Old people health risk prediction method and system based on emotion-physiology-gait three-dimensional coupling

Abstract

The invention discloses an emotion-physiology-gait three-dimensional coupling-based health risk prediction method and system for old people, and the method comprises the steps: collecting the facial image data, voice data, heart rate change and blood pressure data of a user, and the gait data and acceleration data of the user in an advancing process; extracting user emotion features, user physiological health features and user motion features based on the data, establishing an emotion recognition model, inputting a recognition result of the emotion recognition model, collecting physiological indexes and motion stability parameters by combining heart rate, blood pressure and gait sensors, and establishing a risk prediction model; and carrying out multi-modal data fusion in the edge calculation unit, and outputting the health risk index and psychological abnormality early warning of the old people. According to the method, cross-modal deep correlation inference is realized, the problems that traditional single-modal monitoring is unstable, the false alarm rate is high and risks cannot be recognized in advance are solved, and the intelligence and safety of elderly accompanying equipment are remarkably improved.

Description

Technical Field

[0001] This invention discloses a method and system for predicting health risks in the elderly, specifically a method and system for predicting health risks in the elderly based on the three-dimensional coupling of emotion, physiology and gait, belonging to the field of intelligent health monitoring technology. Background Technology

[0002] Falls, emotional instability, and sudden health abnormalities in the elderly are monitored using health monitoring devices, but these devices generally have the following problems:

[0003] 1) Single-modal monitoring is susceptible to interference and has low reliability. Traditional devices typically rely on only: physiological signals such as heart rate / blood pressure; single behavior; or single camera monitoring. However, single-modal signals are greatly affected by posture, lighting, environmental noise, sensor position, etc., resulting in high false alarm and false negative rates.

[0004] 2) Health monitoring devices cannot reflect the comprehensive risks of the elderly's true condition. Studies show that emotional fluctuations can lead to changes in heart rate and blood pressure, and physiological abnormalities can affect gait stability. Gait changes can reflect physiological stress or mental tension. Traditional methods cannot identify this "cross-emotion-physiology-gait" correlation chain, and therefore cannot predict risks in advance.

[0005] 3) Existing technologies lack proactive prediction mechanisms, and most devices can only "alarm after the fact." For example, they only alarm after a fall, only alert when physiological indicators reach dangerous levels, and only alert when emotions are agitated but not associated with health risks, thus lacking early warning capabilities.

[0006] 4) The existing walking aid scenarios have not formed a truly intelligent monitoring closed loop. Existing walking aids mainly focus on the carrier structure or single function enhancement, lacking proactive analysis of the user's status.

[0007] In summary, existing technologies cannot address the lack of exploration of the dynamic coupling relationship between the three types of signals—emotion, physiology, and gait—in health monitoring of the elderly, thus hindering early prediction. Summary of the Invention

[0008] This invention aims to address the problems of unreliable single-modal monitoring and inability to identify health risks in the elderly in advance in existing technologies. It proposes a method and system for predicting health risks in the elderly based on three-dimensional coupling of emotion, physiology and gait. This method can perform deep correlation inference on multimodal signals and achieve continuous monitoring and proactive early warning in the context of walking aids. At the same time, it constructs a corresponding intelligent walking aid system to proactively protect the health status of the elderly.

[0009] This invention achieves the above objective through the following technical solution: a method for predicting health risks in the elderly based on a three-dimensional coupling of emotion, physiology, and gait, comprising the following steps:

[0010] S1: Collect facial image data, voice data, heart rate changes, blood pressure data, gait data, and acceleration data of the user during use;

[0011] S2: Extract user emotional features from facial image data and voice data, extract user physiological health features from heart rate changes and blood pressure data, and extract user motion features from user gait data and acceleration data during the walking process;

[0012] S3: Establish an emotion recognition model by inputting the user's facial image captured by the walker's camera to obtain the user's emotion category and psychological state index; extract tone, speech rate and timbre features from the voice data collected by the microphone and input them into the emotion recognition model to further correct the psychological state judgment;

[0013] S4: Input the recognition results of the emotion recognition model, and combine them with physiological indicators and motion stability parameters collected by heart rate, blood pressure and gait sensors to establish a risk prediction model. Multimodal data fusion is performed in the edge computing unit to output the health risk index and psychological abnormality warning for the elderly.

[0014] S5: When an elderly person's health risk index exceeds the threshold, it will automatically trigger a voice reminder, push a notification to the family's APP, and can be linked to the medical service platform to provide timely health intervention for the user.

[0015] As a further technical solution of the present invention: the emotion recognition model combines convolutional neural networks with speech emotion recognition to improve the accuracy of psychological state analysis, including the following steps:

[0016] S31: Preprocess the user's facial image captured by the camera, including using the MTCNN algorithm to detect faces to locate facial regions, extract key regions such as eyes, eyebrows, and mouth and align key points, and then perform grayscale processing after standardizing the image to a preset size.

[0017] S32: Input the pre-processed facial image into the pre-trained convolutional neural network model, extract facial texture and muscle movement features through convolutional and pooling layers, and output an N-dimensional facial emotion feature vector.

