Wearable vital sign monitoring system capable of synchronously collecting multiple parameters

Abstract

The invention relates to the technical field of wearable vital sign monitoring systems capable of synchronously collecting multiple parameters, and particularly discloses a wearable vital sign monitoring system capable of synchronously collecting multiple parameters. The system comprises a multi-modal physiological sensing array, an edge calculation and fusion module, a psychological state quantification engine and a closed-loop nerve regulation module. According to the invention, a feedback learning and model updating mechanism based on reinforcement learning is introduced, so that the system has continuous self-optimization capability. The system can gradually adapt to physiological and psychological characteristic differences of different users by recording parameters and effects of each intervention and utilizing a strategy gradient method to dynamically optimize a stimulation parameter mapping table. Therefore, an intervention strategy is evolved into a dynamic and personalized scheme from static and universal setting, so that the intervention precision and the comfort and acceptability of the user are continuously improved in long-term use, and technical support is provided for realizing personalized daily maintenance of mental health.

Description

Technical Field

[0001] This invention belongs to the field of medical monitoring technology, specifically relating to a wearable vital sign monitoring system that collects multiple parameters synchronously. Background Technology

[0002] Vital signs monitoring is a crucial foundation in the fields of medical health, sports science, and daily health management. It assesses an individual's health status by continuously and accurately collecting physiological parameters. With the rapid development of sensor technology, microelectronics technology, and wireless communication technology, wearable devices have become the mainstream technology for achieving non-invasive and continuous vital signs monitoring, and are widely used for the daily tracking of basic physiological indicators such as heart rate, blood oxygen, body temperature, and activity level.

[0003] Wearable monitoring systems for mental health assessment and intervention are a current research and application hotspot. These systems aim to indirectly infer the user's mental state, such as stress levels, mood swings, or anxiety levels, by analyzing physiological signals collected by wearable devices, thereby providing objective evidence for early identification and proactive intervention in mental health.

[0004] Current technologies primarily assess psychological state through single or a few physiological parameters, which limits their effectiveness. Existing wearable devices typically rely on post-hoc retrospective questionnaires or alarms based on simple threshold rules. This assessment process is highly subjective and inherently delayed, failing to achieve real-time, objective quantification of psychological fluctuations. Existing systems largely limit themselves to "monitoring" and "early warning," lacking automated adjustment and intervention mechanisms linked to the assessment results, creating an open loop of "diagnosis without treatment." This separation between monitoring and intervention prevents the system from providing timely and effective adjustment measures to alleviate negative states when abnormalities are detected. Consequently, it struggles to achieve a complete closed-loop management system from "state perception" to "proactive adjustment," limiting its practical effectiveness in preventing the worsening of psychological problems and improving daily mental health. Summary of the Invention

[0005] The purpose of this invention is to provide a wearable vital sign monitoring system that collects multiple parameters simultaneously, in order to solve the problems in the prior art that make it impossible to objectively quantify and regulate psychological state in real time due to reliance on a single parameter, delayed assessment, and the separation of monitoring and intervention.

[0006] This invention provides a wearable vital sign monitoring system for simultaneous acquisition of multiple parameters, comprising: A multimodal physiological sensor array is used to acquire users' multidimensional raw physiological signals in parallel with millisecond-level time synchronization accuracy; The edge computing and fusion module is used to preprocess and extract features from the multi-dimensional raw physiological signals collected by the multimodal physiological sensing array, and to perform spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data to generate a fused feature vector. The psychological state quantification engine is used to receive the fusion feature vector generated by the edge computing and fusion module, and to calculate it through the preset psychological state quantification model to output a quantitative index that represents the user's current psychological state. The closed-loop neural modulation module is used to receive the psychological state quantification index output by the psychological state quantification engine, and when the psychological state quantification index exceeds the preset steady-state range, automatically generate and execute non-invasive neural modulation stimulation that matches the deviation direction and amplitude of the psychological state quantification index, so as to guide the physiological index back to the steady-state range.

