Distributed wearable health monitoring system

Abstract

The invention relates to the technical field of wearable health monitoring, and discloses a distributed wearable health monitoring system, which comprises a watch dial and a wearable main body of a watch chain, the distributed sensing module is arranged on the wearable main body and is used for collecting multi-source physiological signals; the data preprocessing and tensor construction module is electrically connected with the distributed sensing module and used for processing the multi-source physiological signals and constructing the multi-source physiological signals into a physiological state tensor structured tensor fusion engine, and the data preprocessing and tensor construction module is used for decomposing the physiological state tensor to extract one or more low-dimensional core features reflecting potential physiological laws. According to the invention, distributed sensing is carried out on the dial plate and the watch chain, and the multi-source signal is constructed into the physiological state tensor, so that the limitation of scattered information of traditional equipment is overcome, and comprehensive and deep monitoring of the health state of the user is realized. The structured data basis makes it possible to capture the coupling relation between different physiological systems, and therefore more accurate evaluation is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of wearable health monitoring, and in particular to a distributed wearable health monitoring system. Background Art

[0002] With rising public health awareness and continuous advancements in microelectronics and sensor technologies, wearable health monitoring devices, such as smartwatches and fitness trackers, have become increasingly popular and are playing a positive role in daily vital sign monitoring, exercise management, and initial screening for chronic diseases. Currently, existing wearable devices typically integrate photoplethysmography (PPG) sensors, accelerometers, and electrocardiogram (ECG) sensors to measure basic physiological parameters such as the user's heart rate, blood oxygen saturation, heart rate variability, and specific heart rhythms.

[0003] However, existing technologies still face multiple significant challenges in achieving in-depth and accurate personalized health assessments: First, at the hardware level, there are limitations due to a single sensor layout and low functional integration. Most devices concentrate sensors on the back of the dial. This design not only affects wearing comfort due to localized skin pressure, but also leads to severe distortion of blood oxygen monitoring data in certain scenarios (such as during sleep due to a loose fit), limiting accuracy. Furthermore, this centralized layout and limited sensor types make it difficult for traditional devices to simultaneously support in-depth, non-invasive analysis of multiple parameters such as blood lipids, liver and kidney function, and monitoring of special conditions such as EEG activity requires users to wear additional head-mounted equipment, making integrated, non-invasive monitoring impossible.

[0004] Secondly, existing devices generally face a sharp conflict between power consumption and accuracy. While introducing high-precision sensors (such as UWB positioning and multispectral sensing) is a promising approach to improving monitoring capabilities, their continuous operation significantly reduces device battery life. Existing batteries are often located within the dial, limiting capacity expansion and forcing users to compromise between high-precision monitoring and extended battery life.

[0005] Thirdly, the lack of a dynamic data calibration mechanism affects the reliability of monitoring results. Physiological data is easily affected by factors such as ambient temperature and user movement, causing drift. Existing devices generally lack a real-time dynamic calibration mechanism to filter out these interferences, leading to doubts about the reliability of the data.

[0006] At the data processing and analysis level, existing technologies typically perform independent or shallow fusion analysis on the different signals collected. Their analysis methods often stop at extracting the isolated features of each signal and simply presenting these indicators to the user, ignoring the profound and dynamic coupling relationships between different physiological subsystems. For example, the complex relationships between subtle changes in the cardiovascular system, the stress response of the nervous system, and the thermodynamic representation of metabolism are difficult to effectively explore and utilize within the existing technological framework. This neglect of the inherent structure of the data and cross-dimensional correlations greatly limits the depth and accuracy of health assessments.

[0007] Furthermore, existing health assessment models often rely on supervised learning algorithms trained on large-scale population data. These approaches not only rely heavily on expensive datasets with precise medical annotations, but also suffer from a fundamental flaw: a one-size-fits-all assessment model fails to adapt to individual physiological uniqueness. Everyone's baseline health and circadian rhythms vary significantly, and a fluctuation in a metric that is normal for a group may signal an early sign of deviation for a specific individual.

[0008] Therefore, the present invention proposes a distributed wearable health monitoring system to address the deficiencies of the prior art. Summary of the Invention

[0009] In response to the shortcomings of the existing technology, the present invention provides a distributed wearable health monitoring system to solve the problem that existing wearable devices are difficult to conduct comprehensive, in-depth and accurate health status assessments of users due to the single dimension of sensor information, shallow data fusion analysis and lack of personalized modeling capabilities.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: a distributed wearable health monitoring system, comprising: The wearing body of the dial and bracelet; A distributed sensing module, located on the wearable body, collects multi-source physiological signals. The dial area features an integrated micro-radar module (60GHz frequency band), a multi-wavelength biophoton sensor (405nm / 940nm / 1300nm), and a metabolic thermal analysis array. The bracelet area incorporates embedded repetitive battery cells (500mAh each), electroencephalographic and myoelectric coupling sensors (EMG), and a pressure-sensitive film. This distributed layout of sensors and batteries reduces dial weight, improves wearing comfort, and enhances signal coverage uniformity. The bracelet's six batteries are hot-swappable, and the wireless charging coil (dual receivers for the dial and strap) allows charging from any location. a data preprocessing and tensor construction module, electrically connected to the distributed sensing module, for processing the multi-source physiological signals and constructing them into physiological state tensors; a structured tensor fusion engine for decomposing the physiological state tensor to extract one or more low-dimensional core features reflecting underlying physiological laws; The health status assessment module assesses the user's health status based on the core features. This module preferentially uses the watch strap battery, while the communication module is powered by the watch dial battery.

