Smart bracelet-based sub-health monitoring and early warning system

Abstract

The application provides a sub-health monitoring and early warning system based on a smart bracelet, and relates to the technical field of intelligent health monitoring and processing. Firstly, the photoelectric pulse wave, inertia and temperature multi-modal data collected by the smart bracelet are preprocessed, the heart rate variability features containing mixed information of metabolic noise and nervous pressure are extracted from the pulse wave, and the mixed factors representing the current body activity and heat metabolism state are constructed by using the inertia and temperature data. Then, based on the orthogonal projection constraint algorithm, the metabolic mixed factors are actively separated from the mixed heart rate variability features in the high-dimensional hidden space to obtain the pure autonomous nervous features after eliminating the environmental and metabolic interference. Finally, the pure features are quantified by the deviation from the individual historical rhythm benchmark, the risk index truly reflecting the psychological or nervous sub-health state is dynamically evaluated, and the graded early warning is triggered.

Description

Technical Field

[0001] This invention belongs to the field of intelligent health monitoring and processing technology, specifically relating to a sub-health monitoring and early warning system based on a smart bracelet. Background Technology

[0002] With the accelerating pace of modern life and the continuous increase in occupational pressure, sub-health has become a significant risk factor affecting public physical and mental health. Sub-health exists in a gray area between health and disease; its early symptoms are insidious and subjectively perceived, and if timely warnings and interventions are not implemented, it can easily develop into irreversible organic diseases such as cardiovascular disease and neurasthenia. Against this backdrop, utilizing wearable devices such as smart bracelets to achieve continuous, non-invasive, and real-time monitoring of the human body's physiological state, and based on this, constructing an early warning mechanism for sub-health, has become an important technological direction in the field of health management. Smart bracelets, with their unique advantages of high portability and long-term wearability, can continuously collect multimodal physiological data such as photoelectric pulse waves, inertial motion, and body surface temperature without interfering with the user's daily life, providing a natural hardware platform for the dynamic assessment of sub-health.

[0003] Currently, existing sub-health monitoring solutions based on smart bracelets mainly rely on static threshold determination of heart rate variability (HRV) indicators to assess users' stress and fatigue levels. However, these solutions have a core flaw in practical applications: they employ a result-oriented, unilinear mapping logic. When a decrease in HRV or an increase in heart rate is detected, the algorithm defaults to mapping it to sympathetic nervous system excitation and directly outputs a high-stress warning for sub-health. In reality, the triggers for sympathetic nervous system excitation are multi-dimensional, including both psychological and neurological stress that truly reflects a sub-health state, and numerous non-targeted interference sources from daily life, such as metabolic heart rate increases caused by gastrointestinal digestion after eating, thermoregulation responses caused by sudden changes in environmental temperature, and metabolic recovery processes after light housework. Because existing algorithms lack contextual awareness of the user's dynamic life scenarios, they cannot effectively decouple physiological metabolic responses from real psychological and neurological stress. These two types of signals are deeply coupled in the feature space, leading to frequent false high-stress warnings generated by the system immediately after a meal or light activity. This high false alarm rate not only severely undermines the practical value and credibility of the sub-health monitoring system, but also causes users to experience trust fatigue due to receiving invalid alarms for a long time, ultimately leading to a loss of compliance with the system.

[0004] Therefore, an optimized sub-health monitoring and early warning system based on smart bracelets is desired. Summary of the Invention

[0005] This invention application provides an adaptive health education and interactive system for patients, comprising: The sensor data preprocessing module is used to filter and preprocess the multimodal sensor data stream collected by the smart bracelet to obtain the wrist pulse wave sequence, wrist inertial sequence and wrist temperature sequence. The multimodal sensor data stream includes photoelectric pulse wave signal, six-axis inertial signal and skin temperature signal. The heart rate variability feature extraction module is used to extract heart rate variability features from wrist pulse wave sequences to obtain a mixed heart rate variability vector. The metabolic state joint encoding module is used to jointly encode the metabolic state of the wrist inertial sequence and the wrist temperature sequence to obtain the metabolic confounding factor tensor. The physiological metabolism and psychoneural feature decoupling module is used to decouple the mixed heart rate variability vector and the metabolic confounding factor tensor based on orthogonal constraints to obtain a pure autonomic neural vector. The sub-health risk quantification module is used to quantify the deviation between the pure autonomic neural vector and the pre-stored individual rhythm benchmark model to obtain the sub-health risk index. The intervention strategy matching module is used to determine the level of the sub-health risk index and match intervention strategies based on the comparison of graded thresholds to obtain graded early warning instructions.

[0006] Compared with existing technologies, this invention proposes an adaptive health education interaction method for patients. First, it preprocesses multimodal data (photoelectric pulse wave, inertia, and temperature) collected by a smart bracelet, extracting heart rate variability features containing mixed information of metabolic noise and neural stress from the pulse wave. Simultaneously, it uses inertia and temperature data to construct confounding factors characterizing current physical activity and metabolic state. Then, based on an orthogonal projection constraint algorithm, it actively removes metabolic confounding factors from the mixed heart rate variability features in a high-dimensional latent space, obtaining pure autonomic neural features free from environmental and metabolic interference. Finally, it quantifies the deviation between this pure feature and an individual's historical rhythm benchmark, dynamically assessing a risk index that truly reflects psychological or neurological sub-health status and triggering tiered early warnings. Attached Figure Description

[0007] Figure 1 This is a system block diagram of a sub-health monitoring and early warning system based on a smart bracelet according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in a sub-health monitoring and early warning system based on a smart bracelet according to an embodiment of this application; Figure 3 This is a block diagram of the physiological metabolism and psychoneural features decoupling module in a sub-health monitoring and early warning system based on a smart bracelet according to an embodiment of this application; Figure 4This is a block diagram of a feature stripping unit in a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Figure 5 This is a block diagram of the sub-health risk quantification module in a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Detailed Implementation

[0008] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0009] Because existing smart bracelet sub-health monitoring solutions cannot distinguish between physiological heart rate variability caused by daily metabolic activities and heart rate variability caused by real psychological and neurological stress when assessing users' stress and fatigue, the two types of signals are deeply coupled at the feature level, causing the system to frequently trigger false warnings when users are in metabolically active scenarios such as post-meal digestion or mild exercise recovery, which seriously reduces the credibility of the monitoring system and user compliance. To address the aforementioned issues, a sub-health monitoring and early warning system based on a smart bracelet is proposed. This system synchronously collects multimodal physiological data from the user using a photoelectric pulse wave sensor, a six-axis inertial sensor, and a temperature sensor built into the smart bracelet. After filtering and attitude calculation preprocessing, pulse wave sequences, inertial sequences, and temperature sequences are obtained, respectively. Subsequently, time-frequency domain heart rate variability analysis is performed on the pulse wave sequences to extract feature vectors containing mixed information. Simultaneously, the inertial and temperature sequences are time-aligned and jointly encoded to generate a mixed factor tensor that can characterize the user's current physical activity level and calorie metabolism level. Based on this, a dual-branch encoder is used to map the above two feature paths to a shared latent space. The projection components of the mixed features in the metabolic mixed direction are calculated using an orthogonal projection constraint algorithm and then stripped to obtain pure autonomic neural features that are completely orthogonal to daily activity metabolism. Finally, the pure features are quantified by Mahalanobis distance deviation from an individual rhythm benchmark model established based on long-term user wear data. The degree of deviation is transformed into a sub-health risk index through nonlinear mapping, and corresponding intervention strategies are matched according to graded thresholds to generate early warning instructions. Through this technological approach, the system can accurately identify sub-health signals that are truly caused by psychological or neurological factors in complex and dynamic life scenarios, effectively suppress false alarms caused by metabolic fluctuations, and achieve high-precision personalized sub-health early warning.

[0010] Figure 1 This is a system block diagram of a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Figure 2This is a schematic diagram of data flow in a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Figure 1 and Figure 2 As shown, the sub-health monitoring and early warning system 100 based on a smart bracelet according to an embodiment of this application includes: a sensor data preprocessing module 110, used to filter and preprocess the multimodal sensor data stream collected by the smart bracelet to obtain a wrist pulse wave sequence, a wrist inertial sequence, and a wrist temperature sequence, wherein the multimodal sensor data stream includes photoelectric pulse wave signals, six-axis inertial signals, and skin temperature signals; a heart rate variability feature extraction module 120, used to extract heart rate variability features from the wrist pulse wave sequence to obtain a mixed heart rate variability vector; and a metabolic state joint encoding module 130, used to process the wrist inertial sequence and wrist temperature... The sequence is jointly encoded with metabolic state to obtain a metabolic confounding factor tensor; the physiological metabolism and psychoneural feature decoupling module 140 is used to decouple the mixed heart rate variability vector and the metabolic confounding factor tensor based on orthogonal constraints to obtain a pure autonomic neural vector; the sub-health risk quantification module 150 is used to quantify the deviation between the pure autonomic neural vector and the pre-stored individual rhythm benchmark model to obtain a sub-health risk index; the intervention strategy matching module 160 is used to determine the level of the sub-health risk index and match intervention strategies based on the graded threshold comparison to obtain graded early warning instructions.

[0011] In the aforementioned sub-health monitoring and early warning system 100 based on a smart bracelet, the sensor data preprocessing module 110 is used to filter and preprocess the multimodal sensor data stream collected by the smart bracelet to obtain a wrist pulse wave sequence, a wrist inertial sequence, and a wrist temperature sequence. The multimodal sensor data stream includes photoelectric pulse wave signals, six-axis inertial signals, and skin temperature signals. It is understandable that the photoelectric pulse wave signal, six-axis inertial signal, and skin temperature signal collected by the smart bracelet during daily wear are all raw analog sensor data. Among them, the photoelectric pulse wave signal inevitably has low-frequency baseline drift noise caused by breathing and changes in skin contact tightness. The six-axis inertial signal inherently contains a constant bias component of the Earth's gravitational acceleration. The skin temperature signal is affected by the sensor's own thermal noise and transient disturbances in ambient temperature, resulting in high-frequency jitter. If the above noise and bias are not eliminated, they will directly pollute the calculation accuracy of subsequent heart rate variability feature extraction and metabolic state encoding, causing the effective physiological information in the feature space to be submerged by noise and unable to be accurately interpreted. Therefore, in the technical solution of this application, the multimodal sensing data stream, including photoelectric pulse wave signals, six-axis inertial signals, and skin temperature signals, collected by the smart bracelet, is filtered and preprocessed with attitude calculation to obtain wrist pulse wave sequences, wrist inertial sequences, and wrist temperature sequences. This process removes interference components from the original sensing signals of each channel and transforms them into standardized time series with clear physical meaning. This ensures that the data input to the subsequent heart rate variability feature extraction module and metabolic state joint encoding module are pure physiological sequences with a high signal-to-noise ratio, fundamentally guaranteeing the computational reliability and assessment accuracy of the entire sub-health monitoring and early warning system at the feature level.