[0018] Let F be the feature map output by the l-th layer convolution operation of the convolutional neural network. l The calculation formula is as follows:

[0019] ;

[0020] Among them, W l F represents the weights of the convolutional kernel in the l-th layer. l-1 b is the input feature map for the previous layer. lHere, σ is the bias term, σ is the ReLU activation function, and (x,y) are the pixel coordinates of the feature map.

[0021] After processing with a fully connected layer and a Softmax activation function, the probability distribution of the emotion category corresponding to the facial expression is obtained;

[0022] S33: Preprocess the voice data collected by the microphone, including using wavelet transform or spectral subtraction to reduce noise, dividing the data into frames at 20-30ms / frame and setting a 50% overlap rate, and enhancing high-frequency signals through pre-emphasis processing;

[0023] S34: Extract the acoustic features of speech, including prosodic features composed of fundamental frequency, speech rate, and sentence duration, spectral features composed of Mel frequency cepstral coefficients and spectral entropy, and energy features composed of short-time energy, and integrate them to form an M-dimensional speech emotion feature vector.

[0024] The m-th Mel frequency cepstral coefficient C m The calculation formula is:

[0025] ;

[0026] Where S(k) is the logarithmic power spectrum of the Mel filter bank output (K is the number of filters), and M is the dimension of the MFCC coefficients;

[0027] Input the speech emotion feature vector into a recurrent neural network or support vector machine, and output a speech emotion probability distribution that is consistent with the facial emotion category.

[0028] S35: An attention mechanism is used to fuse the probability distributions of the two modalities mentioned above. The attention mechanism automatically learns the modal weights in different scenarios through the attention module and dynamically adjusts the fusion ratio.

[0029] Let the facial feature vector be V. f The speech feature vector is V s The formulas for calculating the attention weighting coefficients α (facial) and 1-α (voice) are as follows:

[0030] ;

[0031] Among them, W a The attention learning parameters are optimized through backpropagation.

[0032] S36: Determine the dominant emotion category based on the final emotion probability distribution and map it to a quantitative psychological state indicator, and combine the duration to determine whether it is a psychological abnormality.

[0033] As a further technical solution of the present invention: the health risk index for the elderly includes fall risk prediction, based on gait stability and detection of sudden abnormal acceleration.

[0034] As a further technical solution of the present invention, the health risk index for the elderly includes the following steps:

[0035] Gait stability parameter acquisition: Data is collected through a multi-dimensional gait detection module integrated into the walker. The gait detection module includes a pressure sensor array in the handrail grip area, an inertial measurement unit (IMU), and contact sensors on the inside of the supporting feet. The pressure sensor array collects data on the magnitude and distribution of the user's grip force, with a sampling frequency of 50-100Hz. The inertial measurement unit collects three-dimensional acceleration (X / Y / Z axis) and three-dimensional angular velocity data, with a sampling frequency of 100-200Hz. The contact sensors collect data on the contact state and contact pressure between the supporting feet and the ground, with a sampling frequency of 50Hz.

[0036] Sudden abnormal acceleration data acquisition: The sudden acceleration data of the user's limbs are collected synchronously through the inertial measurement unit. The acceleration acquisition range is set to ±16g to cover the scenarios of elderly users walking and sudden imbalance. At the same time, the tilt angle data of the device is collected with the assistance of the attitude sensor to realize the scenario-based calibration of acceleration data.

[0037] Data cleaning: Outlier removal is performed on the collected gait and acceleration raw data. The 3σ criterion is used to identify and delete isolated abnormal data caused by sensor jitter. Short-term missing values ​​that occur during the data acquisition process are filled by linear interpolation. When the missing value lasts for more than 500ms, the sensor self-test signal is triggered.

[0038] Data standardization and feature enhancement: Gait pressure data, acceleration data, and angular velocity data are normalized to the [0,1] interval to eliminate the influence of differences in the measurement range of different sensors; acceleration data is low-pass filtered using a Butterworth filter with a cutoff frequency of 5Hz to remove high-frequency noise and retain effective signals related to human movement and imbalance; gait data is time-synchronized and aligned using the timestamp of the inertial measurement unit as a reference to unify pressure data and contact sensor data to the same time dimension, with time error controlled within 10ms.

[0039] Basic gait parameter calculation: Based on the preprocessed gait data, calculate the gait cycle (the time interval between two consecutive touches of the same side supporting foot), stride length (the displacement distance during two consecutive touches), gait speed (the ratio of stride length to gait cycle) and support phase ratio (the ratio of the time the supporting foot is in contact with the ground to the gait cycle), forming a basic gait feature set.

[0040] Stability feature quantification: The grip force variation coefficient (the ratio of the standard deviation of grip force to the mean) and the grip force asymmetry of both hands (|average grip force of left hand - average grip force of right hand| / (average grip force of left hand + average grip force of right hand)×100%) are calculated based on pressure sensor array data; the gait acceleration variation coefficient (the ratio of the standard deviation of X / Y / Z axis acceleration to the mean) and the trunk tilt angle fluctuation rate (the standard deviation of the ratio of tilt angle change to time) are calculated based on inertial measurement unit data; the support stability index (the sum of the products of the contact area of ​​the supporting foot and the contact pressure) is calculated based on contact sensor data to form a gait stability feature set.

