the human body and is prevalent in all age groups. It is particularly worthy of attention for elderly people living at home. Fatigue can be divided into two categories: mental fatigue and physical fatigue. The former is mostly caused by prolonged thinking, worry or mental stress, while the latter is caused by excessive physical activity or the accumulation of daily housework. After experiencing a long period of fatigue, the body functions of the elderly will gradually weaken. In severe cases, it may even affect the normal functioning of various organs in the body and make it impossible to maintain the physiological functions required for daily life. However, fatigue is also a self-protection mechanism of the body. It warns us that our current physical and mental state is approaching its limit and that corresponding measures need to be taken to avoid further damage to health or even possible life-threatening situations. For elderly people living at home, this means that when they feel tired, they should actively adjust their pace of life, reduce the intensity of housework or activities, and even consider taking a temporary break or stopping certain activities to give the body sufficient time to recover. In response to the health risks faced by elderly people at home due to fatigue, in order to accurately identify their fatigue status and effectively reduce potential operational risks in daily life, this study innovatively combined the Sparrow Search Algorithm (SSA) and the Back Propagation (BP) neural network to construct an efficient fatigue recognition model.

[0185] 2. Research Methods: First, we designed and implemented a series of fatigue-inducing experiments by simulating common daily activities of homebound elderly individuals. During this process, we used the OpenBCI Cyton development kit as an advanced electrocardiogram (ECG) acquisition tool to ensure data accuracy and real-time performance. Subsequently, using MATLAB, a powerful data processing platform, we meticulously preprocessed the acquired ECG signals, including noise filtering and signal enhancement, to generate an initial sample dataset based on different fatigue levels. This step laid a solid foundation for subsequent analysis and modeling. Next, we introduced the Pan-Tompkins algorithm to perform deep feature extraction on the preprocessed ECG signals. This algorithm accurately identifies key waveforms in the ECG signal, such as the P wave, QRS complex, and T wave, thereby extracting characteristic parameters closely related to fatigue status. To further optimize model performance, we performed Pearson correlation coefficient analysis on the extracted characteristic parameters and screened the most representative optimal indicators through hypothesis testing. These optimal indicators not only simplified the model complexity but also significantly improved the accuracy and efficiency of fatigue identification. Ultimately, the SSA algorithm was used to optimize the parameters of the BP neural network, resulting in the construction of an efficient and accurate fatigue recognition model. This model, fed with ECG signal characteristics, can quickly determine fatigue status in home-based elderly individuals and provide a scientific basis for developing personalized health management plans. By accurately identifying fatigue status and effectively mitigating potential health risks, this research provides strong technical support for health monitoring and intervention for home-based elderly individuals.

[0186] 3. Research Content: The human electrocardiogram (ECG) signal is a relatively weak, low-frequency signal. Time-frequency domain feature analysis can yield rich physiological information. The ECG consists of a series of recurring waves, segments, and intervals, with the signal generally concentrated in the low-frequency range of tens of millivolts. Heart rate information for homebound elderly individuals is crucial for identifying their state. Related physiological indicators include heart rate and heart rate variability. Heart rate variability reflects the constant fluctuation between two consecutive characteristic value beats, manifested as slight differences in the intervals between RR peaks in the human ECG. It reflects both the activity of the autonomic nervous system and the quantification of sympathetic and vagal tone, making it sensitive to mental workload and physical fatigue. The standard deviation of normal to normal (SDNN) of the RR peak interval is a typical indicator of fatigue, and there is a significant correlation between its peak interval and fatigue. Combining the Karolinska Sleepiness Scale with a fatigue level based on daily activities of homebound elderly individuals, and incorporating empirical data, four levels of fatigue are identified: alert, mild fatigue, moderate fatigue, and severe fatigue.