[0012] Specifically, in the embodiments of this application, the sensor data preprocessing module is used to: perform baseline drift filtering and peak detection on the photoelectric pulse wave signal to obtain a wrist pulse wave sequence; perform triaxial acceleration synthesis and gravity bias subtraction on the six-axis inertial signal to obtain a wrist inertial sequence; and perform thermal noise suppression and continuous serialization on the skin temperature signal to obtain a wrist temperature sequence.

[0013] More specifically, in a specific example of this application, the implementation process of filtering and attitude calculation preprocessing for the multimodal sensing data stream is as follows: First, for baseline drift filtering and peak detection processing of the photoelectric pulse wave signal, the system performs analog-to-digital conversion on the voltage analog signal output by the green photoplethysmography (PPG) sensor at the bottom of the smart bracelet at a fixed sampling frequency to obtain a discrete sequence of original sampling points of the photoelectric pulse wave signal. This sequence is superimposed with the low-frequency baseline drift component transmitted to the wrist via venous return due to the periodic changes in thoracic pressure caused by the user's breathing movements, and the slowly varying component of optical coupling intensity caused by the change in wrist posture due to the tightness of the bracelet strap. The system inputs this discrete sampling point sequence into a first-order infinite impulse response high-pass filter with a cutoff frequency set to 0.5 Hz. This filter applies a pole feedback attenuation coefficient to the current filtered output value. The weighted gain scaling value, which is the difference between the original sampled values ​​at the current time and the previous time, is recursively superimposed to attenuate and suppress the baseline drift component with a frequency below 0.5 Hz from the effective pulse waveform point by point, and outputs a baseline-degraded pulse wave signal that retains only the effective waveform fluctuations within the cardiac cycle frequency band. Subsequently, the system applies an adaptive peak detection algorithm to the baseline-degraded pulse wave signal. This algorithm uses the sum of the local mean and standard deviation of the signal amplitude within the sliding time window as a dynamic detection threshold. It scans the baseline-degraded pulse wave signal point by point and marks the sampling points whose amplitude exceeds the dynamic threshold and simultaneously meet the condition that the time interval between the peak and the previous detected peak is greater than the physiological minimum cardiac cycle constraint as the effective systolic main peak. The timestamp coordinates of all marked effective main peaks and their corresponding peak amplitudes are arranged and encapsulated in chronological order to obtain the wrist pulse wave sequence. Next, for the three-axis acceleration synthesis and gravity bias subtraction processing of the six-axis inertial signal, the system synchronously reads the raw acceleration measurements in the three orthogonal directions (X-axis, Y-axis, and Z-axis) from the six-axis microelectromechanical inertial measurement unit built into the smart bracelet. Each of these three axial measurements is superimposed with a constant projection component of the Earth's gravitational field along that axis. The system squares each of the three axial acceleration measurements, sums them, and then performs an arithmetic square root operation on the sum to obtain the signal vector amplitude characterizing the instantaneous synthesized acceleration. This signal vector amplitude contains... The system incorporates the contributions of static gravitational acceleration and the user's dynamic limb movement acceleration. The system subtracts the standard gravitational acceleration constant of 9.81 m / s² from the amplitude of the signal vector to eliminate the constant bias effect of the Earth's gravitational field on acceleration measurement, so that the output only reflects the net acceleration component generated by the user's wrist dynamic movement. The system continuously splices and assembles the net acceleration amplitude calculated at each sampling time step according to equal time intervals to obtain a wrist inertial sequence that can characterize the evolution trend of the intensity of the user's body activity over time.Then, for thermal noise suppression and continuous serialization processing of the epidermal temperature signal, the system reads the raw sampled value of the epidermal temperature signal from the infrared thermal temperature sensor attached to the skin at the bottom of the smart bracelet. This sampled value exhibits high-frequency random fluctuations due to the thermal noise of the sensor's own circuitry and transient temperature fluctuations caused by microcirculatory blood flow pulsations on the skin surface. The system applies an exponentially weighted moving average filter to this raw sampled value. Specifically, the weighted contribution obtained by multiplying a preset smoothing factor by the raw temperature sampled value at the current moment is multiplied by a smoothed temperature value output from the previous moment after subtracting the smoothing factor. The historical inertial contributions are summed to obtain the smoothed temperature output value at the current moment. The smoothing factor is set between 0.1 and 0.3 to ensure sufficient suppression of high-frequency thermal noise without excessively delaying the tracking response to the actual temperature change trend of the body surface. The system continuously serializes and reassembles the smoothed temperature values ​​output by exponential weighted moving average filtering at each sampling time according to the same equal time interval as the wrist inertial sequence, to obtain a wrist temperature sequence that can stably reflect the small heat change trend of the user's body surface caused by basal metabolic heat production and peripheral vasomotor regulation.

[0014] In the aforementioned sub-health monitoring and early warning system 100 based on a smart bracelet, the heart rate variability feature extraction module 120 is used to extract heart rate variability features from the wrist pulse wave sequence to obtain a mixed heart rate variability vector. It should be understood that although the wrist pulse wave sequence obtained after preprocessing has eliminated baseline drift noise and calibrated the effective peak positions, it is still raw time-amplitude waveform data and has not yet been transformed into a quantitative feature representation that can characterize the state of autonomic nervous system regulation. The information on the regulation of heart rhythm by the sympathetic and parasympathetic branches of the autonomic nervous system is precisely implicit in the small fluctuations between adjacent heartbeats. These fluctuations need to be explicitly extracted and quantified through time-domain statistics and frequency-domain power spectrum analysis. Furthermore, the extracted heart rate variability features necessarily include both the autonomic nervous system regulatory response caused by real psychological and neurotic stress and the physiological heart rate fluctuations caused by daily metabolic activities; the two are in a mixed and superimposed state at this feature level. Therefore, in the technical solution of this application, heart rate variability features are further extracted from the wrist pulse wave sequence to obtain a mixed heart rate variability vector, thereby transforming the original waveform data into a high-dimensional quantitative vector representation covering time-domain statistical features and frequency-domain power distribution features. This provides an input feature basis containing complete autonomic neural regulation information for the subsequent orthogonal projection decoupling module, enabling the metabolic hybrid components and neural stress components to have separable mathematical conditions in this feature space, laying a data foundation at the feature level for accurately separating physiological metabolic fluctuations.

[0015] Specifically, the heart rate variability feature extraction module is used to: extract time-frequency domain features of heart rate variability from the wrist pulse wave sequence based on adaptive peak detection and fast Fourier transform to obtain a mixed heart rate variability vector. In a specific example of this application, firstly, the system reads the timestamp coordinates of each calibrated effective systolic main peak from the wrist pulse wave sequence, calculates the time difference between every two adjacent main peak timestamps, and uses this time difference sequence as the RR interval sequence. The value of each element in the RR interval sequence is the length of the time interval between two adjacent heartbeats, recorded in milliseconds. The length of this sequence is equal to the total number of effective main peaks in the wrist pulse wave sequence minus one. Next, the system performs time-domain heart rate variability calculations on the RR interval sequence. Specifically, it calculates the root mean square (RMS) value of the differences between all adjacent RR intervals in the sequence. This is done by taking the difference between each pair of consecutive RR intervals, squaring that difference, summing all the squared differences, dividing by the logarithm of the total differences, and then taking the square root of the quotient. The resulting RMS value of the adjacent interval difference is used as a time-domain characteristic index. This index reflects the short-term regulatory capacity of the parasympathetic nervous system on cardiac rhythm, and its calculation is expressed as follows: in, The root mean square difference between adjacent intervals extracted from the wrist pulse wave sequence is used as a core indicator of temporal heart rate variability. The total number of valid main peaks detected from the wrist pulse wave sequence determines the sample size of the RR interval that can be used for calculation. and , representing the time lengths of the i-th and (i+1)-th adjacent heartbeat intervals in the RR interval sequence, respectively. In sub-health monitoring scenarios, the sum of squared differences of adjacent RR intervals in the numerator of this formula reflects the intensity of heartbeat rhythm fluctuations on a beat-by-beat scale. When a user is in a state of psychological anxiety or nervous fatigue, decreased parasympathetic tone leads to a rigidity in the beat-by-beat intervals, resulting in a decrease in the sum and a corresponding reduction in the root mean square difference between adjacent intervals. Conversely, when a user is in the postprandial digestion period or a mild activity recovery period, metabolic sympathetic activation also suppresses beat-by-beat fluctuations, causing this indicator to decrease. This is a concrete manifestation of the superposition of metabolic noise and neural stress signals in this time-domain feature. Then, the system calculates the frequency domain heart rate variability characteristics of the RR interval sequence. Specifically, it first performs equal-interval resampling interpolation on the RR interval sequence to convert it into a discrete signal with uniform time intervals. Then, it applies a fast Fourier transform to the uniformly sampled signal to convert it from the time domain to the frequency domain, obtaining the power spectral density distribution. The system delineates the low-frequency band from 0.04 Hz to 0.15 Hz in the power spectral density distribution and calculates the integral power value in this band as the low-frequency power characteristic. It also delineates the high-frequency band from 0.15 Hz to 0.4 Hz and calculates the integral power value in this band as the high-frequency power characteristic. The low-frequency power characteristic reflects the joint regulatory activity of the sympathetic and parasympathetic nervous systems, while the high-frequency power characteristic mainly reflects the regulatory activity of the parasympathetic nervous system on respiratory sinus arrhythmia. The system further calculates the ratio of the low-frequency power characteristic to the high-frequency power characteristic as the sympathetic-vagal balance index. Furthermore, the system vectorizes and encapsulates the calculated root mean square interval difference, low-frequency power characteristics, high-frequency power characteristics, and sympathetic-vagal balance index according to a fixed dimensional order to generate a hybrid heart rate variability vector. The feature values ​​of each dimension in this vector simultaneously carry information on autonomic nervous system dysfunction caused by real psychological and neurological sub-health conditions and physiological heart rate fluctuations caused by the user's current physical activity and metabolic state. The two types of information are in a hybrid coupled state in this vector space, waiting for the subsequent orthogonal projection decoupling module to separate and process them.

