Comprehensive health data analysis method and system based on artificial intelligence

Abstract

The invention provides a comprehensive health data analysis method and system based on artificial intelligence, and the method comprises the steps: carrying out the representation learning processing of to-be-analyzed health monitoring data, carrying out the embedded mapping operation of a representation learning result, and obtaining the representation learning result and a monitoring recognition feature vector; and based on the representation learning result, the monitoring identification feature vector and a dynamic event identification feature vector corresponding to the to-be-analyzed health monitoring data, carrying out reduction mapping operation to obtain an event identification result of the to-be-analyzed health monitoring data, so that the event identification result can represent an event in the to-be-analyzed health monitoring data. The dynamic event identification feature vector is determined based on the event identification results of the plurality of preorder monitoring data corresponding to the to-be-analyzed health monitoring data, so that the event identification results determined according to the dynamic event identification feature vector can more accurately represent the events in the to-be-analyzed health monitoring data; therefore, event dynamic tracking can be completed more accurately, and the purpose of real-time dynamic monitoring is achieved.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a comprehensive health data analysis method and system based on artificial intelligence. Background Art

[0002] In today's era of rapid digitalization and intelligence development, the field of health monitoring and analysis faces growing demands and challenges. With the increasing attention paid to people's health and the continuous advancement of medical technology and information technology, how to accurately and efficiently analyze and process health monitoring data to achieve real-time dynamic monitoring and accurate assessment of health status has become a hot and difficult issue in this field.

[0003] In recent years, with the rapid development of artificial intelligence technology, it has gradually become a trend to use machine learning and data analysis methods to process health monitoring data. However, there are still some problems with existing health data analysis methods based on artificial intelligence. In terms of event recognition and dynamic tracking, existing technologies often have difficulty in accurately identifying and tracking various events in health monitoring data, especially those key events that require continuous monitoring. For example, for feasible health events such as abnormal heart rate and blood pressure fluctuations, existing methods may not be able to detect the occurrence of events in a timely and accurate manner, and effectively predict and warn their development trends. This is mainly because existing technologies do not fully utilize the correlation between historical data and events when processing dynamic events, and cannot build a comprehensive and accurate dynamic event recognition model. Summary of the invention

[0004] In view of this, the present application provides a comprehensive health data analysis method and system based on artificial intelligence. The technical solution of the present application is implemented as follows: On the one hand, the present application provides a comprehensive health data analysis method based on artificial intelligence, the method comprising: performing representation learning processing on the health monitoring data to be analyzed to obtain a representation learning result; performing an embedding mapping operation on the representation learning result to obtain a monitoring identification feature vector; performing a restoration mapping operation based on the monitoring identification feature vector and a dynamic event identification feature vector corresponding to the health monitoring data to be analyzed to obtain an event identification result of the health monitoring data to be analyzed; the dynamic event identification feature vector is determined based on the event identification results of multiple preceding monitoring data corresponding to the health monitoring data to be analyzed; the event identification result of the health monitoring data to be analyzed is used to describe one or more health monitoring events in the health monitoring data to be analyzed.

[0005] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, the steps in the above method are implemented. In the embodiment of the present application, for the health monitoring data to be analyzed, a representation learning process is performed on it to obtain a representation learning result, and the representation learning result is made to represent the characteristics of the health monitoring data to be analyzed, and then the representation learning result is embedded and mapped to obtain a monitoring identification feature vector, and a restoration mapping operation is performed based on the monitoring identification feature vector and the dynamic event identification feature vector corresponding to the health monitoring data to be analyzed to obtain an event identification result of the health monitoring data to be analyzed, so that the event identification result can indicate the health monitoring events in the health monitoring data to be analyzed. Since the dynamic event identification feature vector is determined based on the event identification results of multiple preceding monitoring data corresponding to the health monitoring data to be analyzed, the dynamic event identification feature vector can more accurately represent the past states of some events in the health monitoring data to be analyzed, and the event identification result determined based on the dynamic event identification feature vector can more accurately represent the events in the health monitoring data to be analyzed, and more accurately complete the dynamic tracking of events, so as to achieve the purpose of real-time dynamic monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 A schematic diagram of the implementation process of an artificial intelligence-based comprehensive health data analysis method provided in an embodiment of the present application.

[0007] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0009] The embodiment of the present application provides a comprehensive health data analysis method based on artificial intelligence, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer.

[0010] Figure 1 A schematic diagram of an implementation flow of a comprehensive health data analysis method based on artificial intelligence provided in an embodiment of the present application, such as Figure 1 As shown, the method includes: Step S100: performing representation learning processing on the health monitoring data to be analyzed to obtain a representation learning result.

[0011] Representation learning is a machine learning method that aims to learn effective feature representations from raw data. Raw health monitoring data may contain various types of information, such as physiological indicators (such as heart rate, blood pressure, etc.), living habits (such as exercise, diet, etc.), and environmental factors (such as temperature, humidity, etc.). These data are usually high-dimensional, complex, and may contain noise, and direct analysis and processing may face many difficulties. Through representation learning processing, the raw data can be converted into a form that is easier to understand and process, that is, the representation learning result, which can capture the key information and patterns in the data and provide strong support for subsequent analysis.

[0012] When the computer system performs representation learning processing, it can be implemented using principal component analysis (PCA). The computer system first calculates the covariance matrix of the health monitoring data, and then performs eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors. The eigenvalue represents the variance of the data in the direction of the corresponding eigenvector. By selecting the eigenvectors corresponding to the largest eigenvalues, the computer system can project the original data into a low-dimensional space to obtain the representation learning results. For example, suppose there is a set of health monitoring data containing multiple physiological indicators, such as heart rate, blood pressure, body temperature, etc. Through PCA analysis, the computer system can find the correlation between these indicators and extract several main components. These components can be regarded as a simplified representation of the original data, which can reduce the dimension of the data while retaining most of the information. The specific formula is as follows: Calculate the covariance matrix C: , where n is the number of samples, is the i-th sample, is the sample mean.

[0013] Perform eigenvalue decomposition on the covariance matrix C: , where U is the eigenvector matrix, is a diagonal matrix, and the elements on the diagonal are the eigenvalues.

[0014] Select the eigenvectors corresponding to the first k largest eigenvalues ​​to form the projection matrix P: , where u_i is the i-th eigenvector.

[0015] Project the original data X into a low-dimensional space to obtain the representation learning result Y: Y=XP.

[0016] Alternatively, you can also use an autoencoder for representation learning. The autoencoder is a neural network model that consists of two parts: an encoder and a decoder. The encoder compresses the original data into a low-dimensional encoded representation, and the decoder attempts to reconstruct the original data from this encoded representation. The computer system learns an effective representation of the data by training the autoencoder to minimize the reconstruction error. For example, in health monitoring data, the autoencoder can learn the potential relationship between different physiological indicators and encode the original data into a low-dimensional vector, which is the representation learning result. Specifically, the computer system first defines the network structure of the autoencoder, including the number of layers, the number of neurons, and the activation function of the encoder. Then, the autoencoder is trained using health monitoring data, and the weights of the network are updated through the back-propagation algorithm, so that the encoder can accurately encode the original data into a low-dimensional representation, and the decoder can reconstruct the original data as accurately as possible from the low-dimensional representation.