[0187] The original physiological signal collection will inevitably be interfered with, mainly including three aspects of interference, such as myoelectricity, baseline drift caused by breathing or limb movement. In order to deal with the above interference, a band-notch filter, a low-pass filter and a zero-phase shift filter are used for signal preprocessing. Through the observation and analysis of the ECG waveform, it is found that the R wave in the QRS complex has a large amplitude, obvious characteristics and is easy to identify. The R position must be determined before analyzing other waveforms. Therefore, the positioning of the R wave is a key step in ECG identification. The Pan-Tompkins algorithm is used to detect the QRS complex and extract features of the ECG signal data collected from the daily activities of the elderly at home, and to detect the fatigue status of the elderly at home in real time. The algorithm flow is shown in Figure 23 After extracting the QRS wave position, heart rate indicators such as mean heart rate (MeanHR) and representative indicators of heart rate variability such as standard deviation of RR interval (SDNN) and ratio of low-frequency band power (LF / HF) were obtained to conduct a study on the mental state recognition of elderly people at home.

[0188] Physiological signal classification is essentially pattern recognition, and the most commonly used method is currently a neural network. The BP neural network is a multi-layer feedforward neural network consisting of an input layer, a hidden layer, and an output layer. Its learning process involves two phases: forward propagation of signals and backward propagation of errors. It is currently the most widely used neural network model. In the forward propagation phase, the signal is input into the neural network, processed by neurons in the hidden layer, and the actual output is obtained at the output terminal. If the output does not match the expected result or deviates significantly, the backward propagation phase begins, where the weights are modified to achieve the desired result. BP neural networks have excellent nonlinear computational capabilities, generalization and fault tolerance, and are easy to build, making them widely used. However, for complex problems such as fatigue identification in elderly people at home, traditional BP algorithms suffer from limitations such as low learning rates, a lack of effective learning rate selection methods, and the potential for local minima and non-convergence during training. To address these issues, the sparrow search algorithm (SSA) is used to optimize initial weights and thresholds to enhance the robustness of the neural network and improve the accuracy of mental fatigue prediction results. Figure 24 The process of optimizing the BP neural network algorithm by the sparrow search algorithm is shown. Each individual in the population has complete weights and thresholds, and the mean square error is used as the fitness function to obtain the fitness of each individual, and the optimal individual is screened and determined.

[0189] 4. Data Acquisition Plan: The fatigue assessment model for elderly people at home involved in this application requires the use of data: completed through the wearable smart ECG acquisition device and the electrode patch-based ECG acquisition device involved in this application. The human body will produce different differences in ECG signals at different levels of fatigue. These differences can be acquired by the device and analyzed by software and machine learning models. The number of samples of ECG data collected is 400. By designing experiments to measure different levels of fatigue, different fatigue stimulation experiments are designed so that the above 400 data have four levels: awake, mild fatigue, moderate fatigue and severe fatigue. The above levels need to have an equal number of samples at each level, that is, 100 samples need to be divided into each level. After obtaining the above samples, the original data needs to be processed. Due to the limitations of the equipment level, the original physiological signals collected from the wearable smart watch will be subject to various interferences. The main interferences include myoelectric interference, respiratory interference or baseline drift caused by limb movement interference. Certain pre-processing measures are taken for the interference to extract the corresponding waveforms with obvious characteristics. The QRS wave position corresponding to the current sample can be extracted, and the above information combined with the corresponding data to obtain the sample content can be used to obtain the corresponding fatigue prediction model.

[0190] (10) Model 10: Machine learning sports risk level classification assessment model based on various basic physiological indicators and intelligent model evaluation indicators

[0191] 1. Research Purpose: Monitoring activity in older adults is crucial for assessing their mobility and predicting their activity risk. Heart rate variability (HRV), as a measure of autonomic nervous system activity, provides valuable information for balancing activity and rest in older adults, attracting widespread attention and research in the field of exercise science. Heart rate variability, as a metric, can be calculated in a variety of ways: It refers to the variation in the beat-to-beat characteristic value (HRV) between cycles. HRV, derived through machine analysis, is often used as an indicator of autonomic nervous system tone. HRV decreases when sympathetic nervous system activity increases, while it increases when parasympathetic nervous system activity increases. The contact-type PPG watch corresponding to this application can measure PRV, a pulse rate variability indicator, in older adults in different home environments and during exercise. Because heart rate and pulse rate are highly correlated, pulse rate variability can be used to infer heart rate variability (HRV). Heart rate variability is associated with physical adaptability, vascular health, and stress response. Higher heart rate variability indicates greater flexibility and the ability to quickly adjust heart rate in various situations to meet the body's needs. It plays an important role in sports training monitoring.

