An Adaptive Health Threshold Alarm Method for Wearable Devices Based on Millimeter-Wave Radar

Abstract

This invention discloses an adaptive health threshold alarm method for wearable devices based on millimeter-wave radar, comprising: acquiring vital signals from the chest region of a target user using a wearable millimeter-wave radar device and inputting them into an individual baseline health model; preprocessing the vital signals in the individual baseline health model to obtain three-dimensional tensor data through a preprocessing module; extracting statistical features based on the three-dimensional tensor data through a statistical feature extraction module; extracting convolutional features at different time scales based on the three-dimensional tensor data through a three-branch feature extraction module; fusing the statistical features and convolutional features through a feature fusion module to generate a comprehensive feature vector; and dynamically determining the abnormality level based on the comprehensive feature vector through a dynamic decision-making module and triggering corresponding graded alarms. This method can effectively achieve accurate health monitoring and risk warning for different populations in complex indoor and outdoor scenarios, providing technical support for personalized health management.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing, wearable devices and health monitoring and early warning technology, and more specifically to an adaptive health threshold alarm method for wearable devices based on millimeter-wave radar. Background Technology

[0002] With the aging population and increased health awareness, the demand for wearable health monitoring devices has surged. Millimeter-wave radar, with its advantages of non-contact detection, strong privacy protection, good penetration, and all-weather operation, has become one of the core technologies for wearable devices to collect physiological signals. It can obtain key health parameters such as heart rate and respiratory rate without contact with the human body, avoiding the discomfort and usage scenario limitations of traditional contact sensors, while also avoiding the privacy leakage risks of optical cameras.

[0003] However, existing wearable health monitoring technology faces many problems in practical applications, which seriously affect the accuracy and practicality of health monitoring.

[0004] First, the alarm functions of most mainstream wearable health devices are based on fixed health thresholds, such as a default heart rate threshold of 60-100 beats / minute and a respiratory rate of 12-20 breaths / minute. These threshold calculations do not incorporate individual characteristics such as user age, physical condition, underlying diseases, and exercise habits, making alarm accuracy heavily reliant on whether the user meets general standards. Users of different ages, physical conditions, and underlying diseases have fundamentally different definitions of health. For example, the normal resting heart rate range for people over 70 years old should be broadened to 55-105 beats / minute, but current thresholds are still judged according to a uniform adult standard, easily misjudging normal heart rates as abnormal or missing mild abnormalities in the elderly. Furthermore, patients with hypertension and diabetes require stricter heart rate thresholds; hypertensive patients need to control their resting heart rate to <80 beats / minute to avoid increased vascular burden, but current technology lacks disease-related threshold adjustment mechanisms, leading to missed risk assessments. The system cannot incorporate quantitative rules such as raising the upper limit of the heart rate threshold by 5% for every 10 years of age increase and lowering the lower limit of the resting heart rate threshold by 10% for those who exercise ≥5 times per week into the threshold calculation. The thresholds need to be manually modified, which is complicated and cannot dynamically adapt to changes in individual health status.

[0005] Secondly, while millimeter-wave radar's core characteristic is non-contact, penetrating detection, it is also susceptible to environmental noise pollution. Existing signal filtering algorithms can only eliminate some high-frequency noise and cannot distinguish between multipath false signals and genuine physiological signals, easily leading to false alarms. Furthermore, physiological fluctuations in physiological parameters are easily misinterpreted as abnormal. When a user exercises, their heart rate and respiratory rate increase significantly due to physiological compensation mechanisms; during running, the heart rate can reach 120-160 beats per minute, which is within the normal range. However, existing technologies either fail to distinguish between exercise and resting states or use overly coarse classifications of exercise states, resulting in a persistently high false alarm rate in exercise scenarios.

[0006] Furthermore, the alarm triggering logic of existing technologies is generally based on a single physiological parameter exceeding a threshold, without considering the duration, trend, and risk level of parameter deviation. This easily leads to a large number of false alarms due to momentary fluctuations. When a user's emotional excitement causes their heart rate to briefly rise to 130 beats / minute, or when they cough, their respiratory rate briefly rises to 25 breaths / minute, these are physiological momentary fluctuations and do not require an alarm. However, slow deviations pose a serious risk of being missed. When a user's resting heart rate gradually increases from 70 beats / minute to 95 beats / minute over a prolonged period, it may be due to myocardial ischemia. Although the individual measurement may not exceed the threshold, the trend is already abnormal, leading to a missed opportunity for intervention.

[0007] Finally, the thresholds in existing technologies are mostly fixed after being preset at the factory, without iterative updates based on long-term user physiological data. This makes them unable to adapt to the long-term evolution of users' health status. Changes in health habits lead to threshold mismatches; for example, if a user changes from a sedentary lifestyle to exercising more than 5 times a week, their resting heart rate will decrease significantly after 3 months. Furthermore, aging will naturally cause health thresholds to shift, gradually becoming inaccurate after 5 years of use due to user age. In addition, existing technologies generally lack user feedback channels and algorithm self-optimization mechanisms. False alarms and missed alarms cannot be corrected, leading to long-term stagnation in algorithm performance and significantly reducing the device's practicality.