[0017] Alternatively, a deep neural network (DNN) can be used for representation learning. A deep neural network is a neural network model with multiple hidden layers that can automatically learn complex patterns and features in the data. The computer system trains the deep neural network to enable it to classify, predict or generate representation learning results for health monitoring data. For example, in health data analysis, a convolutional neural network (CNN) can be used to process health monitoring data in the form of images (such as medical images) to extract features from the images; a recurrent neural network (RNN) or a long short-term memory network (LSTM) can also be used to process health monitoring data in the form of time series (such as heart rate, blood pressure and other time-varying data) to capture the timing information and long-term dependencies in the data. The specific training process includes defining the network structure, selecting a suitable loss function (such as mean square error, cross entropy, etc.) and an optimization algorithm (such as stochastic gradient descent, Adam, etc.), and then using health monitoring data to train the network and minimize the loss function by continuously adjusting the network weights.

[0018] For example, suppose a computer system receives a set of health monitoring data, including a person's heart rate, blood pressure, body temperature, and number of steps, which are collected continuously over a period of time. The computer system can use the above representation learning method to process this data.

[0019] If the principal component analysis method is used, the computer system first calculates the covariance matrix of these data, and then performs eigenvalue decomposition on the covariance matrix. Assume that two main components are obtained after analysis, one component is related to heart rate and exercise steps, and the other component is related to blood pressure and body temperature. Then the computer system can project the original four-dimensional data into the two-dimensional space composed of these two main components to obtain the representation learning result. Each point in this two-dimensional space represents a simplified representation of the original data. By observing the distribution and changes of these points, the computer system can initially understand some characteristics of the person's health status.

[0020] If the autoencoder method is adopted, the computer system builds an autoencoder network with an appropriate structure, inputs the original health monitoring data into the encoder, and the encoder compresses the data into a low-dimensional encoding vector. For example, this encoding vector may be a vector containing 10 elements. The decoder attempts to reconstruct the original data from this encoding vector. By training the autoencoder, the computer system can enable the encoder to learn the key features in the data, and this encoding vector represents the learning results. The computer system can further use this encoding vector for subsequent analysis, such as determining whether the person's health status is normal.

[0021] If a deep neural network method is used, such as using an LSTM network to process health monitoring data in the form of time series, the computer system inputs data such as heart rate, blood pressure, body temperature, and exercise steps into the LSTM network in chronological order. The LSTM network can capture the timing information and long-term dependencies in the data through its internal memory units and gating mechanisms. After training, the output of the LSTM network can be used as a representation of the learning result. This result can reflect the changing trend of a person's health status over time, providing an important basis for the subsequent identification of health events.

[0022] Step S200: performing an embedding mapping operation on the representation learning result to obtain a monitoring recognition feature vector.

[0023] In an embodiment of the present application, the embedding mapping operation is an encoding process, the purpose of which is to map the representation learning results into a low-dimensional, continuous vector space to obtain a monitoring identification feature vector. Each point in this vector space represents a specific representation of the original data, and the distance or similarity between the vectors can reflect a certain relationship between the data. In the context of health data analysis, through the embedding mapping operation, the representation learning results of complex health monitoring data can be converted into a more compact, easier to calculate and analyze form, thereby facilitating the subsequent event identification and analysis by the computer system. For example, different health monitoring events may have different distribution patterns in this low-dimensional vector space, and the computer system can identify and distinguish these events by analyzing the relationship between the vectors.

[0024] When a computer system performs an embedding mapping operation, it can use an embedding layer to implement the embedding mapping operation. The embedding layer is usually the first layer of a neural network, which receives the representation learning result as input and converts it into a low-dimensional vector representation. For example, in a deep neural network for health event recognition, the input representation learning result may be a vector containing multiple eigenvalues. The computer system first defines the parameters of the embedding layer, including the embedding dimension d and the embedding matrix E. The size of the embedding matrix is ​​n×d, where n is the number of representation learning results and d is the embedding dimension. For the input representation learning result x (assuming x is an integer index representing a specific feature), the embedding layer obtains the corresponding embedding vector e by looking up the xth row of the embedding matrix E. x , that is, e x =E[x, :]. This embedding vector is the monitoring recognition feature vector. During the training of the neural network, the parameters of the embedding matrix E will be updated along with the parameters of other network layers to optimize the performance of the entire network.

[0025] In other implementations, embedding mapping can be performed based on principal component analysis (PCA). The principal component analysis method mentioned in step S100 can not only be used for dimensionality reduction to obtain representation learning results, but can also be further used for embedding mapping operations. After obtaining the results of the principal component analysis, the computer system selects the first k principal components as the embedding dimensions, projects the representation learning results into this k-dimensional space, and obtains the monitoring recognition feature vector. Assume that the eigenvector matrix obtained by the principal component analysis is U, where the first k columns correspond to the first k principal components. For a representation learning result vector x, its corresponding monitoring recognition feature vector y can be calculated by the following formula: ,in is the matrix consisting of the first k columns of U.

[0026] For example, suppose that the computer system obtains a set of representation learning results through principal component analysis in step S100. These results represent the health status characteristics of different individuals. Each result is a vector containing multiple feature values, such as information containing indicators such as heart rate, blood pressure, and body temperature. Now, these representation learning results need to be embedded and mapped.

[0027] Step S300: Perform a restoration mapping operation based on the monitoring identification feature vector and the dynamic event identification feature vector corresponding to the health monitoring data to be analyzed to obtain the event identification result of the health monitoring data to be analyzed; the dynamic event identification feature vector is determined based on the event identification results of multiple preceding monitoring data corresponding to the health monitoring data to be analyzed; the event identification result of the health monitoring data to be analyzed is used to describe one or more health monitoring events in the health monitoring data to be analyzed.

[0028] The restoration mapping operation is the opposite of the embedding mapping operation. Its purpose is to remap the low-dimensional feature vector obtained after the embedding mapping back to a representation closer to the original data semantic space in order to obtain specific information about the events contained in the data. In the health data analysis scenario, the computer system interprets the health monitoring events contained in the health monitoring data to be analyzed from the monitoring identification feature vector and the dynamic event identification feature vector.

[0029] When a computer system performs a restoration mapping operation based on these two feature vectors, there are many ways to implement it. One feasible method is to use a neural network model to construct a restoration mapping relationship. For example, build a multi-layer perceptron (MLP) network that takes a monitoring recognition feature vector and a dynamic event recognition feature vector as input. Assume that the monitoring recognition feature vector is represented by M, with a dimension of m; the dynamic event recognition feature vector is represented by D, with a dimension of n. First, concatenate these two vectors to obtain a new vector V=[M; D] with a dimension of m+n (the semicolon here represents a vector concatenation operation). Then input the new vector V into the multi-layer perceptron network. The multi-layer perceptron network contains multiple hidden layers, and each hidden layer consists of a certain number of neurons. Assume that the i-th hidden layer has neurons, the weight matrix from the input layer to the first hidden layer is W1, with a dimension of (m+n)×h1, and the bias vector is b1, with a dimension of h1. Then the output H1 of the first hidden layer can be obtained by the formula Calculated, where is an activation function, such as the commonly used ReLU function (Rectified Linear Unit), which is defined as .