[0192] 2. Establishing an Exercise Risk Model: Currently, commonly used instruments for measuring heart rate variability include wearable smartwatches developed in this application or specialized equipment such as multi-lead electrocardiograms (ECGs) corresponding to medical-grade life monitors. The choice of instrument can affect the accuracy of heart rate variability parameters, and standardized procedures should be adopted based on actual needs. To extract heart rate variability characteristics, three analysis methods are commonly used: time domain, frequency domain, and nonlinear dynamics. Time domain analysis is the most basic statistical analysis method, often including statistical parameters such as SDNN, RMSSD, and PNN50, which reflect the state and activity of the autonomic nervous system. Frequency domain analysis uses power spectrum analysis to reflect the energy distribution of the heart rate variability signal, with commonly used indicators such as LF, HF, and LF / HF. Nonlinear analysis further reveals the complex dynamic characteristics of heart rate variability. These methods provide multidimensional heart rate variability information, which helps to more comprehensively understand the current physiological status of older adults and can be used to assess their exercise status, predict the risk of excessive exercise, and optimize exercise plans. Excessive exercise is a common problem that exists in the exercise and life of the elderly. It may cause serious physical injuries to the elderly or cause serious problems such as sudden cardiovascular and cerebrovascular events. By using the equipment used in this application to conduct long-term monitoring of the HRV of the elderly who go out for exercise, it is possible to track the activity level of their autonomic nervous system and discover the impact of exercise on the autonomic nervous system, thereby helping to assess the current exercise load and exercise risk of the elderly. When the exercise load increases significantly, the relevant indicators corresponding to HRV will continue to decrease, reflecting that the elderly’s adaptability to exercise has declined rapidly, and the possibility of various risk accidents has increased rapidly. In addition, researchers have begun to explore how to use heart rate variability in combination with other indicators to more comprehensively monitor the physical condition of the elderly. This comprehensive approach helps to provide more accurate information. For example, combining heart rate variability with relevant biochemical indicators can further infer the elderly’s current fatigue state and the risk of subsequent exercise adaptability problems, thereby providing timely and reasonable exercise advice to the elderly. The elderly should avoid exercise that is too intense.

[0193] 3. Data acquisition plan: The main content and objectives of this application are: to establish a machine learning exercise risk level classification assessment model for middle-aged and elderly people based on various basic physiological indicators and intelligent model evaluation indicators. Among them, the most critical indicator for the current project is HRV, namely heart rate variability. Heart rate variability can be achieved through multiple physiological data acquisition devices corresponding to this application: wearable smart watches, smart bracelets and smart rings that measure volume pulse waves, smart electrode patch physiological data acquisition devices and wearable smart watches that can measure electrocardiogram signals. The above devices can be used simultaneously in model prediction and deployment, and several devices that meet the physiological acquisition category conditions can also be used for model prediction and deployment. The data acquisition sample size required in this application is 400 middle-aged and elderly people who frequently exercise outdoors or indoors, and the corresponding electrocardiogram signals and pulse wave signals are collected. Because the pulse rate variability generated by measuring the pulse can replace the heart rate variability to a certain extent, the above physiological indicators can be flexibly switched in subsequent model construction and various actual application scenarios. The exercise risks of the above 400 middle-aged and elderly people were rated as no risk, low risk, medium risk and high risk. Kinematics experts were sent to work with doctors to judge the risk status before and after exercise, and the corresponding exercise risk level was calculated based on heart rate variability.

[0194] (11) Model 11: An intelligent auxiliary diagnosis model for various TCM syndromes that combines artificial intelligence technology with TCM theory.

