A method, system, and device for dynamic fusion of multi-scale health time-series features

Abstract

This invention provides a method, system, and apparatus for dynamic fusion of multi-scale health time-series features, relating to the field of dynamic feature fusion technology. The method includes: collecting multi-source health data from subjects, resampling it onto a time grid to obtain aligned data; extracting single-point feature vectors at each time point in the time grid to form a feature sequence; weighting the single-point feature vectors within a time window according to time decay weights, using a time window as the decreasing unit; calculating the moving average feature across multiple time scales; inputting the moving average feature into an attention scoring network to calculate the relevance score of each time scale within the current time window, and obtaining attention weights through normalization; and weighted fusion of the moving average feature to obtain a dynamic feature vector. This solution can adaptively adjust the focus on short-term or long-term information according to different users, different disease stages, and different scenarios, improving the model's generalization ability.

Description

Technical Field

[0001] This invention relates to the field of feature dynamic fusion technology, and in particular to a method, system and apparatus for dynamic fusion of multi-scale health time-series features. Background Technology

[0002] Currently, with the widespread deployment of wearable devices, home medical equipment, and remote monitoring systems, users' multi-source health data, such as heart rate, blood pressure, blood oxygen, body temperature, respiration, activity level, sleep status, and environmental parameters, can be collected continuously over a long period. However, existing methods for home and community health monitoring still have significant shortcomings:

[0003] Different sampling frequencies and misaligned timestamps among various devices, along with missed sampling, outliers, and drift during the acquisition process, result in highly heterogeneous raw multi-source data on the timeline, making it difficult to form a unified time-series input. Existing monitoring methods are mostly based on single-point data or a single time window, and perform simple statistics. For example, they only analyze short-term heart rate fluctuations or only use long-term average blood pressure to assess chronic disease risk, ignoring the combined role of short-term fluctuations, medium-term changes and long-term baseline in health assessment. Most existing monitoring methods use simple averaging or fixed weighting to splice or merge features from different time periods and scales. They lack adaptability. When users’ lifestyles, routines or disease stages change, fixed weights often fail to reflect new health risk patterns. Existing models have high computational costs for long-term series, making it difficult to achieve real-time inference on edge devices. They also often struggle to interpret "which time period and which time scale the information refers to," resulting in current risk assessment results that are not conducive to being understood and adopted by doctors and users.

[0004] Therefore, there is a need for a method that can process multi-source health time-series data on a unified time axis, simultaneously consider information from multiple time scales, and adaptively fuse multi-scale features using an attention mechanism, so as to simultaneously solve the problems of real-time performance and business interpretability. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and apparatus for dynamic fusion of multi-scale health time-series features to solve at least one of the aforementioned technical problems in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for dynamic fusion of multi-scale health time-series features, comprising the following steps: Step 1: Collect multi-source health data (such as heart rate, blood oxygen, activity level, etc.) from the subjects and resample them uniformly onto a preset time grid to obtain aligned data with a consistent time axis. This maps irregular timestamps and multi-frequency raw values ​​onto equally spaced time grids, solving the problems of different sampling frequencies and timestamp misalignment between different sampling devices (such as smart bracelets, blood pressure monitors, and blood glucose meters), laying a solid foundation for subsequent processing.

[0007] In one feasible implementation, the specific formula for aligning the data in step 1 includes: ; in, Indicators of health In actual time The original value at that time; Indicates to After resampling, in the time grid... Time points The value on; Indicates the time grid interval, such as 5s, 30s, etc.; This represents the resampling operator, used for... The original values ​​within the interval are binned (e.g., aggregating multiple points), missing values ​​are filled (e.g., forward filling LOCF, interpolation), and anomalies are removed or corrected, thereby solving outliers such as missed sampling and jitter, and obtaining a regular time series matrix.

[0008] Step 2: At each time point of the time grid, extract the single-point feature vector of the health data to form a feature sequence with a unified structure, so that subsequent weighting, fusion and modeling can be performed in the time dimension. This solves the problems of high noise in the original data, inconsistent physical quantity dimensions and poor decision reference effect, and obtains feature indicators that are more sensitive to health status and have specific medical / behavioral significance.