[0041] Abnormal acceleration identification: Set acceleration threshold ranges, where the abnormal acceleration threshold for the X-axis (forward direction) is ±2g, the abnormal acceleration threshold for the Y-axis (left-right direction) is ±1.5g, and the abnormal acceleration threshold for the Z-axis (vertical direction) is ±3g; when three consecutive sampling points of the collected acceleration data exceed the corresponding axis threshold, it is determined as a sudden abnormal acceleration event, and the event occurrence time, peak acceleration, and acceleration change rate (the ratio of peak acceleration to the time to reach the peak value) are recorded.

[0042] Abnormal Scene Classification: Combining tilt angle data and support foot contact status, sudden abnormal acceleration events are classified into scenarios, including: unbalanced forward tilt (abnormal acceleration in the positive X-axis direction, tilt angle > 15°, sudden increase in contact pressure of the front support foot), unbalanced side tilt (abnormal acceleration in the Y-axis, tilt angle > 20°, sudden drop in contact pressure of the support foot on one side), and sudden turbulence (abnormal acceleration in the Z-axis, periodic fluctuation of contact pressure of the support foot). Abnormal scene types and corresponding acceleration feature parameters are extracted.

[0043] Feature weight allocation: A hierarchical analysis model is constructed, with gait stability features and sudden abnormal acceleration features as primary indicators, where gait stability features account for 60% of the weight and sudden abnormal acceleration features account for 40% of the weight. Under the primary indicators, gait cycle variation coefficient, grip force asymmetry, peak value of abnormal acceleration on the Y-axis, and support stability index are used as key secondary indicators. Their weights are determined by training and optimization using historical fall case datasets to ensure high sensitivity of key risk features.

[0044] Risk index calculation: The fall risk index is calculated using a weighted summation algorithm. The formula is: Fall Risk Index = Σ (Primary indicator weight × Σ (Secondary indicator feature value × Secondary indicator weight)); where the feature values ​​of each secondary indicator are compared with a standard gait database of healthy elderly people and converted into a quantitative score of 0-100. The standard database is constructed by training on gait data of more than 1,000 healthy elderly people aged 60 and above.

[0045] Risk level determination: Set fall risk level thresholds, where low risk: fall risk index < 30, medium risk: 30 ≤ fall risk index < 60, high risk: fall risk index ≥ 60; when a sudden abnormal acceleration event is detected, a dynamic risk index correction mechanism is triggered, adding a risk increment of 20%-50% to the original calculation result. The increment is determined according to the degree of danger of the abnormal scenario (50% for imbalance and tilt, 20% for sudden turbulence).

[0046] Real-time results output: The fall risk index and its corresponding risk level are synchronized to the edge computing unit and integrated with emotional state indicators and other physiological abnormality indicators to form a comprehensive health risk index; the risk level is displayed in real time on the screen, with low risk displayed in green, medium risk in yellow, and high risk in red;

[0047] Warning Triggering and Linkage: When the fall risk index is ≥60, a local voice reminder is immediately triggered ("Please stand firmly, the current risk of falling is high"), and at the same time, the risk information (including risk index, abnormal characteristics and real-time location) is pushed to the family's APP through the communication module; if a sudden abnormal acceleration is detected and the risk index is ≥80, the medical service platform is linked to upload the user's health record and real-time monitoring data, and request remote intervention support.

[0048] A health risk prediction system for the elderly based on the three-dimensional coupling of emotion, physiology, and gait, comprising:

[0049] The camera module is used to extract facial contour features from acquired facial image data, key part motion features from facial expression dynamic data, and muscle state features from image texture data; it optimizes facial contour, key part motion and muscle state features through image enhancement technology, and fuses them through feature alignment, and then filters the fused features;

[0050] The speech acquisition module is used to extract fundamental frequency features from acquired speech waveform data, speech rate variation features from speech time sequence data, and timbre features from spectrum data; it uses a noise reduction algorithm to clean up the fundamental frequency, speech rate variation, and timbre features, and then fuses them through feature splicing, and regularizes the fused features;

[0051] The physiological signal sensor module is used to extract heart rate variability features from heart rate monitoring data, blood pressure peak features from blood pressure fluctuation data, and physiological rhythm features from signal period data. Adaptive filtering is used to remove interference from heart rate variability, blood pressure peak and physiological rhythm features, and feature normalization is used to fuse them to simplify the fused features.

[0052] The gait detection module is used to extract stride length and speed features from gait cycle data, motion amplitude features from acceleration data, and grip and support force features from pressure distribution data. The module performs smoothing processing to remove jitter from stride length and speed, motion amplitude, and grip and support force features, and then performs feature weighting and fusion to reduce the dimensionality of the fused features.

[0053] The data processing and edge computing unit is used to extract health-related features from the fused features of various modules, risk warning features from real-time monitoring data, and trend change features from historical data; it analyzes the health-related, risk warning, and trend change features through an edge computing model, makes decisions through algorithm fusion, and outputs the decision results in a quantitative manner.

[0054] The cloud-based data analysis platform is used to extract common patterns and features from data uploaded from edge computing, extract group health features from user group data, and extract optimization direction features from model running data. It mines common patterns, group health, and optimization direction features through big data algorithms, upgrades the model through iteration, and feeds the upgrade results back to the edge unit.

[0055] As a further technical solution of the present invention: the camera module, voice acquisition module, physiological signal sensor module, gait detection module, data processing and edge computing unit and cloud data analysis platform are connected to the controller through a bus to realize local real-time analysis and remote synchronization.