[0008] Therefore, how to enable adaptive health threshold alarm methods to integrate individual characteristics, real-time status, and environmental interference, and to balance alarm sensitivity and false alarm rate by dynamically adjusting the threshold, so as to ensure that the health risks of different users in different scenarios are accurately identified, while avoiding interference caused to users by invalid alarms, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] In view of the above problems, the present invention is proposed to provide an adaptive health threshold alarm method for wearable devices based on millimeter-wave radar that overcomes or at least partially solves the above problems.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] This invention provides an adaptive health threshold alarm method for wearable devices based on millimeter-wave radar, comprising the following steps:

[0012] Vital signals in the chest area of ​​the target user are collected using wearable millimeter-wave radar devices.

[0013] The vital signals are input into an individual baseline health model; the individual baseline health model includes a preprocessing module, a statistical feature extraction module, a three-branch feature extraction module, a feature fusion module, and a dynamic decision-making module;

[0014] The life signal is preprocessed using the preprocessing module to obtain three-dimensional tensor data.

[0015] The statistical feature extraction module extracts statistical features based on the three-dimensional tensor data.

[0016] The three-branch feature extraction module extracts convolutional features at different time scales based on the three-dimensional tensor data.

[0017] The feature fusion module fuses the statistical features and the convolutional features at different time scales to generate a comprehensive feature vector.

[0018] The dynamic decision-making module dynamically determines the anomaly level based on the comprehensive feature vector and triggers the corresponding graded alarm.

[0019] Furthermore, the statistical features include time-domain statistical features, periodic statistical features, and interference statistical features;

[0020] The time-domain statistical features include: mean, standard deviation, maximum value, minimum value, kurtosis, and skewness;

[0021] The periodic statistical features include: SDNN and PNN50 of the heart rate RR interval, and the ratio of inspiratory to expiratory time.

[0022] The statistical characteristics of the interference include: noise energy percentage, number of multipath reflections, and number of signal abrupt changes.

[0023] Furthermore, the three-branch feature extraction module includes a short-term feature branch, a medium-term feature branch, and a long-term feature branch;

[0024] The short-term feature branch, the intermediate feature branch, and the long-term feature branch are each configured with convolutional layers with different receptive fields; wherein, the receptive field of the convolutional layer of the long-term feature branch is larger than that of the intermediate feature branch, and the receptive field of the convolutional layer of the intermediate feature branch is larger than that of the short-term feature branch.

[0025] The high-frequency convolutional features characterizing the signal fluctuation level are extracted from the three-dimensional tensor data through the short-term feature branch.

[0026] Through the intermediate feature branch, mid-frequency convolutional features characterizing the periodic changes of the signal are extracted from the three-dimensional tensor data;

[0027] The long-term feature branch is used to extract low-frequency convolutional features that characterize the smoothing trend of the signal from the three-dimensional tensor data.

[0028] Furthermore:

[0029] The short-term feature branch includes: a first convolutional layer, a first batch normalization layer, a second convolutional layer, a second batch normalization layer, and a global average pooling layer connected in sequence;

[0030] The intermediate feature branch includes: a third convolutional layer, a third batch normalization layer, a fourth convolutional layer, a fourth batch normalization layer, and a global max pooling layer connected in sequence;

[0031] The long-term feature branch includes: a fifth convolutional layer, a fifth batch normalization layer, a sixth convolutional layer, a sixth batch normalization layer, and a global adaptive average pooling layer connected in sequence.

[0032] Furthermore, the feature fusion module includes an initial fusion submodule, a mapping submodule, and a final fusion submodule;

[0033] In the initial fusion submodule, the convolutional features of the three scales output by the three-branch feature extraction module are concatenated to obtain the primary fusion features;

[0034] In the mapping submodule, the primary fusion features are nonlinearly mapped into higher-order convolutional features through a fully connected layer;

[0035] In the final fusion submodule, the higher-order convolutional features are concatenated with the statistical features to generate a comprehensive feature vector.

[0036] Furthermore, the dynamic determination of the anomaly level based on the comprehensive feature vector includes:

[0037] Based on the statistical features in the comprehensive feature vector, the actual threshold ranges for multiple health states are generated.

[0038] A fully connected layer maps the high-order convolutional features in the comprehensive feature vector to waveform patterns under the multiple health states, and outputs the confidence scores of each health state.

[0039] For each health state, if the life signal value of the corresponding waveform pattern is outside the actual threshold range and the corresponding confidence level is greater than the preset confidence threshold, then it is determined to belong to that health state.

[0040] Furthermore, the step of generating the actual threshold ranges for multiple health states based on the statistical features in the comprehensive feature vector specifically includes:

[0041] Based on the statistical characteristics of the target users in their initial resting state, the initial health baseline of the target users is determined;

[0042] An exponentially weighted moving average algorithm is used to dynamically update the initial health baseline using the statistical characteristics of the daily resting period, thereby obtaining the dynamic baseline at the current moment and generating baseline threshold ranges for different health states.

[0043] Based on the baseline threshold range, and combined with exercise and environmental factors, the actual threshold ranges for different health states are generated.