[0009] In one feasible implementation, the specific expression for the single-point feature vector in step 2 includes: ; in, Indicates the first Time points The multi-source original index vector at the location, and has: ;in, This represents the total number of health indicators, i.e. The dimension; This represents a feature extraction function used to map raw indicators to higher-level statistical and spectral features, which can include time-domain statistical features, frequency-domain features, and event / state features. Indicates in The single-point feature vector obtained at that time, and has , express The dimension; This indicates the number of time points currently being considered (e.g., the number of points within a time window).

[0010] Step 3: Using a fixed-length time window as the decreasing unit, the single-point feature vectors within the time window are weighted and averaged according to the time decay weight to obtain a window-level fused feature vector (highlighting recent information). This preserves recent information, suppresses older information and single-point noise, and reflects the design idea that data closer to the current moment is more important, providing a foundation for subsequent multi-scale analysis.

[0011] In one feasible implementation, the specific expression for fusing the feature vectors in step 3 includes: ; in, Indicates the first The set of time point indices contained in a time window can be represented as: ; express The time decay weight can be expressed as: ;in, This represents a monotonically decreasing function; For example: ;in, This represents the decay coefficient, used to control the rate at which historical information is "forgotten"; in, express Distance to The time at the end of a time window; Indicates the first The fused feature vector of each time window, that is, the "time-weighted center" of that time window in the feature space, and has ; This means summing all single-point feature vectors within the time window according to their temporal importance. This indicates that the weights are normalized to obtain a weighted average result.

[0012] Step 4: Based on the fused feature vector, calculate the moving average feature at multiple time scales to represent the health status at multiple scales, thereby taking into account both short-term fluctuations and medium- to long-term trends.

[0013] In one feasible implementation, the specific expression for the moving average characteristic in step 4 includes: ; in, Indicates the first The fused feature vector of each time window; Indicates the first At the first time scale, with the first The time window ends at a certain time window, and the sliding window length is [missing information]. The moving average characteristic, and has Time Scale Index The range of values ​​for is: ;in, Indicates the total number of time scales; For example, the time scale may include short-term, medium-term, and long-term scales, and correspondingly It can be: Short-term scale: =6 (last 30 minutes); Mesoscale: =24 (last 2 hours); Long-term scale: =288 (last day).

[0014] In one feasible implementation, the moving average feature in step 4 can also be calculated (efficiently) using a sweeping (sliding) method, with the specific formula including: ; in, Indicates the first At the first time scale, with the first The time window ends at a certain time window, and the sliding window length is [missing information]. The moving average characteristic is used for recursion; Less than the The length between the first and second time windows is determined to ensure the index is effective; Indicates the first Subtract one time window The fused feature vector of the remaining time window, i.e., the oldest window that was moved out of the sliding window during sliding update (the... The fusion feature vectors (within a time window); thus, through "new entry" Old The incremental update strategy enables efficient recursive calculation of the same variable.

[0015] Step 5: Input the moving average features of different time scales into the attention scoring network, calculate the relevance score of each time scale in the current time window, and obtain the attention weight of each time scale through softmax normalization, so as to automatically learn "which time scale should be paid more attention to in the current situation" in order to differentiate between acute event detection and chronic disease risk assessment.

[0016] In one feasible implementation, the specific expression for the attention weight in step 5 includes: ; in, The attention scoring function can be implemented using one or more feedforward neural networks to... The relevance score is mapped to a scalar. Represents an exponential function; Indicates the first The time scale for the first Attention weights are assigned to each time window, thereby transforming the relevance score into a probability distribution. The larger it is, the more likely it is to be the first The time scale for the first The more important the time window, and the more important it is: .

[0017] Step 6: By using attention weights, the moving average features at different time scales are weighted and fused to obtain a dynamic feature vector, so as to simultaneously encode short-term fluctuations, medium-term changes and long-term trends in a unified feature vector, which is beneficial for subsequent use.

[0018] In one feasible implementation, the specific calculation formula for the dynamic feature vector in step 6 includes: ; in, This represents a dynamic feature vector, which is a comprehensive representation obtained by dynamically selecting the time scale based on the current situation, and has the following properties: .

[0019] In one feasible implementation, the multi-scale health time-series feature dynamic fusion method further includes step 7, which evaluates the current risk based on the dynamic feature vector through a health risk assessment model to obtain the evaluation result, thereby combining the multi-scale attention weight to provide an explanation of the current risk at the time scale and time period levels, so as to generate early warning and intervention suggestion information.