[0056] As a further technical solution of the present invention: the data processing and edge computing unit embeds an emotion recognition model and a health assessment algorithm, supporting local early warning and cloud-based model iterative upgrades.

[0057] The beneficial effects of this invention are:

[0058] 1) This invention relates to multimodal perception, cross-modal information fusion and risk prediction, integrating mental health and physiological safety monitoring to meet the comprehensive needs of elderly people living independently, and is used to identify potential falls, emotional instability and health abnormalities in elderly care scenarios in advance.

[0059] 2) This invention adopts a multimodal data fusion algorithm, which improves the accuracy and reduces the false alarm rate compared with single monitoring; it integrates local edge computing and cloud deep learning models, taking into account both real-time performance and personalized analysis; the device has an integrated design that adapts to the form of daily walking aids and has a low barrier to entry. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0061] Figure 2This is a schematic diagram illustrating the principle of the mental health prediction model of the present invention;

[0062] Figure 3 This is a schematic diagram illustrating the principle of the fall risk prediction model of the present invention;

[0063] Figure 4 This is a diagram illustrating the relationship of the three-dimensional coupling risk inference model of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1, as Figure 1 As shown, this embodiment provides a method for predicting health risks in the elderly based on a three-dimensional coupling of emotion, physiology, and gait. This method includes:

[0066] S1: Collect facial image data, voice data, heart rate changes, blood pressure data, gait data, and acceleration data of the user during use;

[0067] S2: Extract user emotional features from facial image data and voice data, extract user physiological health features from heart rate changes and blood pressure data, and extract user motion features from user gait data and acceleration data during the walking process;

[0068] S3: Establish an emotion recognition model by inputting the user's facial image obtained by the camera (walking aid camera) to obtain the user's emotion category and psychological state index, and extracting tone, speech rate and timbre features from the voice data collected by the microphone, and inputting them into the emotion recognition model to further correct the psychological state index.

[0069] S4: Input the recognition results of the emotion recognition model, and combine them with physiological indicators and motion stability parameters collected by heart rate, blood pressure and gait sensors to establish a risk prediction model. Multimodal data fusion is performed in the edge computing unit to output the health risk index and psychological abnormality warning for the elderly.

[0070] S5: When an elderly person's health risk index exceeds the threshold, a voice reminder will be automatically triggered, a push notification will be sent to the family's APP, and the medical service platform will be linked to provide timely health intervention for the user.

[0071] In S3, the emotion recognition model combines convolutional neural networks with speech emotion recognition to improve the accuracy of psychological state analysis. For example, Figure 2 As shown, building an emotion recognition model includes the following steps:

[0072] S31: Preprocess the user's facial image captured by the camera, including using the MTCNN algorithm to detect faces to locate facial regions, extract key regions such as eyes, eyebrows, and mouth and align key points, and then perform grayscale processing after standardizing the image to a preset size.

[0073] S32: Input the pre-processed facial image into the pre-trained convolutional neural network model, extract facial texture and muscle movement features through convolutional and pooling layers, and output an N-dimensional facial emotion feature vector.

[0074] Let F be the feature map output by the l-th layer convolution operation of the convolutional neural network. l The calculation formula is as follows:

[0075] ;

[0076] Among them, W l F represents the weights of the convolutional kernel in the l-th layer. l-1 b is the input feature map for the previous layer. l Here, σ is the bias term, σ is the ReLU activation function, and (x,y) are the pixel coordinates of the feature map.

[0077] After processing with a fully connected layer and a Softmax activation function, the probability distribution of the emotion category corresponding to the facial expression is obtained;

[0078] S33: Preprocess the voice data collected by the microphone, including using wavelet transform or spectral subtraction to reduce noise, dividing the data into frames at 20-30ms / frame and setting a 50% overlap rate, and enhancing high-frequency signals through pre-emphasis processing;

[0079] S34: Extract the acoustic features of speech, including prosodic features composed of fundamental frequency, speech rate, and sentence duration, spectral features composed of Mel frequency cepstral coefficients and spectral entropy, and energy features composed of short-time energy, and integrate them to form an M-dimensional speech emotion feature vector.

[0080] The m-th Mel frequency cepstral coefficient C m The calculation formula is:

[0081] ;

[0082] Where S(k) is the logarithmic power spectrum of the Mel filter bank output (K is the number of filters), and M is the dimension of the MFCC coefficients;

[0083] Input the speech emotion feature vector into a recurrent neural network or support vector machine, and output a speech emotion probability distribution that is consistent with the facial emotion category.

[0084] S35: An attention mechanism is used to fuse the probability distributions of the two modalities mentioned above. The attention mechanism automatically learns the modal weights in different scenarios through the attention module and dynamically adjusts the fusion ratio.

[0085] Let the facial feature vector be V. f The speech feature vector is V s The formulas for calculating the attention weighting coefficients α (facial) and 1-α (voice) are as follows:

[0086] ;

[0087] Among them, W a The attention learning parameters are optimized through backpropagation.

[0088] S36: Determine the dominant emotion category based on the final emotion probability distribution and map it to a quantitative psychological state indicator, and combine the duration to determine whether it is a psychological abnormality.

