Long-term blood pressure early warning method and system based on multi-decision hybrid model
By dynamically allocating physiological signal data to an expert network through a multi-decision hybrid model, the long-term accuracy and personalization issues of blood pressure monitoring in existing technologies are solved, and an efficient and reliable solution for early risk warning and health management is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LIANGAN MEDICAL TECHNOLOGY CO LTD
- Filing Date
- 2025-12-29
- Publication Date
- 2026-05-08
AI Technical Summary
Existing blood pressure monitoring technologies struggle to provide long-term, accurate, and personalized monitoring and early risk alerts. In particular, they have limited generalization capabilities when faced with scenarios involving significant individual differences and large fluctuations in signal quality, making it difficult to effectively capture daily blood pressure fluctuation patterns and early risks.
A multi-decision hybrid model-based approach is adopted. By acquiring users' personal information and physiological signals, the input data is dynamically allocated to expert networks of multiple expert task groups for independent processing using a gating network. Blood pressure scores are generated through fusion and scoring rules to achieve long-term early warning.
It improves the adaptability and accuracy of the model, enabling earlier identification of abnormal blood pressure patterns, achieving multi-angle and multi-granular judgment, reducing false alarms, and providing personalized health management solutions.
Smart Images
Figure CN122000045A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of blood pressure monitoring technology, specifically relating to a long-term blood pressure early warning method and system based on a multi-decision hybrid model. Background Technology
[0002] Hypertension is one of the most common cardiovascular disease risk factors worldwide, and poor long-term control of hypertension is a major cause of serious complications such as heart disease, stroke, and kidney disease. Therefore, effective and continuous monitoring and early warning of blood pressure are crucial.
[0003] Currently, the mainstream blood pressure monitoring technologies mainly rely on occasional blood pressure measurements and ambulatory blood pressure monitoring (ABPM). Occasional blood pressure measurements are limited in scope and frequency, making it difficult to capture the daily fluctuations in blood pressure. While ABPM can provide 24-hour data, the equipment is bulky, uncomfortable to wear, and is mostly used for short-term diagnosis, making it unsuitable as a long-term, daily early warning method.
[0004] With the popularization of wearable devices, non-invasive and continuous blood pressure monitoring technologies based on physiological signals such as photoplethysmography (PPG) and electrocardiogram (ECG) have become a research hotspot. However, these technologies often lack the ability to comprehensively judge the multi-dimensional blood pressure status. When faced with real-world scenarios with large individual differences and large fluctuations in signal quality, their generalization ability is extremely limited, making it impossible to achieve long-term, accurate, and personalized blood pressure monitoring and early risk warning. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, this invention proposes a long-term blood pressure early warning method based on a multi-decision hybrid model, the method comprising: The system acquires the user's personal information and physiological signals, and preprocesses the physiological signals to extract signal features. The personal information and signal characteristics are used as input data and fed into a multi-decision hybrid model that includes a gating network and multiple expert networks. Each expert network is independent of the others and is pre-divided into at least three expert task groups. The expert networks in each expert task group are pre-configured to perform the same type of blood pressure analysis task, and the types of tasks performed by different expert task groups are different. The input data is dynamically allocated to at least one expert network in each of the expert task groups through the gating network, and each of the allocated expert networks processes the input data independently to generate intermediate results. All intermediate results belonging to the same expert task group are merged to obtain the comprehensive processing result of each expert task group. According to the preset scoring rules, the blood pressure score is dynamically adjusted based on the comprehensive processing results, and a warning message is issued when the blood pressure score reaches the preset alarm threshold.
[0006] Specifically, the preprocessing of the physiological signal to extract signal features includes: The physiological signal is periodically segmented to obtain multiple signal segments, and the quality of each signal segment is evaluated to select signal segments that meet the preset quality standards. Waveform feature parameters for characterizing cardiovascular status are extracted from each of the selected signal segments as signal features; the waveform feature parameters cover at least one of the time domain analysis domain, amplitude analysis domain, phase space analysis domain, and frequency domain analysis domain.
[0007] Preferably, the quality assessment of each of the signal segments includes: A multi-index fusion model optimized by a genetic algorithm is used to comprehensively calculate the waveform skewness consistency, sequence similarity between periods, and correlation coefficient consistency between periods for each signal segment. A nonlinear penalty is applied in combination with the number of outlier periods to evaluate the quality of each signal segment.
[0008] Furthermore, the multi-decision hybrid model also includes a general task group composed of several general networks, which is used to process input data not assigned to expert task groups by the gating network, or to process input data that exceeds the preset capacity of any expert task group already assigned to it; the method further includes: Determine the blood pressure analysis task type corresponding to each intermediate result generated by the general task group; Each of the intermediate results is added to the intermediate result set of the expert task group that performed the corresponding blood pressure analysis task, and participates in the step of merging all intermediate results belonging to the same expert task group.
[0009] Optionally, the multi-decision hybrid model includes a first expert task group for performing the blood pressure numerical regression prediction task, a second expert task group for performing the blood pressure status classification task, and a third expert task group for performing the blood pressure classification task. The intermediate results generated by the expert network in the first expert task group include predicted values of systolic and / or diastolic blood pressure; the intermediate results generated by the expert network in the second expert task group include binary labels indicating whether blood pressure is normal or abnormal; and the intermediate results generated by the expert network in the third expert task group include multi-class labels indicating blood pressure levels.