[0044] Furthermore, the motion factor is determined by the number of signal abrupt changes; the environmental factor is jointly determined by the noise energy ratio and the number of multipath reflections.

[0045] Furthermore, the individual baseline health model is trained based on a human chest region vital signal dataset with multiple constitutions, states, and scenarios.

[0046] The multi-body constitution stratification includes age stratification, BMI stratification, and health status stratification;

[0047] The multi-state layering includes resting state and motion state, and is associated with dynamic state tags that record medication and fatigue.

[0048] The multi-scenario layering includes different user postures and different physical environments.

[0049] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an adaptive health threshold alarm method for wearable devices based on millimeter-wave radar, which has the following beneficial effects:

[0050] This method uses millimeter-wave radar to collect the amplitude of micro-movements in the chest cavity to obtain physiological signals such as human heart rate and respiratory rate. It integrates individual user characteristics with real-time environment and current movement status, dynamically adjusts health thresholds and triggers multi-level alarms, and realizes accurate health monitoring and risk warning for different groups in complex indoor and outdoor scenarios, providing technical support for personalized health management. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the adaptive health threshold alarm method for wearable devices based on millimeter-wave radar provided in an embodiment of the present invention.

[0053] Figure 2 This is a schematic diagram of the three-branch feature extraction module architecture provided in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the dynamic decision-making process provided in an embodiment of the present invention. Detailed Implementation

[0055] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] This invention discloses an adaptive health threshold alarm method for wearable devices based on millimeter-wave radar, such as... Figure 1 As shown, it includes the following steps:

[0057] Vital signals in the chest area of ​​the target user are collected using wearable millimeter-wave radar devices.

[0058] Vital signals are input into an individual baseline health model; the individual baseline health model includes a preprocessing module, a statistical feature extraction module, a three-branch feature extraction module, a feature fusion module, and a dynamic decision-making module;

[0059] The life signal is preprocessed using the preprocessing module to obtain three-dimensional tensor data.

[0060] The statistical feature extraction module extracts statistical features based on three-dimensional tensor data.

[0061] The three-branch feature extraction module extracts convolutional features at different time scales based on three-dimensional tensor data.

[0062] The feature fusion module fuses statistical features and convolutional features at different time scales to generate a comprehensive feature vector.

[0063] The dynamic decision-making module dynamically determines the anomaly level based on comprehensive feature vectors and triggers corresponding graded alarms.

[0064] Next, we will provide a detailed explanation of the aforementioned wearable millimeter-wave radar device and individual baseline health model.

[0065] I. Wearable millimeter-wave radar equipment:

[0066] For millimeter-wave radar boards, a configuration of 3 transmitters and 4 receivers is adopted. Beamforming technology is used to reduce environmental interference. Frequency modulated continuous wave (FMCW) is transmitted, preserving the raw I / Q dual-channel data of the millimeter-wave radar. The sampling rate is 100Hz, and the length of a single record = acquisition time × sampling rate. The acquired signal data is made into a dataset and stored as a CSV file. Rows represent the time axis, and columns include location information, transmitted radar frame number, user_id, motion status, and scene label. The specific phase and amplitude can be read from each row and column. The raw radar signal data (radar_raw_data) is associated with the user information table (users) through user_id. The users table contains information such as the age, BMI, and underlying diseases of the corresponding user.

[0067] To ensure the quantity of valid data, data quality was validated. For I / Q data, the 3σ principle was applied: outliers exceeding the mean ± 3 standard deviations were removed. For resting state data, heart rate and respiratory rate were validated to ensure they were within physiologically reasonable ranges (40-180 beats / minute and 9-30 breaths / minute). Data outside these ranges were marked as low quality. Finally, 80% of the dataset was randomly allocated as the training set, 10% as the validation set, and 10% as the test set for subsequent training. The training set was used to directly update and determine the parameters of the preprocessing module, statistical feature extraction module, three-branch feature extraction module, feature fusion module, and dynamic decision-making module in the individual baseline health model, enabling the model to learn a complete mapping relationship. The validation set was used to evaluate and optimize the model's hyperparameters during training, including the receptive field size of the three-branch convolutional layer, the dimension of the fully connected layer, and the baseline update coefficients. α And so on, and to prevent the model from overfitting when jointly learning statistical and convolutional features, the model state with the strongest generalization ability is selected; the test set is used to finally evaluate the overall performance of the trained individual benchmark health model in simulated real-world scenarios, ensuring that the entire process of preprocessing, feature extraction, fusion and dynamic decision-making can still work stably and accurately on unseen data, so as to guarantee the reliability and alarm accuracy of the method of this invention in practical applications.

[0068] II. Training data for the individual baseline health model:

[0069] The individual baseline health model is trained based on a hierarchical dataset of vital signals from the human chest region, encompassing multiple body types, states, and scenarios, ensuring coverage of the entire spectrum of target users. In this embodiment of the invention, 5000 sets of randomly lengthed sample data segments are collected using a wearable millimeter-wave radar device. Among them:

[0070] The multi-body constitution stratification includes age stratification, BMI stratification, and health status stratification. Specifically: for age stratification, it includes 20-30 years old labeled as youth, 31-40 years old labeled as young and middle-aged, 41-50 years old labeled as middle-aged, 51-60 years old labeled as middle-aged and elderly, 61-70 years old labeled as elderly, and 71-80 years old labeled as advanced age; for BMI stratification, it is divided into four categories: underweight, normal, overweight, and obese; for health status stratification, it is divided into healthy people, patients with chronic diseases, and people who exercise regularly.