[0089] In S4, such as Figure 3 As shown, the health risk index for the elderly includes fall risk prediction and detection based on gait stability and sudden abnormal acceleration.

[0090] Specifically, it includes:

[0091] Gait stability parameter acquisition: Data is collected through a multi-dimensional gait detection module integrated into the walker. The gait detection module includes a pressure sensor array located in the grip area of ​​the walker's handrail, an inertial measurement unit (IMU) installed at the connection between the walker's wheels and the frame, and contact sensors arranged on the inside of the walker's support feet. Among them, the pressure sensor array collects data on the magnitude and distribution of the user's grip force, with a sampling frequency of 50-100Hz; the inertial measurement unit collects data on the walker's three-dimensional acceleration (X / Y / Z axes) and three-dimensional angular velocity, with a sampling frequency of 100-200Hz; and the contact sensors collect data on the contact state and contact pressure between the support feet and the ground, with a sampling frequency of 50Hz.

[0092] Sudden abnormal acceleration data acquisition: The above-mentioned inertial measurement unit synchronously collects sudden acceleration data related to the walker and the user's limbs. The acceleration acquisition range is set to ±16g to cover elderly users' walking and sudden imbalance scenarios. At the same time, the posture sensor built into the walker assists in collecting the device's tilt angle data to achieve scenario-based calibration of acceleration data.

[0093] Data cleaning: Outliers are removed from the raw gait and acceleration data. The 3σ criterion is used to identify and delete isolated outliers caused by sensor jitter. Short-term missing values ​​that occur during data acquisition are filled by linear interpolation. When the missing value lasts for more than 500ms, the sensor self-test signal is triggered.

[0094] Data standardization and feature enhancement: Gait pressure data, acceleration data, and angular velocity data are normalized to the [0,1] interval to eliminate the influence of differences in the measurement range of different sensors. Acceleration data is low-pass filtered, and a Butterworth filter with a cutoff frequency of 5Hz is used to filter out high-frequency noise and retain effective signals related to human movement and imbalance. Gait data is time-synchronized and aligned. Based on the timestamp of the inertial measurement unit, pressure data and contact sensor data are unified to the same time dimension, and the time error is controlled within 10ms.

[0095] Basic gait parameter calculation: Based on the preprocessed gait data, calculate the gait cycle (the time interval between two consecutive touches of the same side supporting foot), stride length (the displacement distance of the walker during two consecutive touches), gait speed (the ratio of stride length to gait cycle) and support phase ratio (the ratio of the time the supporting foot is in contact with the ground to the gait cycle), forming a basic gait feature set.

[0096] Stability feature quantification: The grip force variation coefficient (the ratio of the standard deviation of grip force to the mean) and the grip force asymmetry of both hands (|average grip force of left hand - average grip force of right hand| / (average grip force of left hand + average grip force of right hand)×100%) are calculated based on pressure sensor array data; the gait acceleration variation coefficient (the ratio of the standard deviation of X / Y / Z axis acceleration to the mean) and the trunk tilt angle fluctuation rate (the standard deviation of the ratio of tilt angle change to time) are calculated based on inertial measurement unit data; the support stability index (the sum of the products of the contact area of ​​the supporting foot and the contact pressure) is calculated based on contact sensor data to form a gait stability feature set.

[0097] Abnormal acceleration identification: Set acceleration threshold ranges, where the abnormal acceleration threshold for the X-axis (walker forward direction) is ±2g, the abnormal acceleration threshold for the Y-axis (walker left and right direction) is ±1.5g, and the abnormal acceleration threshold for the Z-axis (vertical direction) is ±3g; when the collected acceleration data exceeds the corresponding axis threshold for 3 consecutive sampling points, it is determined as a sudden abnormal acceleration event, and the event occurrence time, peak acceleration, and acceleration change rate (the ratio of peak acceleration to the time to reach the peak value) are recorded.

[0098] Abnormal Scene Classification: Combining the walking aid tilt angle data and the contact status of the support foot, sudden abnormal acceleration events are classified into scenarios, including: unbalanced forward tilt (abnormal acceleration in the positive X-axis direction, tilt angle > 15°, sudden increase in contact pressure of the front support foot), unbalanced side tilt (abnormal acceleration in the Y-axis, tilt angle > 20°, sudden drop in contact pressure of the support foot on one side), and sudden turbulence (abnormal acceleration in the Z-axis, periodic fluctuation of contact pressure of the support foot). Abnormal scene types and corresponding acceleration feature parameters are extracted.

[0099] Feature weight allocation: A hierarchical analysis model is constructed, with gait stability features and sudden abnormal acceleration features as primary indicators. Gait stability features account for 60% of the weight, and sudden abnormal acceleration features account for 40%. Under the primary indicators, gait cycle variation coefficient, grip force asymmetry, Y-axis abnormal acceleration peak and support stability index are used as key secondary indicators. Their weights are determined by training and optimization through historical fall case datasets to ensure high sensitivity of key risk features.

[0100] Risk index calculation: The fall risk index is calculated using a weighted summation algorithm. The formula is: Fall Risk Index = Σ (Primary indicator weight × Σ (Secondary indicator feature value × Secondary indicator weight)). The feature values ​​of each secondary indicator are compared with a standard gait database of healthy elderly people and converted into a quantitative score of 0-100. The standard database is constructed by training on gait data of more than 1,000 healthy elderly people aged 60 and above.