[0020] In one feasible implementation, the specific expression of the health risk assessment model in step 7 includes: ; in, Indicates the first The assessment results corresponding to each time window, such as the risk probability in the 0-1 range or risk levels such as "low risk / medium risk / high risk"; This represents the output layer weight matrix, used to map feature vectors to a score or category space; Indicates the output layer bias term; This represents the activation function. Specifically: for binary risk probabilities, the Sigmoid function can be selected; for multi-level risk stratification, the Softmax function can be selected. In this way, A linear transformation is performed to obtain one or more output dimensions, and then... Mapping to the probability or rating space, we get It can be directly used in practical application scenarios such as threshold judgment, alarm triggering, or hierarchical management.

[0021] In one feasible implementation, in step 7, when it is necessary... When the value is the risk probability, Compared with the first warning threshold When comparing, ≥ At that time, an early warning message will be generated.

[0022] In one feasible implementation, in step 7, when it is necessary... When the value is the risk level, Compared with the first warning threshold and the second warning threshold By comparing these values, the risk level of the current time window can be directly determined. when ≥ At that time, high-risk level information is generated; when ≤ < At that time, medium-risk level information is generated; when < At that time, low-risk level information is generated.

[0023] In one feasible implementation, the multi-scale health time-series feature dynamic fusion method further includes step 8: when the evaluation result reaches the preset intervention action trigger level, a corresponding intervention action is generated; otherwise, the evaluation result is simply stored and its trend is tracked.

[0024] The trend tracking includes: recording and storing the timestamp of each evaluation result; performing sliding statistical analysis on the evaluation results within a preset time span to obtain the trend direction, rate of change, and volatility; and generating tracking strategies such as retesting, increasing monitoring frequency, or upgrading early warning levels based on trend triggering conditions such as continuous exceeding of thresholds, increase in magnitude, or abnormal fluctuations.

[0025] Secondly, based on the same inventive concept, this application also provides a multi-scale health time-series feature dynamic fusion system, including a data acquisition module, a data processing module and a result generation module; The data acquisition module is used to collect multi-source health data of the tested personnel; The data processing module includes a preprocessing unit, a fusion feature vector unit, a moving average feature unit, an attention weight unit, and a dynamic feature vector unit. The preprocessing unit is used to uniformly resample multi-source health data onto a preset time grid to obtain aligned data with a consistent time axis; at each time point of the time grid, single-point feature vectors of health data are extracted to form a feature sequence with a unified structure. The fusion feature vector unit uses a fixed-length time window as the decreasing unit, and performs a weighted average of the single-point feature vectors within the time window according to the time decay weight to obtain the window-level fusion feature vector. The moving average feature unit calculates the moving average features at multiple time scales based on the fused feature vector; The attention weight unit inputs the moving average features of different time scales into the attention scoring network, calculates the relevance score of each time scale in the current time window, and obtains the attention weight of each time scale through softmax normalization. The dynamic feature vector unit obtains a dynamic feature vector by weighting and fusing moving average features at different time scales through attention weights. The result generation module is used to send the results generated by the data processing module to external systems.

[0026] In one feasible implementation, the data processing module further includes an evaluation unit, which, based on dynamic feature vectors, is used to perform an evaluation through a health risk assessment model to obtain an evaluation result.

[0027] In one feasible implementation, the data processing module further includes an intervention unit, which generates a corresponding intervention action when the evaluation result reaches a preset intervention action trigger level; otherwise, it only stores the evaluation result and tracks its trend.

[0028] Thirdly, based on the same inventive concept, this application also provides a multi-scale health time-series feature dynamic fusion device, including a processor, a memory, and a bus. The memory stores instructions and data that can be read by the processor. The processor is used to call the instructions and data in the memory to execute the multi-scale health time-series feature dynamic fusion method as described above. The bus connects the various functional components for transmitting information.

[0029] In one feasible implementation, the multi-scale health time-series feature dynamic fusion device further includes a smart bracelet for collecting the heart rate, blood oxygen, and activity level of the person being tested.

[0030] In one feasible implementation, the multi-scale health time-series feature dynamic fusion device further includes a blood pressure monitor for collecting the blood pressure value of the person being tested.

