Health monitoring system based on smart watch

Abstract

The invention discloses a health monitoring system based on a smart watch, and relates to the technical field of wearable equipment health data analys.The health monitoring system comprises the steps that multi-source physiological signal data is collected, and a perturbation feature vector sequence is constructed; obtaining a health state evolution trajectory based on the multi-scale time perception map; carrying out risk trend discrimination by using a sparse attention network, and generating an individual health risk weight map; self-adaptive adjustment is carried out by combining the environment temperature, the air pressure and the exercise intensity grade, and an intervention node time sequence table is generated; when the risk score in the continuous time period exceeds a threshold value, health abnormality early warning is triggered and fed back to a display interface of the smart watch end; the method can achieve the dynamic perception and intelligent early warning of the health state of the user, has the advantages of high recognition sensitivity, strong intervention timeliness and good adaptability, and is suitable for the scenes of individual health management, high-risk occupational monitoring and the like.

Description

Technical Field

[0001] This invention relates to the field of wearable device health data analysis technology, and more specifically to a health monitoring system based on a smartwatch. Background Technology

[0002] With the rapid development of wearable device technology, smartwatches have been widely used in the field of health monitoring, such as heart rate monitoring, sleep tracking, and exercise recording. However, existing smartwatch health monitoring methods mostly rely on the static collection and analysis of single physiological signals (such as heart rate and blood oxygen saturation), which cannot achieve early identification of abnormal health risks in the context of cross-interference from multiple dynamic data sources. In particular, when weak physiological fluctuations have not yet reached the clinical diagnostic threshold, there is a lack of effective modeling and trend warning mechanisms.

[0003] More seriously, in high-risk occupations or special groups (such as high-altitude workers, polar researchers, and sub-healthy adolescents), due to factors such as environmental noise, electromagnetic interference, and individual differences, traditional smartwatch monitoring methods cannot continuously model the evolution path of micro-states. This makes it difficult to identify and predict sudden health events (such as fainting, transient arrhythmia, and sudden hypoglycemia) in a timely manner, resulting in significant lag and the risk of missed detection. Summary of the Invention

[0004] The purpose of this invention is to provide a health monitoring system based on a smartwatch to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a health monitoring system based on a smartwatch, comprising: Signal acquisition module: Acquires raw multi-source physiological signal data continuously collected by the user while wearing the smartwatch. The multi-source physiological signal data includes heart rate, heart rate variability, skin conductance response, wrist temperature and triaxial acceleration information. Feature extraction module: Performs feature extraction on the multi-source physiological signal data based on a convolutional temporal fusion network to generate a perturbation feature vector sequence representing the short-term physiological fluctuation trend; Trajectory construction module: Constructs a multi-scale time-aware map based on the perturbation feature vector sequence to obtain the weak health state evolution trajectory within the target time period; Trend discrimination module: scores the health risk trend of weak health status evolution trajectory sequence and generates corresponding individual health risk weight map; Intervention and regulation module: Based on the health risk weight map, and combined with ambient temperature, air pressure and exercise intensity level, adaptive adjustments are made to generate an intervention node time series table; Early warning judgment module: If any risk indicator in the health risk weight map exceeds the preset risk threshold continuously, a health abnormality early warning instruction is triggered, and high-risk time nodes in the time series table are marked. Feedback display module: Feeds back the high-risk time points to the display interface on the smartwatch.

[0006] Preferably, the feature extraction module includes: The acquired multi-source physiological signal data were time-axis aligned and divided into several time segments of equal length using a sliding window mechanism. Time-aligned multi-source physiological signals are input into multi-channel convolutional units in a convolutional temporal fusion network, and local temporal variation features of heart rate, skin conductance response, and acceleration signals are extracted by one-dimensional convolutional kernels. The local time variation features are input into the cross-channel feature interaction layer, and a short-term correlation coupling relationship between different physiological signals is established through the inter-channel weighted fusion mechanism to generate a primary fusion feature sequence. The temporal perturbation sensitivity of the primary fusion feature sequence is calculated to extract the small change patterns of physiological fluctuations within a short time period, and finally a perturbation feature vector sequence representing the trend of short-term physiological fluctuations is generated.