[0010] Furthermore, the method for training the multi-decision mixture model includes: For each expert network in the first expert task group, the second expert task group, and the third expert task group, a preset capacity value is set respectively; The gating network dynamically inputs training data into the expert networks of each expert task group for training, and training data exceeding the preset capacity of each expert network is input into the general network of the general task group for training. The mean squared error between the predicted values of the expert network in the first expert task group and the actual blood pressure values is determined as the first loss; the cross-entropy between the classification results of the expert network in the second expert task group and the actual labels is determined as the second loss; the cross-entropy between the classification results of the expert network in the third expert task group and the actual labels is determined as the third loss; and the mean squared error or cross-entropy is selected as the general loss according to the original expert task group type corresponding to the training data processed by the general network in the general task group. Based on the first loss, the second loss, the third loss, and the general loss, the overall loss of the multi-decision hybrid model is calculated, and the parameters of the gating network, each of the expert task groups, and the general task group are jointly optimized according to the overall loss using the backpropagation algorithm.
[0011] Furthermore, the multi-decision hybrid model also includes a balancing network, and the method for obtaining the comprehensive processing results of a set of expert task groups includes: Based on the current input data, the gating network assigns a first weight to the intermediate results generated by each expert network in the expert task group and performs weighted fusion to obtain the expert fusion result. The balancing network dynamically generates balancing parameters for the expert task groups based on the input data, and balances the expert fusion results and intermediate results added by the general task group based on the balancing parameters to obtain the comprehensive processing results of the expert task groups; the balancing parameters generated by the balancing network for each expert task group are independent of each other.
[0012] Preferably, the blood pressure score has a preset lower threshold, and the step of dynamically adjusting the blood pressure score based on the comprehensive processing results according to the preset scoring rules includes: Based on the combined processing results of the first expert task group, the second expert task group, and the third expert task group, a tendency judgment is made on the user's current blood pressure status. If the judgment result tends to be normal, the blood pressure score is reduced accordingly based on the degree of consistency between the comprehensive processing results, until the blood pressure score reaches the lower limit threshold. If the judgment result tends to be abnormal, an adjustment strategy for the blood pressure score is determined based on the degree of consistency between the comprehensive processing results. When the adjustment strategy includes increasing the blood pressure score, the blood pressure score is increased accordingly based on the comprehensive processing results of the first expert task group and the third expert task group.
[0013] This invention also proposes a long-term blood pressure early warning system based on a multi-decision hybrid model, the system comprising: The acquisition module is used to acquire the user's personal information and physiological signals, and to preprocess the physiological signals to extract signal features; An input module is used to input the personal information and signal features as input data into a multi-decision hybrid model that includes a gating network and multiple expert networks. Each expert network is independent of the others and is pre-divided into at least three expert task groups. The expert networks in each expert task group are pre-configured to perform the same type of blood pressure analysis task, and the types of tasks performed by different expert task groups are different. The processing module is used to dynamically allocate the input data to at least one of the expert networks in each of the expert task groups through the gating network, and to enable each of the allocated expert networks to process the input data independently and generate intermediate results. The output module is used to merge all intermediate results belonging to the same expert task group to obtain the comprehensive processing result of each expert task group. The alert module is used to dynamically adjust the blood pressure score based on the comprehensive processing results according to the preset scoring rules, and to issue an early warning message when the blood pressure score reaches the preset alarm threshold.
[0014] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the long-term blood pressure early warning method based on a multi-decision hybrid model as described above.
[0015] The present invention has at least the following beneficial effects: The proposed solution uses a gated network to intelligently select the most relevant expert network for processing based on the user's current input data. This allows the model to flexibly cope with complex situations such as different users and different physiological states, greatly improving the model's adaptability and accuracy. It avoids the problem of insufficient generalization ability of a single model in complex and variable physiological scenarios. The expert network is pre-divided into different task groups, allowing each expert network to analyze the same type of problem from different perspectives. Intra-group fusion enhances the robustness and reliability of the analysis. As a result, the model can capture more subtle and earlier abnormal patterns of physiological signals, which helps to identify risk trends before blood pressure values undergo significant clinical changes and achieve long-term early warning. Furthermore, the proposed solution adopts an adaptive quality assessment model, which can filter out low-quality signal segments caused by motion artifacts, poor equipment contact, etc. In addition to the quality of a single cycle, the stability of the overall signal can also be assessed through the sequence similarity and correlation coefficient between cycles. Outlier cycles are intelligently penalized rather than simply removed, making the assessment more refined, thereby constructing a feature system with comprehensive information. Introducing a general task group can prevent a single expert network from being overloaded with data, provide fallback processing, optimize the training process and resource utilization, and improve the fusion quality of the final decision. The expert task group is divided into three groups, each outputting predicted values, normal / abnormal labels, and blood pressure level labels. Different loss functions are set for different groups and the loss type is dynamically selected for the general network. Finally, the overall loss is calculated and jointly optimized, which can achieve end-to-end collaborative learning and ensure a high degree of automation and balance in the training process. The balancing network set in the model can dynamically generate balancing parameters for each expert task group. It introduces a learnable and adaptive arbitration mechanism, which can effectively adjust the balance between the expert fusion results and the results of the general network. The integral adjustment scheme introduces the degree of consistency between the results of each expert group as the core basis for the adjustment strategy and magnitude, so that the final warning can realize multi-angle and multi-granular judgment and reduce false alarms caused by fluctuations in a single indicator.