[0016] In the aforementioned sub-health monitoring and early warning system 100 based on a smart bracelet, the metabolic state joint encoding module 130 is used to perform metabolic state joint encoding on the wrist inertial sequence and wrist temperature sequence to obtain the metabolic confounding factor tensor. It is understandable that, since the wrist inertial sequence and wrist temperature sequence obtained after preprocessing exist in the form of independent time series, although they each carry information on the user's physical activity intensity and body surface heat metabolism, they differ significantly in sampling frequency, physical dimensions, and data dimensions. The sampling frequency of the wrist inertial sequence is 50 Hz, while the sampling frequency of the wrist temperature sequence is only 1 Hz. Furthermore, the metabolic state synergy between the two is still implicit and has not been explicitly modeled. The subsequent orthogonal projection decoupling module needs a hybrid factor tensor that can fully represent the user's current overall metabolic activity level in a unified high-dimensional feature space as a reference basis for the projection direction. This is to accurately locate and separate the physiological heart rate fluctuation component caused by metabolic activity from the hybrid heart rate variability vector. Without this unified metabolic state fusion representation, the orthogonal projection will lose a clear projection direction and will not be able to achieve effective feature decoupling. Therefore, in the technical solution of this application, the wrist inertial sequence and wrist temperature sequence are further jointly encoded to obtain a metabolic confounding factor tensor. This integrates the two heterogeneous environmental physiological sequences into a unified tensor representation that can fully characterize the user's current physical activity level and calorie metabolism level in a high-dimensional feature space. This provides a metabolic confounding direction basis with clear physical meaning and sufficient characterization capabilities for the subsequent orthogonal projection decoupling module. This allows the system to accurately identify which components in the mixed heart rate variability vector are physiological fluctuations caused by daily metabolic activities during feature stripping, thus ensuring the directional accuracy and thoroughness of the decoupling operation.

[0017] Specifically, in the embodiments of this application, the metabolic state joint encoding module is used to: perform timestamp synchronization alignment and column vector concatenation on the wrist inertial sequence and wrist temperature sequence to obtain an aligned environment sequence; and perform dimensionality reduction and feature cross-coding on the aligned environment sequence to obtain a metabolic confounding factor tensor.

[0018] More specifically, in a specific example of this application, the implementation process of jointly encoding the metabolic state of the wrist inertial sequence and the wrist temperature sequence is as follows: First, the system performs timestamp synchronization and alignment processing on the wrist inertial sequence and the wrist temperature sequence. Since the sampling frequency of the inertial sensor in the smart bracelet is 50 Hz while the sampling frequency of the temperature sensor is 1 Hz, there is a 50-fold difference in the sampling density of the two sequences on the time axis. The system uses the timestamp of the wrist inertial sequence as a unified reference time axis and uses linear interpolation to fill in the missing moments between adjacent sampling points in the wrist temperature sequence. That is, for the 49 missing timestamp positions between any two adjacent valid sampling points in the wrist temperature sequence, the corresponding interpolated temperature values ​​are calculated and filled in one by one according to the linear proportional relationship between the two valid sampling values, so that the wrist temperature sequence has a corresponding temperature value at every timestamp position that is the same as that of the wrist inertial sequence, thereby achieving strict point-by-point alignment of the two sequences in the time dimension. Next, at each aligned timestamp, the system vertically concatenates the net acceleration amplitude of the wrist inertial sequence and the smoothed temperature value of the wrist temperature sequence into a two-dimensional column vector. The two-dimensional column vectors at all timestamps are then arranged and assembled in chronological order to obtain the aligned environment sequence. Each time step data element in this sequence contains both the user's physical activity intensity information and body surface heat metabolism information at that moment. These two heterogeneous environmental physiological indicators are uniformly expressed as a structured multidimensional data stream under the same time coordinate. Then, the system inputs the aligned environment sequence into a pre-trained multilayer perceptron network for dimensionality reduction and feature cross-encoding. The multilayer perceptron consists of three cascaded fully connected layers. The first fully connected layer expands the two-dimensional input vector of each time step in the aligned environment sequence to a 32-dimensional intermediate representation to capture the primary cross-correlation features between body activity and body surface temperature. The second fully connected layer further maps the 32-dimensional intermediate representation to a 64-dimensional deep representation to extract higher-order nonlinear metabolic synergistic patterns between the two signals. The third fully connected layer compresses the 64-dimensional deep representation to the target dimension output space to complete feature refinement and dimensionality reduction. Each fully connected layer is followed by a modified linear unit activation function to introduce nonlinear mapping capabilities, enabling the network to learn the complex nonlinear metabolic coupling relationship between activity intensity and body surface temperature. The final high-dimensional feature representation not only encodes the statistical characteristics of each individual signal but also encodes the covariance and interaction patterns of the two signals in the time dimension.Furthermore, the system performs global average pooling on the time dimension on the high-dimensional feature vector output by the multilayer perceptron network after processing each time step in the aligned environment sequence. This compresses the variable-length time series features into a fixed-dimensional global representation vector, which is encapsulated as a metabolic confounding factor tensor. This tensor encapsulates a comprehensive high-order feature representation of the user's physical activity and calorie metabolism within the current analysis time window. This provides complete metabolic state reference information for the subsequent dual-branch encoder network to map the confounding factor vector and the mixed heart rate variability vector into the latent space for orthogonal projection decoupling.

[0019] In the aforementioned sub-health monitoring and early warning system 100 based on a smart bracelet, the physiological metabolism and psychoneural feature decoupling module 140 is used to decouple the mixed heart rate variability vector and the metabolic confounding factor tensor based on orthogonal constraints to obtain a pure autonomic neural vector. It should be understood that the mixed heart rate variability vector obtained through the aforementioned steps contains two fundamentally different information components: one is the autonomic neural regulatory response component caused by the user's actual psychoneural stress or fatigue state, and the other is the physiological heart rate fluctuation component caused by the user's current physical activity and basal calorie metabolism. These two components are deeply coupled in the original feature space and have complex nonlinear cross-influence relationships. Although the metabolic confounding factor tensor has completely encoded the user's current metabolic activity state information, the two are in different feature spaces with significant differences in dimensionality and distribution characteristics. Therefore, effective component separation operations cannot be performed directly in the original space. An algorithm mechanism is needed that can unify the two heterogeneous features into the same metric space and accurately separate the metabolic contribution component in that space based on the principle of geometric orthogonality. Therefore, in the technical solution of this application, the mixed heart rate variability vector and the metabolic confounding factor tensor are further decoupled based on orthogonal constraints to obtain a pure autonomic neural vector. This allows for the precise identification and complete removal of the metabolically contributed components from the mixed features in the high-dimensional shared latent space using the geometric principle of vector orthogonal projection, retaining only the pure neural regulation feature components that are strictly orthogonal to the metabolic direction. This ensures that the output pure autonomic neural vector is completely independent of the user's current physical activity state and metabolic level, preventing the system from triggering false high-stress warnings due to metabolic heart rate fluctuations when the user is in the postprandial digestion period, mild exercise recovery period, or any metabolically active scenario. This fundamentally solves the core technical problem of high false alarm rates caused by the deep coupling of physiological metabolic fluctuations and psychological neural stress in existing solutions.

[0020] Figure 3 This is a block diagram of the physiological metabolism and psychoneurological characteristics decoupling module in a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Figure 3 As shown, the physiological metabolism and psychoneural feature decoupling module 140 includes: a latent space nonlinear mapping unit 141, used to perform high-dimensional latent space nonlinear mapping processing on the mixed heart rate variability vector and the metabolic confounding factor tensor through a dual-branch encoder network to obtain a latent space mixed vector and a latent space confounding tensor; an orthogonal projection unit 142, used to orthogonally project the latent space mixed vector along the direction of the latent space confounding tensor to obtain a physiological fluctuation baseline vector; and a feature stripping unit 143, used to perform feature stripping on the latent space mixed vector and the physiological fluctuation baseline vector to obtain a pure autonomic neural vector.

[0021] Specifically, the latent space nonlinear mapping unit 141 is used to perform high-dimensional latent space nonlinear mapping processing on the mixed heart rate variability vector and the metabolic confounding factor tensor through a dual-branch encoder network to obtain the latent space mixed vector and the latent space confounding tensor, respectively. It should be understood that since the mixed heart rate variability vector and the metabolic confounding factor tensor originate from different sensing channels and are generated through different feature extraction paths, their dimensionality, numerical dimensions, and distribution characteristics in the original feature space differ significantly. The mixed heart rate variability vector is a low-dimensional dense vector based on time-frequency domain statistics, while the metabolic confounding factor tensor is a high-dimensional nonlinear feature representation encoded by a multilayer perceptron. If orthogonal projection operations are directly performed in these two heterogeneous original feature spaces, since they are not in the same metric space and the interference mechanism of metabolic activity on heart rate variability itself has complex nonlinear coupling characteristics, linear projection will not be able to accurately capture the true contribution direction of metabolic components in the mixed features, leading to directional deviations and incomplete feature stripping in subsequent steps. Therefore, in the technical solution of this application, a high-dimensional latent space nonlinear mapping process is further performed on the mixed heart rate variability vector and the metabolic confounding factor tensor through a dual-branch encoder network to obtain the latent space mixed vector and the latent space confounding tensor. This unifies the two heterogeneous features into the same high-dimensional shared latent space coordinate system, and the complex nonlinear coupling relationship in the original space is unfolded into an approximately linearly separable geometric structure in the latent space through nonlinear transformation. In this way, the subsequent orthogonal projection operation can be performed under the unified metric space and linearized geometric conditions, ensuring the accuracy of the projection direction and the mathematical effectiveness of feature stripping, fundamentally guaranteeing that the metabolic confounding components can be accurately located and completely separated from the mixed features.