[0030] Subsequent hidden layers are calculated in a similar way. The output H of the jth hidden layer (j>1) is jThrough the output H of the previous hidden layer j-1 To calculate, the formula is , where W j is the weight matrix from layer j-1 to layer j, b j is the bias vector of the jth layer.

[0031] Finally, the output layer of the multilayer perceptron network maps the output of the last hidden layer to the space of event recognition results. Assume that the weight matrix of the output layer is W o , the dimension is (h last is the number of neurons in the last hidden layer, k is the number of event categories), the bias vector is b o , with dimension k. Then the event recognition result vector R can be obtained by the formula Here, each element in the event recognition result vector R represents the confidence score of the corresponding event category. For example, if there are three event categories (normal, mild abnormality, severe abnormality), then , r1 represents the confidence that the data belongs to the "normal" category, r2 represents the confidence that the data belongs to the "mild abnormal" category, and r3 represents the confidence that the data belongs to the "severe abnormal" category. By further processing these confidence scores, such as using the softmax function (defined as , where z is the input vector, K is the number of categories, and j represents the jth category) and converts it into a probability distribution. The computer system can then determine the event category to which the health monitoring data to be analyzed is most likely to belong.

[0032] The dynamic event recognition feature vector is determined based on the event recognition results of multiple preceding monitoring data corresponding to the health monitoring data to be analyzed. For example, suppose the health monitoring data to be analyzed is the health data of a person collected today, and the preceding monitoring data includes the data collected yesterday and the day before yesterday. The computer system first extracts some key information from the event recognition results of yesterday's health monitoring data as the first dynamic event recognition result. For example, yesterday's event recognition results show that the person's heart rate has a slight fluctuation in a certain period of time. This fluctuation information and related time points and other information constitute part of the first dynamic event recognition result. Then, based on this first dynamic event recognition result and the event recognition results of the health monitoring data the day before yesterday, the dynamic event recognition feature vector is determined. For example, the data from the day before yesterday show that the person's blood pressure is within the normal range and is relatively stable. Combined with the information of yesterday's heart rate fluctuation, the computer system may obtain a comprehensive dynamic event recognition feature vector by some calculation method, such as weighted summation of different event features. Assuming that the weight of the heart rate fluctuation information is w1, the weight of the blood pressure stability information is w2, the heart rate fluctuation feature vector is F1, and the blood pressure stability feature vector is F2, then the dynamic event recognition feature vector D=w1F1+w2F2 (this is just a simplified example calculation method, the actual situation may be more complicated).

[0033] In this way, the dynamic event recognition feature vector can carry the past state information of some events in the health monitoring data to be analyzed. Back to the restoration mapping operation, the computer system can obtain more accurate event recognition results by combining the monitoring recognition feature vector and the dynamic event recognition feature vector for processing. For example, the health monitoring data to be analyzed collected today shows that the heart rate fluctuates again, but the fluctuation range is slightly larger than yesterday, and the blood pressure also rises slightly. The computer system inputs the monitoring recognition feature vector (containing the processed features of today's heart rate, blood pressure and other data) and the dynamic event recognition feature vector (containing past information such as yesterday's heart rate fluctuations and the day before yesterday's blood pressure stability) into the restoration mapping model (such as the multi-layer perceptron network mentioned above). After calculation, the output event recognition result vector may show that the confidence of the "mild abnormality" category is high, which indicates that the computer system judges that the health status of this person is in a mild abnormal state based on the past data and the current data, and further attention may be needed.

[0034] As an implementation manner, the plurality of preceding monitoring data include first health monitoring data and second health monitoring data; and the dynamic event recognition feature vector is obtained by the following steps: Step S301: extracting a first dynamic event recognition result from the event recognition results of the first health monitoring data; Step S302: Determine a dynamic event recognition feature vector based on the first dynamic event recognition result and the event recognition result of the second health monitoring data.

[0035] When the computer system executes step S301, it is necessary to screen and extract information related to dynamic events from the event recognition results of the first health monitoring data to form the first dynamic event recognition results. The first health monitoring data here is a part of the multiple preceding monitoring data corresponding to the health monitoring data to be analyzed, and its event recognition results contain rich information, but the computer system focuses on key information that can reflect the dynamic changes of events.

[0036] For example, suppose that the first health monitoring data is the health monitoring data of a person the day before, which includes the monitoring values ​​of multiple physiological indicators such as heart rate, blood pressure, body temperature, and the corresponding event recognition results. The event recognition results show that between 2 and 3 pm that day, the person's heart rate briefly increased, exceeding the normal range. When extracting the first dynamic event recognition result, the computer system will pay attention to the event of increased heart rate and extract relevant information, such as the time interval of the event (2 to 3 pm), the magnitude of the increase in heart rate (for example, from the normal 70 beats / minute to 90 beats / minute), etc. This information constitutes part of the first dynamic event recognition result.

[0037] In order to realize this extraction process, the computer system can adopt some implementation methods. One feasible method is rule-based extraction. The computer system predefines some rules to determine which event recognition results belong to dynamic events and determine the information content that needs to be extracted. For example, for the heart rate indicator, a rule can be set: if the change amplitude of the heart rate in a certain time period exceeds a certain threshold (such as 10 times / minute), it is considered a dynamic event, and it is necessary to extract information such as the time period and change amplitude. Another method is a classification model based on machine learning. The computer system can use a trained classification model to classify the event recognition results into dynamic events and non-dynamic events. For the results classified as dynamic events, further extract relevant information. For example, a support vector machine (SVM) model can be used, and the feature vector of the event recognition result can be used as input. The model outputs the judgment result of whether the event is a dynamic event. If it is a dynamic event, the computer system extracts the corresponding information from the result.

[0038] In step S302, after obtaining the first dynamic event recognition result, the computer system determines the dynamic event recognition feature vector in combination with the event recognition result of the second health monitoring data. The second health monitoring data is also the previous monitoring data corresponding to the health monitoring data to be analyzed, and its monitoring time point is before the first health monitoring data.

[0039] For example, suppose the second health monitoring data is the health monitoring data of a person two days ago, and its event recognition result shows that between 8 and 9 pm that day, the person's blood pressure fluctuated slightly, but was still within the normal range. Combined with the first dynamic event recognition result extracted earlier (the heart rate increased between 2 and 3 pm the previous day), the computer system comprehensively considers the information of these two events to determine the dynamic event recognition feature vector.