[0016] Therefore, this invention proposes a long-term blood pressure early warning method and system based on a multi-decision hybrid model. The proposed scheme adopts a multi-task expert network architecture, which can collaboratively analyze blood pressure from different dimensions, overcome the limitations of a single model, and significantly improve the accuracy and reliability of early warning. The gated network can dynamically allocate computing resources according to the user's real-time data characteristics to achieve personalized analysis, making the model more sensitive to individual differences and state changes. By integrating multi-expert decisions and quantifying risk through integration, it can capture early and stable signs of blood pressure abnormalities, achieving true long-term risk early warning and providing an efficient and reliable technical solution for personalized and forward-looking health management. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is an overall schematic diagram of the long-term blood pressure early warning method based on a multi-decision hybrid model provided in Example 1; Figure 2Example graphs showing the results generated using expert task groups and general task groups; Figure 3 This is a schematic diagram illustrating a method for preprocessing physiological signals to extract signal features. Figure 4 A flowchart illustrating the process of preprocessing physiological signals to extract signal features; Figure 5 Example diagram for extracting signal features; Figure 6 This is the phase space diagram of the PPG signal; Figure 7 A schematic diagram illustrating the method for training a multi-decision mixture model; Figure 8 A schematic diagram illustrating the method for obtaining the comprehensive processing results of the expert task group; Figure 9 A diagram illustrating the method for adjusting blood pressure scores; Figure 10 A schematic diagram illustrating the process of blood pressure early warning through unanimous voting; Figure 11 This is a schematic diagram of the module of the long-term blood pressure early warning system based on a multi-decision hybrid model provided in Example 2. Detailed Implementation
[0019] 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 a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0020] Various embodiments of the invention will be described more fully below. The invention may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the invention to the specific embodiments disclosed herein, but rather the invention should be understood to cover all modifications, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of the invention.
[0021] In the following, the terms “comprising” or “may include” as used in various embodiments of the invention indicate the presence of the disclosed functions, operations, or elements, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in various embodiments of the invention, the terms “comprising,” “having,” and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or the possibility of adding one or more combinations of the foregoing.
[0022] In various embodiments of the invention, the expression "or" or "at least one of A and / or B" includes any combination or all combinations of the words listed simultaneously. For example, the expression "A or B" or "at least one of A and / or B" may include A, may include B, or may include both A and B.
[0023] The expressions used in the various embodiments of the present invention (such as "first," "second," etc.) may modify various constituent elements in the various embodiments, but do not limit the corresponding constituent elements. For example, the above expressions do not limit the order and / or importance of the elements. The above expressions are only used for the purpose of distinguishing one element from other elements. For example, a first user device and a second user device refer to different user devices, although both are user devices. For example, a first element may be referred to as a second element without departing from the scope of the various embodiments of the present invention, and similarly, a second element may also be referred to as a first element.
[0024] It should be noted that, in this invention, unless otherwise explicitly specified and defined, terms such as "installation," "connection," and "fixation" 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 according to the specific circumstances.
[0025] In this invention, those skilled in the art should understand that the terms indicating orientation or positional relationship in the text are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the purpose of facilitating the description of this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.
[0026] The terminology used in the various embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. The terms (such as those defined in a generally used dictionary) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.
[0027] Example 1 Please see Figure 1 This embodiment proposes a long-term blood pressure early warning method based on a multi-decision hybrid model. The proposed method utilizes multi-task collaboration and dynamic expert scheduling to achieve long-term, accurate monitoring and early warning of arterial blood pressure. It is applicable to daily health management and clinical decision support. The method specifically includes: S100: Acquires the user's personal information and physiological signals, and preprocesses the physiological signals to extract signal features.
[0028] In this embodiment, the user's personal information obtained in step S100 may include, but is not limited to, the user's height, weight, age, gender, and whether they have a history of hypertension. The obtained physiological signals may include, but are not limited to, electrocardiogram (ECG) signals and photoplethysmography (PPG) signals. The preprocessing of physiological signals may include filtering, downsampling, and normalization, thereby enabling the extraction of physiological signal features while ensuring the quality of the physiological signals. The methods for filtering physiological signals include, but are not limited to, low-pass filtering, high-pass filtering, and band-pass filtering. The downsampling rate can be adjusted according to specific needs, and the formulas used for normalization include: =
[0029] in, Represents the normalized physiological signal. Represents the original signal. Represents the maximum value of the original signal. Represents the minimum value of the original signal; in an optional implementation, and The value can be set by the operator.
[0030] S200: Input personal information and signal characteristics as input data into the most-decision hybrid model.
[0031] In this embodiment, the multi-decision hybrid model includes a gating network and multiple expert networks. Each expert network is independent of the others and is pre-divided into at least three expert task groups. The expert networks in each expert task group are pre-configured to perform blood pressure analysis tasks of the same type, and the types of tasks performed by different expert task groups are different. Optionally, the expert networks may include, but are not limited to, other neural network architectures such as convolutional neural networks (CNN) and recurrent neural networks (RNN).
[0032] S300: The input data is dynamically distributed to at least one expert network in each expert task group through a gating network, and each assigned expert network processes the input data independently to generate intermediate results.