[0022] More specifically, in a specific example of this application, the system first constructs a dual-branch encoder network architecture, which contains two structurally independent encoder branches with the same output dimension, namely the heart rate variability encoder branch and the metabolic state encoder branch. Each branch consists of three fully connected layers cascaded together. After each fully connected layer, a modified linear unit activation function is connected to introduce non-linear mapping capability. The final output layer dimension of both branches is set to 128 dimensions to ensure that the two features are mapped to a shared latent space of the same dimension. Next, the system inputs the mixed heart rate variability vector into the heart rate variability encoder branch. The first fully connected layer of this branch expands the mixed heart rate variability vector from its original dimension to a 64-dimensional intermediate representation. The second fully connected layer further expands the 64-dimensional intermediate representation to a 128-dimensional deep representation. The third fully connected layer performs feature refinement mapping in the 128-dimensional space. The weight matrix of each layer is multiplied with the input vector and the bias vector is superimposed. Then, a nonlinear transformation is performed through the modified linear unit activation function. Finally, a 128-dimensional latent space mixed vector is output. This vector encodes the nonlinear expansion representation of all the autonomic neural regulation information and metabolic coupling information carried by the mixed heart rate variability vector in the high-dimensional latent space. Then, the system inputs the metabolic confounding factor tensor into the metabolic state encoder branch. This branch performs a layer-by-layer nonlinear mapping on the metabolic confounding factor tensor using the same three-layer cascaded fully connected structure as the heart rate variability encoder branch. This mapping transforms the tensor layer by layer from its original feature dimension to a 128-dimensional shared latent space, identical to the latent space mixing vector, outputting a latent space confounding tensor. This tensor encodes the nonlinear high-order feature representation of the user's physical activity intensity and calorie metabolism level carried by the metabolic confounding factor tensor in the shared latent space. Furthermore, since the two encoder branches are jointly optimized during training using a shared loss function, the learned mapping parameters ensure that the latent space mixing vector and the latent space confounding tensor are in the same coordinate system within the same metric space. The geometric relationship between them accurately reflects the direction and intensity of metabolic activity's interference with heart rate variability features, thus providing the mathematical prerequisite for subsequent orthogonal projection units to perform precise directional projection in a unified latent space.

[0023] Specifically, the orthogonal projection unit 142 is used to orthogonally project the latent space mixing vector along the direction of the latent space hybrid tensor to obtain the physiological fluctuation baseline vector. It should be understood that the latent space mixing vector obtained after mapping by the dual-branch encoder network simultaneously contains two fundamentally different information components: one is the autonomic nervous system regulatory response component caused by the user's actual psychological and neurological stress or fatigue state, and the other is the physiological heart rate fluctuation component caused by the user's current physical activity and basal metabolic rate. The directional information of the latter in the latent space has been fully encoded and represented by the latent space hybrid tensor. Therefore, a mathematical operation is needed to accurately calculate the magnitude of the component along the metabolic hybrid direction in the latent space mixing vector, thereby explicitly separating this metabolic contribution component from the mixed features as an independent data object, providing a clear subtraction object for subsequent feature stripping operations. Therefore, in the technical solution of this application, the latent space mixing vector is further orthogonally projected along the direction of the latent space mixing tensor to obtain the physiological fluctuation baseline vector. This allows for the precise quantification of the component contributed by metabolic activity in the latent space mixing vector within the geometric framework of a high-dimensional shared latent space, and the extraction of this component as an independent vector representation. This enables the explicit extraction of metabolic interference components originally hidden in the mixing features, allowing subsequent feature extraction units to completely remove them from the latent space mixing vector through simple vector subtraction operations. This ensures that the final output pure autonomic neural vector reflects only the user's true psychological or neurological sub-health state without being contaminated by daily metabolic activities.

[0024] More specifically, in a concrete example of this application, the process of orthogonally projecting the latent space mixing vector along the direction of the latent space hybrid tensor is as follows. First, the system receives the latent space mixing vector and the latent space hybrid tensor output by the dual-branch encoder network. Both are 128-dimensional high-dimensional vectors and reside in the same shared latent space coordinate system. The system performs an element-wise multiplication operation on the latent space mixing vector and the latent space hybrid tensor, summing all product terms to obtain a scalar value. This scalar value measures the consistency between the projection length and direction of the latent space mixing vector in the metabolic direction indicated by the latent space hybrid tensor. Next, the system performs an element-wise squaring operation on the latent space hybrid tensor itself and sums the results to calculate the square of its L2 norm. This square value characterizes the magnitude scale of the latent space hybrid tensor in high-dimensional space and is used to normalize the projection operation to eliminate the scale influence of the difference in the output amplitude of the metabolic encoder on the projection result. Then, the system divides the scalar of the above vector inner product by the square of the L2 norm to obtain the normalized projection coefficients. These normalized projection coefficients are then multiplied by a vector using a scalar multiplication operation with the latent space hybrid tensor to obtain the physiological fluctuation baseline vector, which is expressed as follows: in, The output physiological fluctuation baseline vector is the explicit vector representation in the latent space of the component contributed by metabolic activity in the latent space mixture vector. This represents the latent space mixing vector output by the heart rate variability encoder branch, which simultaneously contains neural stress components and metabolic fluctuation components. This represents the latent space hybrid tensor output by the metabolic state encoder branch, whose direction encodes the characteristic orientation of metabolic activity in the latent space. This represents the inner product operation of the latent space mixing vector and the latent space hybrid tensor in 128-dimensional space. The inner product value reflects the strength of the metabolic-related components in the mixed features. The L2 norm squared value of the latent space confounding tensor is used as the normalization denominator to eliminate the scaling effect of metabolic encoding amplitude. In the sub-health monitoring scenario, the numerator of this formula, namely the vector inner product, captures the intensity of interference of metabolic activity on heart rate variability features when the user is in the postprandial digestion period or the mild exercise recovery period. The larger the inner product value, the more significant the contribution of metabolic activity to heart rate fluctuation at the current moment. The norm squared in the denominator performs directional normalization on this intensity, ensuring that the projection result only reflects the proportion of the component in the direction and is not affected by the absolute amplitude of the metabolic encoder output. Finally, the normalization coefficient is multiplied by the latent space confounding tensor itself, so that the output physiological fluctuation baseline vector is completely consistent with the direction of metabolic confounding in direction, and its amplitude is exactly equal to the projection length of the latent space mixing vector in that direction, thus extracting the part of the vector component that belongs to metabolic contribution in the mixed features in a geometric sense. Furthermore, the system encapsulates the calculated physiological fluctuation baseline vector and passes it to the feature stripping unit along with the latent space mixing vector, providing an accurate metabolic baseline subtraction object for subsequent vector subtraction operations to obtain pure autonomic neural vectors.

[0025] Specifically, the feature stripping unit 143 is used to strip features from the latent space mixture vector and the physiological fluctuation baseline vector to obtain a pure autonomic neural vector. It should be understood that since the physiological fluctuation baseline vector has already been obtained after orthogonal projection, this vector accurately represents the physiological heart rate fluctuation component caused by the user's daily physical activities and calorie metabolism in the latent space mixture vector within the high-dimensional shared latent space. The remaining part of the latent space mixture vector, excluding this metabolic contribution component, represents the target feature information reflecting the user's true psychological or neurological stress state. An operation is needed to completely remove the identified metabolic interference component from the mixture features, ensuring that the output only retains the pure neural regulation features directly related to the sub-health state. If this stripping operation is not performed, the metabolic component will remain in the feature representation and cause a systematic overestimation of the Mahalanobis distance in subsequent deviation quantification stages, ultimately leading to an artificially high sub-health risk index and false alarms. Therefore, in the technical solution of this application, feature stripping is further performed on the latent space mixture vector and the physiological fluctuation baseline vector to obtain a pure autonomic neural vector. This completely removes the metabolic interference components that have been explicitly extracted by orthogonal projection from the mixture features in the high-dimensional latent space, retaining only the pure neural regulation feature components orthogonal to the metabolic direction. In this way, it can be ensured that the feature data flowing into the subsequent sub-health risk quantification module is completely independent of the user's current physical activity state and metabolic level. This ensures that the system will not trigger false high-stress warnings due to metabolic heart rate fluctuations when the user is in the postprandial digestion period, the mild exercise recovery period, or any metabolically active scenario. This fundamentally solves the core technical problem of high false alarm rate caused by the deep coupling of physiological metabolic fluctuations and psychological neural stress in existing solutions.

[0026] More specifically, in a concrete example of this application, firstly, the system receives the physiological fluctuation baseline vector output by the orthogonal projection unit and the latent space mixing vector retained from the upstream. Both are 128-dimensional high-dimensional vectors located in the same shared latent space coordinate system. The latent space mixing vector carries the mixed information of neural pressure components and metabolic fluctuation components, while the physiological fluctuation baseline vector only carries the precise representation of the metabolic fluctuation component in this space. Next, the system performs element-wise arithmetic subtraction on the latent space mixing vector and the physiological fluctuation baseline vector in each feature dimension of the 128-dimensional space. That is, at each dimension, the value of the latent space mixing vector in that dimension is subtracted from the value of the physiological fluctuation baseline vector in the same dimension, resulting in a 128-dimensional residual vector. This residual vector is geometrically located in a hyperplane strictly orthogonal to the direction of the latent space mixing tensor, and its calculation is expressed as follows: in, This represents the pure autonomic neural vector output, which is a 128-dimensional high-dimensional feature representation that retains only psychological or neurotic stress information after feature stripping. This represents the latent space hybrid vector output by the heart rate variability encoder branch, which simultaneously contains hybrid encoded information of neural stress components and metabolic fluctuation components. This represents the baseline vector of physiological fluctuations calculated from orthogonal projection units, which precisely characterizes the portion of the latent space mixture vector contributed by daily metabolic activities. In the context of sub-health monitoring, the minuend of this formula... This represents the complete autonomic nervous system state encoding of a user at a given moment. It includes both information on excessive sympathetic activation caused by genuine sub-health factors such as work anxiety or lack of sleep, and information on metabolic sympathetic activation caused by the user being in the post-meal digestion period or the recovery period from light housework. (Minus) This precisely corresponds to the latter, i.e., the pure metabolic contribution, whose direction is consistent with the latent space hybrid tensor and whose magnitude is equal to the projection length of the hybrid vector onto the metabolic direction; the difference between the two is... Geometrically orthogonal to the metabolic confounding direction, this means the vector no longer contains any information components related to physical activity or calorie metabolism, retaining only pure feature representations that accurately reflect the user's psychoneurotic sub-health state. The system then performs L2 norm normalization on the calculated 128-dimensional residual vector, scaling it to a unit hypersphere to eliminate inconsistencies in vector magnitude scales caused by physiological differences among users, ensuring cross-individual comparability in the subsequent deviation quantification stage's Mahalanobis distance calculation. Furthermore, the system encapsulates the normalized residual vector into a pure autonomic neural vector. This vector, as the final output of the entire physiological metabolism and psychoneurotic feature decoupling module, is passed to the downstream sub-health risk quantification module for deviation quantification calculation against the individual rhythm benchmark model. Since metabolic interference components have been completely eliminated, the subsequently calculated Mahalanobis distance deviation will only reflect the user's true degree of neurological sub-health deviation, thus ensuring the accuracy of the sub-health risk index assessment and the reliability of early warning triggering.