[0040] One way to determine the dynamic event recognition feature vector is feature fusion. The computer system fuses the relevant features in the first dynamic event recognition result and the event recognition result of the second health monitoring data to form a comprehensive feature vector. For example, the amplitude and occurrence time of the heart rate increase and the amplitude and occurrence time of the blood pressure fluctuation can be spliced ​​to obtain a new vector as the dynamic event recognition feature vector. Assume that the feature vector of the first dynamic event recognition result is expressed as , where t1 represents the time when the heart rate increase event occurs, represents the amplitude of the heart rate increase; the feature vector of the event recognition result of the second health monitoring data is expressed as , where t2 represents the time of occurrence of blood pressure fluctuation event, represents the amplitude of blood pressure fluctuation. Then the dynamic event recognition feature vector D can be obtained by concatenation: .

[0041] Another implementation method is weighted fusion. The computer system assigns different weights to the features in the first dynamic event recognition result and the event recognition result of the second health monitoring data according to the importance or relevance of different events, and then performs weighted summation to obtain the dynamic event recognition feature vector. For example, assuming that the heart rate increase event has a greater impact on the overall health status, the weight is w1=0.6; the weight of the blood pressure fluctuation event is w2=0.4. Then the dynamic event recognition feature vector D can be calculated by weighted summation: .

[0042] In practical applications, the computer system can also use the attention mechanism in deep learning to automatically learn the importance of different event features. The attention mechanism assigns dynamic weights to each feature based on the correlation between events and the degree of contribution to the final result. Specifically, the computer system first inputs the first dynamic event recognition result and the event recognition result of the second health monitoring data into an attention network. The attention network calculates the attention weight of each feature, and then performs weighted summation of the features according to the weight to obtain a dynamic event recognition feature vector.

[0043] Assuming that the input of the attention network is the feature vector F=[F1, F2], and the attention weight calculation function is a(F), it will output a weight vector W=[w1, w2] with the same dimension as the feature vector, where w1 and w2 are the weights of F1 and F2 respectively. Then the dynamic event recognition feature vector D can be obtained by the formula Calculated.

[0044] Suppose there is a health monitoring system that continuously monitors a person's health data. The first health monitoring data is yesterday's data, and the event recognition result shows that between 10 and 11 a.m., the person's exercise steps suddenly increased, 2,000 more than usual; the second health monitoring data is the day before yesterday's data, and the event recognition result shows that between 4 and 5 p.m., the person's sleep time decreased by 1 hour compared to usual.

[0045] The computer system first extracts the first dynamic event recognition result from the event recognition result of the first health monitoring data, and obtains the relevant information of the increase in the number of exercise steps, such as the time of the event (10:00 to 11:00 a.m.) and the increased number of steps (2000 steps), which is represented as the feature vector F1=[10:00-11:00, 2000]. Then, the computer system obtains the information of the reduction in sleep duration from the event recognition result of the second health monitoring data, such as the time of the event (4:00 to 5:00 p.m.) and the reduced duration (1 hour), which is represented as the feature vector F2=[16:00-17:00, -1] (negative numbers are used here to represent reduction).

[0046] If the feature fusion method is adopted, the dynamic event recognition feature vector D can be obtained by concatenation: D = [10:00-11:00, 2000, 16:00-17:00, -1].

[0047] If the weighted fusion method is adopted, assuming that the weight of the event of increased movement steps is w1=0.7, and the weight of the event of reduced sleep duration is w2=0.3, the dynamic event recognition feature vector D can be calculated by weighted summation: D=0.7[10:00-11:00, 2000]+0.3[16:00-17:00, -1]=[0.7×(10:00-11:00)+0.3×(16:00-17:00), 0.7×2000+0.3×(-1)].

[0048] If the attention mechanism is used, the attention network will automatically calculate the attention weight W=[w1, w2] according to the correlation between the two events and the overall health status, and then use the formula The dynamic event recognition feature vector is calculated.

[0049] As an implementation manner, step S302, based on the first dynamic event recognition result and the event recognition result of the second health monitoring data, determines a dynamic event recognition feature vector, including: Step S3021: determining a second dynamic event recognition result based on the event monitoring time point distribution vector in the first dynamic event recognition result and the event content feature vector in the event recognition result of the second health monitoring data; Step S3022: Determine a dynamic event identification feature vector based on the second dynamic event identification result and the first dynamic event identification result; wherein the health monitoring data to be analyzed and multiple preceding monitoring data corresponding to the health monitoring data to be analyzed belong to a monitoring data set of the same monitoring object; the monitoring time point of the health monitoring data to be analyzed is after the monitoring time point of the first health monitoring data; the monitoring time point of the first health monitoring data is after the monitoring time point of the second health monitoring data.

[0050] In step S3021 of the embodiment of the present application, the computer system obtains the event monitoring time point distribution vector from the first dynamic event recognition result, and combines the event content feature vector in the event recognition result of the second health monitoring data to determine the second dynamic event recognition result through a specific implementation method.

[0051] The event monitoring time point distribution vector reflects the distribution of events in the time dimension. For example, suppose the first dynamic event recognition result is about a person's health monitoring data from the previous day, which involves an event of increased heart rate. The computer system extracts the event monitoring time point distribution vector from the result. Suppose this vector is expressed as , where t i Indicates the specific time points of the heart rate increase event in different time periods. For example, t1=14:30 means that the heart rate starts to increase at 2:30 pm, and t2=15:15 means that the heart rate reaches its peak at 3:15 pm.

[0052] The event content feature vector contains specific content features related to the event. For example, the second health monitoring data is the health monitoring data of a person two days ago, and the event content feature vector in the event recognition result is related to the blood pressure fluctuation event. Assume that this event content feature vector is expressed as ,in Indicates different features of blood pressure fluctuation events. For example, c1 may represent the amplitude of blood pressure fluctuation, c2 may represent the duration of blood pressure fluctuation, etc.

[0053] In order to determine the second dynamic event recognition result, the computer system can adopt a variety of implementation methods. One feasible method is a method based on time series analysis. The computer system first analyzes the event monitoring time point distribution vector T and calculates statistical information such as the interval between time points and the duration of the event. For example, the difference between adjacent time points is calculated. , to understand the changing speed of the heart rate increase event. Then, these statistical information are fused with the event content feature vector C. For example, the time interval information and the amplitude information of blood pressure fluctuation can be associated and analyzed. Assume that through a certain association rule or model, it is found that when the time interval of the heart rate increase event is short and the blood pressure fluctuation amplitude is large, it may imply that the body is in a specific health state. The computer system determines the second dynamic event recognition result based on this association relationship.

[0054] Specifically, suppose a correlation function f(T, C) is defined, which outputs a vector representing the second dynamic event recognition result according to the specific values ​​of the event monitoring time point distribution vector T and the event content feature vector C. . For example, the link function can be a simple linear combination function: ,in is a pre-set weight coefficient, and g(T) is the feature vector obtained after processing the event monitoring time point distribution vector T, for example .