[0007] Preferably, the temporal perturbation sensitivity calculation of the primary fusion feature sequence includes: The primary fusion feature sequence is locally sliced ​​according to a preset time window, and the temporal gradient values ​​of adjacent feature points within each time slice are calculated to obtain the basic gradient sequence of feature changes in a short period of time. The basic gradient sequence is input into the perturbation sensitivity evaluation model. By introducing a time-related correlation measurement method based on weight decay, differential sensitivity weights are assigned to gradient changes at different time positions. Based on the sensitivity weights, the basic gradient sequence is reconstructed to obtain a sensitivity-enhanced feature sequence that emphasizes information about small perturbations.

[0008] Preferably, the trajectory construction module includes: The perturbation feature vector sequence is divided into multiple scale groups according to time order. The scale groups include three levels: short-time scale, medium-time scale, and long-time scale. Within each scale group, a time-recursive embedding algorithm is used to map the perturbation features at each time point into node representations, thereby constructing a feature node set based on the time series. A weighted directed graph is constructed based on the temporal relationship of nodes and a similarity threshold. By performing aggregation and convolution operations on maps at various scales, structural evolutionary features of health status changes at different time levels are extracted, and the health status evolution trajectory within the target time period is constructed.

[0009] Preferably, the trend discrimination module includes: The health status evolution trajectory sequence is input into a time series analysis network composed of multiple sparse attention units, and the key time nodes that are most sensitive to risk changes in the trajectory sequence are extracted using the attention allocation mechanism. At each key time point, the corresponding structural evolution features are extracted, and the differences between them and the time points before and after are measured to construct a risk fluctuation vector. Based on the risk fluctuation vector, it is input into the trained risk discrimination and scoring network. The network adopts a multi-layer fully connected structure and a sigmoid activation function to output a continuous health risk trend score. By combining the score with the time location and weight of the node, an individual health risk weight map is constructed.

[0010] Preferably, the corresponding individual health risk weight map is generated, including: The risk score value of each health status evolution trajectory node is bound to its corresponding timestamp information to form a time-annotated risk score sequence. The risk score sequence is then subjected to time-weighted smoothing. The smoothed scoring sequence is locally aggregated according to a preset time window to generate risk aggregate scores for multiple time periods. The aggregated scores for each time period are used to construct a health risk weight map in the form of a time series, with time as the horizontal axis and aggregated risk value as the vertical axis.

[0011] Preferably, the intervention and regulation module includes: The risk aggregation score for each time period in the health risk weight map is fused with the corresponding ambient temperature, ambient air pressure and exercise intensity level to construct a multidimensional intervention influencing factor vector. Based on the input of the aforementioned intervention impact factor vector into the adaptive intervention adjustment model, the model uses a weighted linear regression method to dynamically learn the weight relationship between environmental factors and risk scores; Based on the adjusted risk score results, an intervention trigger threshold range is set, and time periods exceeding the threshold are identified and marked as candidate intervention nodes; All candidate intervention nodes are deduplicated by time and prioritized, and the final intervention node time series table is generated based on node density and interval rules.

[0012] Preferably, the early warning judgment module includes: Sequence analysis is performed on the risk aggregation scores of multiple consecutive time periods in the health risk weight map to identify time periods in which the risk aggregation scores continuously exceed a preset risk threshold. A risk threshold is set as the baseline value. When the aggregate score of any three or more consecutive time periods exceeds the threshold, it is judged as a potential abnormal health state. The system identifies the time point with the highest risk score in the time series and generates a corresponding health anomaly warning instruction.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention achieves deep modeling of the short-term health status evolution trend of users by constructing a perturbation feature extraction method based on multi-source physiological signals and a multi-scale time-aware graph modeling mechanism. This overcomes the problems of insufficient monitoring accuracy and delayed risk response caused by traditional smartwatches relying on single physiological parameters and static analysis methods. By introducing temporal perturbation sensitivity calculation, sparse attention trend discrimination network, and risk weight graph construction, this invention can accurately identify weak but continuous health risk signals, enabling early warning and dynamic intervention planning for potential health abnormalities, significantly improving the intelligence level and risk prediction capability of wearable health monitoring systems.