[0033] Please see Figure 2 In this embodiment, the multi-decision hybrid model includes a first expert task group for performing the blood pressure numerical regression prediction task, a second expert task group for performing the blood pressure status classification task, and a third expert task group for performing the blood pressure classification task. The intermediate results generated by the expert network in the first expert task group include predicted values for systolic and / or diastolic blood pressure. The intermediate results generated by the expert network in the second expert task group include binary labels indicating whether blood pressure is normal or abnormal. The intermediate results generated by the expert network in the third expert task group include multi-class labels indicating blood pressure levels. It should be noted that the multi-class labels generated by the third expert task group may include classification labels determined based on blood pressure levels in the Chinese Guidelines for the Prevention and Treatment of Hypertension 2024.
[0034] In an optional implementation, the multi-decision hybrid model may further include a fourth expert task group for performing the blood oxygen prediction task and a fifth expert task group for performing the blood glucose prediction task.
[0035] Preferably, the specific expert network in the same expert task group can be targeted to a certain extent. For example, the first expert task group may include an expert network that is good at performing regression prediction tasks for blood pressure values in adolescents, an expert network that is good at performing regression prediction tasks for blood pressure values in middle-aged people, and an expert network that is good at performing regression prediction tasks for blood pressure values in the elderly.
[0036] S400: Merge all intermediate results belonging to the same expert task group to obtain the comprehensive processing result of each expert task group.
[0037] Furthermore, the multi-decision hybrid model also includes a general task group composed of several general networks. The general task group is used to process input data that has not been assigned to expert task groups by the gating network, or to process input data that has been assigned to any expert task group but exceeds the preset capacity of that expert task group. The method proposed in this embodiment can also determine the blood pressure analysis task type corresponding to each intermediate result generated by the general task group after the general task group generates intermediate results, so as to add each intermediate result to the intermediate result set of the expert task group that performs the corresponding blood pressure analysis task, and participate in the step of fusing all intermediate results belonging to the same expert task group.
[0038] S500: Based on the preset scoring rules and the results of each comprehensive processing, the blood pressure score is dynamically adjusted, and a warning message is issued when the blood pressure score reaches the preset alarm threshold.
[0039] Specifically, please see Figures 3-4 The preprocessing of physiological signals to extract signal features in step S100 includes: S110: Periodically segment the physiological signal to obtain multiple signal segments, evaluate the quality of each signal segment, and then select the signal segments that meet the preset quality standards.
[0040] Preferably, step S110 can comprehensively calculate the waveform skewness consistency, sequence similarity between periods, and correlation coefficient consistency between periods of each signal segment using a multi-index fusion model optimized by a genetic algorithm, and apply a nonlinear penalty in combination with the number of outlier periods to evaluate the quality of each signal segment.
[0041] In this embodiment, the skewness consistency is calculated ( This includes calculating the skewness of a single period and calculating the skewness consistency index, for the first period... One effective period signal The skewness can be defined as:
[0042] in, Representing the The first cycle One sampling point, Represents periodic signals The mean, Represents standard deviation, This represents the number of valid data points (only those ≥ 3 are counted), and is truncated. Represents periodic signals skewness, when It can be forcibly set to To prevent interference from extreme values.
[0043] The preserved non-outlier periodic set Skewness consistency can be defined as:
[0044] in, Represents a non-outlier periodic set. Represents standard deviation, Represents the maximum value. Represents the minimum value. , The value is Used to prevent zeroing, skewness consistency index It can measure the degree of centrality of all effective periodicity skewness distributions. The closer the value is to 1, the more consistent the skewness.
[0045] Please refer to Table 1 and Figure 5 In this embodiment, sequence similarity is calculated ( This includes signal preprocessing, calculating dynamic time warping (DTW) distance, calculating DTW similarity (normalized), and calculating average DTW similarity; Table 1 Summary of parameters and thresholds for evaluating the quality of signal periodic sets
[0046] The preprocessing of the signal includes processing each period of the signal. Zero-mean L2 normalization is performed before DTW calculations to ensure the signal has zero mean and the signal is normalized to the unit L2 norm, standardizing each period of the signal to a uniform, unbiased reference frame to eliminate amplitude and baseline effects. The formula is as follows:
[0047] DTW distance is calculated based on two given normalized sequences. , The calculation is performed using dynamic programming with window constraints, where... Representative sequence Length, Representative sequence length Through size The window only calculates those that satisfy The path, and the formula for the recursive relationship, include:
[0048] Among them, only when Time calculation , Representing the DTW dynamic programming matrix The value of the location, the final distance. And through the formula To implement the length difference penalty, in the formula This represents the final DTW distance after applying the length difference penalty. 0.1 is the length difference sensitivity, i.e., an empirical value; 0.5 is the maximum penalty limit, which can prevent long sequences from being over-penalized.
[0049] For two normalized sequences and The formula for calculating the normalized DTW distance is:
[0050] DTW similarity can be obtained using the reciprocal sigmoid function, with the following formula:
[0051] It should be noted that, It can eliminate the influence of cycle length. The range of values is , The smaller the value, the higher the similarity.
[0052] For non-outlier periodic sets (size is) ), calculate all Yes The average DTW similarity is obtained by taking the mean, and the formula is:
[0053] For any two unnormalized original signals If the signal lengths are the same, that is At that time, signal The correlation coefficient between them is:
[0054] in, Represents the sequence number of the sampling point; If the signal lengths are different, truncate both signals to the same length; that is, take the smaller of the two lengths. The correlation coefficient is calculated based on the common length. .