[0055] For example, suppose the event monitoring time point distribution vector T=[14:30, 15:15, 15:45] in the first dynamic event recognition result, indicating that the heart rate increase event started at 2:30 pm, reached a peak at 3:15 pm, and gradually returned to normal at 3:45 pm. The event content feature vector C=[10, 30] in the event recognition result of the second health monitoring data, where c1=10 indicates that the amplitude of the blood pressure fluctuation is 10 mmHg, and c2=30 indicates that the duration of the blood pressure fluctuation is 30 minutes. Assuming the weight coefficient =0.4, =0.6, and the event monitoring time point distribution vector T is processed to obtain g(T)=[45, 30] (representing the intervals between adjacent time points in minutes). Then the second dynamic event recognition result R2 is calculated according to the correlation function: R2=0.4×[45, 30]+0.6×[10, 30]=[0.4×45+0.6×10, 0.4×30+0.6×30]=[24, 30].

[0056] In step S3022, after obtaining the second dynamic event recognition result, the computer system combines it with the first dynamic event recognition result to determine a final dynamic event recognition feature vector.

[0057] The first dynamic event recognition result contains rich information. In addition to the event monitoring time point distribution vector mentioned above, it may also contain other event-related features. For example, for a heart rate increase event, the first dynamic event recognition result may also include information such as the maximum amplitude of the heart rate increase and the time when the heart rate returns to normal. Assume that the first dynamic event recognition result is expressed as ,in Represents different features of the first dynamic event recognition result.

[0058] When determining the dynamic event recognition feature vector, the computer system can use feature fusion to achieve the goal. A simple feature fusion method is to directly splice the second dynamic event recognition result R2 and the first dynamic event recognition result R1. For example, assuming , then the dynamic event recognition feature vector D can be obtained by concatenation: .

[0059] Another feature fusion method is weighted fusion. The computer system assigns corresponding weights to each feature according to the importance of different features in the second dynamic event recognition result and the first dynamic event recognition result, and then performs weighted summation to obtain the dynamic event recognition feature vector. Assume that the weight vector is assigned to the features in the first dynamic event recognition result R1. , assigns a weight vector to the features in the second dynamic event recognition result R2 , then the dynamic event recognition feature vector D can be calculated by weighted summation: .

[0060] For example, suppose the first dynamic event recognition result R1=[15, 80, 16:00], where r 11 =15 means the maximum increase in heart rate is 15 beats / minute, r 12 =80 means the highest heart rate is 80 beats / minute. 13 =16:00 means that the heart rate returns to normal at 4 pm. The second dynamic event recognition result R2=[24, 30], as calculated above. Assuming that the weight vector W1=[0.3, 0.4, 0.3], W2=[0.5, 0.5], the dynamic event recognition feature vector D is: D=0.3×15+0.4×80+0.3×16:00+0.5×24+0.5×30 (the time value here may need to be appropriately converted in actual calculations, such as converted to minutes, etc.).

[0061] As an implementation mode, step S300 performs a restoration mapping operation based on the monitoring identification feature vector and the dynamic event identification feature vector corresponding to the health monitoring data to be analyzed to obtain an event identification result of the health monitoring data to be analyzed, including: Step S310: determining a to-be-processed identification feature vector based on the monitoring identification feature vector and the dynamic event identification feature vector; Step S320: performing a restoration mapping operation based on the identification feature vector to be processed to obtain an event identification result of the health monitoring data to be analyzed.

[0062] In step S310, when the computer system executes this step, it effectively merges the monitoring identification feature vector and the dynamic event identification feature vector to generate a to-be-processed identification feature vector. The key to this process is how to reasonably merge two different but interrelated feature vectors so that the merged vector can contain enough information to provide strong support for the subsequent restoration mapping operation.

[0063] The monitoring identification feature vector is obtained by performing representation learning on the health monitoring data to be analyzed and then undergoing an embedding mapping operation. It contains the current feature information of the health monitoring data to be analyzed. For example, assuming that the monitoring identification feature vector is represented as , where m i Represents different feature dimensions, which may correspond to the encoding representation of health monitoring indicators such as heart rate, blood pressure, body temperature, etc. at the current moment.

[0064] The dynamic event recognition feature vector is determined based on the event recognition results of multiple previous monitoring data corresponding to the health monitoring data to be analyzed, which reflects the past state information of the event. For example, the dynamic event recognition feature vector is expressed as , where d j Represents features related to the dynamic changes of events, such as the fluctuation of heart rate in a previous period of time, the changing trend of blood pressure, etc.

[0065] A common implementation method for a computer system to determine the identification feature vector to be processed is vector concatenation. That is, the monitoring identification feature vector M and the dynamic event identification feature vector D are directly concatenated together to form a new vector as the identification feature vector to be processed. Assuming that the identification feature vector to be processed is represented by P, the following is obtained by vector concatenation: For example, if the monitoring recognition feature vector M=[0.2, 0.3, 0.4] represents a certain coding feature of the current heart rate, blood pressure, and body temperature, and the dynamic event recognition feature vector D=[0.1, 0.05] represents the trend characteristics of the heart rate and blood pressure over the past period of time, then the recognition feature vector to be processed obtained by vector concatenation is P=[0.2, 0.3, 0.4, 0.1, 0.05].

[0066] Another implementation method is weighted fusion. The computer system assigns corresponding weights to each feature according to the importance of different features in the monitoring recognition feature vector and the dynamic event recognition feature vector, and then performs weighted summation to obtain the recognition feature vector to be processed. Assume that the weight vector is assigned to the features in the monitoring recognition feature vector M. , assign weight vectors to the features in the dynamic event recognition feature vector D , then the feature vector P to be processed can be calculated by weighted summation: For example, if M = [0.2, 0.3, 0.4], W M =[0.4, 0.3, 0.3], D=[0.1, 0.05], W D =[0.5, 0.5], then P=0.4×0.2+0.3×0.3+0.3×0.4+0.5×0.1+0.5×0.05=0.335.

[0067] In addition, a fusion method based on a neural network can also be used. A fusion network is constructed, and the monitoring identification feature vector and the dynamic event identification feature vector are used as input. Through the learning and training of the neural network, the features are automatically extracted and fused, and the identification feature vector to be processed is output. For example, the fusion network can contain multiple hidden layers, each hidden layer processes the input features through nonlinear transformation, and finally outputs the identification feature vector to be processed.

[0068] In step S320, after obtaining the identification feature vector to be processed, the computer system performs a restoration mapping operation on it to obtain the event recognition result of the health monitoring data to be analyzed. The purpose of the restoration mapping operation is to remap the identification feature vector to be processed back to a representation closer to the original data semantic space, so as to obtain specific information about the event contained in the data, such as the category confidence corresponding to the event.

[0069] One feasible way to implement the restoration mapping operation is to use a multilayer perceptron (MLP) network. A multilayer perceptron network is a feedforward neural network consisting of an input layer, multiple hidden layers, and an output layer. The computer system takes the feature vector P to be processed as the input of the multilayer perceptron network. Assume that the dimension of the feature vector P to be processed is p, and the number of neurons in the input layer is also p. In the hidden layer, each neuron is connected to all neurons in the previous layer and connected through weights. Assume that the lth hidden layer has h l neurons, the weight matrix from the input layer to the first hidden layer is W1, with a dimension of p×h1, and the bias vector is b1, with a dimension of h1. Then the output H1 of the first hidden layer can be obtained by the formula Calculated, where is an activation function, such as the commonly used ReLU function ( (x)=max(0, x)).