[0195] 1. Research purpose: China's pulse diagnosis is a human body information diagnosis system. Pulse information comprehensively reflects the process of human body metabolism of matter, energy and information, and is a program for regulating human body functions. Pulse diagnosis is the most distinctive diagnostic method in traditional Chinese medicine. It has a long history and rich content. It is the embodiment and application of the basic spirit of "holistic concept" and "differentiation and treatment" of traditional Chinese medicine, and is also an indispensable part of the theoretical system of traditional Chinese medicine. The pulse carries a wealth of information about the health status of the human body. Since the publication of the earliest pulse theory monograph "Pulse Classic" in my country in the third century AD, pulse theory has been continuously enriched and improved, and has had a great impact on the development of medicine at home and abroad. At present, with the development of traditional Chinese medicine, pulse diagnosis, as a means and method of non-invasive detection, has been appreciated and paid attention to by people at home and abroad. This application attempts to objectify traditional Chinese medicine pulse diagnosis, use the corresponding equipment and equipment of the project to collect corresponding physiological data and information, and then establish an auxiliary diagnosis model for various syndromes of traditional Chinese medicine.

[0196] 2. Introduction to Chinese medicine pulse diagram: Pulse is actually the image of the pulse, that is, the shape of the pulse. Pulse wave PPG is the main manifestation of the pulse signal. The pulse waveform contains the physiological information of the pulse. The study of pulse mainly takes the pulse wave as the research object. The pulse waveform is closely related to the functional state, physiological and pathological changes of the eigenvalue 1 blood vessels. The human body's blood circulation system is an elastic pipe system filled with blood. The human body's blood circulation process reflects the changes in two physical fields, hemodynamics and hemorheology. With the contraction and relaxation of the eigenvalue 1 organs, the pulsation of the blood vessels can be touched and felt in the superficial part of the arteries. Its period is the same as the eigenvalue 1 beat, which is called a pulse. Every time the eigenvalue 1 organ contracts and relaxes, the arterial system produces a change in pressure and blood flow, that is, a pulse wave. Such as Figure 25 As shown, the amplitude and corresponding characteristics of the pulse wave graph are as follows:

[0197] Time-domain analysis primarily analyzes the relationship between pulse wave amplitude and pulse phase. The time-domain parameters of the pulse graph and their corresponding physiological significance are as follows: h1: Main wave amplitude, defined as the height from the main wave peak to the pulse wave baseline (the baseline is parallel to the time axis). This primarily reflects the ejection function of the left ventricle and the compliance of the aorta. h2: Main wave gorge amplitude, defined as the trough between the main wave and the pre-dicrotic wave. Its physiological significance is the same as h3 and can be ignored in pulse graph analysis. h3: Pre-dicrotic wave amplitude, defined as the height from the pre-dicrotic wave peak to the pulse wave baseline. This primarily reflects arterial elasticity and peripheral resistance. h4: Descending mid-valley amplitude, defined as the height from the bottom of the descending mid-valley to the pulse waveform baseline. This primarily reflects the magnitude of arterial peripheral resistance. h5: Dicrotic wave amplitude, defined as the height between the dicrotic wave peak and the bottom of the descending mid-valley parallel to the baseline. This reflects the elasticity (compliance) of the aorta. t1: The time value from the starting point of the pulse wave graph to the main wave peak, corresponding to the rapid ejection period of the left eigenvalue 1 chamber. t4: The time value from the starting point of the pulse wave graph to the descending isthmus, corresponding to the systolic period of the left eigenvalue 1 chamber. t6: The time value from the descending isthmus to the end point of the pulse wave graph, corresponding to the diastolic period of the left eigenvalue 1 chamber. t: The time value from the starting point to the end point of the pulse wave, corresponding to one eigenvalue 1 cycle of the left eigenvalue 1 chamber. w: The width of 1 / 3 of the main wave, equivalent to the time the high pressure level in the artery is maintained. In the statistics of various physiological parameters, for the amplitude parameter, the absolute value of the amplitude h is directly taken only when comparing the differences in pulse force between test groups. To better reflect the pulse characteristics and vascular status, the relative ratios of various amplitude parameters are generally used, such as h3 / hl, h4 / hl, h5 / hl, (hl-h3) / hl, t1 / t, t1 / t4, (t4-t1) / t, t5 / t4, and w / t. By analyzing the pulse amplitude and duration, we can understand the pulse frequency and rhythm, pulse strength, pulse strength, and pulse morphology.