[0014] 2. This invention also incorporates external interference factors such as ambient temperature, air pressure, and exercise intensity levels into the intervention and regulation logic. Based on weighted linear modeling and dynamic threshold screening, it generates a personalized intervention node time series table, further enhancing the accuracy and adaptability of the intervention strategy. Furthermore, by feeding high-risk information back to the smartwatch in real time, combined with graphical display and user interaction mechanisms, the immediacy and operability of health risk alerts are ensured, comprehensively improving users' risk perception capabilities and proactive health management levels. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 As shown in this embodiment, a health monitoring system based on a smartwatch includes: Signal acquisition module: Acquires raw multi-source physiological signal data continuously collected by the user while wearing the smartwatch. The multi-source physiological signal data includes heart rate, heart rate variability, skin conductance response, wrist temperature and triaxial acceleration information.

[0019] In this invention, the signal acquisition module is used to continuously and stably acquire raw multi-source physiological signal data while the user is wearing the smartwatch, providing basic data support for subsequent feature extraction and health risk analysis. The multi-source physiological signal data includes, but is not limited to, the following categories: Heart rate (HR) signal: Collected by the photoplethysmography (PPG) sensor built into the smartwatch, it reflects the user's current heart rate and can be used to determine basic physiological states such as tachycardia and bradycardia.

[0020] Heart rate variability (HRV) signal: Based on the acquisition of continuous heart rate signals, statistical change parameters of the heartbeat interval are calculated, such as SDNN, RMSSD, LF / HF ratio, etc., to assess the dynamic regulatory capacity of the autonomic nervous system. It is often used to reflect stress load and fatigue state.

[0021] Electrodermal conductance (EDA) signal: Changes in the skin conductivity of the wrist are obtained through electrodes on the back of the smartwatch, reflecting the level of sympathetic nerve activity. It can be used to identify the user's physiological response under emotional fluctuations such as anxiety and tension.

[0022] Wrist temperature signal: The built-in temperature sensor measures the skin surface temperature of the wearing area in real time to analyze changes in metabolic status, low fever / high fever trends, and help identify problems such as fever or hypothermia.

[0023] Triaxial Accelerometer (ACC): Collects motion data of the user's wrist in the X, Y, and Z axes using a triaxial accelerometer, which is used to analyze dynamic behavioral patterns such as motion state, fall events, and sleep cycles.

[0024] The signal acquisition module can sample data at fixed time intervals (such as once per second or once every 5 seconds) and store historical data within a certain period of time through a set data buffer mechanism to meet the input requirements of subsequent time series feature analysis.

[0025] Feature extraction module: Performs feature extraction on the multi-source physiological signal data based on a convolutional temporal fusion network to generate a perturbation feature vector sequence representing the short-term physiological fluctuation trend.

[0026] The multi-source physiological signal data collected by the smartwatch were preprocessed. First, various signals (heart rate, skin conductance response, triaxial acceleration, etc.) were aligned along a unified time axis. Specifically, with a sampling frequency of fs, low-frequency channels (such as skin conductance response) were time-pasted using linear interpolation, while high-frequency channels were downsampled to align all signals on the same time series. Subsequently, time series were segmented using a fixed-length sliding window, with each window length set to Tw=10 seconds and a sliding step size of Ts=2 seconds, resulting in several time series segments of equal length, which served as the basic units for subsequent feature extraction.

[0027] Each sliding window segment is input into a multi-channel one-dimensional convolutional unit in the constructed convolutional temporal fusion network. A one-dimensional convolutional channel is independently configured for each physiological signal; for example, the heart rate channel uses a kernel size of k=3 and a stride of s=1 to extract local heart rate fluctuation features. The skin conductance response channel and the triaxial acceleration channel use convolutional kernels with different parameter configurations to enhance their respective receptive fields. All the above convolutional units use the ReLU activation function and are normalized using Batch Normalization to output local temporal variation feature tensors for each physiological channel.

[0028] The local temporal variation features extracted from the different physiological channels are input into the cross-channel feature interaction layer to construct short-term dependencies between signals. The interaction layer is implemented using a three-dimensional tensor attention weighted structure, with a defined channel attention matrix. Where C is the number of physiological signal channels, the coupling weight between each channel is calculated using the following formula: ;in i and j represents the embedding vectors of the i-th and j-th channels, respectively. The weights A are applied to the feature tensors of each channel through matrix multiplication to generate a fused primary feature sequence, which serves as the output describing the cooperative change pattern of multiple signals.