[0031] In one feasible implementation, the multi-scale health time-series feature dynamic fusion device further includes a blood glucose meter for collecting blood glucose values ​​of the tested person.

[0032] By adopting the above technical solution, the present invention has the following beneficial effects: This invention provides a method, system, and device for dynamic fusion of multi-scale health time-series features. By resampling and time alignment, it unifies multi-source health data from different devices, with different sampling frequencies, and containing missing data onto the same time grid, providing regular and clean time-series input for subsequent feature extraction and modeling. By performing time-decay weighted fusion of single-point features within a time window, both the temporal continuity within the window and the importance of recent data are considered. Sliding window features are constructed across multiple time scales, including short-term, medium-term, and long-term, enabling the model to simultaneously perceive short-term fluctuations and long-term trends. By introducing a multi-scale attention mechanism, the weights of features at different time scales are automatically learned, eliminating the reliance on human experience to set fixed ratios. This allows for adaptive adjustment of attention to short-term or long-term information based on different users, disease stages, and scenarios, thereby improving the model's generalization ability. In the process of constructing multi-scale sliding window features, the sweeping (sliding recursion) algorithm is used for incremental updates. Historical statistics are updated only by the difference between the "newly entered window" and the "sliding out window", which significantly reduces the computational complexity. This allows the solution to run in real time on edge devices or terminals with limited resources. On common embedded processors or home gateways, the evaluation time for a single time window is within about 100-300ms, and the end-to-end evaluation latency is less than 3-5s. By combining multi-scale attention weights and the contribution of features at each time scale, it is also possible to explain which time period and time scale the current risk score is mainly affected by, providing doctors and users with a more visualized and in-depth basis for decision-making, and improving the understandability and trustworthiness of the results.

[0033] This solution can also improve the accuracy of health risk identification, reduce false alarms and false negatives, and enhance the overall effectiveness of the home health monitoring system while ensuring real-time performance. Through experimental comparison, compared with methods that use only a single time scale or fixed weighted fusion, the accuracy of health event identification in this scheme is improved from about 85% to about 92% (an overall improvement of about 7 percentage points). For events such as abnormal heart rate and abnormal breathing, the average early warning time has been extended from about 10 minutes to about 25 minutes, an increase of about 15 minutes, which is significantly beneficial for users and doctors to intervene in advance. The overall false alarm rate was reduced by about 30% and the false negative rate was reduced by about 20% compared with the baseline method. While maintaining or improving the recognition rate, it effectively reduced the interference burden on users and medical staff.

[0034] In simulated follow-up and questionnaire assessments, the risk report of this protocol showed an average improvement of about 20-30% in subjective scores such as "ease of understanding of results" and "value for decision support" compared to traditional black-box models, which helps to make personalized adjustments and interventions based on clinical experience and lifestyle. Attached Figure Description

[0035] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 A flowchart of a multi-scale health time-series feature dynamic fusion method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the sliding window provided in an embodiment of the present invention; Figure 3 Flowchart of another multi-scale health time-series feature dynamic fusion method provided in this embodiment of the invention; Figure 4 Flowchart of another method for dynamic fusion of multi-scale health time-series features provided in this embodiment of the invention; Figure 5 A diagram of a multi-scale health time-series feature dynamic fusion system provided in an embodiment of the present invention; Figure 6This is a diagram of another multi-scale health time-series feature dynamic fusion system provided in an embodiment of the present invention; Figure 7 This is another diagram of a multi-scale health time-series feature dynamic fusion system provided in the embodiments of the present invention. Detailed Implementation

[0037] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0038] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0039] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0040] The present invention will be further explained below with reference to specific embodiments.

[0041] It should also be noted that the specific embodiments or implementation methods described below are a series of optimized settings listed by the present invention to further explain the specific content of the invention, and these settings can be combined or used in conjunction with each other.

[0042] Example 1: like Figure 1 As shown in the figure, the multi-scale health time-series feature dynamic fusion method provided in this embodiment includes the following steps: Step 1: Collect multi-source health data (such as heart rate, blood oxygen, activity level, etc.) from the subjects and resample them uniformly onto a preset time grid to obtain aligned data with a consistent time axis. This maps irregular timestamps and multi-frequency raw values ​​onto equally spaced time grids, solving the problems of different sampling frequencies and timestamp misalignment between different sampling devices (such as smart bracelets, blood pressure monitors, and blood glucose meters), laying a solid foundation for subsequent processing.