[0011] Preferably, the distributed sensing module includes: at least one first sensing unit disposed on the dial, configured to collect physiological signals related to the user's core circulatory system; at least one second sensing unit disposed on the bracelet, configured to collect physiological signals related to a user's peripheral physiological system; The first sensing unit and the second sensing unit work together to capture multi-dimensional information of different physiological systems.

[0012] Preferably, the first sensing unit includes at least one selected from the following: The micro-radar sensor unit is used to collect signals reflecting the mechanical activity of the cardiovascular system. It uses a 60GHz millimeter-wave radar to capture microvascular pulsations at a depth of 0.5-2mm below the skin. Combined with a multi-wavelength photon sensor, it analyzes the oxygenation status of hemoglobin and enables dynamic blood pressure calculation (systolic / diastolic pressure error ≤±3mmHg). A multi-wavelength optical sensing unit for collecting signals reflecting the optical characteristics of blood; Furthermore, the second sensing unit includes at least one selected from the following: The metabolic thermal analysis unit is used to collect temperature field signals reflecting the local metabolic state. The 8×8 thermopile array on the back of the dial continuously maps the wrist temperature distribution. The AI ​​model (LSTM network) non-invasively estimates the blood lipid metabolism rate and liver and kidney function indicators (such as ALT / AST trend warning) based on the trend of thermal field changes. The neuromuscular coupling sensing unit is used to collect electromyographic signals that reflect the activity state of the nervous system. It captures electromyographic signals of abnormal EEG conduction (such as pre-epileptic myoclonus) through the EMG sensor on the strap, and combines it with the micro-tremor frequency (0.5-30Hz) monitored by radar to indirectly reconstruct the characteristics of EEG activity and warn of neural abnormalities.

[0013] Preferably, the physiological state tensor It is a fourth-order tensor whose dimensions are composed of time, mode, feature and physical channel, expressed as: ; in, Represent the lengths of time, modality, feature and physical channel dimensions respectively.

[0014] Preferably, the structured tensor fusion engine transforms the physiological state tensor into Decomposition, the decomposition form is expressed as: ; in, is the core tensor, , , , are the factor matrices of time, modality, feature, and channel dimensions respectively.

[0015] Preferably, the execution of Tucker decomposition is achieved by solving a constrained optimization objective containing physiological prior knowledge, wherein the constrained optimization objective is: ; in, represents the reconstruction error term; is the time smoothing constraint; is the modal sparsity constraint term; is the difference matrix; and are the hyperparameters that control the strength of the corresponding constraints.

[0016] Preferably, the health status assessment module includes a personalized deep neural network model trained by a self-supervised learning paradigm, and the self-supervised learning paradigm enables the personalized deep neural network model to learn the user's physiological dynamic baseline by performing tasks such as masking and reconstructing the physiological state tensor.

[0017] Preferably, the masking and reconstruction tasks are performed by minimizing a loss function that only calculates the reconstruction error of the masked position. To achieve this, the loss function is: ; in, is the mask tensor; Reconstructed physiological state tensors for personalized deep neural network models; is the original physiological state tensor, is a tensor related to the physiological state A tensor of exactly the same shape, with all elements containing 1. For Hadamard; are all learnable parameters within the personalized deep neural network model.

[0018] Preferably, the health status assessment module quantifies the degree to which the current physiological state deviates from the user's physiological dynamic baseline by calculating the residual between the physiological state tensor collected in real time and the reconstructed physiological state tensor output by the personalized deep neural network model.

[0019] The present invention also provides a distributed wearable health monitoring method, which includes the following steps: S1. Collecting multi-source physiological signals through the distributed sensing module; S2. Processing the multi-source physiological signals and constructing them into a physiological state tensor through the data preprocessing and tensor construction module; S3. Decomposing the physiological state tensor by the structured tensor fusion engine to extract one or more low-dimensional core features reflecting potential physiological laws; S4. Evaluate the user's health status based on the core features through the health status evaluation module.

[0020] The present invention provides a distributed wearable health monitoring system. It has the following beneficial effects: 1. This invention achieves comprehensive and in-depth monitoring of the user's health status by deploying distributed sensors on the dial and bracelet of the wearable device and innovatively constructing the collected multi-source heterogeneous signals into a unified physiological state tensor. This method overcomes the limitation of traditional wearable devices that can only provide isolated and fragmented physiological indicators. By structurally integrating multiple information reflecting cardiovascular mechanical activity, blood optical properties, metabolic thermal field distribution, and neuromuscular coupling status, it forms a physiological snapshot that is far richer and more three-dimensional than a single data source. This comprehensive data foundation enables subsequent analysis to capture the complex coupling relationships between different physiological subsystems, thereby making a more accurate and in-depth assessment of the user's overall health status.