[0055] For non-outlier sets The formula for calculating the average correlation coefficient is:
[0056] in, Representative set The absolute value of the number of signals in the waveform emphasizes the consistency of the waveform shape, rather than the phase or polarity.
[0057] In this embodiment, the MAD method is used for discrete value detection, and the formula for calculating the summative similarity score used for outlier detection is as follows:
[0058] in, The similarity conversion value representing the DTW distance; in this embodiment, the DTW similarity uses a value specifically for outlier detection. The final metric, clean_dtw_sim, uses .
[0059] The formula for standardizing MAD is:
[0060] Where 1.4826 represents the conversion factor between MAD and standard deviation under a normal distribution ( A threshold of 2.5 corresponds to approximately a 99% confidence level (two-sided), which is a commonly used statistical outlier criterion. If so, it can be determined as an outlier. The formulas for calculating the basic consistency score include:
[0061] ,
[0062] , All values are ∈ [0,1], ensuring that the linear weighting is effective. When the number of outliers When the outlier penalty formula is:
[0063]
[0064] in, , representing the steepness of the penalty curve for GA optimization; , representing the maximum penalty value for GA optimization, with the center point fixed at 3. When the outlier count = 3, the penalty = This reflects that a small number of outliers are tolerable, while a large number of outliers severely degrade quality; in this embodiment, the sigmoid penalty function is optimized by GA in the range of [0.005,0.5]×[0.05,0.5] to match expert annotations.
[0065] In this embodiment, the fragment quality is based on consistency_metric. The data fragments marked as good have consistency_metric > consistency_threshold (1), the data fragments marked as fair have consistency_metric ≥ consistency_threshold (2), and the remaining data fragments are marked as poor.
[0066] After data processing, the method proposed in this embodiment will preferentially select data segments labeled as good quality for subsequent signal feature extraction. If there are no segments with good quality, segments labeled as fair quality will be used. If there are no segments with good or fair quality, the signal segment will be skipped.
[0067] S120: Extract waveform feature parameters used to characterize cardiovascular status from each selected signal segment as signal features.
[0068] Please refer to Table 2. In this embodiment, the waveform feature parameters may include, but are not limited to, the original waveform signal and the pulse wave propagation time. The waveform feature parameters extracted in step S120 cover at least one of the time domain analysis domain, amplitude analysis domain, phase space analysis domain, and frequency domain analysis domain.
[0069] Table 2. Definitions of Some Important and Representative Characteristic Parameters
[0070] exist Figure 5 In the text, R-peaks represent R-peaks, Pulse foot represents the trough of the pulse wave, Systolic peak represents the systolic peak, Dicrotic Notch represents the dicrotic notch, and Diastolic peak represents the diastolic peak. And in Figure 6 In the diagram, the X-axis represents the PPG signal delay, and the Y-axis represents the PPG signal. Figure 7 The figure shows the case where the PPG signal delay is 20ms. This method divides the graph into N×N small squares. The squares through which the PPG trajectory passes can be identified as black boxes. B_all represents the number of black boxes in the whole graph. The whole graph is divided into the following 9 regions, and B_1 is the number of black boxes in region 1.
[0071] In this embodiment, steps S200-S400 are all implemented based on the completed multi-decision mixture model. Please refer to [link / reference]. Figure 7Methods for training multi-decision mixture models include: S610: Set a preset capacity value for each expert network in each expert task group.
[0072] In this embodiment, the preset capacity value can be set as the amount of data in a single batch or the number of experts during training, multiplied by a coefficient between 1 and 2. The value of the coefficient can be adjusted according to the actual situation, and in most cases, it is preferably 1.25 or 1.5.
[0073] S620: The training data is dynamically input into the expert networks of each expert task group for training through a gating network, and the training data exceeding the preset capacity of each expert network is changed to be input into the general network of the general task group for training.
[0074] In this embodiment, the input format of each expert task group during the training process is the same: physiological signal features and user information. The output of the first expert task group is the value of systolic blood pressure and diastolic blood pressure, the output of the second expert task group is the two-class blood pressure result, and the output of the third expert task group is the five-class blood pressure result.
[0075] It should be noted that the method proposed in this embodiment only inputs training data exceeding the preset capacity of each expert network into the general network in the general task group, and does not input it into other expert networks in the expert task group. The general task group can process all data with overflow capacity values and, as a connection of a residual network, provide higher performance for the multi-decision hybrid model.
[0076] S630: Determine the losses of each expert task group and the general task group separately.
[0077] In this embodiment, the expert network in the first expert task group is trained using mean squared error (MSE) as the loss function, while the expert networks in the second and third expert task groups are trained using cross entropy as the loss function. Specifically, step S630 can determine the mean square error between the predicted value and the actual blood pressure value of the expert network in the first expert task group as the first loss, the cross entropy between the classification result and the actual label of the expert network in the second expert task group as the second loss, the cross entropy between the classification result and the actual label of the expert network in the third expert task group as the third loss, and select the mean square error or cross entropy as the general loss according to the original expert task group type corresponding to the training data processed by the general network in the general task group.