[0027] Specifically, in the sub-health monitoring and early warning scenario based on smart bracelets, the feature stripping unit uses latent space orthogonal subtraction to achieve feature decoupling. Its core calculation logic is: the pure autonomic nervous regulation feature vector equals the latent space mixed vector minus the physiological fluctuation baseline vector. This linear subtraction operation implicitly assumes that metabolic and neural components are completely linearly separable in the latent space, meaning there is no cross-influence between them. However, in the actual physiological regulation system of the human body, the relationship between metabolic activity and psychological stress is not a simple linear superposition, but rather exhibits a significant nonlinear cross-coupling effect. Specifically, when the bracelet wearer is in a metabolically active state (such as the post-meal digestion period or the recovery period after light housework) while also experiencing psychological stress (such as work anxiety or emotional tension), the sympathetic nervous system amplifies the response gain to psychological stress due to the metabolic load. That is, metabolic activity multiplicatively modulates or distorts the expression intensity of neural stress in the heart rate variability feature space. This modulation effect manifests as a multiplicative interaction term in the latent space, rather than an additive component. Therefore, the true latent space mixture vector actually consists of three parts: a pure neural component, a pure metabolic baseline component, and a metabolic-neural coupling interaction component. The simple linear subtraction in the first embodiment only strips away the pure metabolic baseline component, leaving the coupling interaction component as residual noise contamination in the output pure vector.

[0028] In typical sub-health monitoring application scenarios, such as when a user faces work pressure immediately after completing a period of light exercise, the physiological response during the metabolic recovery period is deeply intertwined with the actual psychological and neurological stress. The residual coupling interaction can lead to a systematic overestimation of the deviation of the Mahalanobis distance in subsequent calculations, resulting in an artificially high comprehensive index of sub-health risk and false alarms of high stress after exercise. This seriously weakens the credibility of the early warning system and user compliance.

[0029] To address the deficiency of incomplete feature decoupling caused by the residual metabolic-neural nonlinear coupling interaction mentioned above, a second embodiment is proposed. More specifically, in another specific example of this application, Figure 4 This is a block diagram of the feature stripping unit in a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Figure 4As shown, the feature stripping unit 143 includes: a metabolic-neural nonlinear coupling interaction subunit 1431, used to perform metabolic-neural nonlinear coupling interaction on each feature dimension of the latent space mixture vector and the physiological fluctuation baseline vector to obtain a metabolic-neural coupling interaction vector; an activity intensity adaptive mapping subunit 1432, used to perform activity intensity adaptive mapping processing on the concatenation of the physiological fluctuation baseline vector and the metabolic-neural coupling interaction vector to obtain an adaptive coupling gating coefficient; and a joint feature stripping subunit 1433, used to perform joint feature stripping processing on the latent space mixture vector and the physiological fluctuation baseline vector based on the adaptive coupling gating coefficient to obtain a pure autonomic neural vector.

[0030] Accordingly, the metabolic-neural nonlinear coupling interaction subunit 1431 is used to perform metabolic-neural nonlinear coupling interaction on each feature dimension of the latent space mixing vector and the physiological fluctuation baseline vector to obtain the metabolic-neural coupling interaction vector. It is understandable that, in the real physiological regulation system of the human body, the relationship between metabolic activity and psychological stress is not a simple linear superposition, but rather a significant nonlinear cross-coupling effect. Specifically, when a smart bracelet wearer is in a metabolically active state while also experiencing psychological stress, the sympathetic nervous system will amplify the response gain to psychological stress due to the presence of metabolic load. That is, metabolic activity will enhance or distort the expression intensity of nerve stress in the heart rate variability feature space in a multiplicative modulation manner. This modulation effect manifests as a dimensional multiplicative coupling pattern rather than an additive component in the latent space. Therefore, the real latent space mixture vector actually consists of three parts: a pure neural component, a pure metabolic baseline component, and a metabolic-neural coupling interaction component. The simple linear subtraction in the first embodiment mentioned above only stripped the pure metabolic baseline component, leaving the coupling interaction component as residual noise pollution in the output pure vector. Since the modulation effect of metabolic activity on the expression of nerve stress presents as a dimensional multiplicative coupling pattern in the feature space, it is necessary to first explicitly capture and quantify this latent interaction as an independent data object before it can be selectively stripped in subsequent steps. Therefore, in the technical solution of this application, metabolic-neural nonlinear coupling interaction is further performed on the latent space mixture vector and the physiological fluctuation baseline vector in each feature dimension to obtain the metabolic-neural coupling interaction vector. This explicitly extracts the nonlinear multiplicative coupling pattern originally hidden in the mixture features into an operable independent vector representation. In this way, a clear target object can be provided for subsequent adaptive stripping, enabling the system to not only remove the linear additive metabolic baseline component, but also further identify and eliminate the nonlinear modulation coupling residues generated by metabolic activity on neural stress expression. Thus, in the sub-health assessment scenario after high metabolic activity, the problem of systematically large Mahalanobis distance caused by coupling residues is effectively eliminated, and the false alarm rate of the sub-health risk index is reduced.

[0031] In a specific example of this application, the implementation process of metabolic-neural nonlinear coupling interaction of the latent space mixture vector and the physiological fluctuation baseline vector across various feature dimensions is as follows. First, the system receives the latent space mixture vector output by the dual-branch encoder network and the physiological fluctuation baseline vector calculated by the orthogonal projection unit. Both are 128-dimensional high-dimensional vectors located in the same shared latent space coordinate system. The latent space mixture vector carries the mixed encoding information of the neural pressure component, the metabolic fluctuation component, and their coupling interaction components. The physiological fluctuation baseline vector represents the response intensity distribution of pure metabolic activity across various feature dimensions in this space. Next, the system performs element-wise Hadamard product operations on the latent space mixture vector and the physiological fluctuation baseline vector. That is, at each dimension position in the 128-dimensional feature space, the value of the latent space mixture vector in that dimension is multiplied by the value of the physiological fluctuation baseline vector in the same dimension position. The element-wise product results at all 128 dimension positions are arranged and assembled according to the original dimensional order to generate the metabolic-neural coupling interaction vector, the calculation of which is expressed as follows: in, The output metabolic-neural coupling interaction vector represents an explicit quantitative representation of the multiplicative modulation coupling pattern of metabolic activity on neural stress expression captured across each feature dimension. This represents a latent space mixture vector, which simultaneously contains pure neural components, pure metabolic baseline components, and nonlinear coupling interaction components of the two. The vector represents the baseline vector of physiological fluctuations, with each dimension reflecting the response intensity of metabolic activity in the corresponding feature channels. The symbol ⊙ represents the element-wise Hadamard product operator, which performs scalar multiplication on two equal-dimensional vectors at each corresponding dimension. In the context of sub-health monitoring, the core physical meaning of this formula lies in accurately capturing the channel-selective modulation effect of metabolic activity on neural signals through dimension-wise multiplicative interaction operations. Specifically, the Hadamard product operation ensures that a non-zero coupling interaction value is generated only when both the metabolic response (i.e., the baseline vector of physiological fluctuations) and the mixed signal (i.e., the latent space mixed vector) are non-zero in a certain feature dimension. This precisely corresponds to the multiplicative coupling mechanism in physiology where metabolic activity modulates neural stress expression: neural signals are only amplified or suppressed in metabolically active feature channels, while there is no coupling interference in metabolically quiescent channels. For example, when a user faces work stress immediately after completing a period of light exercise, the physiological response during the metabolic recovery period causes the baseline vector of physiological fluctuations to exhibit large values ​​in feature dimensions related to heart rate acceleration and vasodilation. Meanwhile, the latent space mixture vector also exhibits a high response in these same dimensions due to the co-activation of metabolism and nerves. The element-wise product of the two produces significant non-zero coupling interaction values ​​in these dimensions, accurately quantifying the amplification and modulation degree of neural stress expression by metabolic recovery. However, in feature dimensions related to sleep rhythm or cognitive load, since the response of metabolic activity in these channels is close to zero, even if neural stress signals exist in the latent space mixture vector in these dimensions, the Hadamard product result is close to zero, indicating that there is no metabolic-neural coupling interference in these channels. Through this operation, the nonlinear coupling patterns originally hidden in the mixture features are explicitly extracted as operable independent vectors, providing a clear target for subsequent adaptive stripping.

[0032] Then, the system performs data integrity verification on the calculated metabolic neural coupling interaction vector, verifying that its dimension is consistent with the latent space mixture vector and the physiological fluctuation baseline vector, both being 128-dimensional, ensuring dimensionality matching for subsequent gating coefficient generation and joint feature stripping operations. Furthermore, the system passes the metabolic neural coupling interaction vector and the physiological fluctuation baseline vector together to the downstream activity intensity adaptive mapping subunit as joint input for generating adaptive coupling gating coefficients. This allows the subsequent stripping intensity to be dynamically adjusted according to the actual intensity of the current coupling interaction, avoiding the introduction of new distortions due to overcorrection in resting scenarios, while ensuring sufficient elimination of coupling residues in high metabolic activity scenarios.