[0101] Risk level determination: Set fall risk level thresholds, where low risk: fall risk index < 30, medium risk: 30 ≤ fall risk index < 60, high risk: fall risk index ≥ 60; when a sudden abnormal acceleration event is detected, a dynamic risk index correction mechanism is triggered, adding a risk increment of 20%-50% to the original calculation result. The increment is determined according to the degree of danger of the abnormal scenario (50% for imbalance and tilt, 20% for sudden turbulence).

[0102] Real-time results output: The fall risk index and its corresponding risk level are synchronized to the edge computing unit of the walker and integrated with emotional state indicators and other physiological abnormality indicators to form a comprehensive health risk index; the risk level is displayed in real time on the display screen at the handrail of the walker, with green for low risk, yellow for medium risk and red for high risk.

[0103] Warning Triggering and Linkage: When the fall risk index is ≥60, a local voice reminder is immediately triggered ("Please stand firmly, the current risk of falling is high"). At the same time, the risk information (including risk index, abnormal characteristics and real-time location) is pushed to the family's APP through the walker's built-in communication module. If a sudden abnormal acceleration is detected and the risk index is ≥80, the medical service platform is linked to upload the user's health record and real-time monitoring data and request remote intervention support.

[0104] Example 2: This example provides a health risk prediction system for the elderly based on a three-dimensional coupling of emotion, physiology, and gait. This health risk prediction system is used to implement the health risk prediction method for the elderly described in Example 1, and specifically includes:

[0105] The camera module is used to extract facial contour features from acquired facial image data, key part motion features from facial expression dynamic data, and muscle state features from image texture data; it optimizes facial contour, key part motion and muscle state features through image enhancement technology, and fuses them through feature alignment, and then filters the fused features;

[0106] The speech acquisition module is used to extract fundamental frequency features from acquired speech waveform data, speech rate variation features from speech time sequence data, and timbre features from spectrum data; it uses a noise reduction algorithm to clean up the fundamental frequency, speech rate variation, and timbre features, and then fuses them through feature splicing, and regularizes the fused features;

[0107] The physiological signal sensor module is used to extract heart rate variability features from heart rate monitoring data, blood pressure peak features from blood pressure fluctuation data, and physiological rhythm features from signal period data. Adaptive filtering is used to remove interference from heart rate variability, blood pressure peak and physiological rhythm features, and feature normalization is used to fuse them to simplify the fused features.

[0108] The gait detection module is used to extract stride length and speed features from gait cycle data, motion amplitude features from acceleration data, and grip and support force features from pressure distribution data. The module performs smoothing processing to remove jitter from stride length and speed, motion amplitude, and grip and support force features, and then performs feature weighting and fusion to reduce the dimensionality of the fused features.

[0109] The data processing and edge computing unit is used to extract health-related features from the fused features of various modules, risk warning features from real-time monitoring data, and trend change features from historical data; it analyzes the health-related, risk warning, and trend change features through an edge computing model, makes decisions through algorithm fusion, and outputs the decision results in a quantitative manner.

[0110] The cloud-based data analysis platform is used to extract common patterns and features from data uploaded from edge computing, extract group health features from user group data, and extract optimization direction features from model running data. It mines common patterns, group health, and optimization direction features through big data algorithms, upgrades the model through iteration, and feeds the upgrade results back to the edge unit.

[0111] The camera module, voice acquisition module, physiological signal sensor module, gait detection module, data processing and edge computing unit, and cloud data analysis platform are connected to the controller via a bus to achieve local real-time analysis and remote synchronization. The data processing and edge computing unit has an embedded emotion recognition model and health assessment algorithm, which supports local early warning and cloud model iterative upgrades.

[0112] like Figure 4 As shown, this application also proposes a three-dimensional coupled risk inference model, including:

[0113] Single-dimensional feature quantification: Quantify the results of psychological state prediction model, physiological state detection data and gait stability risk prediction model to facilitate multi-dimensional coupling analysis.

[0114] Based on three coupling models—psycho-gait coupling, psycho-physiological coupling, and physiological-gait coupling—feature weights and score corrections are assigned.

[0115] A three-dimensional fusion calculation is performed to couple psychological prediction data, physiological data, and gait risk prediction data to obtain a three-dimensional coupled health risk prediction result.

[0116] Working process and principle: By collecting facial images, voice, physiological signals and gait behavior data, emotional state, emotional fluctuation trend, physiological stability index and gait stability index are generated respectively; then, based on the dynamic coupling relationship between emotional changes, physiological load changes and gait stability, a joint risk inference model is constructed. This model can identify health risks of the elderly in advance and generate early warnings at the edge, realizing deep correlation inference across modalities. It solves the problems of instability, high false alarm rate and inability to identify risks in advance in traditional single-modal monitoring, and significantly improves the intelligence and safety of elderly care equipment.