[0078] S640: Calculate the overall loss of the multi-decision hybrid model based on the losses of each expert task group and the general task group, and jointly optimize the parameters of the gating network, each expert task group, and the general task group according to the overall loss through the backpropagation algorithm.
[0079] Further, please see Figure 8 Multi-decision hybrid models also include balancing networks, and methods for obtaining the comprehensive processing results of a set of expert task groups include: S410: Based on the current input data, the gating network assigns the first weight to the intermediate results generated by each expert network in the expert task group and performs weighted fusion to obtain the expert fusion result.
[0080] S420: The balance network dynamically generates balance parameters for the expert task group based on the input data, and balances the expert fusion results and the intermediate results added by the general task group based on the balance parameters to obtain the comprehensive processing results of the expert task group.
[0081] In this embodiment, the balancing network is a single-layer neural network that can learn the optimal balancing parameters through neural network learning, and the balancing parameters generated by each expert task group are independent of each other. Specifically, the formulas for obtaining the comprehensive processing results include:
[0082] in, Represents the equilibrium parameter. Represents input data, Represents a gating network. Representative of the expert task force. This represents the number of the expert network within the expert task group.
[0083] Preferably, the blood pressure score has a preset lower threshold; please refer to [link / reference]. Figures 9-10 The step S500, which involves dynamically adjusting the blood pressure score based on the preset scoring rules and the results of each comprehensive processing step, specifically includes: S510: Based on the combined processing results of the first expert task group, the second expert task group, and the third expert task group, make a tendency judgment on the user's current blood pressure status.
[0084] In this embodiment, if the combined processing results of at least two of the first expert task group, the second expert task group, and the third expert task group indicate that the user's blood pressure is normal, then the user's current blood pressure is determined to be tending towards a normal state; if the combined processing results of at least two of the first expert task group, the second expert task group, and the third expert task group indicate that the user's blood pressure is abnormal, then the user's current blood pressure is determined to be tending towards an abnormal state.
[0085] S520: If the judgment result tends to be normal, the blood pressure score is reduced accordingly based on the consistency between the comprehensive processing results until the blood pressure score reaches the lower limit threshold.
[0086] S530: If the judgment result tends to be abnormal, determine the adjustment strategy for the blood pressure score based on the consistency between the comprehensive processing results. When the adjustment strategy includes increasing the blood pressure score, increase the blood pressure score accordingly based on the comprehensive processing results of the first expert task group and the third expert task group.
[0087] For example, the lower limit threshold for the blood pressure score can be set to 0 points, and the alarm threshold can be set to 6 points. When the processing results of all three expert task groups indicate that the user's blood pressure is normal, the blood pressure score is reduced by 3 points; when the processing results of two of the expert task groups indicate that the user's blood pressure is normal, and the processing result of the other expert task group indicates that the user's blood pressure is abnormal, the blood pressure score is reduced by 1 point; when the processing results of two of the expert task groups indicate that the user's blood pressure is abnormal, and the processing result of the other expert task group indicates that the user's blood pressure is normal, the blood pressure score is increased by 1 point. When the results from all three expert task groups indicate that the user's blood pressure is abnormal, if the combined results from the first and third expert task groups indicate that the user's blood pressure is hypertension, the blood pressure score will be increased by 3 points; if the combined results from the first and third expert task groups indicate that the user's blood pressure is moderate hypertension, the blood pressure score will be increased by 4 points; if the combined results from the first and third expert task groups indicate that the user's blood pressure is severe hypertension, the blood pressure score will be increased by 6 points; and when the blood pressure score reaches 6 points, a warning message will be issued.
[0088] Example 2 Please see Figure 11 This embodiment proposes a long-term blood pressure early warning system based on a multi-decision hybrid model to implement the long-term blood pressure early warning method based on a multi-decision hybrid model proposed in Embodiment 1. The system specifically includes: The acquisition module 10 is used to acquire the user's personal information and physiological signals, and to preprocess the physiological signals to extract signal features; Input module 20 is used to input personal information and signal characteristics as input data into a multi-decision hybrid model that includes a gating network and multiple expert networks; The processing module 30 is used to dynamically allocate input data to at least one expert network in each expert task group through a gating network, and to enable each allocated expert network to process the input data independently and generate intermediate results. Output module 40 is used to merge all intermediate results belonging to the same expert task group to obtain the comprehensive processing result of each expert task group. The prompt module 50 is used to dynamically adjust the blood pressure score based on the comprehensive processing results according to the preset scoring rules, and issue a warning message when the blood pressure score reaches the preset alarm threshold.
[0089] In this embodiment, the user personal information acquired by the acquisition module 10 may include, but is not limited to, the user's height, weight, age, gender, and whether they have a history of hypertension. The acquired physiological signals may include, but are not limited to, electrocardiogram (ECG) signals and photoplethysmography (PPG) signals. The preprocessing of the physiological signals may include filtering, downsampling, and normalization, thereby enabling the extraction of physiological signal features while ensuring the quality of the physiological signals. The multi-decision hybrid model includes a gating network and multiple expert networks. Each expert network is independent of the others and is pre-divided into at least three expert task groups. The expert networks in each expert task group are pre-configured to perform the same type of blood pressure analysis task, and the types of tasks performed by different expert task groups are different. In this embodiment, the expert task group includes a first expert task group for performing blood pressure numerical regression prediction, a second expert task group for performing blood pressure status classification, and a third expert task group for performing blood pressure grading classification. The intermediate results generated by the expert network in the first expert task group include predicted values for systolic and / or diastolic blood pressure; the intermediate results generated by the expert network in the second expert task group include binary labels indicating whether blood pressure is normal or abnormal; and the intermediate results generated by the expert network in the third expert task group include multi-class labels indicating blood pressure levels.