[0071] The multi-state stratification includes resting state and movement state, and records dynamic state tags related to medication and fatigue. Specifically, 30 minutes of continuous data collection is collected in the resting state as the individual's initial baseline, and dynamic states such as whether medication is being taken and whether fatigue is present are recorded simultaneously. In subsequent data collection, each person is collected twice a week, each time including resting and movement states, for a period of 8 weeks.

[0072] Multi-scenario layering includes different user postures and different physical environments. Specifically, to distinguish between physiological fluctuations and pathological abnormalities and avoid false alarms during movement, test subjects were in sitting, lying, and standing positions, with several signal segments of different lengths collected for each posture. Additionally, signal segments of different lengths were collected for each of the following movement states: walking, jogging, running, and climbing stairs. Secondly, to provide samples for the anti-interference design in subsequent feature extraction, signal segments of different lengths were collected from indoor locations such as living rooms, kitchens, game rooms, reading rooms, and farmers' markets, as well as outdoor locations such as community parks and buses. This addresses the problem of poor model generalization due to a single sample and ensures the completeness of the samples.

[0073] The dataset's comprehensive design, covering multiple physical conditions, states, and scenarios, enables the algorithm to accurately learn changes in the core vital sign characteristics of wearable devices. Parameter adjustments are more aligned with actual needs, providing complete sample support from feature learning to parameter mapping for the adaptive dynamic adjustment of alarm threshold algorithms.

[0074] III. Structure of Individual Baseline Health Model:

[0075] This individual baseline health model includes a preprocessing module, a statistical feature extraction module, a three-branch feature extraction module, a feature fusion module, and a dynamic decision-making module. Based on the frequency characteristics of physiological signals—heart rate 0.5-1.25Hz, respiration 0.15-0.5Hz, and motion artifacts 2-5Hz—it calculates statistical features using fixed mathematical formulas in the statistical feature extraction module and designs a three-branch convolutional network. The receptive field is adjusted by kernel size and expansion rate to cover different signal scales, accurately extracting convolutional features to determine multi-scale physiological characteristics. Specifically:

[0076] 1. Preprocessing module:

[0077] In the input preprocessing module, the raw I / Q data collected by the wearable millimeter-wave radar device, i.e., the collected vital signals, are in time series format. The channels are adjusted to a 3D tensor of 1×time series×2, and bandpass filtering (0.15-1.25Hz) and Z-score normalization are performed on each channel. The output data is [1,T,2], where T is the time series.

[0078] 2. Statistical Feature Extraction Module:

[0079] The statistical feature extraction module extracts statistical features based on three-dimensional tensor data. These statistical features are 12-dimensional, covering time-domain statistical features, periodic statistical features, and interference statistical features. Specifically:

[0080] (1) Time-domain statistical characteristics:

[0081] The time-domain statistical characteristics include: mean, standard deviation, maximum value, minimum value, kurtosis, and skewness. Among them, the mean is the average level of the signal amplitude, reflecting the overall energy intensity, as shown in formula (1); the standard deviation reflects the degree of signal fluctuation, as shown in formula (2); the maximum value is the peak value of the signal amplitude, reflecting the strongest energy point; the minimum value is the trough value of the signal amplitude, reflecting the weakest energy point; kurtosis is the sharpness of the signal amplitude distribution. A value greater than 0 indicates a peaked distribution, while a value less than 0 indicates a flat distribution. High kurtosis indicates that the signal has more sharp peaks, as shown in formula (3); skewness indicates the degree of asymmetry in the signal amplitude distribution. A value greater than 0 indicates right skewness, and a value less than 0 indicates left skewness. For example, breathing may be skewed due to asymmetry in inhalation / exhalation, as shown in formula (4).

[0082] (1)

[0083] (2)

[0084] (3)

[0085] (4)

[0086] in, H Rmean The mean; T The length of the time series; x t For the first t Signal amplitude at any given moment; H Rstd The standard deviation of the signal amplitude; N The total number of samples; H Ri For the first i Each signal sample value; Kurtosis For kurtosis; Skewness Skewness;

[0087] (2) Periodic statistical characteristics:

[0088] The periodic statistical features include: SDNN and PNN50 of the heart rate RR interval, and the ratio of inspiratory to expiratory time; among them, heart rate variability is mainly reflected by calculating the standard deviation of adjacent heartbeat intervals to reflect autonomic nerve function, see formula (5); PNN50 reflects short-term heart rate variability, and the higher the value, the stronger the parasympathetic nerve activity, see formula (6); the ratio of inspiratory to expiratory time is calculated by first calculating the inspiratory and expiratory time of a single breath and then taking the average value within 60 seconds to determine the symmetry of the respiratory rhythm;