[0198] 3. Introduction to the ten major syndromes in traditional Chinese medicine: Based on the sub-health physiological phenomena and diseases that often occur in elderly people at home, and combined with the actual needs of this application, the syndromes obtained from traditional Chinese medicine pulse are divided into: liver qi stagnation syndrome, liver qi stagnation characteristic value 5 deficiency syndrome, characteristic value 1 characteristic value 5 deficiency syndrome, liver and kidney yin deficiency syndrome, characteristic value 4 characteristic value 5 qi deficiency syndrome, characteristic value 5 deficiency and dampness obstruction syndrome, phlegm-heat internal disturbance syndrome, characteristic value 1 kidney disharmony syndrome, qi and blood deficiency syndrome, and damp-heat accumulation syndrome.

[0199] 4. Pulse Feature Extraction: In addition to the aforementioned time-domain features corresponding to pulse patterns, this application also extracts frequency-domain features for the classification of TCM syndrome types. Frequency-domain feature data primarily involves converting temporary waveform data into frequency-domain signals using wavelet and Fourier transforms. The wavelet transform is a signal analysis tool that represents a signal by breaking it down into a series of wavelet basis functions. Unlike the Fourier transform, which decomposes a signal into a series of sine waves, the wavelet basis functions used in the wavelet transform have finite durations and varying frequencies, making them more suitable for analyzing unstable and transient signals. This characteristic allows the wavelet transform to better capture transient changes and frequency information in the signal, facilitating more accurate analysis of the characteristics of non-stationary signals. The collected pulse waveform is subjected to a three-layer wavelet decomposition, resulting in a four-layer wavelet decomposition result. From this result, the time-frequency feature data that can be extracted include: 1. Wavelet energy ratio, which is the sum of the last three signals from the three-layer wavelet decomposition to calculate the ratio and obtain the corresponding ratio value. 2. Wavelet Shannon entropy: The Shannon entropy effectively characterizes the uncertainty of the wavelet coefficient distribution.

[0200] 5. Pulse syndrome feature recognition model: Based on the above situation, the overall system structure flow chart of TCM pulse syndrome recognition is as follows: Figure 26 As shown:

[0201] As can be seen from the figure, to effectively identify pulse syndromes, the following basic steps must be included: 1. Preprocessing and feature extraction of the pulse signal, i.e., extracting parameters that can effectively characterize the individual characteristics of the TCM syndrome type. Currently, there is no technology that can accurately extract the individual characteristic parameters of the TCM pulse syndrome type. 2. Modeling of the TCM pulse syndrome type and training of the model parameters. This includes the representation of the model structure and the parameter estimation algorithm. 3. Calculation of the matching distance between the pulse signal to be tested and the TCM syndrome type model. In fact, the TCM syndrome type model structure corresponds to different methods for pulse signal identification. How to define the measure of similarity so that the calculation of similarity is both simple and reliable is a further problem to be solved. 4. Identification or judgment strategy. Based on the calculation result of the matching distance, it is judged whether the pulse signal to be tested is the claimed TCM syndrome type or which type of TCM syndrome type the pulse signal to be tested belongs to. After obtaining the corresponding pulse characteristics, this application uses a Gaussian mixture model (GMM) to model the input data. The probability density function of any shape can be approximated by a GMM with multiple mixed numbers. When GMM is used for pulse signal recognition, the probability density function of the short-time spectrum feature vector of the pulse signal of different syndrome types is modeled, and then training is performed to obtain the mixed Gaussian probability density function of each syndrome type, which is used as the template for each TCM syndrome type. During recognition, the observed feature vector sequence is substituted into the template of each syndrome type, and the maximum posterior probability calculated corresponds to the identified TCM syndrome type. The commonly used parameter estimation of the GMM model is based on the maximum likelihood criterion. Maximum likelihood estimation treats the quantity to be estimated as a fixed but unknown quantity, and then finds the parameter value that maximizes the probability of the learning sample appearing, and uses it as the parameter estimate.