[0090] In an optional implementation, the expert task group may further include a fourth expert task group for performing the blood oxygen prediction task and a fifth expert task group for performing the blood glucose prediction task.
[0091] Furthermore, the multi-decision hybrid model also includes a general task group composed of several general networks. The general task group is used to process input data that has not been assigned to the expert task group by the gating network, or to process input data that has been assigned to any expert task group but exceeds the preset capacity of that expert task group. After the general task group generates intermediate results, the output module 40 can also determine the blood pressure analysis task type corresponding to each intermediate result generated by the general task group, so as to add each intermediate result to the intermediate result set of the expert task group that performs the corresponding blood pressure analysis task, and participate in the step of fusing all intermediate results belonging to the same expert task group.
[0092] Specifically, the acquisition module 10 can periodically segment the physiological signal to obtain multiple signal segments, evaluate the quality of each signal segment, select signal segments that meet the preset quality standards, and extract waveform feature parameters for characterizing the cardiovascular state from each selected signal segment as signal features.
[0093] The output module 40 can use a gating network to assign first weights to the intermediate results generated by each expert network in the expert task group based on the current input data and perform weighted fusion to obtain the expert fusion result. It can also use the balancing network in the multi-decision hybrid model to dynamically generate balancing parameters for the expert task group based on the input data, and use the balancing parameters to balance the expert fusion result and the intermediate results added by the general task group to obtain the comprehensive processing result of the expert task group.
[0094] The prompt module 50 can make a tendency judgment on the user's current blood pressure status based on the comprehensive processing results of the first expert task group, the second expert task group and the third expert task group. When the judgment result tends to be normal, the blood pressure score is reduced accordingly based on the consistency between the comprehensive processing results until the blood pressure score reaches the lower limit threshold. Furthermore, when the judgment result tends to be abnormal, the adjustment strategy for the blood pressure score can be determined based on the degree of consistency between the comprehensive processing results. When the adjustment strategy includes increasing the blood pressure score, the blood pressure score is increased accordingly based on the comprehensive processing results of the first expert task group and the third expert task group.
[0095] The system proposed in this embodiment also includes: The setting module 61 is used to set a preset capacity value for each expert network in each expert task group; The allocation module 62 is used to dynamically input training data into the expert networks of each expert task group for training through the gating network, and to change the training data exceeding the preset capacity of each expert network to be input into the general network of the general task group for training. Calculation module 63 is used to determine the losses of each expert task group and the general task group respectively; The optimization module 64 is used to calculate the overall loss of the multi-decision hybrid model based on the losses of each expert task group and the general task group, and to jointly optimize the parameters of the gating network, each expert task group and the general task group according to the overall loss through the backpropagation algorithm.
[0096] Example 3 The present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the long-term blood pressure early warning method based on a multi-decision hybrid model as proposed in Example 1.
[0097] It should be noted that computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0098] In summary, this invention proposes a long-term blood pressure early warning method and system based on a multi-decision hybrid model. The proposed scheme adopts a multi-task expert network architecture, which can collaboratively analyze blood pressure from different dimensions, overcoming the limitations of a single model and significantly improving the accuracy and reliability of early warning. The gated network can dynamically allocate computing resources according to the user's real-time data characteristics to achieve personalized analysis, making the model more sensitive to individual differences and state changes. By integrating multi-expert decisions and quantifying risk through integration, it can capture early and stable signs of blood pressure abnormalities, achieving true long-term risk early warning and providing an efficient and reliable technical solution for personalized and forward-looking health management.
[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A long-term blood pressure early warning method based on a multi-decision mixture model, characterized in that, The method includes: The system acquires the user's personal information and physiological signals, and preprocesses the physiological signals to extract signal features. The personal information and signal characteristics are used as input data and fed into a multi-decision hybrid model that includes a gating network and multiple expert networks. Each expert network is independent of the others and is pre-divided into at least three expert task groups. The expert networks in each expert task group are pre-configured to perform the same type of blood pressure analysis task, and the types of tasks performed by different expert task groups are different. The input data is dynamically allocated to at least one expert network in each of the expert task groups through the gating network, and each of the allocated expert networks processes the input data independently to generate intermediate results. All intermediate results belonging to the same expert task group are merged to obtain the comprehensive processing result of each expert task group. According to the preset scoring rules, the blood pressure score is dynamically adjusted based on the comprehensive processing results, and a warning message is issued when the blood pressure score reaches the preset alarm threshold.
2. The long-term blood pressure early warning method based on a multi-decision hybrid model according to claim 1, characterized in that, The preprocessing of the physiological signal to extract signal features includes: The physiological signal is periodically segmented to obtain multiple signal segments, and the quality of each signal segment is evaluated to select signal segments that meet the preset quality standards. Waveform feature parameters for characterizing cardiovascular status are extracted from each of the selected signal segments as signal features; the waveform feature parameters cover at least one of the time domain analysis domain, amplitude analysis domain, phase space analysis domain, and frequency domain analysis domain.