[0089] (5)

[0090] (6)

[0091] in, SDNN The overall standard deviation of the time interval between consecutive normal heartbeats (R waves) (RR interval); RR i For the first i The heartbeat cycle of each signal; PNN50 is the percentage of adjacent RR differences exceeding 50ms;

[0092] (3) Statistical characteristics of interference:

[0093] Interference statistical characteristics include: noise energy ratio, number of multipath reflections, and number of signal abrupt changes. Noise energy ratio is the ratio of noise frequency band energy to total energy, reflecting electromagnetic and environmental noise interference. The correlation coefficient r(t) is calculated by performing autocorrelation analysis on the original signal. The number of peaks where r(t) > 0.5 and t < 0.5 seconds or t > 10 seconds is considered the number of multipath reflections. The number of signal abrupt changes is the number of sharp jumps in signal amplitude. d t > T hThe number of times reflects the interference of motion or equipment vibration, see formula (7).

[0094] (7)

[0095] in, d t This represents the absolute change in signal amplitude between time t and time t+1. Th This is the threshold for mutation detection; std ( d t )for d t The standard deviation of the absolute change in the amplitude of a sequence signal.

[0096] 3. Three-branch feature extraction module:

[0097] A three-branch feature extraction module extracts convolutional features at different time scales based on 3D tensor data. This module includes a short-term feature branch, a medium-term feature branch, and a long-term feature branch. Each of these branches is configured with convolutional layers with different receptive fields; the receptive field of the long-term feature branch is larger than that of the medium-term branch, and the receptive field of the medium-term branch is larger than that of the short-term branch. The short-term feature branch extracts high-frequency convolutional features representing signal fluctuation levels from the 3D tensor data; the medium-term feature branch extracts mid-frequency convolutional features representing periodic changes in the signal; and the long-term feature branch extracts low-frequency convolutional features representing the smoothing trend of the signal. Specifically... Figure 2 As shown:

[0098] (1) Short-term characteristic branches:

[0099] The short-term feature branch consists of a first convolutional layer, a first batch normalization layer, a second convolutional layer, a second batch normalization layer, and a global average pooling layer, connected sequentially. The first convolutional layer has a kernel size of 5 (kernel_size=5), a dilation rate of 1 (dilation=1), a stride of 1 (stride=1), and padding of 2 (padding=2). It is batch normalized using the first batch normalization layer, resulting in 2 input channels and 16 output channels. The second convolutional layer has a kernel size of 3 (kernel_size=3), a dilation rate of 1 (dilation=1), a stride of 1 (stride=1), and padding of 1 (padding=1). It is batch normalized using the second batch normalization layer, resulting in 16 output channels equal to 16 input channels. Finally, global average pooling (avg_pool) is performed to calculate the average value of the entire time series, compressing it into fixed-dimensional features to capture the overall fluctuation level of the high-frequency signal.

[0100] (2) Mid-term characteristic branches:

[0101] The intermediate feature branch consists of a third convolutional layer, a third batch normalization layer, a fourth convolutional layer, a fourth batch normalization layer, and a global max pooling layer, connected sequentially. The third convolutional layer has a kernel size of 15, a dilation rate of 2, a stride of 1, and padding of 14. It is batch normalized using the third batch normalization layer, resulting in 2 input channels and 16 output channels. The fourth convolutional layer has a kernel size of 5, a dilation rate of 2, a stride of 1, and padding of 4. It is batch normalized using the fourth batch normalization layer, resulting in 16 output channels equal to 16 input channels. Finally, global max pooling is performed, taking the maximum value of the entire time series. The most prominent peak in this signal represents the whole, capturing the strongest periodic features of the intermediate frequency signal.

[0102] (3) Long-term characteristic branches:

[0103] The long-term feature branch consists of a fifth convolutional layer, a fifth batch normalization layer, a sixth convolutional layer, a sixth batch normalization layer, and a global adaptive average pooling layer, connected sequentially. The fifth convolutional layer has a kernel size of 31, a dilation rate of 4, a stride of 1, and padding of 60. It is batch normalized using the fifth batch normalization layer, resulting in 2 input channels and 16 output channels. The sixth convolutional layer has a kernel size of 7, a dilation rate of 4, a stride of 1, and padding of 12. It is batch normalized using the sixth batch normalization layer, resulting in 16 output channels equal to 16 input channels. Finally, global adaptive average pooling (adaptive_avg_pool) is performed, automatically adjusting the pooling window size to ensure a fixed output dimension, adapting to inputs of different lengths to capture the smoothing trend of low-frequency signals.

[0104] The embodiments of the present invention specifically designed a lightweight convolutional neural network and a single fully connected layer for wearable devices, which, while meeting the requirements of real-time monitoring, ensures low power consumption and greatly improves the device's battery life.

[0105] 4. Feature Fusion Module:

[0106] The feature fusion module fuses statistical features and convolutional features at different time scales to generate a comprehensive feature vector. The feature fusion module includes an initial fusion submodule, a mapping submodule, and a final fusion submodule.