[0033] Accordingly, the activity intensity adaptive mapping subunit 1432 is used to perform activity intensity adaptive mapping processing on the concatenation of the physiological fluctuation baseline vector and the metabolic neural coupling interaction vector to obtain an adaptive coupling gating coefficient. It should be understood that, due to the significant differences in the intensity of the metabolic-neural coupling effect under different activity intensities, the coupling residue after high-intensity exercise is much greater than the weak coupling in the resting state. If a uniform fixed stripping force is applied to all scenarios, new distortions will be introduced due to overcorrection in the resting scenario, while noise will remain due to insufficient correction in the high-activity scenario. Therefore, a gating mechanism that can sense the current activity intensity and dynamically adjust the stripping force accordingly is needed, enabling the system to adaptively determine the stripping ratio of the metabolic neural coupling interaction vector based on the user's current metabolic activity level. Therefore, in the technical solution of this application, the concatenation of the physiological fluctuation baseline vector and the metabolic neural coupling interaction vector is further processed by activity intensity adaptive mapping to obtain an adaptive coupling gating coefficient, thereby generating a dynamic scaling factor with a value between zero and one. This factor can automatically adjust the magnitude of the subsequent coupling residue stripping force according to the actual intensity of the current metabolic activity. In this way, the system can achieve optimal decoupling accuracy across the entire activity intensity spectrum from rest to intense activity. In high metabolic activity scenarios, it can remove coupling residues with maximum force to eliminate false alarms, and in resting scenarios, it can automatically degenerate into the original simple linear subtraction mode to avoid overcorrecting the already weak signal and introducing new distortions, thus achieving adaptive feature purification with on-demand compensation.

[0034] In a specific example of this application, the implementation process of adaptive mapping of activity intensity on the concatenated body of the physiological fluctuation baseline vector and the metabolic-neural coupling interaction vector is as follows. First, the system receives the physiological fluctuation baseline vector output by the orthogonal projection unit and the metabolic-neural coupling interaction vector output by the metabolic-neural nonlinear coupling interaction subunit. Both are 128-dimensional high-dimensional vectors. The system performs a vector concatenation operation on the physiological fluctuation baseline vector and the metabolic-neural coupling interaction vector along the feature dimension direction, that is, the 128-dimensional elements of the physiological fluctuation baseline vector and the 128-dimensional elements of the metabolic-neural coupling interaction vector are concatenated end to end to generate a 256-dimensional joint input vector. The first 128 dimensions of this joint input vector carry the direct representation information of the user's physical activity intensity and calorie metabolism level at the current moment, and the last 128 dimensions carry the dimension-multiplicative modulation coupling strength information of metabolic activity on neural stress expression. The combination of the two parts of information provides the gating network with the complete contextual basis required to determine the current coupling stripping strength. Next, the system inputs the 256-dimensional joint input vector into a pre-trained fully connected gating network. This gating network consists of a single fully connected layer with a weight matrix of 128 rows and 256 columns, mapping the 256-dimensional input to a 128-dimensional output. The weight matrix and the joint input vector are multiplied together, and then a pre-trained bias vector is superimposed to obtain a 128-dimensional linear transformation result. Then, the system applies a saturated nonlinear activation function to this 128-dimensional linear transformation result, compressing and mapping the real values ​​in each dimension to an open interval between zero and one, generating adaptive coupling gating coefficients, the calculation of which is expressed as follows: in, The output adaptive coupling gating coefficients are 128-dimensional vectors, with each dimension's value constrained to an open interval between zero and one. They are used to control the strength of subsequent coupling residue stripping on a dimension-by-dimensional basis. This represents a saturated nonlinear activation function that maps any real number to an open interval between zero and one, ensuring that the output values ​​of the gating coefficients have clear physical boundary constraints. This represents the pre-trained weight matrix in the gating network, with dimensions of 128 rows and 256 columns. Through training, it learns a nonlinear mapping relationship between the joint representation of activity intensity and coupling intensity and the optimal stripping force. This represents a 256-dimensional joint input vector obtained by concatenating the physiological fluctuation baseline vector and the metabolic neural coupling interaction vector along the feature dimension; This represents the pre-trained bias vector in the gating network, a 128-dimensional vector used to perform translation correction on the gating output to enhance the model's generalization ability across different individuals. In the sub-health monitoring scenario, the core physical meaning of this formula lies in achieving an adaptive response of stripping force to activity intensity. Specifically, the joint input in the formula... Simultaneously, two layers of key information are encoded: the physiological fluctuation baseline vector reflects the absolute intensity level of metabolic activity at the current moment, and the metabolic-neural coupling interaction vector reflects the actual modulation and interference degree of this metabolic activity on neural signals. The combination of these two elements provides the gating network with the complete decision-making basis needed to determine the stripping intensity. The weight matrix WgWg learns a complex nonlinear mapping relationship from this joint representation to the optimal stripping ratio through pre-training, while the saturated activation function... This strictly constrains the mapping result to a bounded interval between zero and one, giving the gating coefficient a clear physical semantic meaning, namely the stripping ratio coefficient.

[0035] It should be noted that when a user is in a high-intensity metabolic state (such as the recovery period after running), the amplitude of the physiological fluctuation baseline vector is large, and the coupling interaction vector is also strong. The gating network output approaches one, meaning the system will strip away coupling residues with maximum force. Conversely, when a user is in a resting state (such as during quiet reading or sleep), the amplitude of the physiological fluctuation baseline vector is extremely small, the coupling effect is weak, and the gating output approaches zero. At this time, the system automatically degenerates into the original simple linear subtraction mode, avoiding over-correction of the already weak signal and the introduction of new distortions. This adaptive characteristic of on-demand compensation ensures optimal decoupling accuracy across the entire activity intensity spectrum, from rest to after strenuous exercise.

[0036] Furthermore, the system encapsulates the calculated 128-dimensional adaptive coupling gating coefficients and transmits them, along with the latent space mixing vector, the physiological fluctuation baseline vector, and the metabolic neural coupling interaction vector, to the downstream joint feature stripping subunit. This serves as a dynamic weight parameter to control the stripping intensity of nonlinear coupling residues, enabling subsequent dual-path compensation and purification operations to perform coupling residue elimination at the optimal scaling factor in each feature dimension, balancing robustness across the entire activity spectrum with computational efficiency under limited computing power at the wristband.

[0037] Accordingly, the joint feature stripping subunit 1433 is used to perform joint feature stripping processing on the latent space mixture vector and the physiological fluctuation baseline vector based on an adaptive coupling gating coefficient to obtain a pure autonomic neural vector. It should be understood that after obtaining the explicit metabolic-neural coupling interaction vector and the adaptive stripping strength coefficient, it is necessary to integrate the linear metabolic baseline stripping and the nonlinear coupling interaction stripping into a single complete purification operation. If only a single-path linear subtraction is performed, only the pure metabolic baseline component can be removed, leaving coupling interaction residues. If only coupling interaction stripping is performed while ignoring the linear baseline, the pure metabolic component will still contaminate the output result. Therefore, it is necessary to simultaneously perform dual-path subtraction from the latent space mixture vector to achieve a one-time complete purification of the linear additive baseline component and the nonlinear multiplicative coupling interaction residue component. Therefore, in the technical solution of this application, joint feature stripping processing is performed on the latent space mixture vector and the physiological fluctuation baseline vector based on an adaptive coupling gating coefficient to obtain a pure autonomic neural vector, thereby simultaneously completing the linear stripping of the pure metabolic baseline and the adaptive nonlinear stripping of the metabolic-neural coupling interaction residue in a single unified purification operation. In this way, the output pure autonomic neural vector can be mathematically orthogonal to both the metabolic baseline subspace and the coupling interaction subspace. Physiologically, this completely eliminates the root cause of misjudgment of pseudo-high stress after exercise, so that the feature data flowing into the subsequent Mahalanobis distance deviation calculation only reflects the user's true psychological or neurological sub-health state.

[0038] In a specific example of this application, the system first receives a latent space mixing vector output by a dual-branch encoder network, a physiological fluctuation baseline vector calculated by an orthogonal projection unit, a metabolic-neural coupling interaction vector output by a metabolic-neural nonlinear coupling interaction subunit, and an adaptive coupling gating coefficient output by an activity intensity adaptive mapping subunit. All four are 128-dimensional high-dimensional vectors located in the same shared latent space coordinate system. The latent space mixing vector carries mixed encoded information of the pure neural component, the pure metabolic baseline component, and the metabolic-neural coupling interaction component. The physiological fluctuation baseline vector represents the pure metabolic baseline component, the metabolic-neural coupling interaction vector represents the nonlinear coupling interaction component, and the adaptive coupling gating coefficient encodes the optimal stripping ratio of the coupling residue under the current activity intensity. Next, the system performs a first path subtraction operation, subtracting the physiological fluctuation baseline vector element-by-element from the latent space mixing vector. This path completes the linear stripping of the pure metabolic baseline, maintaining consistency with the function of the original mechanism in the first embodiment, and eliminating the additive heart rate fluctuation component directly caused by daily physical activity and calorie metabolism. Then, the system performs a second-path subtraction operation, which involves performing an element-wise Hadamard product operation on the adaptive coupling gating coefficient and the metabolic neural coupling interaction vector to obtain a gated weighted coupling residual stripping amount. The magnitude of this stripping amount in each feature dimension is precisely controlled by the value of the adaptive coupling gating coefficient in that dimension. Subsequently, this gated weighted coupling residual stripping amount is further subtracted element-wise from the result of the first-path subtraction. The second path accurately and adaptively strips away the nonlinear modulation coupling residuals generated by the expression of neural stress caused by metabolic activity, ultimately generating a pure autonomic neural vector, the calculation of which is expressed as follows: in, This represents the final output of a pure autonomic nervous system regulation feature vector, which is a 128-dimensional high-dimensional vector that only carries feature information related to the user's true psychological or neurological sub-health state. This represents a latent space mixture vector, which simultaneously contains pure neural components, pure metabolic baseline components, and nonlinear coupling interaction components of the two. This represents the baseline vector of physiological fluctuations, and characterizes the linear additive component in the latent space mixture vector that is directly contributed by daily metabolic activities. The adaptive coupling gating coefficient is a 128-dimensional vector, with each dimension taking values ​​in an open interval from zero to one. It is used to dynamically control the stripping force of coupling residues dimension by dimension. The metabolic-neural coupling interaction vector represents the dimension-multiplicative modulation coupling pattern of metabolic activity on neural stress expression. This represents the element-wise Hadamard product operator, ensuring that the weighted control of the gating coefficients on the coupled interaction vector is executed independently on each feature dimension. In the sub-health monitoring scenario, the core physical meaning of this formula lies in achieving a complete one-time purification of two fundamentally different metabolic interference components in the mixed features through dual-path collaborative subtraction. Specifically, the first subtraction term in the formula... Linear stripping of the pure metabolic baseline was achieved, maintaining functional consistency with the original mechanism and eliminating additive heart rate variability fluctuations directly caused by metabolic activity during the postprandial digestion period or after light housework; the second subtraction term It precisely and adaptively strips away the nonlinear modulation coupling residues caused by metabolic activity on neural stress expression. The adaptive coupling gating coefficient λ ensures that the stripping force of this term approaches the maximum value in high metabolic activity scenarios to fully eliminate coupling residues, while in resting scenarios, this term approaches zero, causing the system to automatically degenerate into the original mode that only performs linear subtraction of the first path.