[0202] 6. Data acquisition plan: The main content and objectives of this application are: an intelligent auxiliary diagnosis model for various TCM syndromes that combines artificial intelligence technology with TCM theory, combines TCM syndromes with various basic physiological indicators, and derives corresponding TCM syndromes based on the corresponding indicators. The physiological data used in this application is volume pulse wave PPG, and relevant physiological monitoring equipment can be used. In the actual data acquisition process, the project also used wearable smart watches, wearable smart bracelets, and wearable smart rings to complete the extraction of pulse waves. This application is based on the diagnosis and patient resources of TCM hospitals, and extracts 500 people as a pulse collection sample data set, where each corresponding syndrome sample collects actual pulse data from nighttime sleep for analysis. Because the syndrome evaluation involves 10 common syndromes divided by TCM, when extracting TCM diagnostic data, it is necessary to perform stratified and uniform extraction. 50 people are extracted for each syndrome for experimental verification, and pulse wave data for each person is collected for 10 days. The above-mentioned feature extraction and machine learning methods are used for data modeling.

[0203] 3.3.2.2 Technical Features

[0204] This application has built an intelligent monitoring device for the elderly at home, which is divided into an acquisition subsystem, a transmission subsystem, a server subsystem, a visualization application subsystem, and an AI algorithm processing subsystem. From data acquisition, data transmission, data storage to data visualization, this application has formed a relatively complete product ecosystem.

[0205] The products developed in this application are used in medical monitoring or surveillance scenarios. Some of them are disposable consumables, and most of them are consumable products, which are prone to repeated purchases. In addition, due to the universality of the equipment, any social scenario that requires the collection of vital signs is suitable for the use of this system, that is, it has strong applicability.

[0206] We are backed by the Key Laboratory of Geriatrics and Life Support of the Ministry of Education. Under the guidance of a number of experienced and accomplished clinical experts, we have formed a highly skilled and powerful medical and engineering cross-team. The technical advantages of this application are mainly concentrated in two key aspects: the calibration algorithm of vital signs data and the communication protocol of the wireless vital signs medical database. In terms of the calibration algorithm of vital signs data, when using front-end acquisition circuits such as ADS1292, max30205, and max30102 for data acquisition, it is often affected by various factors such as power frequency noise and electromyographic interference. However, by applying the algorithm developed in this application, the adverse effects of these interferences on the patient's physiological signals can be significantly reduced, and the accuracy of the patient's signals can be greatly improved. For example, in practical applications, motion artifact suppression algorithms can reduce muscle tremors and changes in vascular filling state caused by the subject's movement. These conditions are random and unpredictable. This system proposes a new adaptive motion artifact elimination method. This method does not require prior knowledge of the statistical characteristics of the interfering noise. It utilizes the uncorrelated nature of the interfering signal and the detected signal to adaptively adjust the filter's transmission characteristics, minimizing the impact of motion artifacts. When the statistical characteristics of the input electrophysiological signal are unknown or change, the adaptive filter can automatically and iteratively adjust its own filter parameters to meet certain criteria, thereby achieving optimal filtering.

[0207] Regarding the communication protocol for the wireless vital signs medical database, leveraging advanced 5G transmission technology and building upon the TLV transmission protocol framework, a unique transmission protocol specifically designed for vital signs data transmission has been developed. This innovative approach significantly enhances the security and stability of data transmission. For example, by employing encryption technology and optimizing data encapsulation formats, the confidentiality and integrity of vital signs data during transmission are ensured, effectively mitigating the risk of data leakage and loss, and providing a solid foundation for the secure transmission of medical data.

[0208] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A elderly care service robot, comprising a base (1), characterized in that: A robot housing (2) is fixedly mounted on the top of the base (1), a main controller (3) is fixedly mounted on the top of the robot housing (2), a placement slot (4) is provided on the front of the robot housing (2), a wristband (5) is placed inside the placement slot (4), a protective component (6) is installed inside the robot housing (2), a camera (7) is fixedly mounted on the top of the main controller (3), an alarm (8) is fixedly mounted on the top of the robot housing (2), and a positioning module (9), a processing module (10), a wireless module (11) and a storage module (12) are fixedly mounted inside the robot housing (2); A vital sign data acquisition module (501) is fixedly installed inside the wristband (5); The protective component (6) comprises a heat dissipation slot (601) provided on the right outer surface of the robot housing (2); two U-shaped frames (602) are fixedly mounted on the right outer surface of the robot housing (2); a dustproof plate (603) is slidably mounted between the two U-shaped frames (602); connecting strips (604) are fixedly mounted on the left and right sides of the dustproof plate (603); and a heat dissipation fan (605) is fixedly mounted on the left inner wall of the robot housing (2).