[0107] In the initial fusion submodule, the convolutional features of the three scales output by the three-branch feature extraction module are concatenated to obtain the primary fusion feature, which integrates the abstract features of the three scales. Specifically, this step uses the torch.cat function to concatenate in the channel dimension (dim=1), and the resulting primary fusion feature is 48-dimensional.

[0108] In the mapping submodule, the primary fusion features are nonlinearly mapped to higher-order convolutional features through a fully connected layer. Specifically, the fc_conv function is used to map the 48-dimensional primary fusion features to 64-dimensional features to improve nonlinear expressive power. The ReLU activation function and BatchNormalization are used for batch normalization to stabilize the output of the fully connected layer.

[0109] In the final fusion submodule, high-order convolutional features and statistical features are concatenated to generate a comprehensive feature vector. Specifically, the `torch.cat` function concatenates 64-dimensional high-order convolutional features with 12-dimensional statistical features, resulting in a 76-dimensional final feature vector. The statistical features provide specific numerical values ​​as interpretable parameters, complementing the detailed patterns of the convolutional features and providing comprehensive input for subsequent dynamic adaptive threshold calculation and tiered alarms. It's important to note that the concatenation of high-order convolutional features and statistical features is crucial because they both originate from raw millimeter-wave radar signals collected by the same user within the same time period. Storing these two types of features separately would require additional alignment of timestamps and user IDs during real-time monitoring, and even minor delays in signal transmission or processing could lead to alignment failures or errors. More importantly, anomaly detection typically sets thresholds based on statistical features from a specific time period (e.g., time period A), necessitating verification using convolutional features corresponding to the same time period to ensure consistency and accuracy. Furthermore, the device must support instant alarms; the concatenated features avoid multiple data calls and associations during runtime, significantly improving processing efficiency and meeting real-time requirements.

[0110] 5. Dynamic Decision Module:

[0111] The dynamic decision-making module dynamically determines the anomaly level based on comprehensive feature vectors and triggers corresponding tiered alarms; specifically, as follows: Figure 3 As shown:

[0112] (1) Dynamically determine the anomaly level based on the comprehensive feature vector:

[0113] 1) Based on the statistical features in the comprehensive feature vector, generate the actual threshold ranges for multiple health states; specifically including:

[0114] Based on the statistical characteristics of the target user in the initial resting state, the initial health baseline of the target user is determined. The initial health baseline includes the resting heart rate baseline, resting heart rate fluctuation, resting respiratory baseline, and resting respiratory fluctuation. Among them, the resting heart rate baseline is the mean heart rate characteristic during the resting period, the resting heart rate fluctuation is the standard deviation characteristic of the heart rate during the resting period, the resting respiratory baseline is the mean respiratory rate characteristic during the resting period, and the resting respiratory fluctuation is the standard deviation characteristic of the respiratory rate during the resting period.

[0115] The exponentially weighted moving average algorithm is used to dynamically update the initial health baseline using the statistical characteristics of the daily resting period to obtain the dynamic baseline at the current moment; see formula (8). If there is no effective resting data on the day, the baseline remains unchanged to avoid interference from abnormal conditions.

[0116] (8)

[0117] in, μ t This serves as the dynamic baseline for the current moment. μ t-1 This is the dynamic baseline of the previous moment; μ 今日静息 The baseline values ​​obtained through statistical characteristics during the daily resting period; weighting coefficients. α This is a key hyperparameter, and its optimal value of 0.8 was determined after evaluating the overall performance of the dynamic decision-making module using the validation set. Specifically, during the model training phase, the direct parameters determined from the training set are evaluated using the validation set to assess different... α The effect of dynamic baseline updates under different values, and their impact on the precision and recall of the final graded alarms. Experimental verification shows that when... α When the value is 0.8, the dynamic baseline can maintain sufficient stability through the previous dynamic baseline to smooth daily physiological fluctuations, and can also respond in a timely manner to trend shifts caused by long-term changes in the user's state. This makes the real-time threshold range generated by the dynamic decision module based on the comprehensive feature vector more reasonable, effectively balancing the sensitive identification of slow health risks with the robustness to instantaneous interference.

[0118] Based on the current dynamic baseline, a two-factor coupling model is used to generate baseline threshold ranges for different health states; these health states include normal, abnormal heart rate, apnea, and arrhythmia; for example, the heart rate baseline threshold is... The respiratory baseline threshold is ; for t The mean of the heart rate signal at any given time; for t Standard deviation of heart rate signal at any given time; fort The average respiratory signal at any given time; for t Standard deviation of respiratory signals at any given time;

[0119] Based on the baseline threshold range, combined with the exercise factor and environmental factor, the actual threshold range under different health conditions is generated; among them, the exercise factor S is mainly determined by the number of signal mutations, and the higher its value, the more intense the exercise, as shown in formula (9); the environmental factor E is mainly determined by the noise energy ratio and the number of multipath reflections, and the higher its value, the more severe the environmental interference, as shown in formula (10).

[0120] (9)

[0121] (10)

[0122] The actual threshold range is shown in formula (11):

[0123] (11)

[0124] (12)

[0125] in, This is the upper limit of the baseline threshold. This is the lower limit of the baseline threshold.