[0043] Furthermore, the specific formula for aligning the data in step 1 includes: ; in, Indicators of health In actual time The original value at that time; Indicates to After resampling, in the time grid... Time points The value on; Indicates the time grid interval, such as 5s, 30s, etc.; This represents the resampling operator, used for... The original values ​​within the interval are binned (e.g., aggregating multiple points), missing values ​​are filled (e.g., forward filling LOCF, interpolation), and anomalies are removed or corrected, thereby solving outliers such as missed sampling and jitter, and obtaining a regular time series matrix.

[0044] Step 2: At each time point of the time grid, extract the single-point feature vector of the health data to form a feature sequence with a unified structure, so that subsequent weighting, fusion and modeling can be performed in the time dimension. This solves the problems of high noise in the original data, inconsistent physical quantity dimensions and poor decision reference effect, and obtains feature indicators that are more sensitive to health status and have specific medical / behavioral significance.

[0045] Furthermore, the specific expression for the single-point feature vector in step 2 includes: ; in, Indicates the first Time points The multi-source original index vector at the location, and has: ;in, This represents the total number of health indicators, i.e. The dimension; This represents a feature extraction function used to map raw indicators to higher-level statistical and spectral features, which may include time-domain statistical features (such as short-time mean, variance, slope, and RMSSD), frequency-domain features (such as power in a specific frequency band, dominant frequency, and bandpass energy), and event / state features (such as encoding sleep stages, activity types, and medication events). Indicates in The single-point feature vector obtained at that time, and has , express The dimension; This indicates the number of time points currently being considered (e.g., the number of points within a time window).

[0046] Step 3: Using a fixed-length time window as the decreasing unit, the single-point feature vectors within the time window are weighted and averaged according to the time decay weight to obtain a window-level fused feature vector (highlighting recent information). This preserves recent information, suppresses older information and single-point noise, and reflects the design idea that data closer to the current moment is more important, providing a foundation for subsequent multi-scale analysis.

[0047] Furthermore, the specific expression for fusing the feature vectors in step 3 includes: ; in, Indicates the first The set of time point indices contained in a time window can be represented as: ; express The time decay weight can be expressed as: ;in, This represents a monotonically decreasing function; For example: ;in, This represents the decay coefficient, used to control the rate at which historical information is "forgotten"; in, express Distance to The time at the end of a time window; Indicates the first The fused feature vector of each time window, that is, the "time-weighted center" of that time window in the feature space, and has ; This means summing all single-point feature vectors within the time window according to their temporal importance. This indicates that the weights are normalized to obtain a weighted average result.

[0048] Step 4: Based on the fused feature vector, calculate the moving average feature at multiple time scales to represent the health status at multiple scales, thereby taking into account both short-term fluctuations (such as a sudden increase in heart rate) and medium- to long-term trends (such as long-term hypertension).

[0049] Furthermore, such as Figure 2 As shown, the specific expression for the moving average characteristic in step 4 includes: ; in, Indicates the first The fused feature vector of each time window; Indicates the first At the first time scale, with the first The time window ends at a certain time window, and the sliding window length is [missing information]. The moving average characteristic, and has Time Scale Index The range of values ​​for is: ;in, Indicates the total number of time scales; For example, the time scale may include short-term, medium-term, and long-term scales, and correspondingly It can be: Short-term scale: =6 (last 30 minutes); Mesoscale: =24 (last 2 hours); Long-term scale: =288 (last day).

[0050] Furthermore, the moving average feature in step 4 can also be calculated (efficiently) using a sweeping (sliding) method, with specific formulas including: ; in, Indicates the first At the first time scale, with the first The time window ends at a certain time window, and the sliding window length is [missing information]. The moving average characteristic is used for recursion; Less than the The length between the first and second time windows is determined to ensure the index is effective; Indicates the first Subtract one time window The fused feature vector of the remaining time window, i.e., the oldest window that was moved out of the sliding window during sliding update (the... The fusion feature vectors (within a time window); thus, through "new entry" Old The incremental update strategy enables efficient recursive calculation of the same variable.