2. The elderly care service robot according to claim 1, characterized in that: The two U-shaped frames (602) are located on the left and right sides of the heat dissipation slot (601), and the two connecting bars (604) are both slidably connected to the inside of the U-shaped frames (602); A ventilation hole is provided on the left outer surface of the robot housing (2), and a cooling fan (605) is fixedly installed inside the ventilation hole. A steel mesh is fixedly connected inside the ventilation hole and on the left side of the cooling fan (605).

3. The elderly care service robot according to claim 1, characterized in that: The main controller (3) is electrically connected to the vital sign data acquisition module (501), the wireless module (11), the positioning module (9), the camera (7), the processing module (10), the alarm (8) and the storage module (12); The wireless module (11) includes a 4G module, a 5G module, a WIFI module and a Bluetooth module, and the positioning module (9) includes a GPS module; The vital sign data acquisition module (501) is electrically connected to the processing module (10), and the processing module (10) is electrically connected to the alarm (8).

4. The elderly care service robot according to claim 1, characterized in that: Also includes: The home-based elderly emotion recognition model is configured to classify the elderly's emotions into six types: anger, disgust, fear, happiness, sadness, and surprise. A learning model is built based on the data of the above six types, and the currently collected ECG data is used to determine whether the elderly at home are in one of the above types. The pulse recognition and classification model is configured to process the pulse wave signal, filter out noise interference to obtain a high-quality PPG wave signal, and then extract signal features to provide a sufficiently accurate data basis for the data set required by the pulse classification algorithm; use a feature selection algorithm combining mRMR and SVM-RFE to select features from the fused feature set to obtain the optimal feature subset, and use an artificial bee colony algorithm to optimize the parameters of the support vector machine; The sub-health status classification and assessment model is configured as follows: using a 116-dimensional pulse feature dataset, a 135-dimensional tongue feature dataset, and a 6-dimensional human body feature dataset, 75% of the data samples are selected as training sets and the remaining 25% as test sets, and four algorithms, SVM, RF, LightGBM, and CatBoost, are used for classification and recognition research. In terms of feature selection, LR12 and SVMRE are used for feature selection.

5. The elderly care service robot according to claim 4, characterized in that: Also includes: A multi-dimensional fall state classification and body posture recognition model is configured to: analyze the changing trends of acceleration signals, identify the different data signal changes presented by different postures, obtain different postures, determine the characteristic vectors of walking, running, climbing stairs, and descending stairs, and determine whether they meet the characteristics of falling; The classification prediction and identification models for various cardiovascular and cerebrovascular diseases and dangerous conditions are configured as follows: including the logistic classification model, the decision tree classification model, and the random forest algorithm. After preprocessing the data, the relevant classification prediction and identification models for various cardiovascular and cerebrovascular diseases and dangerous conditions based on machine learning methods are established according to the selected features.

6. The elderly care service robot according to claim 4, characterized in that: Also includes: The infection and body inflammation risk identification and classification model is configured as follows: baseline characteristics of elderly patients are obtained, including gender, current smoker, ASA score, comorbidities, and blood measurements; surgical-related data include surgical procedures, anesthesia techniques, laparoscopic surgery, cancer surgery, and surgical severity; surgical checklists are obtained; postoperative infections include urinary tract infection, bloodstream infection, superficial surgical site infection, deep surgical site infection, body cavity infection, and characteristic value 4 inflammation; and infection is assessed according to the definition of infection. All data are guaranteed to be desensitized; Based on the obtained data of elderly patients, independent risk factors for postoperative infection were identified by using the inverse probability weighting method, and the correlations derived from these risk factors and the IP weights for postoperative infection were used to construct a traditional logistic regression prediction model; The sleep state and effect assessment classification and regression model, as well as the sleep apnea syndrome identification model, are configured as follows: Blood oxygen saturation signals for seven hours of sleep are extracted from sleep monitoring data, combined with vital sign data extracted from medical diagnosis reports as the raw data for training the model. For model training, the training data for each sample represents the subject's vital sign data and blood oxygen saturation, and the label of each sample is its SAS severity. The convolutional neural network is trained based on the raw data. A convolutional neural network is used to automatically extract the blood oxygen saturation characteristics and basic vital signs of SAS subjects with different degrees of severity throughout the night. The blood oxygen saturation signal of each record monitored for 7 hours at night is input into the first input of the model. The blood oxygen saturation features are extracted through three convolutional layers. The vital sign data of each record is then input into the second input of the model. The vital sign features are extracted through one convolutional layer. Feature fusion is then performed, and classification prediction is performed based on the subject's blood oxygen saturation characteristics and vital sign characteristics. The predicted SAS severity is finally output.