[0029] The initial fused feature sequence is further divided into local windows along the time axis, with each time slice having a length of L = 5 seconds. Within each slice, the gradient of feature changes between adjacent time points is calculated, denoted as the temporal gradient sequence G = {g1, g2, ..., gn}, where... , where xt represents the fused feature value at time point t. This gradient sequence is used to reflect the changing trend of minute physiological fluctuations over a short period of time, providing a basis for subsequent perturbation sensitivity assessment.

[0030] The basic gradient sequence is input into a custom perturbation sensitivity assessment model to determine the degree of fluctuation at different times. This model uses a weighted time decay function to calculate the perturbation weight at each time point. Let the weight function be: ; where α is the normalization coefficient, β is the time decay coefficient, n is the current time slice length, t∈[1,n]. This function assigns higher weight to feature changes closer to the current time point and lower weight to earlier fluctuations, thus capturing short-term abnormal changes more sensitively.

[0031] Applying the aforementioned time weight wt to each gradient value in the base gradient sequence, a sensitivity-enhanced feature sequence is constructed: The sequence G′ is then normalized to ensure its values ​​are within the range of [0,1], thus preventing interference from differences in physiological signal amplitudes in subsequent modeling. The final output sensitivity-enhanced feature sequence is the perturbation feature vector sequence, used to describe the user's short-term physiological fluctuation trend within the current time window and as input for subsequent health trajectory modeling.

[0032] Trajectory construction module: Constructs a multi-scale time-aware map based on the perturbation feature vector sequence to obtain the weak health state evolution trajectory within the target time period.

[0033] First, the generated perturbation feature vector sequences are grouped chronologically to construct a multi-scale temporal structure, enhancing the ability to model health status trends at different time levels. Specifically, based on the real-time requirements of the application scenario and the data sampling frequency, time is divided into three scale levels: Short timescale: Represents a continuous time period of 1 to 5 minutes, used to capture rapidly changing short-term physiological responses; Mesochronous scale: Represents a continuous time period of 5 to 20 minutes, used to characterize physiological regulatory changes in a medium cycle; Long timescale: refers to a continuous time period of 20 to 60 minutes, used to depict slower trends in the evolution of physical characteristics.

[0034] Each time scale corresponds to a set of continuous perturbation feature vector subsequences, and subsequent map construction and feature extraction operations will be performed at each scale.

[0035] Within each scale group, a time-recursive embedding algorithm is used to process the perturbation feature vectors at each time point, mapping them to node representations with temporal context information. This algorithm constructs time-sensitive embedding vectors by introducing a time decay coefficient and a historical state propagation method. The specific implementation is as follows: For each time point t, its corresponding perturbation feature vector is used as the initial input; A historical state vector is introduced as a recursive variable, and the degree of fusion between the current feature and the historical state is controlled by a gating structure. At the same time, a time decay factor is introduced, the value of which is calculated based on the interval between the current time point and the previous time point. The larger the interval, the stronger the influence of the current feature on the final embedding. The final output node embedding vector takes into account the current physiological state features and previous temporal information, forming a structured set of feature nodes.

[0036] For each scale's generated set of feature nodes, a graph structure is constructed according to temporal order and similarity relationships. Each feature node is a node in the graph, and directed connections exist between two nodes, established only if the following two conditions are met: The temporal sequence satisfies causality, that is, an earlier point in time points to a later point in time; If the similarity of feature vectors between nodes exceeds a set threshold, it indicates that there is continuity or a trend of change between the two states.

[0037] Similarity is obtained by calculating Euclidean distance or cosine similarity, and a time interval decay factor is introduced to adjust the edge weights. The weight calculation process is as follows: first, the feature similarity between two nodes is standardized, and then multiplied by an exponential function value. This function takes the time interval as the independent variable and the exponential decay coefficient is a constant, indicating that the edge weight decreases as the time interval increases.

[0038] The final constructed graph is a directed graph structure, where the edge weights reflect the trend strength and temporal tightness of feature evolution.