[0126] 2) A fully connected layer maps the high-order convolutional features in the comprehensive feature vector to multiple health state waveforms under different health states, and outputs the confidence scores of each health state. Specifically, the trained conv_head fully connected layer maps the 64-dimensional convolutional features to 4 waveform modes (normal / abnormal heart rate / apnea / arrhythmia), and outputs the confidence scores of each category as soft mode verification.

[0127] 3) For each health status, if the corresponding waveform pattern's vital signal value is outside the actual threshold range and the corresponding confidence level is greater than the preset confidence threshold, then it is determined to belong to that health status. For example, for abnormal heart rate, both conditions must be met simultaneously: the real-time heart rate exceeds the corrected threshold and the confidence level of the abnormal heart rate waveform is >0.85 before an abnormality can be confirmed.

[0128] (2) Trigger the corresponding graded alarm:

[0129] The reliability of the above anomaly determination is evaluated by using a graded alarm response with a collaborative confidence level C. The higher the value, the less motion and environmental interference there is, and the more reliable the anomaly determination is, as shown in formula (13).

[0130] (13)

[0131] Different levels of anomalies will trigger corresponding alarm mechanisms. The alarm mechanisms in this embodiment of the invention are shown in Table 1 below:

[0132] Table 1: Three-Level Alarm Mechanism

[0133]

[0134] The three-level alarm system includes: Level 1, which is a slight deviation (exceeding 1.2 times the threshold) and requires a confidence level C of ≥0.7 to trigger a 5-10 second device vibration and app notification; Level 2, which is a moderate deviation (between 1.2 and 2 times the threshold) and requires a confidence level C of ≥0.5 to trigger a 10-30 second audible and visual notification and SMS notification to the designated emergency contact; and Level 3, which is a severe deviation (exceeding 2 times the threshold or respiratory arrest exceeding 20 seconds), and considering the priority of life, requires only a confidence level of 0.3 to trigger a high-decibel alarm exceeding 30 seconds, automatically dialing emergency services and pushing current physiological data.

[0135] In addition, an interface for user feedback and correction is provided. If a user marks a false alarm, the system analyzes the specific reason. If it is due to baseline deviation, the historical weight will be reduced in the next update to accelerate baseline adaptation. If it is due to interference misjudgment, the two-factor synergy coefficient will be adjusted to enhance or weaken the interference threshold correction.

[0136] In the aforementioned dynamic adaptive threshold and hierarchical alarm, the individual threshold baseline is slowly updated according to the user's health status through personalized baseline and iterative sliding weighted average (EWMA), solving the problem that fixed thresholds cannot adapt to individual differences in health status. Secondly, environmental factors determined by temperature and noise, and motion factors adjusted by exercise intensity are introduced to adjust the threshold, solving the problem of false alarms and missed alarms caused by scene and motion interference, allowing the threshold to evolve from a general group value to an individual's real-time optimal value. Traditional alarm logic relies on triggering an alarm when a single physiological indicator exceeds the threshold, which is essentially an isolated and instantaneous judgment. This results in false alarms due to instantaneous fluctuations being judged as abnormal, and missed alarms due to slow deviations that do not trigger a single exceedance. Therefore, this invention designs a multi-dimensional reconstructed hierarchical alarm mechanism. It avoids false alarms due to small fluctuations by classifying deviation magnitudes, filters instantaneous interference by using duration constraints, and verifies abnormalities by combining heart rate and respiration, providing different alarm levels. A user feedback entry is designed: after an alarm, the APP pops up a feedback window where users can mark false alarms, missed alarms, or accurate alarms and fill in the reason. The system analyzes the feedback data monthly and automatically adjusts key parameters, overcoming the limitation of traditional systems being fixed at the factory and unable to evolve. Traditional alarms often rely on a single method of high-decibel ringing and flashing bright lights, failing to consider the scenario, user status, and alarm level. This leads to insufficient warnings when necessary and excessive interference when unnecessary. This invention employs a tiered alarm system to achieve scenario adaptation: the audible and visual alarms automatically shut down during sleep, using only vibration alerts; alarm volume is reduced during meetings; level adaptation means that minor anomalies trigger only device vibration and app text notifications; moderate anomalies trigger audible and visual alarms and emergency contact SMS messages; and severe anomalies automatically dial emergency contacts and push real-time location information. Privacy protection ensures that emergency contacts can only receive risk alerts and location information, not complete user health data; location information is only sent during severe alarms, and its accuracy can be preset by the user. This invention adheres to a user-centric philosophy, enabling millimeter-wave radar wearable devices to truly provide practical value for precise health protection.