7. The elderly care service robot according to claim 4, characterized in that: Also includes: An intelligent risk assessment model for heatstroke or heatstroke among elderly people at home is configured as follows: 10% of the data are randomly selected from the original dataset for model testing and validation, the model performance is analyzed using a linear fitting method, the fitness of the model is measured using linear R2, the mean difference between the observed and predicted results is calculated using the Bland-Altman method, the degree of consistency between the prediction and observation is measured using the mean difference and the corresponding 95% confidence limit, the training data is changed to 80% and 70% randomly selected from the original data, and the model performance is evaluated for sensitivity analysis.

8. The elderly care service robot according to claim 4, characterized in that: Also includes: The intelligent evaluation model is configured as follows: using the OpenBCI Cyton development kit as an advanced ECG signal acquisition tool, and using the MATLAB data processing platform to preprocess the collected ECG signals, including noise filtering and signal enhancement, to form an initial sample data set based on different fatigue levels; using the Pan-Tompkins algorithm to perform deep feature extraction on the preprocessed ECG signals, performing Pearson correlation coefficient analysis on the extracted feature parameters, and screening out the most representative preferred indicators through hypothesis testing, and using the SSA algorithm to optimize the parameters of the BP neural network to construct a fatigue recognition model. By inputting the ECG signal characteristics of elderly people at home, their fatigue status can be quickly judged.

9. The elderly care service robot according to claim 4, characterized in that: Also includes: The machine learning exercise risk level classification assessment model is configured as follows: the exercise risk of the elderly is rated as no risk, low risk, medium risk and high risk, the risk status before and after exercise is judged, and the corresponding exercise risk level is calculated in combination with heart rate variability. The HRV of the elderly who exercise outdoors is monitored over a long period of time, the activity level of their autonomic nervous system is tracked, the impact of exercise on the autonomic nervous system is discovered, and the current exercise load and exercise risk of the elderly are assessed. When the exercise load increases significantly, the relevant indicators corresponding to HRV will continue to decrease, reflecting that the elderly's adaptability to exercise is rapidly declining and the possibility of various risk accidents is rapidly increasing.

10. The elderly care service robot according to claim 4, characterized in that: Also includes: The intelligent auxiliary diagnosis model for various TCM syndromes is configured as follows: extracting parameters that can effectively characterize the individual characteristics of TCM syndromes; establishing a TCM pulse syndrome model and training model parameters; calculating the matching distance between the pulse signal to be tested and the TCM syndrome model; judging whether the pulse signal to be tested is the claimed TCM syndrome or which type of TCM syndrome the pulse signal is based on the calculation result of the matching distance. After obtaining the corresponding pulse characteristics, the Gaussian mixture model is used to model the input data. The probability density function of any shape can be approximated by multiple mixed numbers. GMM is used for pulse signals. During recognition, the probability density function of the short-time spectrum feature vector of the pulse signal of different syndrome types is modeled, and then training is performed to obtain the mixed Gaussian probability density function of each syndrome type, which is used as the template of each TCM syndrome type; during identification, the observed feature vector sequence is substituted into the template of each syndrome type, and the calculated maximum posterior probability corresponds to the identified TCM syndrome type. The commonly used parameter estimation of the GMM model is based on the maximum likelihood criterion. The maximum likelihood estimation is to regard the quantity to be estimated as a fixed but unknown quantity, and to find the parameter value that can maximize the probability of the learning sample appearing, and use it as the parameter estimate.

Images