[0051] Step 5: Input the moving average features of different time scales into the attention scoring network, calculate the relevance score of each time scale in the current time window, and obtain the attention weight of each time scale through softmax normalization, so as to automatically learn "which time scale should be paid more attention to in the current situation" in order to differentiate between acute event detection and chronic disease risk assessment.

[0052] Furthermore, the specific expression for the attention weight in step 5 includes: ; in, The attention scoring function can be implemented using one or more feedforward neural networks to... The relevance score is mapped to a scalar. Represents an exponential function; Indicates the first The time scale for the first Attention weights are assigned to each time window, thereby transforming the relevance score into a probability distribution. The larger it is, the more likely it is to be the first The time scale for the first The more important the time window, and the more important it is: .

[0053] Step 6: By using attention weights, the moving average features at different time scales are weighted and fused to obtain a dynamic feature vector, so as to simultaneously encode short-term fluctuations, medium-term changes and long-term trends in a unified feature vector, which is beneficial for subsequent use.

[0054] Furthermore, the specific calculation formula for the dynamic feature vector in step 6 includes: ; in, This represents a dynamic feature vector, which is a comprehensive representation obtained by dynamically selecting the time scale based on the current situation, and has the following properties: .

[0055] Furthermore, such as Figure 3As shown, the multi-scale health time-series feature dynamic fusion method further includes step 7, which evaluates the current risk based on the dynamic feature vector through a health risk assessment model to obtain the evaluation result. Then, combined with the multi-scale attention weight, it provides an explanation of the current risk at the time scale and time period levels, so as to generate early warning and intervention suggestion information.

[0056] Furthermore, the specific expression of the health risk assessment model in step 7 includes: ; in, Indicates the first The assessment results corresponding to each time window, such as the risk probability in the 0-1 range or risk levels such as "low risk / medium risk / high risk"; This represents the output layer weight matrix, used to map feature vectors to a score or category space; Indicates the output layer bias term; This represents the activation function. Specifically: for binary risk probabilities, the Sigmoid function can be selected; for multi-level risk stratification, the Softmax function can be selected. In this way, A linear transformation is performed to obtain one or more output dimensions, and then... Mapping to the probability or rating space, we get It can be directly used in practical application scenarios such as threshold judgment, alarm triggering or hierarchical management; For example, when When choosing the Sigmoid function, an output of 0.82 indicates high risk; when... When choosing the Softmax function, the probability of the "high-risk" level is used as the basis for subsequent decisions.

[0057] Furthermore, in step 7, when it is necessary... When the value is the risk probability, Compared with the first warning threshold When comparing, ≥ At that time, an early warning message will be generated.

[0058] Furthermore, in step 7, when it is necessary... When the value (Score) is the risk level, Compared with the first warning threshold and the second warning threshold By comparing these values, the risk level of the current time window can be directly determined. when ≥ At that time, high-risk level information is generated; when ≤ < At that time, medium-risk level information is generated; when < At that time, low-risk level information is generated.

[0059] Furthermore, such as Figure 4 As shown, the multi-scale health time-series feature dynamic fusion method further includes step 8, where when the evaluation result reaches the preset intervention action trigger level, a corresponding intervention action is generated (such as pushing medical advice information, notifying family members or doctors, adjusting the monitoring frequency, etc.); otherwise, the evaluation result is only stored and its trend is tracked.

[0060] Example 2: This embodiment applies the method from Embodiment 1 to a comparative experiment on health event early warning. The specific experimental conditions, process, and results are as follows: In a home-based health monitoring scenario, multi-source health data (including heart rate, respiratory rate, blood oxygen saturation, body temperature, and activity / sleep status) were collected from 100 participants. Each participant was monitored continuously for 30 days, with a 1-minute time window as the basis for data extraction and fusion. In the set of sliding window lengths across multiple time scales Next, calculate Furthermore, attention weights at different time scales are dynamically fused through an attention mechanism.

[0061] The events labeled include those such as abnormal heart rate and abnormal breathing. Start and end times of these events are labeled using conventional methods such as doctor annotation, rule-based annotation, and log alignment. The dataset is divided into training and testing sets according to the subject dimension, with a ratio of 8:2. This division is repeated three times, and the average is taken after each division.

[0062] Baseline comparison methods include: Baseline-1 Single Time Scale Method (using only single time scale features). Baseline-2 fixed-weighted fusion method (fixed-weight multi-scale fusion).