[0039] For the short-, medium-, and long-scale graphs constructed above, graph convolutional neural networks are used for structural feature extraction. The graph at each scale is input into a multi-layer graph convolutional network structure. Each layer of the network aggregates the features of neighboring nodes based on weighted summaries to generate updated node representations. The aggregation process includes the following operations: For each node, iterate through all its neighboring nodes; The features of adjacent nodes are weighted and summed based on the edge weights to obtain the structural context information of the node. The structural information is then nonlinearly combined with the features of the current node (e.g., through linear transformation and activation function processing) to generate the node representation for the next layer.

[0040] After repeating the above process multiple times, the final embedded representation of node features at each scale is output. Subsequently, the node embedding results at the three scales are concatenated or weighted and fused in chronological order to obtain the complete health state evolution trajectory.

[0041] Trend discrimination module: scores the health risk trend of the weak health status evolution trajectory sequence and generates the corresponding individual health risk weight map.

[0042] First, the health status evolution trajectory sequence is input into the constructed time-series analysis network. This network is based on a multi-layer sparse attention unit structure and is used to model the degree of risk correlation at different time points in the trajectory sequence. Each sparse attention unit introduces a restrictive mask to control the attention range to focus only on adjacent nodes within a local time window, thereby improving the model's responsiveness to short-term risk trend changes.

[0043] After training, the network outputs a set of time-related weight coefficients, which represent the importance of each trajectory node in the overall sequence. The nodes are sorted according to their weights, and several high-weight nodes are selected as key time nodes. These nodes are considered major turning points or highly sensitive moments in the trend of health risk fluctuations.

[0044] At each identified key time point, the corresponding structural evolution features are extracted, and a difference analysis is performed between them and the structural features of adjacent time points.

[0045] Let the current critical time node be the t-th node, and its structural feature vector be represented by Ft. Calculate the feature differences with neighboring nodes, using methods such as Euclidean distance or cosine difference as a metric. Combine this difference result with the current node's own features to form a complete risk fluctuation vector. This vector quantifies the degree of state fluctuation around the critical time node and serves as a direct input for risk trend scoring.

[0046] The risk fluctuation vector corresponding to each of the aforementioned key time points is input into the constructed risk discrimination and scoring network. This scoring network is a feedforward neural network composed of multiple fully connected neurons. Its last layer uses the sigmoid activation function, limiting the output value to the range of 0 to 1, representing a continuous probability score of health risk trends. The specific implementation method is as follows: The first layer is a linear fully connected layer with ReLU activation function; The second layer introduces a Dropout layer to avoid overfitting; The third layer is the output layer, which uses the Sigmoid activation function to output a single score value. The network training objective function is a binary cross-entropy loss function, and supervised learning optimization is performed using labeled risk trend samples.

[0047] The closer the score is to 1, the closer the physiological state at the current time point is to a high-risk pattern.

[0048] To transform the risk scoring results into a risk evolution map with a time structure, each score value is first bound to its corresponding time node to form a time-stamped risk score sequence, represented as an ordered pair (ti, si), where ti is the time node and si is the corresponding risk score value.

[0049] The scoring sequence is then subjected to time-weighted smoothing, introducing a time decay coefficient αi. Weights are assigned based on the relative distance of nodes from the current time point, with weights increasing as they approach the current time point. Gaussian or exponential decay functions are commonly used to generate these weights. The weighted scores are obtained as follows: For each rating value si, its smoothing value is All smoothed score values ​​are arranged in chronological order to form a new smoothed risk score sequence.

[0050] Next, the smoothed scoring sequence is divided into multiple time periods, with the length of each time period set according to the actual monitoring frequency (e.g., every 5 minutes is a time period). The average or weighted sum of all scores within each time period is calculated as the risk aggregation score for that time period.

[0051] Finally, the aggregated scores for each time period are arranged chronologically to construct a complete individual health risk weight map. The horizontal axis of this map represents time, and the vertical axis represents the aggregated risk score. The map continuously depicts the evolution trend of a user's health risk within a target time range, visually revealing high-risk periods and providing a basis for intervention decisions.

[0052] Intervention and regulation module: Based on the health risk weight map, and combined with ambient temperature, air pressure and exercise intensity level, adaptive adjustments are made to generate an intervention node time series table.