[0007] Preferably, the multimodal physiological sensing array integrates at least four types of physiological sensors, including a photoplethysmography (PPG) sensor, a skin conductance sensor, a triaxial accelerometer and gyroscope combination sensor, and an infrared thermal radiation sensor. The photoplethysmography (PPG) sensor uses a light-emitting diode as a light source and a photodetector to receive the light intensity signal reflected or transmitted through human tissue, which is used to extract the time-domain and frequency-domain features of heart rate and heart rate variability. The skin electrical activity sensor uses a constant voltage measurement circuit and contacts the user's skin area through a pair of silver silver chloride electrodes to measure the skin conductivity level and its fluctuation frequency. The triaxial accelerometer and gyroscope combined sensor is used to continuously monitor the user's limb movement acceleration and angular velocity, and eliminates the interference of motion artifacts on the photoplethysmography signal through the built-in attitude calculation algorithm, while extracting motion features that characterize the user's activity intensity and behavior pattern. The infrared thermal radiation sensor is aimed at the user's body surface area in a non-contact manner, and inverts the body surface temperature and its dynamic change trend by measuring the intensity of the received infrared radiation.

[0008] Preferably, the edge computing and fusion module includes a signal preprocessing unit, a feature extraction unit, and a data fusion unit; The signal preprocessing unit is used to perform bandpass filtering, power frequency interference notch filtering, and baseline drift correction on each raw physiological signal. The feature extraction unit is used to extract at least five heart rate variability features from the preprocessed photoplethysmography (PPG) signal, including the standard deviation of adjacent heartbeat intervals and the ratio of low-frequency power to high-frequency power; to extract at least three features from the skin conductance signal, including the average skin conductance and the number of nonspecific fluctuations; to extract at least two features from the motion signal, including the activity count and the percentage of resting time; and to extract at least two features from the body surface temperature signal, including the mean temperature and the slope of temperature change. The data fusion unit is used to establish a unified timestamp system, align feature data from different sensors to the same time reference; it uses a kernel function-based canonical correlation analysis method to calculate the nonlinear correlation weights between different modal features; and it performs a weighted summation of all features based on the calculated correlation weights to generate a high-dimensional fusion feature vector.

[0009] Preferably, the canonical correlation analysis method based on kernel functions is as follows: physiological feature vectors from different modalities are mapped to a high-dimensional regenerative kernel Hilbert space; the projection direction vector that maximizes the correlation between the projections of feature vectors from different modalities is solved in the high-dimensional regenerative kernel Hilbert space; and the contribution weight of each modal feature in the fusion space is calculated based on the solved projection direction vector.

[0010] Preferably, the psychological state quantification engine has a built-in psychological state quantification model that is a deep neural network model, which takes the fused feature vector as input. The deep neural network model contains three hidden layers, with 128, 64 and 32 neurons in each layer, respectively, and the activation function is a modified linear unit. The output layer of the deep neural network model consists of two neurons, corresponding to two dimensions of the psychological state quantification index: arousal and valence. Arousal represents the level of psychological activation, and valence represents the positive or negative orientation of emotions.

[0011] Preferably, the closed-loop neural modulation module includes a stimulation decision unit and a transcranial alternating current stimulation unit; The stimulus decision unit is pre-set with a two-dimensional psychological state steady-state interval, which is defined as an elliptical region on the two-dimensional plane formed by arousal and valence. The stimulus decision unit is used to compare the positional relationship between the quantitative index points output by the psychological state quantification engine and the boundary of the steady-state interval in real time. When the quantitative index point is inside the steady-state interval, no output is generated. When the quantitative index point is outside the steady-state interval, the unit calculates the vector from the quantitative index point to the center point of the steady-state interval. The direction and length of the vector define the target direction and intensity requirement of the adjustment, respectively. Based on the vector, the unit queries and generates a set of stimulus parameter instructions from a preset stimulus parameter mapping table. The set of stimulus parameter instructions includes at least the stimulus frequency, stimulus intensity, and stimulus duration. The transcranial alternating current stimulation unit is used to receive stimulation parameter instructions from the stimulation decision unit and generate and execute corresponding non-invasive neuromodulation stimulation.

[0012] Preferably, the transcranial alternating current stimulation unit includes a waveform generator, a constant current source circuit, and a pair of wearable stimulation electrodes; The waveform generator is used to receive stimulus parameter instructions from the stimulus decision unit and generate an AC signal with corresponding frequency and waveform. The constant current source circuit is used to ensure that the current intensity output to the human body is accurately and stably stable at the value required by the instruction, and its maximum output current is no more than 2 mA. The pair of wearable stimulation electrodes are made of conductive silicone and are positioned and attached to the user's scalp according to the international 10-20 EEG electrode placement system. They are used to safely apply weak alternating current signals to the functional network of the cerebral cortex.