[0070] The calculation method of subsequent hidden layers is similar. The output H_m of the mth hidden layer (m>1) is calculated by the output H_m of the previous hidden layer. m-1 To calculate, the formula is ,in is the weight matrix from the m-1th layer to the mth layer, b m is the bias vector of the mth layer.

[0071] Finally, the output layer maps the output of the last hidden layer to the space of event recognition results. Assume that the weight matrix of the output layer is W o , the dimension is (h last is the number of neurons in the last hidden layer, c is the number of event categories), the bias vector is b o , with a dimension of c. Then the event recognition result vector R can be obtained by the formula Here, each element in the event recognition result vector R represents the confidence score of the corresponding event category. For example, if there are three event categories (normal, slightly abnormal, and severely abnormal), then R=[r1, r2, r3], where r1 represents the confidence that the data belongs to the "normal" category, r2 represents the confidence that the data belongs to the "mildly abnormal" category, and r3 represents the confidence that the data belongs to the "severely abnormal" category.

[0072] In order to obtain a more reasonable probability distribution, the event recognition result vector R can be normalized by applying the softmax function.

[0073] Another way to implement the restoration mapping operation is a classification method based on a support vector machine (SVM). The computer system first trains an SVM classifier, using health monitoring data with known category labels as training samples, and inputs the feature vector P to be processed into the trained SVM classifier. The classifier classifies the feature vector into different event categories based on its characteristics, and outputs the corresponding category label and confidence.

[0074] For example, suppose that the trained SVM classifier can identify three event categories: "normal", "mild abnormality" and "severe abnormality". When the feature vector P to be processed is input, the classifier determines that it belongs to the "mild abnormality" category based on the characteristics of the feature vector, and outputs the category label "mild abnormality" and the confidence level 0.8, indicating that the confidence level of the data belonging to the "mild abnormality" category is 0.8.

[0075] As an implementation method, the event recognition result of the health monitoring data to be analyzed is determined by a previously debugged converged reduction mapping model; the reduction mapping model includes one or more reduction mapping networks, the internal attention operator in the reduction mapping network is constructed by a first selective attention filter, and the first selective attention filter is used to block the information exchange between the corresponding identification feature vectors belonging to the same event under different previous monitoring data; the one or more events include at least one event to be continuously monitored; the one or more reduction mapping networks include a first reduction mapping network and a second reduction mapping network; the reduction mapping model also includes a feature vector integration network; the feature vector integration network is configured to receive the information input by the first reduction mapping network. The dynamic event recognition data outputted are integrated to obtain dynamic event recognition integrated data; the dynamic event recognition data is determined based on the first restoration mapping network and the dynamic event recognition feature vector; the dynamic event recognition data includes the processed recognition feature vectors obtained by the first restoration mapping network for each event to be continuously monitored; the dynamic event recognition integrated data includes the integrated recognition feature vectors outputted by the feature vector integration network for each processed recognition feature vector; the second restoration mapping network is configured to process the dynamic event recognition integrated data and the event recognition data outputted by the first restoration mapping network; the event recognition data is determined based on the first restoration mapping network and the monitoring recognition feature vector.

[0076] In the embodiment of the present application, the restoration mapping model is a model that has been debugged and converged in advance, and its purpose is to accurately restore the event recognition results of the health monitoring data to be analyzed based on the relevant feature vectors input. The model consists of one or more restoration mapping networks and feature vector integration networks. This architectural design enables the model to fully process and integrate information from different sources, thereby more accurately identifying health monitoring events.

[0077] The restoration mapping network is one of the core components of the restoration mapping model. The internal attention operator is constructed through the first selective attention filter (attention mask). The role of this first selective attention filter is crucial. It can block the information exchange between the corresponding identification feature vectors of the same event under different previous monitoring data. For example, assuming that in health monitoring, "abnormal heart rate" is an event to be continuously monitored, there may be identification feature vectors about "abnormal heart rate" in the previous monitoring data collected at different time points. If not blocked, unnecessary interference may occur between these feature vectors, affecting the accurate judgment of the "abnormal heart rate" event in the current health monitoring data to be analyzed. Through the first selective attention filter, the computer system can selectively focus on and process relevant information to avoid such interference.

[0078] Specifically, suppose that in the previous monitoring data at a certain moment, the identification feature vector of the "abnormal heart rate" event is X1, and in the previous monitoring data at another moment, the identification feature vector of the same event is X2. The first selective attention filter can achieve information blocking by setting a mask matrix M. When calculating the attention weight, for the feature vectors under different previous monitoring data belonging to the same event, the corresponding elements in the mask matrix M will be set to 0, so that when calculating the attention score, the information exchange between these feature vectors is blocked. For example, the attention score can be calculated by the formula to achieve this, where Q and K are the query vector and key vector respectively. is the dimension of the key vector, Represents an element-by-element multiplication operation. In this way, the computer system can more accurately capture the unique information related to the event in each previous monitoring data.

[0079] The one or more events include at least one event to be continuously monitored, that is, an event that needs to be dynamically tracked, such as the aforementioned "abnormal heart rate", etc. In this example, the restoration mapping network includes a first restoration mapping network and a second restoration mapping network.

[0080] The first restoration mapping network is responsible for processing data related to the dynamic event identification feature vector. It receives the dynamic event identification feature vector as input, and performs processing operations on each event to be continuously monitored to obtain the processed identification feature vector, which constitutes the dynamic event identification data. For example, for the "abnormal heart rate" event to be continuously monitored, the first restoration mapping network will process the dynamic event identification feature vector according to the relevant information, such as the changes in heart rate in different time periods, the fluctuation range, etc., to obtain a processed identification feature vector that can more accurately reflect the characteristics of the event.

[0081] The second restoration mapping network mainly processes the dynamic event recognition integration data and the event recognition data output by the first restoration mapping network. The event recognition data is determined based on the first restoration mapping network and the monitoring recognition feature vector. For example, the monitoring recognition feature vector may contain information on various health monitoring indicators such as heart rate and blood pressure at the current moment. The first restoration mapping network will combine this information for processing to obtain event recognition data. The second restoration mapping network will comprehensively consider the dynamic event recognition integration data and event recognition data, and further mine and extract the information therein to obtain more accurate event recognition results.

[0082] The feature vector integration network is configured to integrate the dynamic event recognition data output by the first restoration mapping network to obtain dynamic event recognition integrated data. The dynamic event recognition integrated data includes the integrated recognition feature vectors respectively output by the feature vector integration network for each processed recognition feature vector.

[0083] The feature vector integration network can be implemented in a variety of ways to achieve the integration function. One feasible method is to use weighted summation. Assume that the dynamic event recognition data output by the first reduction mapping network contains n processed recognition feature vectors, which are represented as , the feature vector integration network assigns a weight to each feature vector , then dynamic event recognition integrates data The formula can be For example, for the processed identification feature vectors V1 and V2 corresponding to the two events to be continuously monitored, namely, “abnormal heart rate” and “abnormal blood pressure”, if the assigned weights are w1=0.6 and w2=0.4 respectively, the dynamic event recognition integrated data .