[0021] 2. The present invention proposes a structured tensor fusion engine based on physiological prior knowledge, which significantly improves the robustness and interpretability of the extracted features. In the process of decomposing high-dimensional physiological state tensors, by imposing constraints such as time smoothing and modal sparsity, the engine can actively filter out random noise and automatically identify the key sensor modal combinations that dominate specific physiological patterns. This makes the low-dimensional core features it extracts no longer a pure mathematical dimensionality reduction result, but a purer and more stable pattern with clear physiological significance. This method makes the analysis of health status no longer a "black box" process, but is based on a set of understandable and traceable structured features, thereby improving the reliability of the evaluation results.

[0022] 3. The present invention adopts a self-supervised learning paradigm to construct a personalized health assessment model, which solves the problem of traditional methods' dependence on large-scale labeled medical data and realizes personalized monitoring in a true sense. By allowing the model to perform masking and reconstruction training on the user's own unlabeled data, the system can autonomously learn and internalize the user's unique physiological dynamic baseline. Therefore, the assessment of health status is no longer compared with the broad group average level, but is compared with the user's own historical normal state. This highly personalized method can more sensitively capture individual, subtle deviations in physiological status, providing strong technical support for early risk warning and refined health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a system architecture diagram of the present invention; Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] Please see the attached Figure 1 , an embodiment of the present invention provides a distributed wearable health monitoring system, comprising: The wearing body of the dial and bracelet; A distributed sensing module is arranged on the wearable body and is used to collect multi-source physiological signals; a data preprocessing and tensor construction module, electrically connected to the distributed sensing module, for processing the multi-source physiological signals and constructing them into physiological state tensors; a structured tensor fusion engine for decomposing the physiological state tensor to extract one or more low-dimensional core features reflecting underlying physiological laws; The health status assessment module assesses the user's health status based on the core features.

[0026] In this embodiment, a wearable body of a distributed wearable health monitoring system is designed in terms of physical structure and functional layout to serve as a solid foundation for all subsequent data processing and algorithm operation.

[0027] The wearable device consists of a dial, the core of the system, and a connected flexible bracelet with integrated sensing capabilities. This design eliminates the limitations of traditional smartwatches, which concentrate all functional units on a single dial. By distributing sensing capabilities across space, it builds a collaborative multi-node perception network.

[0028] The dial, serving as the main control center of the system, not only houses the main processor, power management module and wireless communication module in its internal shell, but also integrates a set of key core sensing units.

[0029] The main processor is configured as the computing core of the system. It is not only responsible for scheduling the synchronization of all sensor units, but more importantly, it carries the execution tasks of all subsequent core algorithms of the present invention. Specifically, it runs the data preprocessing and tensor construction modules, aligns and normalizes the collected multi-source signals, and finally constructs a well-structured fourth-order physiological state tensor. .

[0030] Furthermore, the main processor performs the function of the structured tensor fusion engine. In this process, it performs the constructed physiological state tensor A Tucker decomposition based on physiological prior knowledge constraints is performed. The objective function of this decomposition is as follows: ; in, represents the reconstruction error term; is the time smoothing constraint; is the modal sparsity constraint term; is the difference matrix; and are the hyperparameters that control the strength of the corresponding constraints.

[0031] By solving this optimization problem, the processor can extract more physiologically meaningful low-dimensional core features, namely core tensors, from complex raw data. and a set of factor matrices .

[0032] In addition, the main processor also carries the training and inference tasks of the self-supervised physiological dynamics model. During the training phase, it optimizes a deep neural network model by performing masking and reconstruction-based learning tasks, and its loss function is: ; in, is the mask tensor, Reconstructed physiological state tensors for personalized deep neural network models, is the original physiological state tensor, is a tensor related to the physiological state A tensor of exactly the same shape, with all elements containing the value 1, For Hadamard, are all learnable parameters within the personalized deep neural network model.

[0033] In daily use, the trained model is used for real-time inference to assess the user's health status.

[0034] The bottom of the dial, the side that comes into direct contact with the user's wrist skin, is preferably integrated with a first sensing unit. This unit comprises a miniature millimeter-wave radar sensor unit and a multi-wavelength optical sensor unit. Placing these two units side by side in densely vascular areas, such as the radial artery at the wrist, allows for the coordinated capture of deep and superficial information directly related to the cardiovascular system. The radar unit detects minute vibrations of the blood vessel wall caused by blood flow pulsation, while the optical unit monitors changes in optical properties such as blood oxygen and hemoglobin, complementing each other in functionality.

[0035] The structure of the bracelet is different from that of traditional passive straps. It integrates multiple sensor units through an embedded flexible printed circuit board (FPC) and realizes the electrical signal connection between the sensor units and the main processor of the dial.