[0013] Preferably, the system also includes a feedback learning and model update module; The feedback learning and model update module is used to continuously record the parameters of each stimulus executed by the closed-loop neural modulation module, the quantitative indicators of the psychological state before execution, and the trajectory of changes in the quantitative indicators of the psychological state within a preset time window after execution; calculate the effectiveness score of this stimulus intervention through the built-in stimulus effect evaluation algorithm; store all intervention records and effectiveness scores in the local intervention log database; and periodically fine-tune and optimize the stimulus parameter mapping table in the stimulus decision unit using the policy gradient method in the reinforcement learning algorithm.

[0014] Preferably, the calculation process of the stimulus effect evaluation algorithm is as follows: obtain the expected adjustment direction vector formed by the quantitative index point of the psychological state before the stimulus is executed and the center point of the steady-state interval; obtain the actual movement vector formed by the mean of the quantitative index of the psychological state within a preset time window after the stimulus is executed and the center point of the steady-state interval; calculate the cosine similarity between the actual movement vector and the expected adjustment direction vector, and multiply the cosine similarity by the ratio of the magnitude of the actual movement vector to the magnitude of the expected adjustment vector to obtain an effectiveness score between -1 and 1.

[0015] Preferably, the system operates according to a hierarchical decision-making and control architecture; The hierarchical decision-making and control architecture includes a bottom-level sensing and control layer, a middle-level state assessment and decision-making layer, and a high-level intervention execution and learning layer. The underlying sensing control layer operates at a sampling rate of 100 Hz and is responsible for driving the multimodal physiological sensing array, acquiring raw signals, and performing preliminary filtering. The middle-level state assessment and decision-making layer operates at a frequency of 1 Hz, performing feature extraction, data fusion, psychological state quantification calculation, steady-state interval comparison, and stimulus decision-making. The high-level intervention execution and learning layer operates in an event-driven manner, activating the closed-loop neural modulation module only when a stimulus command is received, and running the update process of the feedback learning and model update module on a daily or weekly basis.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates a multimodal physiological sensor array and employs feature-level fusion technology based on kernel canonical correlation analysis to achieve millisecond-level synchronous acquisition and deep correlation analysis of multi-dimensional physiological parameters such as heart rate variability, skin conductance, movement patterns, and body surface temperature. This multi-parameter fusion sensing method overcomes the problems of one-sided information and susceptibility to interference when assessing psychological state using a single parameter. It provides a high-dimensional, complementary, and robust data foundation for the objective quantification of psychological state, improving the accuracy and reliability of state assessment.

[0017] 2. This invention constructs a complete technical closed loop, from multi-parameter synchronous perception to psychological state quantification, and then to automated neural regulation. The system uses a psychological state quantification engine to map fused physiological characteristics into quantifiable arousal and valence indicators, and sets a clear two-dimensional steady-state range as the regulation target. When an indicator deviates from the steady-state range, the closed-loop neural regulation module can automatically generate and apply transcranial alternating current stimulation for intervention. This integrated design of perception, assessment, and intervention changes the existing open-loop model of "monitoring without intervention," achieving real-time detection and immediate proactive regulation of psychological fluctuations, and shortening the response time from abnormality detection to intervention.

[0018] 3. This invention introduces a reinforcement learning-based feedback learning and model update mechanism, enabling the system to continuously self-optimize. By recording the parameters and effects of each intervention and dynamically optimizing the stimulus parameter mapping table using a policy gradient method, the system can gradually adapt to the differences in the physiological and psychological characteristics of different users. This allows intervention strategies to evolve from static, universal settings to dynamic, personalized solutions, thereby continuously improving the accuracy of interventions and the comfort and acceptance of users in long-term use, providing technical support for achieving personalized daily mental health maintenance. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the present invention based on spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data; Figure 3 This is a logical flowchart of the process of quantifying physiological signals into psychological states in this invention. Figure 4 This is a diagram of the stimulation decision and execution logic framework of the closed-loop neural modulation module in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationships and data flow of the hierarchical decision-making and control architecture in this invention. Detailed Implementation

[0020] The overall technical architecture of the wearable vital signs monitoring system with multi-parameter synchronous acquisition proposed in this invention is shown in the attached figure. Figure 1 As shown in the figures, the system comprises a multimodal physiological sensor array, an edge computing and fusion module, a psychological state quantification engine, a closed-loop neural modulation module, and a feedback learning and model update module. These components operate collaboratively within a hierarchical decision-making and control architecture to achieve real-time perception, quantitative assessment, and closed-loop intervention of the user's psychological state. The specific implementation methods of each component will be described below with reference to the accompanying figures.