3. The long-term blood pressure early warning method based on a multi-decision hybrid model according to claim 2, characterized in that, The quality assessment of each of the signal segments includes: A multi-index fusion model optimized by a genetic algorithm is used to comprehensively calculate the waveform skewness consistency, sequence similarity between periods, and correlation coefficient consistency between periods for each signal segment. A nonlinear penalty is applied in combination with the number of outlier periods to evaluate the quality of each signal segment.
4. The long-term blood pressure early warning method based on a multi-decision hybrid model according to any one of claims 1-3, characterized in that, The multi-decision hybrid model further includes a general task group composed of several general networks. The general task group is used to process input data not assigned to expert task groups by the gating network, or to process input data that exceeds the preset capacity of any expert task group already assigned to it. The method further includes: Determine the blood pressure analysis task type corresponding to each intermediate result generated by the general task group; Each of the intermediate results is added to the intermediate result set of the expert task group that performed the corresponding blood pressure analysis task, and participates in the step of merging all intermediate results belonging to the same expert task group.
5. The long-term blood pressure early warning method based on a multi-decision hybrid model according to claim 4, characterized in that, The multi-decision hybrid model includes a first expert task group for performing the numerical regression prediction task of blood pressure, a second expert task group for performing the blood pressure status classification task, and a third expert task group for performing the blood pressure classification task. The intermediate results generated by the expert network in the first expert task group include predicted values of systolic and / or diastolic blood pressure; the intermediate results generated by the expert network in the second expert task group include binary labels indicating whether blood pressure is normal or abnormal; and the intermediate results generated by the expert network in the third expert task group include multi-class labels indicating blood pressure levels.
6. The long-term blood pressure early warning method based on a multi-decision hybrid model according to claim 5, characterized in that, The methods for training the multi-decision mixture model include: For each expert network in the first expert task group, the second expert task group, and the third expert task group, a preset capacity value is set respectively; The gating network dynamically inputs training data into the expert networks of each expert task group for training, and training data exceeding the preset capacity of each expert network is input into the general network of the general task group for training. The mean squared error between the predicted values of the expert network in the first expert task group and the actual blood pressure values is determined as the first loss; the cross-entropy between the classification results of the expert network in the second expert task group and the actual labels is determined as the second loss; the cross-entropy between the classification results of the expert network in the third expert task group and the actual labels is determined as the third loss; and the mean squared error or cross-entropy is selected as the general loss according to the original expert task group type corresponding to the training data processed by the general network in the general task group. Based on the first loss, the second loss, the third loss, and the general loss, the overall loss of the multi-decision hybrid model is calculated, and the parameters of the gating network, each of the expert task groups, and the general task group are jointly optimized according to the overall loss using the backpropagation algorithm.
7. The long-term blood pressure early warning method based on a multi-decision hybrid model according to claim 6, characterized in that, The multi-decision hybrid model also includes a balancing network, and the method for obtaining the comprehensive processing results of a set of expert task groups includes: Based on the current input data, the gating network assigns a first weight to the intermediate results generated by each expert network in the expert task group and performs weighted fusion to obtain the expert fusion result. The balancing network dynamically generates balancing parameters for the expert task groups based on the input data, and balances the expert fusion results and intermediate results added by the general task group based on the balancing parameters to obtain the comprehensive processing results of the expert task groups; the balancing parameters generated by the balancing network for each expert task group are independent of each other.
8. The long-term blood pressure early warning method based on a multi-decision hybrid model according to claim 5, characterized in that, The blood pressure score has a preset lower threshold. The step of dynamically adjusting the blood pressure score based on the preset scoring rules and the comprehensive processing results includes: Based on the combined processing results of the first expert task group, the second expert task group, and the third expert task group, a tendency judgment is made on the user's current blood pressure status. If the judgment result tends to be normal, the blood pressure score is reduced accordingly based on the degree of consistency between the comprehensive processing results, until the blood pressure score reaches the lower limit threshold. If the judgment result tends to be abnormal, an adjustment strategy for the blood pressure score is determined based on the degree of consistency between the comprehensive processing results. When the adjustment strategy includes increasing the blood pressure score, the blood pressure score is increased accordingly based on the comprehensive processing results of the first expert task group and the third expert task group.
9. A long-term blood pressure early warning system based on a multi-decision hybrid model, characterized in that, The system includes: The acquisition module is used to acquire the user's personal information and physiological signals, and to preprocess the physiological signals to extract signal features; An input module is used to input the personal information and signal features as input data into a multi-decision hybrid model that includes a gating network and multiple expert networks. Each expert network is independent of the others and is pre-divided into at least three expert task groups. The expert networks in each expert task group are pre-configured to perform the same type of blood pressure analysis task, and the types of tasks performed by different expert task groups are different. The processing module is used to dynamically allocate the input data to at least one of the expert networks in each of the expert task groups through the gating network, and to enable each of the allocated expert networks to process the input data independently and generate intermediate results. The output module is used to merge all intermediate results belonging to the same expert task group to obtain the comprehensive processing result of each expert task group. The alert module is used to dynamically adjust the blood pressure score based on the comprehensive processing results according to the preset scoring rules, and to issue an early warning message when the blood pressure score reaches the preset alarm threshold.
10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the long-term blood pressure early warning method based on a multi-decision hybrid model as described in any one of claims 1-8.