[0036] Preferably, a second sensor unit is arranged on the bracelet, which includes a metabolic heat analysis unit and a neuromuscular coupling sensor unit.

[0037] Metabolic thermal analysis units are distributed in an array along the inside of the bracelet. This array design is key to achieving "temperature field" monitoring, which aims to capture subtle temperature gradients in different areas of the wrist caused by differences in blood perfusion and metabolic activity, which is impossible to detect with a single point temperature sensor.

[0038] The neuromuscular coupling sensor unit is also integrated into the watch strap in the form of a flexible electrode array. Its placement on the strap allows it to cover a wider area of ​​the forearm muscles, allowing for more stable capture of the extremely weak surface electromyographic signals generated by central nervous system activity in the limb extremities.

[0039] Ultimately, this distributed physical layout of the dial and bracelet is a prerequisite for the effective operation of this invention. It is precisely because of the simultaneous possession of the core circulatory system information provided by the dial sensor unit and the metabolic and nervous system peripheral information provided by the bracelet sensor unit that the system can construct a tensor with sufficient information dimensions to fully reflect the physiological state. This richness enables subsequent structured tensor decomposition to discover meaningful coupling relationships between different physiological subsystems, and also provides the self-supervised learning model with sufficient contextual information to learn and reconstruct masked physiological data, thereby achieving in-depth and personalized assessment of the user's health status.

[0040] In this embodiment, a distributed sensing module is integrated into the watch face and bracelet of the wearable device. Its core purpose is to collect multi-source heterogeneous signals that comprehensively reflect the user's physiological state through a collaborative multi-point sensing strategy. The design concept of this module is based on the inherent information limitations of any single physiological signal source. By distributing multiple sensors with complementary functions in physical space, a perception foundation with richer information dimensions and stronger anti-interference capabilities can be established.

[0041] Specifically, the distributed sensing module is physically divided into a first sensing unit arranged on the dial and a second sensing unit arranged on the bracelet, and the two realize electrical signal connection and synchronization through a flexible circuit inside the wearable body.

[0042] The first sensing unit is preferably arranged at the bottom of the dial, that is, in direct contact with the user's wrist radial artery and other blood vessel-dense areas. This layout is intended to maximize the capture of core physiological signals directly related to the cardiovascular system.

[0043] This first sensing unit further includes a miniature radar sensing unit. This unit operates based on frequency-modulated continuous wave (FMCW) radar technology. It incorporates a low-power millimeter-wave transceiver, which transmits high-frequency electromagnetic waves to the subcutaneous tissue of the wrist. These electromagnetic waves are reflected by the walls of blood vessels, which periodically expand and contract due to the heartbeat. Due to the Doppler effect, the echo signals caused by these minute displacements of the vessel walls carry subtle phase variations that are perfectly synchronized with the cardiac cycle. The unit performs high-speed sampling and digital signal processing on the echo signals to demodulate the time series of these phase variations, thereby non-invasively acquiring raw data reflecting the mechanical state of the cardiovascular system.

[0044] The first sensing unit also includes a multi-wavelength optical sensing unit. This unit utilizes photoplethysmography (PPG) technology, but its innovation lies in the use of multiple light sources with different wavelengths. Preferably, the unit integrates multiple discrete LEDs, whose wavelengths range from visible light to near-infrared light. Within a very short sampling period, the system uses time-division multiplexing technology to rapidly and sequentially illuminate each LED. When light of a specific wavelength strikes the subcutaneous tissue, a highly sensitive photodetector simultaneously records the intensity of light reflected or transmitted by the blood and tissue. Because different substances (such as oxygenated and deoxygenated hemoglobin) have significant differences in their absorption rates at different wavelengths, the unit is able to simultaneously acquire multiple independent PPG time-series signals, each of which contains specific information about blood composition and hemodynamics.

[0045] The second sensing unit is innovatively integrated into a flexible bracelet, extending the sensing capability from the core dial area to the entire wrist circumference, aiming to capture peripheral signals reflecting a wider range of physiological systems.

[0046] This second sensing unit includes a metabolic thermal analysis unit. This unit abandons traditional single-point temperature measurement and consists of a two-dimensional array of thermal sensors. For example, a 4x4 matrix of high-precision MEMS thermal sensors is encapsulated on a flexible substrate and evenly distributed along the inside of the bracelet. This array can synchronously collect skin surface temperatures at different wrist regions in real time. Its output is no longer a single temperature value, but a two-dimensional temperature field distribution matrix that varies over time. This design aims to indirectly infer local blood perfusion, metabolic rate, and even potential inflammatory responses by analyzing the temperature field's gradients, dynamic changes, and hot spot patterns—information that single-point temperature measurement cannot reveal.