[0021] A multimodal physiological sensor array is the physical foundation for this system to achieve synchronous acquisition of high-dimensional physiological information. This array is integrated into the wearable device and typically adheres to the user's skin as a flexible wristband, headband-style ring, or patch-like vest, ensuring good sensor contact with the skin without interfering with daily activities. The array comprises four types of sensors: a photoplethysmography (PPG) sensor, a skin conductance sensor, a triaxial accelerometer and gyroscope combination sensor, and an infrared thermal radiation sensor. All sensors are controlled by a unified clock source, synchronously initiating data acquisition at a 100 Hz sampling rate, ensuring millisecond-level alignment accuracy of all raw signals in the time dimension.

[0022] The photoplethysmography (PPG) sensor uses dual-wavelength LEDs as its light source. One beam is green light (525 nm) to penetrate superficial tissue and acquire a high signal-to-noise ratio pulse wave signal; the other beam is infrared light (850 nm) to compensate for ambient light interference and motion artifacts. A photodetector receives the light intensity signal reflected by the tissue and outputs an analog voltage signal. This analog voltage signal is amplified by a preamplifier and then sent to an analog-to-digital converter for digital processing to form the original PPG sequence. The electrodermal activity (EDA) sensor uses a constant-voltage measurement circuit. A pair of silver-silver chloride electrodes are in contact with the skin on the inside of the user's palm or wrist. A 0.5V DC bias voltage is applied, and the weak current flowing through the skin is measured to invert the skin's conductivity level. This constant-voltage measurement circuit features high input impedance and low noise characteristics, effectively suppressing common-mode interference and ensuring the stability of the EDA signal.

[0023] A triaxial accelerometer and gyroscope combined sensor continuously monitors the linear acceleration and angular velocity of the user's limbs in a six-degree-of-freedom manner. The accelerometer has a range of ±2g, and the gyroscope has a range of ±2000 degrees / second; both output raw data at a frequency of 100 Hz. This motion data is used not only for subsequent behavioral pattern recognition but also as a reference signal for motion artifact elimination. An infrared thermal radiation sensor uses a non-contact thermopile array, aimed at the user's forehead or radial artery area of ​​the wrist. By detecting the infrared radiation energy emitted by the human body, it inverts the body surface temperature according to the Stefan-Boltzmann law. This infrared thermal radiation sensor has a resolution of 0.1 degrees Celsius and an accuracy of ±0.3 degrees Celsius, capable of capturing subtle dynamic changes in body temperature.

[0024] The four types of raw signals mentioned above are acquired by a multimodal physiological sensor array and immediately transmitted to the edge computing and fusion module. This edge computing and fusion module is deployed within a low-power embedded processor inside the wearable device. Its hardware platform is typically based on an ARM Cortex-M7 or RISC-V architecture and equipped with a dedicated digital signal processing instruction set. The edge computing and fusion module is internally divided into three functional units: a signal preprocessing unit, a feature extraction unit, and a data fusion unit. The workflow is shown in the attached figure. Figure 2 As shown.

[0025] The signal preprocessing unit independently performs standardization processing on each raw signal. For photoplethysmography (PPG) signals, a 0.5 Hz to 10 Hz bandpass filter is used to remove DC components and high-frequency noise, while a 50 Hz notch filter is applied to suppress power frequency interference, and baseline drift is corrected using a moving average method. For electrodermal activity (EDA) signals, a 0.05 Hz high-pass filter removes slow drift, followed by a low-pass filter to retain the effective frequency band from 0.01 Hz to 5 Hz. Motion signals are smoothed by a low-pass filter before being used for attitude calculation. Infrared temperature signals are filtered using a sliding window midpoint filter to eliminate transient jumps.

[0026] Based on the preprocessed signal, the feature extraction unit extracts key features with discriminative power regarding psychological states. From the photoplethysmography (PPG) signal, it calculates the interval sequence between adjacent heartbeats and further extracts at least five heart rate variability features, including: standard deviation of adjacent heartbeat intervals, root mean square of the difference between adjacent heartbeat intervals, the ratio of low-frequency power (0.04–0.15 Hz) to high-frequency power (0.15–0.4 Hz), total power, and the standard deviation of the short axis of the Poincaré plot. From the electrodermal activity (EDA) signal, it extracts the average skin conductivity, the number of nonspecific EDA responses per unit time, and the standard deviation of skin conductivity fluctuations. From the motion signal, it calculates the activity count by integrating the acceleration amplitude and statistically analyzes the duration of the resting state (acceleration amplitude less than 0.1g). From the infrared temperature signal, it calculates the mean temperature within a sliding window and the slope of the temperature change obtained from linear fitting.