[0053] First, obtain the aggregated risk score for each time period in the health risk weighting graph. This score reflects the risk trend of the user's overall health status within the current time period. Simultaneously, obtain the environmental parameter data corresponding to that time period, including ambient temperature, ambient air pressure, and the user's current exercise intensity level.

[0054] The risk aggregation score is fused with three types of environmental factors to construct a four-dimensional vector, defined as the intervention impact factor vector. Wherein: The first dimension represents the risk aggregation score; The second dimension represents the ambient temperature value (in degrees Celsius). The third dimension represents the air pressure value (in kilopascals). The fourth dimension represents the exercise intensity level (calculated based on the user's acceleration change and step frequency per unit time, divided into low intensity, medium intensity, and high intensity, corresponding to values ​​1, 2, and 3 respectively).

[0055] Each time period corresponds to an intervention impact factor vector, which is used to comprehensively characterize the interaction between current health risks and external environmental conditions.

[0056] The vectors of the aforementioned intervention influencing factors are input into the adaptive intervention adjustment model. This model employs a weighted linear regression method to learn the degree of influence of different factors on intervention priority and dynamically update the weight relationships between factors.

[0057] Let the vector of intervention impact factors be X = [x1, x2, x3, x4], where x1 is the risk aggregation score and the other three are environmental factors. The intervention priority score is defined as y. Then the model form is: Where w1~w4 are adjustable weight parameters learned during training, and b is the bias term. The model is trained in a supervised manner using historical intervention effect data, and the loss function is optimized to mean squared error to ensure that the model can output intervention priority scores that reflect individual state characteristics.

[0058] After obtaining the intervention priority score for each time period, an intervention trigger threshold range is set to screen time points with high risk and significant environmental impact. This threshold can be set as a dynamic value based on the statistical analysis results of long-term user data, initially set at 0.75 (within the range of 0 to 1), indicating that time periods with scores higher than 0.75 should be given priority for intervention.

[0059] Iterate through all time period scoring results. Any time period with an intervention priority score greater than the threshold is marked as a candidate intervention node, and its corresponding timestamp and score are recorded for subsequent processing.

[0060] For all identified candidate intervention nodes, time deduplication and priority sorting operations are performed to generate the final intervention node time series table.

[0061] First, the execution time is deduplicated. For multiple consecutive high-scoring time periods, only the time period with the highest score is retained as a representative node to avoid frequent interventions that burden users. Then, the time periods are sorted in descending order based on intervention priority scores to ensure that high-risk nodes appear at the top of the time series table.

[0062] Finally, based on the intervention frequency and minimum time interval requirements (e.g., no less than 10 minutes between any two intervention nodes), the nodes are filtered and sorted to form the final intervention node time series table that meets the strategy requirements. This time series table records the user's intervention time point suggestions in the following monitoring period, providing input basis for the subsequent display module and early warning module.

[0063] Early warning judgment module: If any risk indicator in the health risk weight map exceeds the preset risk threshold continuously, a health abnormality early warning instruction is triggered, and the high-risk time node in the time series table is marked.

[0064] First, a time-series scan analysis was performed on the constructed health risk weight map. This map uses time as the horizontal axis and aggregated risk score as the vertical axis to represent the user's health risk trend over different time periods.

[0065] In this step, the risk aggregation scores of each time period in the graph are extracted into an ordered numerical sequence in chronological order, denoted as: S={s1,s2,s3,…,sn}; where si represents the risk aggregation score of the i-th time period, and n is the total number of time periods.

[0066] The system uses a sliding window algorithm to scan sequence S segment by segment to determine whether there are multiple consecutive scores that exceed the risk threshold, which is used for subsequent early warning judgment.

[0067] A threshold T is set for the risk aggregation score, which is used to define the criteria for judging abnormal risks. This threshold can be set to a fixed value based on the statistical results of population data, preferably 0.8. That is, when the aggregated risk score for a certain period of time is greater than 0.8, it indicates that there is a high health risk during that period of time.

[0068] During the sliding window analysis, it is checked whether there are cases where the risk aggregation scores of three or more consecutive time periods all exceed the threshold. These are denoted as the subsequence S′={sk,sk+1,sk+2,...,sk+m} that meet the conditions, where m≥2, and for all sj∈S′, sj>T.