[0047] The second sensing unit also includes a neuromuscular coupling (NMC) sensor. This unit is also integrated into the inner side of the bracelet in the form of a flexible electrode array, preferably using dry electrodes that do not require conductive gel. Its placement on the bracelet allows its coverage to extend to the upper wrist, closer to the extensor and flexor muscles of the forearm. Its operating principle is that many physiological changes in the human body (such as stress and fatigue) are transmitted to the extremities of the limbs in the form of weak electrical signals through the central nervous system, triggering involuntary changes in muscle fiber tension. The electrode array is extremely sensitive, capable of capturing these surface electromyography (sEMG) signals, which are typically in the microvolt range. Through multi-channel acquisition, the system analyzes the energy, frequency, and inter-channel correlation of these signals, thereby obtaining a window into the state of nervous system activity.

[0048] In summary, the distributed sensing module in this invention, through its coordinated physical layout on the dial and bracelet, achieves the simultaneous acquisition of signals from four different physical dimensions (mechanical vibration, optical properties, thermal field distribution, and biopotential). The multi-source physiological signals it outputs are used to construct a fourth-order physiological state tensor that can comprehensively characterize the user's status. It provides necessary, rich and complementary raw data input, which is the fundamental premise and information guarantee for the entire system to achieve in-depth and personalized health analysis.

[0049] In this embodiment, a data preprocessing and tensor construction module physically receives the raw data stream from the distributed sensing module via an electrical connection and functionally serves as the data input source for the subsequent structured tensor fusion engine. The core mission of this module is to transform the multi-source, heterogeneous, and irregularly time-domain and amplitude-modulated physiological signals collected by the front-end through a series of standardized signal processing and data structuring processes into a unified, advanced, high-dimensional data representation that can be understood by multiple linear algebra methods, namely the physiological state tensor.

[0050] This module is essential because it addresses the fundamental problem of subsequent advanced algorithms being unable to directly process raw heterogeneous data. Raw data exhibits significant variations in sampling rate, physical dimensions, numerical range, and noise characteristics. Direct fusion analysis without processing will result in information overwhelm and model failure. Therefore, this module's function is to lay a solid and reliable data foundation for the entire system's intelligent analysis process.

[0051] Specifically, the operation of the data preprocessing and tensor construction module can be further refined into a series of logically progressive functional units.

[0052] First, the module includes a multi-source signal timing alignment unit. This unit receives multiple parallel data streams from radar, optical, thermal analysis and neuromuscular coupling sensor units. The technical problem to be solved is that the sampling frequencies of the original data of each sensor unit are not completely consistent due to their different physical characteristics and working modes. The unit uses the high-precision clock provided by the system main processor as a unified time base, and preferably uses the B-spline interpolation algorithm to resample all data streams of non-base frequencies. The reason for using B-spline interpolation is that compared with simple linear interpolation, it can generate smoother curves in all orders of derivatives. This feature is more in line with the inherent continuous and slowly changing characteristics of most physiological signals, so that the original dynamic information of the signal can be retained to the maximum extent during the alignment process, and finally a set of signals that are strictly aligned to a unified system frequency on the time axis is output. Multi-channel time series.

[0053] Next, the module includes a data normalization unit. After timing alignment, the signal channels are synchronized in time, but their physical dimensions and numerical ranges are still very different. For example, the output of the thermal analysis unit is a floating point number in degrees Celsius, while the output of the neuromuscular coupling unit is a voltage value in microvolts. In order to eliminate this dimensional difference and avoid bias in subsequent algorithms due to numerical values, the unit performs Z-score normalization on each independent signal channel. Specifically, for any signal channel, the unit calculates its mean within a sliding time window. and standard deviation , and for each data point in the window according to the following formula Perform the transformation: ; After this processing, the data of all signal channels were converted into dimensionless values ​​with a mean of 0 and a standard deviation of 1, making them comparable in the subsequent joint analysis.

[0054] Finally, the module includes a core physiological state tensor construction unit. This unit is the final implementer of the module's functions and is responsible for integrating multiple one-dimensional time series that have been processed and standardized into a single, multi-dimensional structured data entity. This unit constructs a fourth-order tensor structure in memory. , each dimension of which has a clear physical meaning: The first dimension, the time dimension (indexed by ), its length Corresponds to the number of sampling points within an analysis time window.

[0055] The second dimension, the modal dimension (indexed by ), its length Corresponds to the total number of sensor types used.

[0056] The third dimension, the feature dimension (index is ), its length Corresponding to the different characteristic channel numbers within a single mode, such as different wavelengths in an optical mode.

[0057] The fourth dimension, the physical channel dimension (index is ), its length Corresponds to the number of physical probes in an array sensor, such as a thermal array or electrode array.

[0058] This unit accurately fills the normalized data into the corresponding coordinate positions in the fourth-order tensor structure according to its time point, mode, feature type and physical channel source.

[0059] After completing the above process, the final output of this module is a physiological state tensor with complete content and regular structure. This tensor not only completely preserves all the original signal information, but more importantly, through its inherent multidimensional structure, it explicitly expresses the intrinsic correlations between signals in time, space, modality, and channels. This physiological state tensor will be passed as the only and necessary input to the subsequent structured tensor fusion engine. Its data quality and structural integrity directly determine the effectiveness and reliability of all subsequent analysis steps of the entire invention scheme.