[0027] The data fusion unit is responsible for integrating the aforementioned heterogeneous features into a unified representation. First, a unified timestamp system is established with the system startup time as the origin. All features are aligned according to a 1-second time window, forming a time-synchronized feature matrix. Then, a kernel-based canonical correlation analysis method is used to calculate the nonlinear correlation strength between different modal features. This method maps the original features to a high-dimensional regenerative kernel Hilbert space, and solves for the projection direction that maximizes the correlation within this high-dimensional regenerative kernel Hilbert space. For the heart rate variability feature vector (5-dimensional), skin conductance activity feature vector (3-dimensional), motion feature vector (2-dimensional), and body surface temperature feature vector (2-dimensional), they are mapped to the high-dimensional space using their respective kernel mapping functions. The maximum correlation of each type of feature projection vector is then solved to obtain the projection weight vectors of each feature, thereby determining the contribution weight of each modal feature in the fusion space. Finally, all 12-dimensional features are weighted and summed based on these weights to generate a high-dimensional fusion feature vector, which serves as the input to the mental state quantification engine.

[0028] The mental state quantization engine is deployed in the same embedded processor and includes a deep neural network model, the structure of which is shown in the attached figure. Figure 3 As shown, this deep neural network model consists of an input layer, three hidden layers, and an output layer. The input layer receives a high-dimensional fused feature vector. The first hidden layer contains 128 neurons, the second hidden layer contains 64, and the third hidden layer contains 32. Each layer uses a modified linear unit (MRU) as its activation function. The output layer contains two neurons, corresponding to two orthogonal dimensions of psychological state: arousal and valence. Arousal ranges from 0 to 10, with higher values ​​indicating stronger psychological activation. Valence ranges from -5 to +5, with positive values ​​representing positive emotions and negative values ​​representing negative emotions.

[0029] The deep neural network model was trained offline before the device left the factory. Training data came from a large-scale multimodal physiological signal and psychological state label pairing database. This database contained physiological signals and self-reported or expert-annotated psychological state labels synchronously recorded by more than 10,000 subjects in standardized emotion-inducing experiments (such as watching emotional videos or completing stress tasks). During training, the mean squared error loss function was used to measure the deviation between the model's predicted values ​​and the true labels, and the network parameters were updated using the backpropagation algorithm and the Adam optimizer until the validation set loss converged. After training, the model parameters were permanently stored in the device's flash memory for real-time inference.

[0030] When the intermediate state assessment and decision-making layer triggers the psychological state quantification engine at a frequency of 1 Hz, the engine loads the current fused feature vector, performs forward propagation calculation, and outputs the arousal and valence at the current moment, forming a two-dimensional psychological state quantification index point.

[0031] The two-dimensional psychological state quantification index points are then sent to the closed-loop neural regulation module for decision-making. (See attached...) Figure 4 As shown, the closed-loop neural modulation module consists of a stimulation decision unit and a transcranial alternating current stimulation unit. The stimulation decision unit has a pre-defined two-dimensional steady-state region of mental state. This region is defined as an elliptical area on the arousal-valence plane, with its geometric center located at (arousal 5.0, valence 0.0). The major semi-axis along the arousal direction is 2.0, and the minor semi-axis along the valence direction is 1.5. The criterion for determining the boundary of this ellipse is: the ratio of the square of the arousal deviation from the center value (5.0) to the square of the arousal direction semi-axis (2.0), plus the ratio of the square of the valence deviation from the center value (0.0) to the square of the valence direction semi-axis (1.5), the result of which is not greater than 1.

[0032] The stimulus decision unit determines in real time whether the current indicator point meets the above criteria. If it does, the user is considered to be in a stable psychological state, and no intervention is triggered. If it does not meet the criteria, the deviation vector from the point to the center of the ellipse (5.0, 0.0) is calculated. The direction of this deviation vector indicates the main dimension of the psychological state deviating from the stable state (e.g., high arousal negative valence corresponds to anxiety), and its magnitude represents the degree of deviation, i.e., the need for regulatory intensity.