[0069] When a continuous subsequence that meets the above conditions exists, the system determines that the time segment is a potential health abnormality segment and proceeds to the next step of processing.

[0070] Within a series of high-risk segments deemed abnormal, the time point with the highest aggregated risk score within that segment is extracted and designated as the peak time point tmax of the abnormal state, with its corresponding score smax. Using this peak time point as a warning reference, a health anomaly warning instruction is generated, which includes the following information: Anomaly type: Continuous high-risk trend; Time stamp: tmax; Risk level: Classified according to the value of smax, such as between 0.8 and 0.9 is medium risk, and above 0.9 is high risk; Warning and suggestion: Remind users to pay attention to their own status and guide them to take the next step of intervention.

[0071] This warning instruction will be displayed to the user as a high-priority event and can be simultaneously transmitted to the remote health management platform for subsequent intervention arrangements and abnormal record analysis.

[0072] Feedback display module: Feeds back the high-risk time points to the display interface on the smartwatch.

[0073] When the early warning judgment module detects that the risk aggregation score of multiple consecutive time periods exceeds the preset threshold and marks the high-risk time nodes, the system encapsulates the time node and its related data to form a risk event data packet.

[0074] The data packet includes, but is not limited to, the following information fields: Timestamp information: Indicates the time of the current high-risk node; Risk level: Classified as medium risk or high risk based on risk score results; Abnormal type identifiers: such as "continuous increase in risk" or "severe short-term fluctuations"; Intervention suggestion codes: These correspond to subsequent user behavior prompts, such as "Recommend resting", "Recommend drinking water", or "Recommend contacting a doctor".

[0075] The above information will be compressed and packaged in a structured data format to reduce resource consumption during transmission and processing on the smartwatch.

[0076] The encapsulated risk event data packet is sent to the smartwatch via a wireless communication link (such as Bluetooth or Low Energy Wi-Fi). To ensure the timeliness and integrity of information transmission, a frame check and acknowledgment mechanism is adopted. Once the smartwatch successfully receives the complete data packet, it automatically enters the risk information parsing process.

[0077] After receiving and parsing the data packet, the smartwatch launches a dedicated health risk display interface and uses its built-in graphics rendering interface to display information. The displayed content includes, but is not limited to: Current time and risk level; Risk trend icons (such as color gradients, warning icons); Clear risk description text, such as: "Abnormal heart rate fluctuations were continuously detected, risk level: high"; Dynamic prompts, such as: "Please rest as soon as possible and avoid strenuous exercise."

[0078] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A health monitoring system based on a smartwatch, characterized in that: include: Signal acquisition module: Acquires raw multi-source physiological signal data continuously collected by the user while wearing the smartwatch. The multi-source physiological signal data includes heart rate, heart rate variability, skin conductance response, wrist temperature and triaxial acceleration information. Feature extraction module: Performs feature extraction on the multi-source physiological signal data based on a convolutional temporal fusion network to generate a perturbation feature vector sequence representing the short-term physiological fluctuation trend; Trajectory construction module: Constructs a multi-scale time-aware map based on the perturbation feature vector sequence to obtain the weak health state evolution trajectory within the target time period; Trend discrimination module: scores the health risk trend of weak health status evolution trajectory sequence and generates corresponding individual health risk weight map; Intervention and regulation module: Based on the health risk weight map, and combined with ambient temperature, air pressure and exercise intensity level, adaptive adjustments are made to generate an intervention node time series table; Early warning judgment module: If any risk indicator in the health risk weight map exceeds the preset risk threshold continuously, a health abnormality early warning instruction is triggered, and high-risk time nodes in the time series table are marked. Feedback display module: Feeds back the high-risk time points to the display interface on the smartwatch.

2. The health monitoring system based on a smartwatch according to claim 1, characterized in that: The feature extraction module includes: The acquired multi-source physiological signal data were time-axis aligned and divided into several time segments of equal length using a sliding window mechanism. Time-aligned multi-source physiological signals are input into multi-channel convolutional units in a convolutional temporal fusion network, and local temporal variation features of heart rate, skin conductance response, and acceleration signals are extracted by one-dimensional convolutional kernels. The local time variation features are input into the cross-channel feature interaction layer, and a short-term correlation coupling relationship between different physiological signals is established through the inter-channel weighted fusion mechanism to generate a primary fusion feature sequence. The temporal perturbation sensitivity of the primary fusion feature sequence is calculated to extract the small change patterns of physiological fluctuations within a short time period, and finally a perturbation feature vector sequence representing the trend of short-term physiological fluctuations is generated.