[0039] Under the synergistic effect of the dual-pathway, the output pure autonomic neural vector is mathematically orthogonal to both the metabolic baseline subspace and the coupling interaction subspace. Physiologically, it completely eliminates the root cause of misjudgment due to pseudo-high stress after exercise, ensuring that the feature data flowing into the subsequent Mahalanobis distance deviation calculation only reflects the user's true psychological or neurological sub-health state. For example, when a user faces work pressure immediately after completing a period of light exercise, the first path removes the baseline component of heart rate acceleration directly caused by metabolic activity during the exercise recovery period. The second path, under the regulation of a gating coefficient approaching one, further removes the amplification and modulation effect of metabolic recovery on neural stress expression. The final output pure autonomic neural vector only retains the user's true neurological stress characteristics caused by work anxiety, avoiding false alarms caused by inflated sub-health risk indices due to coupling residues.

[0040] Furthermore, the system performs L2 norm normalization on the 128-dimensional pure autonomic neural vectors obtained after dual-path joint feature stripping and encapsulates the data. This data is then passed to the downstream sub-health risk quantification module as the final output of the entire physiological metabolism and psychoneural feature decoupling module. By introducing a cascaded architecture of Hadamard product coupling modeling, saturated activation function gating adaptation, and dual-path compensation purification, the nonlinear multiplicative modulation coupling residues of metabolic activity on neural stress expression are explicitly captured and adaptively stripped. This improves feature decoupling from simply stripping the linear additive baseline to a dual purification that simultaneously strips both the linear baseline and nonlinear coupling interactions. In sub-health assessment scenarios following high metabolic activity, this effectively eliminates the systematic overestimation of Mahalanobis distance caused by coupling residues, reduces the false alarm rate of the sub-health risk comprehensive index, and improves the accuracy of tiered warning command triggering and users' long-term trust in the warning system. Simultaneously, the adaptive degradation characteristics of the gating coefficient ensure that no additional computational distortion is introduced in resting scenarios, balancing robustness across the entire activity spectrum with computational efficiency under limited computing power at the wristband.

[0041] In the aforementioned sub-health monitoring and early warning system 100 based on a smart bracelet, the sub-health risk quantification module 150 is used to quantify the deviation between the pure autonomic neural vector and the pre-stored individual rhythm baseline model to obtain a sub-health risk index. It should be understood that although the pure autonomic neural vector obtained after orthogonal projection decoupling has completely eliminated metabolic interference components and retained only feature information related to the user's true psychological or neurological state, the vector itself is still a high-dimensional abstract feature representation and has not yet been transformed into a quantitative risk assessment result that can directly indicate whether the user is currently deviating from a healthy state. Furthermore, due to differences in individual physiological foundations, the baseline of autonomic neural characteristics in the healthy state of different users exhibits significant inter-individual differences. The health baseline of the same user also shows periodic fluctuations at different times of the day due to circadian rhythm regulation. If a uniform fixed threshold is used to judge all users at all times, a large number of misjudgments due to individual and rhythm differences will inevitably occur. Therefore, an assessment mechanism that can combine the user's individual historical data and the rhythm characteristics of the current time period for personalized deviation quantification is needed. Therefore, in the technical solution of this application, the deviation between the pure autonomic neural vector and the pre-stored individual rhythm benchmark model is further quantified to obtain a sub-health risk index. This allows for precise quantification of the degree of deviation of the user's current neuromodulation state from their historical health baseline, taking into full account individual differences and diurnal rhythm fluctuations, and converting this deviation into an intuitive percentage-based risk score. This enables personalized sub-health assessments tailored to each individual, avoiding misjudgments caused by differences in individual physiological foundations or normal diurnal rhythm fluctuations. It ensures that the sub-health risk index only shows an abnormal increase when the user truly deviates from their normal health state, providing a reliable quantitative decision-making basis for the accurate triggering of subsequent graded early warning commands.

[0042] Figure 5 This is a block diagram of the sub-health risk quantification module in a sub-health monitoring and early warning system based on a smart bracelet, according to an embodiment of this application. Figure 5 As shown, the sub-health risk quantification module 150 includes: a fatigue correlation dimension focusing unit 151, used to perform fatigue correlation dimension weighted focusing on pure autonomic neural vectors to obtain a current fatigue state representation; a rhythm period retrieval and distribution parameter extraction unit 152, used to perform rhythm period retrieval and distribution parameter extraction on the individual rhythm benchmark model based on the current timestamp information to obtain a historical normal representation; and a multidimensional deviation quantification unit 153, used to perform multidimensional deviation quantification on the current fatigue state representation and the historical normal representation to obtain a sub-health risk index.

[0043] Specifically, in a specific example of this application, the system first inputs the pure autonomic neural vector into the fatigue-related dimension focusing unit for fatigue-related dimension weighted focusing processing. This unit uses a self-attention mechanism layer to dynamically assign weights to the 128 feature dimensions of the pure autonomic neural vector. By mapping the pure autonomic neural vector through three independent linear projection weight matrices to generate a query matrix, a key matrix, and a value matrix, the dot product of the query matrix and the transpose of the key matrix is ​​calculated and divided by the square root of the key vector dimension. Then, the attention weight distribution is obtained by normalizing it using the Softmax function. This attention weight distribution is then multiplied by the value matrix, so that the feature dimensions most related to fatigue and stress receive higher focusing weights, while dimensions with weaker correlation to sub-health assessment are correspondingly suppressed. The current fatigue state representation after dimension focusing is output. This representation encodes the user's real neural fatigue and stress state information at the current moment in a more concise and targeted manner. Next, the system obtains the current precise timestamp information, including hourly segment identifiers and weekday or rest day labels. Based on this timestamp information, it performs rhythmic period retrieval and distribution parameter extraction operations in the individual rhythm baseline model pre-stored in the smart bracelet's flash memory. This individual rhythm baseline model is a personalized physiological rhythm memory model built based on long-term user wear data through continuous learning and updating via a long short-term memory network. It divides a 24-hour day into multiple rhythmic periods, storing the user's multidimensional physiological baseline mean vector and historical baseline covariance matrix in a healthy state for each period. After matching the corresponding rhythmic period based on the current timestamp, the system extracts the baseline mean vector and covariance matrix for that period, packaging them together to generate a historical normal representation. This representation fully describes the expected distribution center and variance and correlation structure of the user's autonomic neural characteristics when in a healthy state during the current same period. Then, the system performs a multidimensional deviation metric operation on the current fatigue state representation and the historical normal representation, specifically calculating the Mahalanobis distance between the baseline mean vectors in the current fatigue state representation and the historical normal representation. The calculation expression is as follows: in, The calculated Mahalanobis distance is used to scientifically measure the degree of multidimensional deviation of the current fatigue state from the historical individual's normal state. This represents the current fatigue state characterization output by the fatigue-related dimension focusing unit; This represents the historical baseline mean vector for the current time period retrieved from the individual rhythm baseline model, which represents the user's expected physiological value in their health state during that time period. This represents the inverse matrix of the historical benchmark covariance matrix for the current time period retrieved from the individual rhythm benchmark model, used to eliminate the influence of dimensional differences and correlations between different feature dimensions; This represents the difference vector between the current fatigue state representation and the historical baseline mean vector, reflecting the original deviation of the current state from the healthy baseline in each feature dimension; superscript This represents the vector transpose operation. In the sub-health monitoring scenario, the core physical meaning of this formula lies in introducing the inverse of the covariance matrix as a metric weight. This allows the deviation calculation to automatically adapt to the inherent variance of each feature dimension and the correlation structure between dimensions. Smaller deviation weights are assigned to feature dimensions where the user's health fluctuates significantly, while larger deviation weights are assigned to feature dimensions that are highly stable in a healthy state. This avoids the problem of artificially inflated deviations caused by naturally large fluctuations in certain dimensions, achieving a scientific and personalized measurement of the degree of sub-health deviation in a multi-dimensional feature space. Furthermore, the system inputs the calculated Mahalanobis distance into the Sigmoid nonlinear mapping function, converting the dimensionless deviation value into a percentage risk score ranging from zero to one hundred. The calculation is expressed as follows: in, This indicates that the final output sub-health risk index has a stable value within a closed range of zero to one hundred. represents the natural constant; k represents the preset mapping curve scaling factor, which is used to control the sensitivity of the risk index growth caused by state abrupt changes. The larger the value, the steeper the jump of the risk index near the threshold. This indicates the deviation of the Mahalanobis distance calculated in the previous step; This represents the set deviation risk threshold bias constant. When the Mahalanobis distance exceeds this threshold, the risk index begins to show a significant non-linear and sharp upward trend. In the sub-health monitoring scenario, this formula maps continuous Mahalanobis distance values ​​to a percentage score with clear health semantics through an S-shaped curve. When the user's neuromodulation state is highly consistent with their historical health baseline, the Mahalanobis distance is small and below the threshold bias constant, and the Sigmoid function output approaches zero, keeping the sub-health risk index at a low level. When the user's neuromodulation state deviates significantly from the health baseline due to continuous work anxiety or long-term sleep deprivation, the Mahalanobis distance exceeds the threshold bias constant, and the Sigmoid function output rises sharply, causing the sub-health risk index to climb rapidly to a high level. This non-linear mapping characteristic matches the threshold triggering effect of the sub-health state from quantitative to qualitative change in medicine. The system encapsulates the mapped percentage score into a sub-health risk index, which is passed as the final output of the sub-health risk quantification module to the downstream intervention strategy matching module, providing accurate quantitative decision input for the generation of graded early warning instructions.