[0063] Under the same classification / prediction model (health risk assessment model) structure and training parameters, only the fusion strategy is replaced to ensure fair comparison.

[0064] Evaluation metrics include event recognition accuracy (Acc), false alarm rate (FAR), false negative rate (MR), and average early warning time. (For each event, calculate the difference between the time of the first warning trigger and the time of the event occurrence, and take the average).

[0065] The experimental results are shown in Table 1: Table 1

[0066] Therefore, it is evident that this solution has better early warning accuracy and stronger early warning capabilities.

[0067] Example 3: like Figure 5 As shown, this embodiment provides a multi-scale health time-series feature dynamic fusion system, including a data acquisition module, a data processing module, and a result generation module; The data acquisition module is used to collect multi-source health data of the tested personnel; The data processing module includes a preprocessing unit, a fusion feature vector unit, a moving average feature unit, an attention weight unit, and a dynamic feature vector unit. The preprocessing unit is used to uniformly resample multi-source health data onto a preset time grid to obtain aligned data with a consistent time axis; at each time point of the time grid, single-point feature vectors of health data are extracted to form a feature sequence with a unified structure. The fusion feature vector unit uses a fixed-length time window as the decreasing unit, and performs a weighted average of the single-point feature vectors within the time window according to the time decay weight to obtain the window-level fusion feature vector. The moving average feature unit calculates the moving average features at multiple time scales based on the fused feature vector; The attention weight unit inputs the moving average features of different time scales into the attention scoring network, calculates the relevance score of each time scale in the current time window, and obtains the attention weight of each time scale through softmax normalization. The dynamic feature vector unit obtains a dynamic feature vector by weighting and fusing moving average features at different time scales through attention weights. The result generation module is used to send the results generated by the data processing module to external systems.

[0068] Furthermore, such as Figure 6 As shown, the data processing module also includes an evaluation unit, which, based on dynamic feature vectors, is used to perform evaluations through a health risk assessment model to obtain evaluation results.

[0069] Furthermore, such as Figure 7 As shown, the data processing module also includes an intervention unit, which generates a corresponding intervention action when the evaluation result reaches a preset intervention action trigger level; otherwise, it only stores the evaluation result and tracks its trend.

[0070] Example 4: This embodiment provides a multi-scale health time-series feature dynamic fusion device, including a processor, a memory, and a bus. The memory stores instructions and data that can be read by the processor. The processor is used to call the instructions and data in the memory to execute the multi-scale health time-series feature dynamic fusion method as described above. The bus connects the various functional components for information transmission.

[0071] Furthermore, the multi-scale health time-series feature dynamic fusion device also includes a smart bracelet for collecting the heart rate, blood oxygen, and activity level of the tested person.

[0072] Furthermore, the multi-scale health time-series feature dynamic fusion device also includes a blood pressure monitor for collecting the blood pressure values ​​of the person being tested.

[0073] Furthermore, the multi-scale health time-series feature dynamic fusion device also includes a blood glucose meter for collecting blood glucose values ​​of the tested individuals.

[0074] In another embodiment, this solution can also be implemented using an integrated device, which may include corresponding modules that perform one or more steps in the various embodiments described above. A module may be one or more hardware modules specifically configured to perform the corresponding step, or implemented by a processor configured to perform the corresponding step, or stored in a computer-readable medium for implementation by a processor, or implemented through some combination thereof.

[0075] The processor executes the various methods and processes described above. For example, the method implementations in this scheme can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some implementations, part or all of the software program can be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above can be performed. Alternatively, in other implementations, the processor can be configured to execute one of the methods described above by any other suitable means (e.g., by means of firmware).

[0076] This device can be implemented using a bus architecture. A bus architecture can include any number of interconnect buses and bridges, depending on the specific application of the hardware and overall design constraints. The bus connects various circuits, including one or more processors, memory, and / or hardware modules. The bus can also connect various other circuits such as peripherals, voltage regulators, power management circuitry, external antennas, etc.