3. A health monitoring system based on a smartwatch according to claim 2, characterized in that: The temporal perturbation sensitivity calculation of the primary fusion feature sequence includes: The primary fusion feature sequence is locally sliced ​​according to a preset time window, and the temporal gradient values ​​of adjacent feature points within each time slice are calculated to obtain the basic gradient sequence of feature changes in a short period of time. The basic gradient sequence is input into the perturbation sensitivity evaluation model. By introducing a time-related correlation measurement method based on weight decay, differential sensitivity weights are assigned to gradient changes at different time positions. Based on the sensitivity weights, the basic gradient sequence is reconstructed to obtain a sensitivity-enhanced feature sequence that emphasizes information about small perturbations.

4. A health monitoring system based on a smartwatch according to claim 1, characterized in that: The trajectory construction module includes: The perturbation feature vector sequence is divided into multiple scale groups according to time order. The scale groups include three levels: short-time scale, medium-time scale, and long-time scale. Within each scale group, a time-recursive embedding algorithm is used to map the perturbation features at each time point into node representations, thereby constructing a feature node set based on the time series. A weighted directed graph is constructed based on the temporal relationship of nodes and a similarity threshold. By performing aggregation and convolution operations on maps at various scales, structural evolutionary features of health status changes at different time levels are extracted, and the health status evolution trajectory within the target time period is constructed.

5. A health monitoring system based on a smartwatch according to claim 1, characterized in that: The trend discrimination module includes: The health status evolution trajectory sequence is input into a time series analysis network composed of multiple sparse attention units, and the key time nodes that are most sensitive to risk changes in the trajectory sequence are extracted using the attention allocation mechanism. At each key time point, the corresponding structural evolution features are extracted, and the differences between them and the time points before and after are measured to construct a risk fluctuation vector. Based on the risk fluctuation vector, it is input into the trained risk discrimination and scoring network. The network adopts a multi-layer fully connected structure and a sigmoid activation function to output a continuous health risk trend score. By combining the score with the time location and weight of the node, an individual health risk weight map is constructed.

6. A health monitoring system based on a smartwatch according to claim 5, characterized in that: Generate the corresponding individual health risk weight map, including: The risk score value of each health status evolution trajectory node is bound to its corresponding timestamp information to form a time-annotated risk score sequence. The risk score sequence is then subjected to time-weighted smoothing. The smoothed scoring sequence is locally aggregated according to a preset time window to generate risk aggregate scores for multiple time periods. The aggregated scores for each time period are used to construct a health risk weight map in the form of a time series, with time as the horizontal axis and aggregated risk value as the vertical axis.

7. A health monitoring system based on a smartwatch according to claim 1, characterized in that: The intervention and regulation module includes: The risk aggregation score for each time period in the health risk weight map is fused with the corresponding ambient temperature, ambient air pressure and exercise intensity level to construct a multidimensional intervention influencing factor vector. Based on the input of the aforementioned intervention impact factor vector into the adaptive intervention adjustment model, the model uses a weighted linear regression method to dynamically learn the weight relationship between environmental factors and risk scores; Based on the adjusted risk score results, an intervention trigger threshold range is set, and time periods exceeding the threshold are identified and marked as candidate intervention nodes; All candidate intervention nodes are deduplicated by time and prioritized, and the final intervention node time series table is generated based on node density and interval rules.

8. A health monitoring system based on a smartwatch according to claim 1, characterized in that: The early warning judgment module includes: Sequence analysis is performed on the risk aggregation scores of multiple consecutive time periods in the health risk weight map to identify time periods in which the risk aggregation scores continuously exceed a preset risk threshold. A risk threshold is set as the baseline value. When the aggregate score of any three or more consecutive time periods exceeds the threshold, it is judged as a potential abnormal health state. The system identifies the time point with the highest risk score in the time series and generates a corresponding health anomaly warning instruction.

Images