[0084] Another approach is integration based on the attention mechanism. The feature vector integration network can dynamically determine the importance of each processed recognition feature vector by calculating the attention weight of each feature vector. For example, by calculating the similarity score between each feature vector and a query vector Q (where d is the dimension of the feature vector), and then apply the softmax function to convert the similarity scores into attention weights Finally, dynamic event recognition integrates data The formula can be Calculated.

[0085] Through the integration operation of the feature vector integration network, the computer system can effectively fuse the processed identification feature vectors of different events to be continuously monitored, so that the final dynamic event identification integration data can more comprehensively and accurately reflect the event information in the health monitoring data.

[0086] As an implementation method, the feature vector integration network includes a feature reduction subcomponent and a feature enhancement subcomponent; the dynamic event recognition integration data is obtained by the following steps, including: Step S10: Processing the dynamic event recognition data through the feature reduction subcomponent to obtain a retained feature vector; Step S20: Processing the dynamic event recognition data through the feature enhancement subcomponent to obtain a pseudo-enhanced feature vector; Step S30: Determine dynamic event recognition integrated data based on the retained feature vectors and the feature vectors to be enhanced.

[0087] In step S10 of the embodiment of the present application, the computer system processes the dynamic event recognition data using the feature reduction subcomponent, with the purpose of selecting the most representative and valuable features from the original dynamic event recognition data to form a retained feature vector. This process involves multiple specific operation steps to achieve effective reduction of features.

[0088] First, the time series position extraction of the dynamic event recognition data is performed to obtain the first time series position. Time series position extraction refers to extracting time-related information from the dynamic event recognition data, which can reflect the distribution and changes of events in the time series. For example, suppose that the dynamic event recognition data is about the health monitoring data of a person over a period of time, which contains the values ​​of indicators such as heart rate and blood pressure at different time points. The computer system extracts the specific time point of each event from these data through specific algorithms or rules, and these time points constitute the first time series position. For example, for abnormal heart rate events, the computer system may record the specific time when each abnormal heart rate occurs, such as t1=10:30, t2=11:15, etc. These moments constitute the first time series position set T1={t1, t2,…}.

[0089] Next, the computer system performs internal weight focusing (that is, self-attention processing) on ​​the first time series position and the dynamic event recognition data to obtain an internal weight focusing result. Internal weight focusing is an operation based on the self-attention mechanism, which can assign corresponding weights to each element according to the correlation between different elements in the data, thereby highlighting important information. In this process, the computer system calculates the correlation score between the first time series position and other features in the dynamic event recognition data. Taking the abnormal heart rate event as an example, it is assumed that in addition to time information, the dynamic event recognition data also contains features such as the amplitude and duration of heart rate changes. The computer system calculates the correlation between each time point and these features, for example, by calculating the similarity or correlation index between them. Assuming that cosine similarity is used to calculate the correlation score, for the time point and a certain eigenvector , relevance score The formula can be Calculated, where represents the vector dot product, Represents the modulus of the vector. Then, the relevance score is converted into a weight through the softmax function Finally, the dynamic event recognition data is weighted and summed according to these weights to obtain the internal weight focusing result. For example, for a certain feature vector x j , the internal weight focuses on the result y j The formula can be Calculated.

[0090] Then, the computer system performs global information integration (e.g., full connection) on the internal weight focusing results to obtain the global information integration results. The purpose of global information integration is to further fuse and abstract the local information after the internal weight focusing process to obtain a more comprehensive feature representation. Full connection is a feasible global information integration method, which uses a fully connected layer to linearly transform and nonlinearly activate the internal weight focusing results to obtain the global information integration results. Assume that the internal weight focusing result is , the weight matrix of the fully connected layer is W, the bias vector is b, and the activation function is , then the global information integration result z can be obtained by formula Calculated.

[0091] Based on the global information integration result, the computer system determines the feature vector to be filtered. The feature vector to be filtered is the feature vector to be filtered out according to the importance or relevance of each feature in the global information integration result. For example, the computer system can set a threshold. When the weight or score of a feature in the global information integration result is lower than the threshold, it is considered that the feature contributes less to the final result and can be used as the feature vector to be filtered.

[0092] Finally, based on the quasi-filtered feature vector, the computer system simplifies the dynamic event recognition data to obtain the retained feature vector. Specifically, the computer system removes the features corresponding to the quasi-filtered feature vector from the dynamic event recognition data and retains the remaining features, which constitute the retained feature vector. For example, assuming that the dynamic event recognition data contains 10 features, and after the previous processing, it is determined that the vectors corresponding to 3 of the features are quasi-filtered feature vectors, then the computer system removes these 3 features, and the remaining 7 features constitute the retained feature vector.

[0093] In step S20, the computer system processes the dynamic event recognition data using the feature enhancement subcomponent to obtain a proposed enhanced feature vector, thereby further highlighting and strengthening the key information in the data.

[0094] First, the computer system extracts the time series position of the dynamic event recognition data to obtain the second time series position. Similar to the time series position extraction in step S10, the second time series position here also extracts time-related information from the dynamic event recognition data, but may adopt different extraction methods or focus on different time features. For example, in step S10, the focus is on the specific time point when the event occurs, while here more attention may be paid to information such as the duration or time interval of the event. Assume that for a person's health monitoring data, the second time series position extracted by the computer system is the duration of each abnormal heart rate event, such as d1=15 minutes, d2=20 minutes, etc. These durations constitute the second time series position set T2={d1, d2,…}.

[0095] Then, the computer system performs internal weight focusing on the second time series position and the dynamic event recognition data to obtain a quasi-enhanced feature vector. This process is similar to the internal weight focusing in step S10, which is also performed by calculating the correlation score between the second time series position and other features in the dynamic event recognition data, converting the score into a weight, and then weighted summing the data according to the weight. The difference is that the purpose here is to enhance certain features in the data to make the key information more prominent. For example, for the duration d of the abnormal heart rate event i and other eigenvectors x k , calculate the correlation score S ik , and converted to weight w ik , and then get the quasi-enhanced feature vector e k , the calculation formula is similar to the internal weight focusing in step S10, that is, .

[0096] In step S30, after obtaining the retained feature vector and the proposed enhanced feature vector, the computer system combines the two to determine the dynamic event recognition integrated data. The key to this process is how to effectively fuse the two feature vectors to make full use of the information they contain. One feasible method is direct splicing. The computer system directly splices the retained feature vector and the proposed enhanced feature vector together to form a new vector as the dynamic event recognition integrated data. Another method is weighted fusion. The computer system assigns corresponding weights to each feature according to the importance of different features in the retained feature vector and the proposed enhanced feature vector, and then performs weighted summation to obtain the dynamic event recognition integrated data. Through the above steps S10, S20 and S30, the computer system can filter out important retained feature vectors from the dynamic event recognition data, enhance key information to obtain the proposed enhanced feature vector, and finally effectively integrate the two to obtain dynamic event recognition integrated data. This data integration method can more comprehensively and accurately reflect the event information in the health monitoring data, and provide strong support for subsequent health event recognition and analysis. For example, in actual health monitoring applications, by analyzing the integrated data of dynamic event recognition, the computer system can more accurately determine whether a person's health status is abnormal, as well as the type and degree of the abnormality, thereby providing a basis for taking appropriate medical measures in a timely manner.