[0044] In the aforementioned sub-health monitoring and early warning system 100 based on a smart bracelet, the intervention strategy matching module 160 is used to determine the level of the sub-health risk index and match intervention strategies based on a graded threshold comparison to obtain graded early warning instructions. It should be understood that although the sub-health risk index obtained after the aforementioned deviation quantification has accurately quantified the degree to which the user's current neuromodulation state deviates from their own health baseline in a percentage format, this value itself is only a continuous scalar score and has not yet been converted into a discrete risk level with clear medical grading semantics. Furthermore, the intervention measures corresponding to different risk levels differ fundamentally in type and intensity. Mild sub-health only requires gentle relaxation guidance, while severe sub-health requires strong warnings and immediate intervention suggestions. If the system only outputs an abstract numerical score without converting it into a multimodal early warning action that the user can perceive, the entire monitoring link will stop at the data level and will not be able to achieve a substantial closed-loop intervention on the user's behavior. Therefore, in the technical solution of this application, the sub-health risk index is further classified and matched with intervention strategies based on a graded threshold comparison to obtain graded early warning instructions. This transforms continuous risk assessment values ​​into discrete risk levels and matches corresponding intensity intervention measures accordingly. Finally, hardware-level execution instructions that can directly drive the smart bracelet screen display and vibration motor are compiled and generated. In this way, a closed-loop system can be achieved from physiological data collection, feature extraction, decoupled computation, risk quantification to terminal electromechanical feedback. This allows users to perceive early warning information and intervention suggestions matching their current sub-health risk level through both visual and tactile channels in real time, prompting users to take timely adjustment measures to prevent the sub-health state from evolving into a disease.

[0045] Specifically, in this embodiment, the intervention strategy matching module is used to: determine the numerical range of the sub-health risk index based on preset mild and severe thresholds to obtain the warning level; use the warning level as a hash index pointer to perform row vector addressing and data capture on the strategy mapping matrix pre-stored in non-volatile memory to obtain intervention strategy data; and perform parallel compilation and splicing of pixel stream and vibration electrical signal on the intervention strategy data and the warning level through display rendering encoding and pulse width modulation driven synthesis to obtain graded warning instructions.

[0046] More specifically, in a specific example of this application, the system first receives the sub-health risk index output by the sub-health risk quantification module, and compares it with the mild and severe thresholds preset in the system firmware to perform a numerical range classification process. The mild threshold is set to 40 points and the severe threshold is set to 75 points. When the sub-health risk index is less than or equal to the mild threshold, the system determines the current state as healthy and outputs a warning level of 0. When the sub-health risk index is greater than the mild threshold and less than or equal to the severe threshold, the system determines the current state as mild sub-health and outputs a warning level of 1. When the sub-health risk index is strictly greater than the severe threshold, the system determines the current state as severe sub-health with high risk and outputs a warning level of 2. Through this segmented judgment logic, the continuous percentage risk score is discretized into three warning level labels with clear medical classification meaning. Next, the system uses the determined warning level as a hash index pointer to perform row vector addressing and data retrieval operations on the strategy mapping matrix pre-stored in the non-volatile memory of the smart bracelet. This strategy mapping matrix is ​​a two-dimensional data table structure with three rows and multiple columns. Each row corresponds to a warning level, and each row stores the complete set of intervention strategy parameters corresponding to that level. The system uses the value of the warning level as the row address index to perform zero-latency random access addressing on the matrix, retrieves all the strategy parameter data stored in the corresponding row, and encapsulates it into intervention strategy data. The intervention strategy data retrieved when the warning level is 0 includes a green status indicator icon and a no-vibration indicator. The intervention strategy data retrieved when the warning level is 1 includes a yellow warning icon, deep breathing guidance animation parameters, and mild vibration frequency parameters. The intervention strategy data retrieved when the warning level is 2 includes a red high-risk warning icon, immediate rest suggestion text, medical reminder information, and strong vibration frequency and duration parameters. Then, the system performs parallel compilation and splicing processing of pixel streams and vibration electrical signals on the intervention strategy data and warning levels. Specifically, the screen instruction encoder extracts UI display-related parameters from the intervention strategy data and converts them into an RGB pixel matrix data stream that can be directly refreshed by the smart bracelet's OLED screen after display rendering encoding. This includes the pixel array of warning icons of corresponding colors, the character dot matrix rendering sequence of suggested text, and the frame sequence data of the guiding animation. Simultaneously, the motor controller extracts vibration-related parameters from the warning levels and intervention strategy data. Through pulse width modulation drive synthesis, it converts the vibration intensity corresponding to the warning level into a PWM square wave electrical signal sequence with a specific duty cycle and frequency. No vibration signal is generated when the warning level is 0, a mild vibration signal with a duty cycle of 30% and a frequency of 100 Hz is generated when the warning level is 1, and a strong vibration signal with a duty cycle of 80% and a frequency of 200 Hz is generated when the warning level is 2.Furthermore, the system compiles and splices the screen pixel data stream and the motor PWM vibration electrical signal sequence in parallel according to the frame format of the hardware bus communication protocol. The two heterogeneous hardware drive data are encapsulated into a unified hierarchical warning instruction. This instruction is distributed to the screen driver integrated circuit and the motor drive circuit for synchronous execution through the hardware abstraction layer inside the smart bracelet. The screen renders and displays the warning interface and intervention suggestions corresponding to the current sub-health risk level in real time. The motor outputs tactile vibration feedback of intensity matched with the risk level in real time. Through the synergistic effect of visual and tactile dual channels, the system ensures that users can perceive the warning information in a timely manner in any usage situation. This completes the entire technical link of the sub-health monitoring and warning system from sensor data acquisition to terminal electromechanical closed-loop feedback.

[0047] The above-described embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent for the embodiments of this disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this disclosure, and these all fall within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the appended claims. As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and detail can be made without departing from the spirit and scope of the present invention as defined in the appended claims.

[0048] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A sub-health monitoring and early warning system based on a smart bracelet, characterized in that, include: The sensor data preprocessing module is used to filter and preprocess the multimodal sensor data stream collected by the smart bracelet to obtain the wrist pulse wave sequence, wrist inertial sequence and wrist temperature sequence. The multimodal sensor data stream includes photoelectric pulse wave signal, six-axis inertial signal and skin temperature signal. The heart rate variability feature extraction module is used to extract heart rate variability features from wrist pulse wave sequences to obtain a mixed heart rate variability vector. The metabolic state joint encoding module is used to jointly encode the metabolic state of the wrist inertial sequence and the wrist temperature sequence to obtain the metabolic confounding factor tensor. The physiological metabolism and psychoneural feature decoupling module is used to decouple the mixed heart rate variability vector and the metabolic confounding factor tensor based on orthogonal constraints to obtain a pure autonomic neural vector. The sub-health risk quantification module is used to quantify the deviation between the pure autonomic neural vector and the pre-stored individual rhythm benchmark model to obtain the sub-health risk index. The intervention strategy matching module is used to determine the level of the sub-health risk index and match intervention strategies based on the comparison of graded thresholds to obtain graded early warning instructions.

2. The sub-health monitoring and early warning system based on a smart bracelet according to claim 1, characterized in that, The sensor data preprocessing module is used for: Baseline drift filtering and peak detection were performed on the photoelectric pulse wave signal to obtain the wrist pulse wave sequence; The wrist inertial sequence was obtained by synthesizing triaxial acceleration and subtracting gravity bias from the six-axis inertial signal. Thermal noise suppression and serialization were performed on the skin temperature signal to obtain the wrist temperature sequence.

3. The sub-health monitoring and early warning system based on a smart bracelet according to claim 1, characterized in that, The heart rate variability feature extraction module is used to: extract heart rate variability time-frequency domain features from the wrist pulse wave sequence based on adaptive peak detection and fast Fourier transform to obtain a hybrid heart rate variability vector.

4. The sub-health monitoring and early warning system based on a smart bracelet according to claim 1, characterized in that, The metabolic state joint encoding module is used for: The wrist inertial sequence and wrist temperature sequence are synchronized and aligned using timestamps and their column vectors are concatenated to obtain the aligned environment sequence. The aligned environment sequence is dimensionality reduced and feature cross-encoded to obtain the metabolic confounding factor tensor.

5. The sub-health monitoring and early warning system based on a smart bracelet according to claim 1, characterized in that, The physiological metabolism and psychoneurological feature decoupling module includes: The latent space nonlinear mapping unit is used to perform high-dimensional latent space nonlinear mapping on the mixed heart rate variability vector and the metabolic confounding factor tensor through a dual-branch encoder network to obtain the latent space mixed vector and the latent space confounding tensor. Orthogonal projection unit is used to orthogonally project the latent space mixing vector along the direction of the latent space mixing tensor to obtain the physiological fluctuation baseline vector; The feature stripping unit is used to strip features from the latent space mixture vector and the physiological fluctuation baseline vector to obtain a pure autonomic neural vector.

6. The sub-health monitoring and early warning system based on a smart bracelet according to claim 1, characterized in that, The sub-health risk quantification module includes: The fatigue association dimension focusing unit is used to perform fatigue association dimension weighted focusing on pure autonomic neural vectors to obtain a representation of the current fatigue state. The rhythm period retrieval and distribution parameter extraction unit is used to retrieve rhythm period and extract distribution parameters for individual rhythm baseline models based on the current timestamp information in order to obtain historical normal characteristics. The multidimensional deviation measurement unit is used to measure the deviation between the current fatigue state and the historical normal state to obtain the sub-health risk index.

7. The sub-health monitoring and early warning system based on a smart bracelet according to claim 1, characterized in that, The intervention strategy matching module is used for: Based on preset mild and severe thresholds, the sub-health risk index is processed to determine the numerical range of the risk index in order to obtain the warning level. Based on the warning level as a hash index pointer, row vector addressing and data retrieval are performed on the policy mapping matrix pre-stored in non-volatile memory to obtain intervention policy data; By display rendering encoding and pulse width modulation driven synthesis, the intervention strategy data and early warning level are processed and spliced ​​in parallel using pixel streams and vibration electrical signals to obtain graded early warning instructions.

8. The sub-health monitoring and early warning system based on a smart bracelet according to claim 5, characterized in that, The feature stripping unit includes: The metabolic-neural nonlinear coupling interaction subunit is used to perform metabolic-neural nonlinear coupling interaction on each feature dimension of the latent space mixture vector and the physiological fluctuation baseline vector to obtain the metabolic-neural coupling interaction vector. The activity intensity adaptive mapping subunit is used to perform activity intensity adaptive mapping on the concatenation of the physiological fluctuation baseline vector and the metabolic neural coupling interaction vector to obtain the adaptive coupling gating coefficient. The joint feature stripping subunit is used to perform joint feature stripping on the latent space mixture vector and the physiological fluctuation baseline vector based on the adaptive coupling gating coefficient to obtain a pure autonomic neural vector.

Images