[0033] Based on the deviation vector, the stimulus decision unit queries the built-in stimulus parameter mapping table. This stimulus parameter mapping table is a multidimensional lookup table, whose index dimensions include the direction quadrant and intensity level of the deviation vector, divided into three levels: low, medium, and high, corresponding to deviation vector magnitudes less than 1.0, 1.0 to 2.0, and greater than 2.0. Each entry contains a set of stimulus parameter instructions, including at least the stimulus frequency, ranging from 4 Hz to 40 Hz; the stimulus intensity, ranging from 0.5 mA to 2.0 mA; and the stimulus duration, ranging from 5 minutes to 20 minutes. For example, when the deviation vector points to the second quadrant (high arousal, negative valence) and the intensity is medium, the mapping table returns the instructions: frequency 10 Hz, corresponding to the alpha band, used to reduce arousal; intensity 1.2 mA; duration 10 minutes.

[0034] The transcranial alternating current stimulation unit activates upon receiving the command from the mapping table. Its internal waveform generator, based on direct digital frequency synthesis (DFD), generates an AC signal with a specified frequency and sinusoidal waveform. This AC signal is amplified by a constant current source circuit to ensure the output current is precisely and stably maintained at the milliampere level required by the command, with the maximum output current strictly limited to within 2 milliamperes, complying with international safety standards. A pair of wearable stimulation electrodes are made of medical-grade conductive silicone, coated with a silver chloride layer to reduce contact resistance. The electrodes are positioned according to the international 10-20 EEG electrode placement system, typically selecting F3 / F4 (corresponding to the left / right prefrontal cortex) or P3 / P4 (corresponding to the parietal lobe) as stimulation targets. The specific location is dynamically selected based on the psychological dimension being modulated. For example, for negative valence (depressive tendency), the left prefrontal cortex (F3) is stimulated first; for high arousal (anxiety), the parietal lobe (P3 / P4) is stimulated first. The electrodes are securely attached to the scalp via elastic straps or magnetic structures, ensuring that the stimulating current effectively penetrates the skull and acts on the target cortical area, regulating its neural oscillation activity and thus guiding the psychological state back to a stable range.

[0035] The entire system operates as follows: Figure 5 The hierarchical decision-making and control architecture is shown. The bottom-level sensing and control layer operates continuously at a frequency of 100 Hz, responsible for sensor driving, raw signal acquisition, and preliminary filtering to ensure high fidelity and temporal consistency of the data source. The middle-level state assessment and decision-making layer is periodically activated at a frequency of 1 Hz, performing feature extraction, data fusion, psychological state quantification, and stimulus decision-making, balancing computational load and real-time response. The high-level intervention execution and learning layer adopts an event-driven mechanism, activating the transcranial AC stimulation unit only when a stimulus command is received, and automatically triggering feedback learning and model update processes on a daily or weekly basis.

[0036] The feedback learning and model update module continuously records the complete context of each intervention event: including pre-stimulation psychological state indicators; applied stimulus parameters, frequency, intensity, duration, and electrode location; and the trajectory of psychological state indicator changes within a 15-minute time window after the stimulus. Based on this data, the module's built-in stimulus effect evaluation algorithm calculates the effectiveness score of the intervention. This effectiveness score is obtained by multiplying the cosine similarity between the actual movement vector and the expected adjustment direction vector by the ratio of the actual adjustment amplitude to the expected adjustment amplitude (with a maximum value of 1). The final result ranges from -1 to 1, with positive values ​​indicating effective intervention in the correct direction and negative values ​​indicating ineffective intervention or even reversal and deterioration. The expected adjustment direction is the direction in which the pre-stimulation indicator points towards the steady-state center, and the actual movement vector is the vector of the mean psychological state indicator within 15 minutes after the stimulus pointing towards the steady-state center.

[0037] All intervention records and effectiveness scores are encrypted and stored in a local intervention log database. Every Sunday at 2:00 AM, the system automatically initiates a model update process. This automatic model update process employs the policy gradient method from reinforcement learning, treating the stimulus parameter mapping table as a policy function, using psychological state bias as input, and stimulus parameter combinations as output actions. The goal is to minimize the long-term accumulated negative effectiveness scores. By calculating the policy gradient, the parameters of corresponding entries in the mapping table are fine-tuned. For example, if a certain type of stimulus has a low effectiveness score across multiple records, its intensity or frequency is appropriately reduced. This process requires no user intervention and is entirely completed on the device, ensuring privacy and security while allowing the system to gradually adapt to the user's individual physiological and psychological response characteristics.