[0117] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0118] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting health risks in the elderly based on a three-dimensional coupling of emotion, physiology, and gait, characterized in that, The methods for predicting health risks in the elderly include: S1: Collects user facial image data, voice data, heart rate changes, blood pressure data, as well as user gait data and acceleration data during the walking process; S2: Extract user emotional features from facial image data and voice data, extract user physiological health features from heart rate changes and blood pressure data, and extract user motion features from gait data and acceleration data during walking. S3: Establish an emotion recognition model, input the user's facial image captured by the camera, obtain the user's emotion category and psychological state index, and extract the tone, speech rate and timbre features from the voice data collected by the microphone, and input them into the emotion recognition model to further correct the psychological state index; S4: Input the recognition results of the emotion recognition model, and combine them with physiological indicators and motion stability parameters collected by heart rate, blood pressure and gait sensors to establish a risk prediction model. Perform multimodal data fusion in the edge computing unit to output the health risk index and psychological abnormality warning for the elderly. S5: When an elderly person's health risk index exceeds the threshold, a voice reminder will be automatically triggered, a push notification will be sent to the family's APP, and the medical service platform will be linked to provide timely health intervention for the user.

2. The method for predicting health risks in the elderly according to claim 1, characterized in that, In S3, building an emotion recognition model specifically includes the following steps: S31: Preprocess the user's facial image captured by the camera, including using the MTCNN algorithm to detect faces to locate facial regions, extracting key regions such as eyes, eyebrows, and mouth and aligning key points, and standardizing the user's facial image to a preset size before grayscale processing. S32: Input the pre-processed facial image into the pre-trained convolutional neural network model, extract facial texture and muscle movement features through convolutional and pooling layers, and output an N-dimensional facial emotion feature vector. Let F be the feature map output by the l-th layer convolution operation of the convolutional neural network. l The calculation formula is as follows: ； Among them, W l F represents the weights of the convolutional kernel in the l-th layer. l-1 b is the input feature map for the previous layer. l Here, σ is the bias term, σ is the ReLU activation function, and (x,y) are the pixel coordinates of the feature map. After processing with a fully connected layer and a Softmax activation function, the probability distribution of the emotion category corresponding to the facial expression is obtained; S33: Preprocess the voice data collected by the microphone, including using wavelet transform or spectral subtraction to reduce noise, dividing the data into frames at 20-30ms / frame and setting a 50% overlap rate, and enhancing high-frequency signals through pre-emphasis processing; S34: Extract the acoustic features of speech, including prosodic features composed of fundamental frequency, speech rate, and sentence duration, spectral features composed of Mel frequency cepstral coefficients and spectral entropy, and energy features composed of short-time energy, and integrate them to form an M-dimensional speech emotion feature vector. The m-th Mel frequency cepstral coefficient C m The calculation formula is: ； Where S(k) is the logarithmic power spectrum of the Mel filter bank output, and M is the dimension of the MFCC coefficients; Input the speech emotion feature vector into a recurrent neural network or support vector machine, and output a speech emotion probability distribution that is consistent with the facial emotion category. S35: An attention mechanism is used to fuse the probability distributions of the two modalities mentioned above. The attention mechanism automatically learns the modal weights in different scenarios through the attention module and dynamically adjusts the fusion ratio. Let the facial feature vector be V. f The speech feature vector is V s The formulas for calculating the attention weight coefficients α and 1-α are: ； Among them, W a For attention learning parameters; S36: Determine the dominant emotion category based on the final emotion probability distribution and map it to a quantitative psychological state indicator, and combine the duration to determine whether it is a psychological abnormality.

3. The method for predicting health risks in the elderly according to claim 1, characterized in that: In S4, the health risk index for the elderly includes fall risk prediction and detection based on gait stability and sudden abnormal acceleration.