[0060] In this embodiment, a structured tensor fusion engine is used to process the data preprocessing and tensor building module outputs, i.e., the physiological state tensor. , and is responsible for performing the core data fusion and feature extraction tasks of this invention. The fundamental purpose of this engine is not simply to compress high-dimensional data, but to discover and isolate low-dimensional core features that reflect the user's underlying physiological state and dynamic patterns from the complex, noisy, and redundant raw data through a decomposition process guided by physiological priors.

[0061] The technical solution of this engine is based on a core insight: directly analyzing high-dimensional physiological state tensors It is difficult to understand the physiological patterns within it, as they are often overwhelmed by a large amount of redundant information and random noise. This engine uses a constrained tensor decomposition method to break down the complex original tensor into a set of more understandable "basic building blocks" (i.e., factor matrices) with clear physical meaning, and a "blueprint" (i.e., core tensor) that describes the core interactions between these building blocks.

[0062] Specifically, the structured tensor fusion engine preferably adopts a constrained Tucker decomposition model. This model aims to transform the input fourth-order physiological state tensor It can be approximately expressed as the following multilinear product form: ; in, is the core tensor, , , , are factor matrices of time, modality, feature, and channel dimensions, respectively. Their column vectors can be considered as the "basis modes" that constitute the original data space of the corresponding dimensions. Core Tensors The dimension is much smaller than the original tensor , which captures the core interaction intensity and coupling relationship between these “base modes” and can be regarded as a highly condensed representation of the user’s potential physiological state.

[0063] The innovation of this engine lies in that its decomposition process is not completed by solving a simple least squares problem, but by solving a complex constrained optimization objective that incorporates physiological prior knowledge as a regularization term. The optimization objective function The definition is as follows: ; The objective function consists of three key parts: The first part is the reconstruction error term This item is a data fidelity item, and its function is to ensure that the core tensor and factor matrix obtained by decomposition can reconstruct the original tensor to the greatest extent possible, ensuring that the decomposition process does not lose too much effective information.

[0064] The second part is the time smoothing constraint This is the first physiological prior introduced by this engine. The mechanism behind it is that most physiological processes (such as changes in heart rate and body temperature) are continuous and smooth on a macroscopic scale, without any meaningless drastic jumps. This constraint is preferably defined as ,in Is a difference matrix. By adjusting the time factor matrix during the optimization process The gradient of the constraint is penalized. This constraint guides the engine to learn a smooth temporal basis pattern, which effectively filters out high-frequency noise and makes the extracted temporal pattern more physiologically interpretable. is a hyperparameter that controls the strength of this constraint.

[0065] The third part is the modal sparsity constraint term This is the second physiological prior introduced by this engine. Its mechanism is that many specific physiological events or state changes are often led by a few key physiological subsystems (corresponding to specific sensor modalities), rather than all systems participating equally. This constraint is preferably defined as the L1 norm form During the optimization process, this L1 norm drives the modal factor matrix Many elements that do not contribute much to the current pattern become exactly zero. As a result, the engine can automatically discover and highlight the key sensor modality combinations that dominate specific physiological patterns, realize automatic feature selection, and greatly enhance the interpretability of the decomposition results. is a hyperparameter that controls the sparsity strength.

[0066] To solve this constrained optimization problem, the engine uses an iterative Alternating Least Squares (ALS) algorithm. In each iteration, the algorithm fixes all variables except one (e.g., a factor matrix) and then solves a subproblem with a corresponding regularization term for that variable, alternating between them until the overall objective function converges.

[0067] Ultimately, the output of the structured tensor fusion engine is a set of low-dimensional core features that have been deeply fused, denoised, and contain clear physiological meanings, namely the core tensor and a set of structured factor matrices Compared to the original physiological state tensor, this set of features is not only lower dimensional and easier to process, but also has a clearer internal structure and less noise. It will serve as a key input for the subsequent personalized physiological dynamics modeling module, providing a high-quality, concise feature foundation for learning the user's health status baseline and dynamic changes.

[0068] In this embodiment, a health status assessment module, implemented as the top-level application of the monitoring system described herein, is responsible for converting the information processed and extracted by the preceding modules into meaningful health status assessments, risk warnings, and trend analysis for users. This module's design philosophy is to achieve a comprehensive assessment of a user's health status, combining both immediate and long-term insights, through a dual-track approach that combines deep learning models with structured feature analysis.

[0069] Functionally, this module uses a personalized trained deep neural network model to dynamically evaluate and detect anomalies in real-time physiological data. On the other hand, it performs long-term health trend analysis and pattern recognition based on the low-dimensional core features extracted by the structured tensor fusion engine.

[0070] Specifically, the core of the health status assessment module is a deep neural network model trained through a self-supervised learning paradigm. The model's training process is the foundation for building personalized assessment capabilities. Its purpose is to enable the model to autonomously learn and internalize the user's unique physiological data distribution patterns, which serve as their health baseline, without the need for any external medical labels.