[0137] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0138] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A wearable device adaptive health threshold alarm method based on millimeter-wave radar, characterized in that, Includes the following steps: Vital signals in the chest area of ​​the target user are collected using wearable millimeter-wave radar devices. The vital signals are input into an individual baseline health model; the individual baseline health model includes a preprocessing module, a statistical feature extraction module, a three-branch feature extraction module, a feature fusion module, and a dynamic decision-making module; The life signal is preprocessed using the preprocessing module to obtain three-dimensional tensor data. The statistical feature extraction module extracts statistical features based on the three-dimensional tensor data. The three-branch feature extraction module extracts convolutional features at different time scales based on the three-dimensional tensor data. The feature fusion module fuses the statistical features and the convolutional features at different time scales to generate a comprehensive feature vector. The dynamic decision-making module dynamically determines the anomaly level based on the comprehensive feature vector and triggers the corresponding graded alarm. The dynamic determination of the anomaly level based on the comprehensive feature vector includes: Based on the statistical features in the comprehensive feature vector, the actual threshold ranges for multiple health states are generated. A fully connected layer maps the high-order convolutional features in the comprehensive feature vector to waveform patterns under the multiple health states, and outputs the confidence scores of each health state. For each health state, if the life signal value of the corresponding waveform pattern is outside the actual threshold range and the corresponding confidence level is greater than the preset confidence level threshold, then it is determined to belong to that health state. The step of generating actual threshold ranges for multiple health states based on statistical features in the comprehensive feature vector specifically includes: Based on the statistical characteristics of the target users in their initial resting state, the initial health baseline of the target users is determined; An exponentially weighted moving average algorithm is used to dynamically update the initial health baseline using the statistical characteristics of the daily resting period, thereby obtaining the dynamic baseline at the current moment and generating baseline threshold ranges for different health states. Based on the baseline threshold range, and combined with exercise and environmental factors, the actual threshold ranges under different health states are generated; The dynamic baseline at the current moment is represented as: ;in, μ t This serves as the dynamic baseline for the current moment. μ t-1 This is the dynamic baseline of the previous moment; μ 今日静息 The baseline value is obtained through statistical characteristics during the daily resting period; 0.8 is the weighting coefficient.

2. The adaptive health threshold alarm method for wearable devices based on millimeter-wave radar as described in claim 1, characterized in that, The statistical features include time-domain statistical features, periodic statistical features, and interference statistical features; The time-domain statistical features include: mean, standard deviation, maximum value, minimum value, kurtosis, and skewness; The periodic statistical features include: SDNN and PNN50 of the heart rate RR interval, and the ratio of inspiratory to expiratory time. The statistical characteristics of the interference include: noise energy percentage, number of multipath reflections, and number of signal abrupt changes.

3. The adaptive health threshold alarm method for wearable devices based on millimeter-wave radar as described in claim 1, characterized in that, The three-branch feature extraction module includes a short-term feature branch, a medium-term feature branch, and a long-term feature branch; The short-term feature branch, the intermediate feature branch, and the long-term feature branch are each configured with convolutional layers with different receptive fields; wherein, the receptive field of the convolutional layer of the long-term feature branch is larger than that of the intermediate feature branch, and the receptive field of the convolutional layer of the intermediate feature branch is larger than that of the short-term feature branch. The high-frequency convolutional features characterizing the signal fluctuation level are extracted from the three-dimensional tensor data through the short-term feature branch. Through the intermediate feature branch, mid-frequency convolutional features characterizing the periodic changes of the signal are extracted from the three-dimensional tensor data; The long-term feature branch is used to extract low-frequency convolutional features that characterize the smoothing trend of the signal from the three-dimensional tensor data.

4. The adaptive health threshold alarm method for wearable devices based on millimeter-wave radar as described in claim 3, characterized in that: The short-term feature branch includes: a first convolutional layer, a first batch normalization layer, a second convolutional layer, a second batch normalization layer, and a global average pooling layer connected in sequence; The intermediate feature branch includes: a third convolutional layer, a third batch normalization layer, a fourth convolutional layer, a fourth batch normalization layer, and a global max pooling layer connected in sequence; The long-term feature branch includes: a fifth convolutional layer, a fifth batch normalization layer, a sixth convolutional layer, a sixth batch normalization layer, and a global adaptive average pooling layer connected in sequence.

5. The adaptive health threshold alarm method for wearable devices based on millimeter-wave radar as described in claim 1, characterized in that, The feature fusion module includes an initial fusion submodule, a mapping submodule, and a final fusion submodule; In the initial fusion submodule, the convolutional features of the three scales output by the three-branch feature extraction module are concatenated to obtain the primary fusion features; In the mapping submodule, the primary fusion features are nonlinearly mapped into higher-order convolutional features through a fully connected layer; In the final fusion submodule, the higher-order convolutional features are concatenated with the statistical features to generate a comprehensive feature vector.

6. The adaptive health threshold alarm method for wearable devices based on millimeter-wave radar as described in claim 1, characterized in that, The motion factor is determined by the number of signal abrupt changes; the environmental factor is jointly determined by the noise energy ratio and the number of multipath reflections.

7. The adaptive health threshold alarm method for wearable devices based on millimeter-wave radar as described in claim 1, characterized in that, The individual baseline health model is trained based on a multi-physical, multi-state, and multi-scenario hierarchical human chest cavity region vital signal dataset. The multi-body constitution stratification includes age stratification, BMI stratification, and health status stratification; The multi-state layering includes resting state and motion state, and is associated with dynamic state tags that record medication and fatigue. The multi-scenario layering includes different user postures and different physical environments.

Images