4. The method for predicting health risks in the elderly according to claim 3, characterized in that, The health risk index for older adults includes the following steps: Gait stability parameter acquisition: Data is acquired through a multi-dimensional gait detection module, which includes a pressure sensor array, an inertial measurement unit, and contact sensors. The pressure sensor array acquires data on the magnitude and distribution of the user's grip force, with a sampling frequency of 50-100Hz. The inertial measurement unit acquires data on the user's three-dimensional acceleration and three-dimensional angular velocity, with a sampling frequency of 100-200Hz. The contact sensors acquire data on the contact state and contact pressure between the supporting foot and the ground, with a sampling frequency of 50Hz. Sudden abnormal acceleration data acquisition: The sudden acceleration data of the user's limbs are collected synchronously through the inertial measurement unit. The acceleration acquisition range is set to ±16g to cover the scenarios of elderly users walking and sudden imbalance. At the same time, the tilt angle data of the device is collected with the assistance of the attitude sensor to realize the scenario-based calibration of acceleration data. Data cleaning: Outliers are removed from the collected gait stability parameters and sudden abnormal acceleration data. The 3σ criterion is used to identify and delete isolated abnormal data caused by sensor jitter. Short-term missing values ​​that occur during the data acquisition process are filled by linear interpolation. When the missing value lasts for more than 500ms, the sensor self-test signal is triggered. Data standardization and feature enhancement: Gait pressure data, acceleration data, and angular velocity data are normalized to the [0,1] interval to eliminate the influence of differences in the range of different sensors. Acceleration data is low-pass filtered, and a Butterworth filter with a cutoff frequency of 5Hz is used to filter out high-frequency noise and retain effective signals related to human movement and imbalance. Gait data is time-synchronized and aligned, using the timestamp of the inertial measurement unit as the reference, and pressure data and contact sensor data are unified to the same time dimension. Basic gait parameter calculation: Based on the preprocessed gait data, gait cycle, stride length, stride speed and support phase ratio are calculated to form a basic gait feature set; Stability feature quantification: The grip force variation coefficient and the asymmetry of grip force between the two hands are calculated using data from the pressure sensor array; the gait acceleration variation coefficient and trunk tilt angle fluctuation rate are calculated based on data from the inertial measurement unit; and the support stability index is calculated using data from the contact sensor to form a gait stability feature set. Abnormal acceleration identification: Set acceleration threshold ranges, where the abnormal acceleration threshold for the X-axis is ±2g, the abnormal acceleration threshold for the Y-axis is ±1.5g, and the abnormal acceleration threshold for the Z-axis is ±3g. When the collected acceleration data exceeds the corresponding axis threshold for 3 consecutive sampling points, it is determined to be a sudden abnormal acceleration event. Record the event occurrence time, peak acceleration, and acceleration change rate. Abnormal scene classification: Combining tilt angle data and support foot contact status, sudden abnormal acceleration events are classified into scenarios, including: unbalanced forward tilt, unbalanced side tilt, and sudden turbulence. Abnormal scene types and corresponding acceleration feature parameters are extracted. Feature weight allocation: A hierarchical analysis model is constructed, with gait stability features and sudden abnormal acceleration features as primary indicators, where gait stability features account for 60% of the weight and sudden abnormal acceleration features account for 40% of the weight. Under the primary indicators, gait cycle variation coefficient, grip force asymmetry, peak value of abnormal acceleration on the Y-axis, and support stability index are used as key secondary indicators. Their weights are determined by training and optimization using historical fall case datasets to ensure high sensitivity of key risk features. Risk index calculation: The fall risk index is calculated using a weighted summation algorithm. The formula is: Fall risk index = Σ (first-level indicator weight × Σ (second-level indicator feature value × second-level indicator weight)). The feature values ​​of each second-level indicator are compared with the standard gait database of healthy elderly people and converted into a quantitative score of 0-100. Risk level determination: Set fall risk level thresholds, where low risk: fall risk index < 30, medium risk: 30 ≤ fall risk index < 60, high risk: fall risk index ≥ 60; Real-time results output: The fall risk index and its corresponding risk level are synchronized to the edge computing unit and integrated with emotional state indicators and other physiological abnormality indicators to form a comprehensive health risk index; the risk level is displayed in real time on the screen, with low risk displayed in green, medium risk in yellow, and high risk in red; Warning Triggering and Linkage: When the fall risk index is ≥60, a local voice reminder is immediately triggered, and the communication module pushes the risk information to the family's APP. If a sudden abnormal acceleration is detected and the risk index is ≥80, the medical service platform is linked to upload the user's health record and real-time monitoring data, and request remote intervention support.

5. A health risk prediction system for the elderly for implementing the health risk prediction method for the elderly according to any one of claims 1 to 4, characterized in that, The elderly health risk prediction system includes: The camera module is used to extract facial contour features from acquired facial image data, key part motion features from facial expression dynamic data, and muscle state features from image texture data. At the same time, it optimizes facial contour, key part motion and muscle state features through image enhancement technology, and fuses them through feature alignment, and filters the fused features. The speech acquisition module is used to extract fundamental frequency features from acquired speech waveform data, speech rate variation features from speech time sequence data, and timbre features from spectrum data. It also uses a noise reduction algorithm to clean up the fundamental frequency, speech rate variation, and timbre features, and performs feature splicing to fuse them and regularize the fused features. The physiological signal sensor module is used to extract heart rate variability features from heart rate monitoring data, blood pressure peak features from blood pressure fluctuation data, and physiological rhythm features from signal period data. At the same time, it uses adaptive filtering to remove interference from heart rate variability, blood pressure peak and physiological rhythm features, and fuses them through feature normalization to simplify the fused features. The gait detection module is used to extract stride length and speed features from gait cycle data, motion amplitude features from acceleration data, and grip and support force features from pressure distribution data. At the same time, it performs smoothing processing to remove jitter from stride length, speed, motion amplitude, and grip and support force features, and performs feature weighting and fusion to reduce the dimensionality of the fused features. The data processing and edge computing unit is used to extract health-related features from the fused features of each module, risk warning features from real-time monitoring data, and trend change features from historical data. At the same time, it analyzes the health-related, risk warning, and trend change features through the edge computing model, makes decisions through algorithm fusion, and outputs the decision results in a quantitative manner. The cloud-based data analysis platform is used to extract common patterns and features from data uploaded from edge computing, extract group health features from user group data, and extract optimization direction features from model running data. At the same time, it uses big data algorithms to mine common patterns, group health, and optimization direction features, and upgrades the model through iteration, feeding the upgrade results back to the edge unit.

6. The health risk prediction system for the elderly according to claim 5, characterized in that: The camera module, voice acquisition module, physiological signal sensor module, gait detection module, data processing and edge computing unit, and cloud data analysis platform are connected to the controller via a bus to achieve local real-time analysis and remote synchronization.

7. The health risk prediction system for the elderly according to claim 5, characterized in that: The data processing and edge computing unit embeds an emotion recognition model and a health assessment algorithm, supporting local early warning and cloud-based model iterative upgrades.

Images