[0071] The self-supervised learning paradigm preferably adopts a learning task based on masking and reconstruction. During the training phase, the system obtains the physiological state tensor from the user's historical data. As a training sample. Through a randomly generated binary mask tensor , for the input Perform partial masking to generate incomplete tensors This incomplete tensor is then fed into the model , and requires the model to output a reconstruction of the original complete tensor The training goal of the model is not to minimize the overall reconstruction error, but to focus on minimizing the reconstruction error at the masked position. Its loss function is The definition is as follows: ; in, is the mask tensor, Reconstructed physiological state tensors for personalized deep neural network models, is the original physiological state tensor, is a tensor related to the physiological state A tensor of exactly the same shape, with all elements containing the value 1, For Hadamard, All learnable parameters within the personalized deep neural network model. In this way, the model is forced to learn the deep internal connections between different modalities, different channels, and different time points, rather than simply copying visible data. After continuous training on the user's personal data, the model It will converge into a personalized generative model that can accurately characterize the user's regular physiological dynamics After obtaining the trained model, the health status assessment module performs real-time status assessment. When the new real-time data stream is preprocessed, a new physiological state tensor is formed. After that, the tensor is input into the model Perform a forward inference on the model to obtain an "ideal" reconstruction based on the healthy baseline understood by the model. By calculating the residual tensor between the observed value and the reconstructed value The system quantifies the extent to which the current physiological state deviates from the user's personal baseline. A significantly increased residual indicates that the current physiological data pattern is "unseen" by the model within its learned healthy baseline, potentially indicating the emergence of an abnormal physiological state. Based on the size, duration, and modal range of this residual, the module can implement risk stratification and early warning.

[0072] At the same time, the health status assessment module also performs long-term health status assessment and trend analysis based on the core features extracted by the structured tensor fusion engine. and the factor matrix , which is a concise expression of the original high-dimensional data after structured denoising and fusion.

[0073] This module stores these core features generated periodically by the fusion engine for a long time. By analyzing the evolution of these low-dimensional features over time (for example, in units of days or weeks), it is possible to reveal the slow and long-term changing trends of the user's health status. For example, by analyzing the modal factor matrix By tracking the changes in the core tensor, we can observe whether the weight of a specific physiological subsystem (such as the cardiovascular system or the nervous system) in the overall physiological activity has changed over the long term. The evolution of specific elements in the data provides insights into the changes in the coupling strength between different physiological patterns. This core feature-based analysis provides data support for generating in-depth and highly interpretable periodic health reports.

[0074] In summary, the health status assessment module of the present invention combines real-time anomaly detection based on a self-supervised learning model with long-term trend analysis based on structured core features, achieving a multi-dimensional, multi-timescale health assessment paradigm. It not only enables rapid response to immediate status deviations that may indicate acute risks, but also provides in-depth insight into long-term changes that may reflect the development of sub-health or chronic disease risks, thereby providing users with a comprehensive, dynamic, and highly personalized health monitoring and assessment service.

[0075] Non-invasive lipid monitoring: After the user wears the watch, the thermopile array continuously monitors the thermal field distribution in the radial artery area of ​​the wrist (sampling rate 1Hz).

[0076] Metabolic thermal feature extraction: The temperature gradient in the lipid metabolism active area increases by 0.3-0.5°C. The AI ​​model combines the user's age and BMI data to output the blood lipid trend index.

[0077] Verification and comparison: The correlation with the hospital's biochemical test results reached R²=0.87 (N=100 samples).

[0078] Sleep apnea warning: A micro-radar monitors the amplitude of chest micro-movements (with an accuracy of 0.1mm), and a multi-wavelength sensor synchronously collects blood oxygen saturation.

[0079] When breathing interruption is detected for more than 10 seconds and the blood oxygen drop rate is ≥4% / minute, gentle vibration is triggered to wake the user.

[0080] A sleep report is generated the next morning: number of respiratory events, percentage of deep sleep, and intervention effectiveness statistics.

[0081] Distributed hardware architecture: Breaking through the space limitations of the dial, enabling repeated arrangement of sensors / battery straps to improve comfort and accuracy.

[0082] Non-invasive multi-physiological parameter monitoring: micro radar + photon sensing + thermal field analysis replaces traditional invasive testing (such as blood drawing to measure blood lipids).

[0083] Indirect monitoring of EEG activity: Reconstructing EEG features through electromyographic signals to avoid the need for additional EEG wearable devices.

[0084] Dynamic environmental compensation mechanism: AI calibration eliminates motion / temperature drift interference to ensure clinical-grade data reliability.

[0085] Dynamic monitoring of body temperature field (sampling rate 10Hz, accuracy +0.1C); Psychological stress index (based on HRV nonlinear analysis) medical-grade blood pressure monitoring (error <±2mmHg); cardiac electrical activity monitoring (equivalent to single-lead ECG) non-invasive brain electrical activity monitoring (through V-band electromyography inversion).