[0038] In summary, this embodiment achieves millisecond-level synchronous data acquisition through a multimodal physiological sensor array, constructs high-dimensional physiological representations using feature-level fusion technology based on nuclear canonical correlation analysis, maps physiological features to quantifiable arousal and valence indices using a deep neural network model, and sets a clear two-dimensional steady-state interval as the regulation target. When a deviation from the steady-state psychological state is detected, the system automatically generates a personalized transcranial alternating current stimulation protocol for closed-loop intervention, and continuously optimizes the intervention strategy through a reinforcement learning-based feedback mechanism. This approach achieves a complete technical closed loop of "perception-assessment-intervention-learning" on a resource-constrained wearable platform, providing a reliable, accurate, and adaptive technical means for the proactive daily maintenance of mental health.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wearable vital sign monitoring system with multi-parameter synchronous acquisition, characterized in that, include: A multimodal physiological sensor array is used to acquire users' multidimensional raw physiological signals in parallel with millisecond-level time synchronization accuracy; The edge computing and fusion module is used to preprocess and extract features from the multi-dimensional raw physiological signals collected by the multimodal physiological sensing array, and to perform spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data to generate a fused feature vector. The psychological state quantification engine is used to receive the fusion feature vector generated by the edge computing and fusion module, and to calculate it through the preset psychological state quantification model to output a quantitative index that represents the user's current psychological state. The closed-loop neural modulation module is used to receive the psychological state quantification index output by the psychological state quantification engine, and when the psychological state quantification index exceeds the preset steady-state range, automatically generate and execute non-invasive neural modulation stimulation that matches the deviation direction and amplitude of the psychological state quantification index, so as to guide the physiological index back to the steady-state range.

2. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 1, characterized in that, The multimodal physiological sensing array integrates at least four types of physiological sensors, including a photoplethysmography pulse wave sensor, a skin conductance sensor, a triaxial accelerometer and gyroscope combination sensor, and an infrared thermal radiation sensor. The photoplethysmography (PPG) sensor uses a light-emitting diode as a light source and a photodetector to receive the light intensity signal reflected or transmitted through human tissue, which is used to extract the time-domain and frequency-domain features of heart rate and heart rate variability. The skin electrical activity sensor uses a constant voltage measurement circuit and contacts the user's skin area through a pair of silver silver chloride electrodes to measure the skin conductivity level and its fluctuation frequency. The triaxial accelerometer and gyroscope combined sensor is used to continuously monitor the user's limb movement acceleration and angular velocity, and eliminates the interference of motion artifacts on the photoplethysmography signal through the built-in attitude calculation algorithm, while extracting motion features that characterize the user's activity intensity and behavior pattern. The infrared thermal radiation sensor is aimed at the user's body surface area in a non-contact manner, and inverts the body surface temperature and its dynamic change trend by measuring the intensity of the received infrared radiation.

3. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 2, characterized in that, The edge computing and fusion module includes a signal preprocessing unit, a feature extraction unit, and a data fusion unit; The signal preprocessing unit is used to perform bandpass filtering, power frequency interference notch filtering, and baseline drift correction on each raw physiological signal. The feature extraction unit is used to extract at least five heart rate variability features from the preprocessed photoplethysmography pulse wave signal, including the standard deviation of adjacent heartbeat intervals and the ratio of low-frequency power to high-frequency power. Extract at least three features from the skin electrical activity signal, including the average skin conductivity and the number of nonspecific fluctuations; extract at least two features from the motion signal, including the activity count and the percentage of rest time; and extract at least two features from the body surface temperature signal, including the mean temperature and the slope of temperature change. The data fusion unit is used to establish a unified timestamp system, align feature data from different sensors to the same time reference, and use a kernel-based canonical correlation analysis method to calculate the nonlinear correlation weights between different modal features. The calculated correlation weights are used to sum all features in a weighted manner to generate a high-dimensional fusion feature vector.

4. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 3, characterized in that, The process of the canonical correlation analysis method based on kernel function is as follows: physiological feature vectors from different modalities are mapped to a high-dimensional regenerative kernel Hilbert space; the projection direction vector that maximizes the correlation between the projections of feature vectors from different modalities is solved in the high-dimensional regenerative kernel Hilbert space; and the contribution weight of each modal feature in the fusion space is calculated based on the solved projection direction vector.

5. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 4, characterized in that, The psychological state quantification engine has a built-in psychological state quantification model that is a deep neural network model, which takes the fused feature vector as input. The deep neural network model contains three hidden layers, with 128, 64 and 32 neurons in each layer, respectively, and the activation function is a modified linear unit. The output layer of the deep neural network model consists of two neurons, corresponding to two dimensions of the psychological state quantification index: arousal and valence. Arousal represents the level of psychological activation, and valence represents the positive or negative orientation of emotions.

6. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 5, characterized in that, The closed-loop neuromodulation module includes a stimulation decision unit and a transcranial alternating current stimulation unit. The stimulus decision unit is pre-set with a two-dimensional psychological state steady-state interval, which is defined as an elliptical region on the two-dimensional plane formed by arousal and valence. The stimulus decision unit is used to compare the positional relationship between the quantitative index points output by the psychological state quantification engine and the boundary of the steady-state interval in real time. When the quantitative index point is inside the steady-state interval, no output is generated. When the quantitative index point is outside the steady-state interval, the unit calculates the vector from the quantitative index point to the center point of the steady-state interval. The direction and length of the vector define the target direction and intensity requirement of the adjustment, respectively. Based on the vector, the unit queries and generates a set of stimulus parameter instructions from a preset stimulus parameter mapping table. The set of stimulus parameter instructions includes at least the stimulus frequency, stimulus intensity, and stimulus duration. The transcranial alternating current stimulation unit is used to receive stimulation parameter instructions from the stimulation decision unit and generate and execute corresponding non-invasive neuromodulation stimulation.

7. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 6, characterized in that, The transcranial alternating current stimulation unit includes a waveform generator, a constant current source circuit, and a pair of wearable stimulation electrodes. The waveform generator is used to receive stimulus parameter instructions from the stimulus decision unit and generate an AC signal with corresponding frequency and waveform. The constant current source circuit is used to ensure that the current intensity output to the human body is accurately and stably stable at the value required by the instruction, and its maximum output current is no more than 2 mA. The pair of wearable stimulation electrodes are made of conductive silicone and are positioned and attached to the user's scalp according to the international 10-20 EEG electrode placement system. They are used to safely apply weak alternating current signals to the functional network of the cerebral cortex.

8. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 7, characterized in that, The system also includes a feedback learning and model update module; The feedback learning and model update module is used to continuously record the parameters of each stimulus executed by the closed-loop neural modulation module, the quantitative indicators of the psychological state before execution, and the trajectory of changes in the quantitative indicators of the psychological state within a preset time window after execution; calculate the effectiveness score of this stimulus intervention through the built-in stimulus effect evaluation algorithm; store all intervention records and effectiveness scores in the local intervention log database; and periodically fine-tune and optimize the stimulus parameter mapping table in the stimulus decision unit using the policy gradient method in the reinforcement learning algorithm.

9. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 8, characterized in that, The calculation process of the stimulus effect evaluation algorithm is as follows: obtain the expected adjustment direction vector formed by the quantitative index point of the psychological state before the stimulus is executed and the center point of the steady-state interval; obtain the actual movement vector formed by the mean of the quantitative index of the psychological state within the preset time window after the stimulus is executed and the center point of the steady-state interval; calculate the cosine similarity between the actual movement vector and the expected adjustment direction vector, and multiply the cosine similarity by the ratio of the magnitude of the actual movement vector to the magnitude of the expected adjustment vector to obtain an effectiveness score between -1 and 1.

10. The wearable vital signs monitoring system with multi-parameter synchronous acquisition according to claim 9, characterized in that, The system operates according to a hierarchical decision-making and control architecture. The hierarchical decision-making and control architecture includes a bottom-level sensing and control layer, a middle-level state assessment and decision-making layer, and a high-level intervention execution and learning layer. The underlying sensing control layer operates at a sampling rate of 100 Hz and is responsible for driving the multimodal physiological sensing array, acquiring raw signals, and performing preliminary filtering. The middle-level state assessment and decision-making layer operates at a frequency of 1 Hz, performing feature extraction, data fusion, psychological state quantification calculation, steady-state interval comparison, and stimulus decision-making. The high-level intervention execution and learning layer operates in an event-driven manner, activating the closed-loop neural modulation module only when a stimulus command is received, and running the update process of the feedback learning and model update module on a daily or weekly basis.

Images