[0077] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Component (EISA) buses, etc. Buses can be divided into address buses, data buses, control buses, etc.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic fusion of multi-scale health time-series features, characterized in that, include: Step 1: Collect multi-source health data of the subjects and resample them uniformly to a preset time grid to obtain aligned data with a consistent time axis; Step 2: At each time point in the time grid, extract single-point feature vectors of health data to form a feature sequence with a uniform structure; Step 3: Using a fixed-length time window as the decreasing unit, the single-point feature vectors within the time window are weighted and averaged according to the time decay weight to obtain the window-level fused feature vector. Step 4: Based on the fused feature vector, calculate the moving average feature across multiple time scales; Step 5: Input the moving average features at different time scales into the attention scoring network, calculate the relevance score of each time scale in the current time window, and obtain the attention weight of each time scale by normalization through softmax. Step 6: By using attention weights, the moving average features at different time scales are weighted and fused to obtain a dynamic feature vector.

2. The dynamic fusion method according to claim 1, characterized in that, The specific formula for aligning the data in step 1 includes: ； in, Indicators of health In actual time The original value at that time; Indicates to After resampling, in the time grid... Time points The value on; Indicates the time grid interval; This represents the resampling operator.

3. The dynamic fusion method according to claim 2, characterized in that, The specific expression for the single-point feature vector in step 2 includes: ； in, Indicates the first Time points The multi-source original index vector at the location, and has: ;in, This represents the total number of health indicators; Represents the feature extraction function; Indicates in The single-point feature vector obtained at that time, and has , express The dimension; This indicates the number of time points currently being considered.

4. The dynamic fusion method according to claim 3, characterized in that, The specific expression for fusing the feature vectors in step 3 includes: ； in, Indicates the first The set of time point indices contained in each time window is specifically represented as follows: ； express The time decay weight is specifically expressed as follows: ;in, This represents a monotonically decreasing function; express Distance to The time at the end of a time window; Indicates the first The fused feature vectors of each time window, and have .

5. The dynamic fusion method according to claim 4, characterized in that, The specific expression for the moving average characteristic in step 4 includes: ； in, Indicates the first The fused feature vector of each time window; Indicates the first At the first time scale, with the first The time window ends at a certain time window, and the sliding window length is [missing information]. The moving average characteristic, and has .

6. The dynamic fusion method according to claim 5, characterized in that, The moving average feature in step 4 is calculated using the sweeping method, and the specific formula includes: ； in, Indicates the first At the first time scale, with the first The time window ends at a certain time window, and the sliding window length is [missing information]. The characteristics of the moving average; Less than the The length between the first time window and the second time window; Indicates the first Subtract one time window The fused feature vector of the remaining time window.

7. The dynamic fusion method according to claim 6, characterized in that, The specific expression for the attention weight in step 5 includes: ； in, This represents the attention scoring function; Represents an exponential function; Indicates the first The time scale for the first Attention weights for each time window, Let represent the total number of time scales, and we have: 。 8. The dynamic fusion method according to claim 7, characterized in that, The specific calculation formula for the dynamic feature vector in step 6 includes: ； in, Let represent a dynamic feature vector, and we have: .

9. A multi-scale health time-series feature dynamic fusion system, characterized in that, It includes a data acquisition module, a data processing module, and a result generation module; The data acquisition module is used to collect multi-source health data of the tested personnel; The data processing module includes a preprocessing unit, a fusion feature vector unit, a moving average feature unit, an attention weight unit, and a dynamic feature vector unit. The preprocessing unit is used to uniformly resample multi-source health data onto a preset time grid to obtain aligned data with a consistent time axis; at each time point of the time grid, single-point feature vectors of health data are extracted to form a feature sequence with a unified structure. The fusion feature vector unit uses a fixed-length time window as the decreasing unit, and performs a weighted average of the single-point feature vectors within the time window according to the time decay weight to obtain the window-level fusion feature vector. The moving average feature unit calculates the moving average features at multiple time scales based on the fused feature vector; The attention weight unit inputs the moving average features of different time scales into the attention scoring network, calculates the relevance score of each time scale in the current time window, and obtains the attention weight of each time scale through softmax normalization. The dynamic feature vector unit obtains a dynamic feature vector by weighting and fusing moving average features at different time scales through attention weights. The result generation module is used to send the results generated by the data processing module to external systems.

10. A multi-scale health time-series feature dynamic fusion device, characterized in that, It includes a processor, a memory, and a bus. The memory stores instructions and data read by the processor. The processor is used to call the instructions and data in the memory to execute the dynamic fusion method as described in any one of claims 1-8. The bus connects the functional components for transmitting information.

Images