[0086] Chronic disease management: fusion analysis of multi-source data on blood sugar, diet and exercise in patients with diabetes; Emergency warning: Myocardial infarction risk prediction (combined with ECG abnormalities, sudden increase in blood pressure, and user's subjective symptoms); Rehabilitation monitoring: Postoperative recovery trajectory tracking (comparing historical data with rehabilitation standards).

[0087] See also Figure 2 The present invention also provides a distributed wearable health monitoring method, which includes the following steps: S1. Collecting multi-source physiological signals through the distributed sensing module; S2. Processing the multi-source physiological signals and constructing them into a physiological state tensor through the data preprocessing and tensor construction module; S3. Decomposing the physiological state tensor by the structured tensor fusion engine to extract one or more low-dimensional core features reflecting potential physiological laws; S4. Evaluate the user's health status based on the core features through the health status evaluation module.

[0088] The method of this embodiment can be used to execute the above system embodiment. Its principles and technical effects are similar and will not be described in detail here.

[0089] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A distributed wearable health monitoring system, characterized in that: include: The wearing body of the dial and bracelet; A distributed sensing module is arranged on the wearable body and is used to collect multi-source physiological signals; a data preprocessing and tensor construction module, electrically connected to the distributed sensing module, for processing the multi-source physiological signals and constructing them into physiological state tensors; a structured tensor fusion engine for decomposing the physiological state tensor to extract one or more low-dimensional core features reflecting underlying physiological laws; The health status assessment module assesses the user's health status based on the core features.

2. A distributed wearable health monitoring system according to claim 1, characterized in that: The distributed sensing module includes: at least one first sensing unit disposed on the dial, configured to collect physiological signals related to the user's core circulatory system; at least one second sensing unit disposed on the bracelet, configured to collect physiological signals related to a user's peripheral physiological system; The first sensing unit and the second sensing unit work together to capture multi-dimensional information of different physiological systems.

3. A distributed wearable health monitoring system according to claim 2, characterized in that: The first sensing unit includes at least one selected from the following: A miniature radar sensor unit, used to collect signals reflecting cardiovascular mechanical activity status; A multi-wavelength optical sensing unit for collecting signals reflecting the optical characteristics of blood; Furthermore, the second sensing unit includes at least one selected from the following: Metabolic thermal analysis unit, used to collect temperature field signals reflecting the local metabolic state; The neuromuscular coupling sensing unit is used to collect electromyographic signals that reflect the activity state of the nervous system.

4. A distributed wearable health monitoring system according to claim 1, characterized in that: The physiological state tensor It is a fourth-order tensor whose dimensions are composed of time, mode, feature and physical channel, expressed as: ； in, Represent the lengths of time, modality, feature and physical channel dimensions respectively.

5. A distributed wearable health monitoring system according to claim 1, characterized in that: The structured tensor fusion engine transforms the physiological state tensor by performing Tucker decomposition Decomposition, the decomposition form is expressed as: ； in, is the core tensor, , , , are the factor matrices of time, modality, feature, and channel dimensions respectively.

6. A distributed wearable health monitoring system according to claim 5, characterized in that: The Tucker decomposition is performed by solving a constrained optimization objective that includes physiological prior knowledge. The constrained optimization objective is: ； in, represents the reconstruction error term; is the time smoothing constraint; is the modal sparsity constraint term; is the difference matrix; and are the hyperparameters that control the strength of the corresponding constraints.

7. A distributed wearable health monitoring system according to claim 1, characterized in that: The health status assessment module includes a personalized deep neural network model trained through a self-supervised learning paradigm, which enables the personalized deep neural network model to learn the user's physiological dynamic baseline by performing masking and reconstructing tasks on the physiological state tensor.

8. A distributed wearable health monitoring system according to claim 7, characterized in that: The mask and reconstruction task is to minimize a loss function that only calculates the reconstruction error of the masked position. To achieve this, the loss function is: ； in, is the mask tensor; Reconstructed physiological state tensors for personalized deep neural network models; is the original physiological state tensor, is a tensor related to the physiological state A tensor of exactly the same shape, with all elements containing 1. For Hadamard; are all learnable parameters within the personalized deep neural network model.

9. A distributed wearable health monitoring system according to claim 8, characterized in that: The health status assessment module quantifies the degree to which the current physiological state deviates from the user's physiological dynamic baseline by calculating the residual between the physiological state tensor collected in real time and the reconstructed physiological state tensor output by the personalized deep neural network model.

10. A distributed wearable health monitoring method, applied to the system according to any one of claims 1 to 9, characterized in that: The method comprises the following steps: S1. Collecting multi-source physiological signals through the distributed sensing module; S2. Processing the multi-source physiological signals and constructing them into a physiological state tensor through the data preprocessing and tensor construction module; S3. Decomposing the physiological state tensor by the structured tensor fusion engine to extract one or more low-dimensional core features reflecting potential physiological laws; S4. Evaluate the user's health status based on the core features through the health status evaluation module.

Images