[0097] In the embodiment of the present application, the retained feature vector is obtained by the following steps: Step S11: extracting the time sequence position of the dynamic event recognition data to obtain a first time sequence position; Step S12: performing internal weight focusing on the first time series position and the dynamic event recognition data to obtain an internal weight focusing result; Step S13: performing global information integration on the internal weight focusing results to obtain a global information integration result; Step S14: determining a feature vector to be filtered based on the global information integration result; Step S15: Based on the proposed filtering feature vector, the dynamic event recognition data is simplified to obtain a retained feature vector.

[0098] In the embodiment of the present application, the proposed enhanced feature vector is obtained by the following steps: Step S21: extracting the time sequence position of the dynamic event recognition data to obtain a second time sequence position; Step S22: performing internal weight focusing on the second temporal position and the dynamic event recognition data to obtain a pseudo-enhanced feature vector.

[0099] In step S11, when processing dynamic event recognition data, the computer system performs a time series position extraction operation to obtain a first time series position. Time series position extraction refers to extracting specific time-related information from dynamic event recognition data, which can reflect the distribution and change patterns of events in time series. For example, in health monitoring data, dynamic event recognition data may contain multiple physiological indicator data of a person over a period of time, such as heart rate, blood pressure, body temperature, etc., as well as related information on abnormal changes in these indicators. The computer system pays attention to the specific time points when each abnormal event occurs, and uses these time points as the first time series position.

[0100] One implementation method is to traverse and filter the dynamic event recognition data. The computer system checks the data records one by one in chronological order, and when a data point that meets a specific abnormal condition (such as a heart rate exceeding the normal range) is detected, the time information corresponding to the data point is recorded as the first time sequence position.

[0101] In step S12, after obtaining the first temporal position, the computer system performs an internal weight focusing operation on the first temporal position and the dynamic event recognition data, which is actually a self-attention processing mechanism. The purpose of the self-attention mechanism is to assign corresponding weights to each element according to the correlation between different elements in the data, thereby highlighting important information.

[0102] Specifically, the computer system calculates the correlation score between the first time sequence position and other features in the dynamic event recognition data. Taking the abnormal heart rate event as an example, in addition to time information, the dynamic event recognition data may also contain features such as the heart rate change amplitude and duration. For each time point t at the first time sequence position i and other eigenvectors x j , the computer system can calculate the similarity between them to get the correlation score S ij One possible similarity calculation method is cosine similarity.

[0103] After obtaining the relevance scores, the computer system converts these scores into weights through the softmax function. In this way, each feature vector will get a corresponding weight. Finally, the dynamic event recognition data is weighted and summed according to these weights to obtain the internal weight focusing result. For a certain feature vector.

[0104] In step S13, after obtaining the internal weight focusing result, the computer system performs global information integration on it to obtain a more comprehensive feature representation. Full connection is a commonly used global information integration method, which linearly transforms and nonlinearly activates the internal weight focusing result through a fully connected layer to obtain a global information integration result.

[0105] Assume that the internal weight focusing result is , the weight matrix of the fully connected layer is W, with a dimension of m×n (where m is the dimension of the output of the fully connected layer and n is the dimension of the internal weight focusing result), the bias vector is b, with a dimension of m, and the activation function is \sigma (for example, the ReLU function, ). Then the global information integration result z can be calculated by the following formula: .

[0106] For example, assuming the internal weight focusing result y=[0.5, 0.3, 0.2], the weight matrix of the fully connected layer is , bias vector b=[0.1, 0.2], then: ; After applying the ReLU activation function, the global information integration result .

[0107] In step S14, based on the global information integration result, the computer system determines the feature vector to be filtered. The feature vector to be filtered is the feature vector to be filtered out according to the importance or relevance of each feature in the global information integration result.

[0108] One feasible method is to set a threshold. When the weight or score of a feature in the global information integration result is lower than the threshold, the feature is considered to contribute less to the final result and can be used as a quasi-filter feature vector. For example, assuming that the threshold is set to 0.3, in the global information integration result z=[0.27, 0.67], 0.27 is lower than the threshold, then the corresponding feature vector can be determined as a quasi-filter feature vector.

[0109] In step S15, after determining the feature vectors to be filtered, the computer system simplifies the dynamic event recognition data according to these vectors to obtain the retained feature vectors. Specifically, the computer system removes the features corresponding to the feature vectors to be filtered from the dynamic event recognition data and retains the remaining features, which constitute the retained feature vectors.

[0110] For example, assuming that the dynamic event recognition data contains the feature vector X=[x1, x2, x3, x4, x5], after the previous processing, it is determined that the vector corresponding to x3 is the quasi-filtered feature vector, then the computer system removes x3, and the remaining feature vectors [x1, x2, x4, x5] constitute the retained feature vector.

[0111] Step S21 is similar to step S11. When the computer system obtains the feature vector to be enhanced, it also needs to extract the time sequence position of the dynamic event recognition data, but the second time sequence position is obtained here. The second time sequence position also extracts time-related information from the dynamic event recognition data, but may focus on different time features.

[0112] For example, in health monitoring data, in addition to paying attention to the specific time point when an abnormal event occurs, the computer system may also pay attention to the duration of the event. For the abnormal heart rate events mentioned above, in addition to recording the starting time point, the duration of each abnormal event is also recorded. Suppose the abnormal heart rate event on Monday starts at 10:30 and ends at 11:00, with a duration of 30 minutes; the abnormal heart rate event on Tuesday starts at 14:15 and ends at 14:45, with a duration of 30 minutes; the abnormal heart rate event on Wednesday starts at 13:20 and ends at 13:50, with a duration of 30 minutes. The computer system extracts these duration information as the second time series position to form a set T2={30 周一 , 30 周二 , 30 周三}.

[0113] One implementation method is to calculate the time interval of each abnormal event in the dynamic event recognition data. After the computer system detects the start time and end time of the abnormal event, it calculates the difference between the two to obtain the duration of the event and uses it as the second time sequence position.

[0114] In step S22, after obtaining the second temporal position, the computer system performs an internal weighted focusing operation on the second temporal position and the dynamic event recognition data to obtain a pseudo-enhanced feature vector. This process is similar to the internal weighted focusing in step S12, and is also performed by calculating the correlation score between the second temporal position and other features in the dynamic event recognition data, converting the score into a weight, and then performing a weighted summation on the data according to the weight.

[0115] For example, for the duration 30 in the second temporal position 周一 and other eigenvectors x k (such as the heart rate variation amplitude feature vector), the computer system first calculates the correlation score S between them ik , a cosine similarity calculation method similar to that in step S12 can be used. Then, the correlation score is converted into a weight w by the softmax function ik , and finally get the quasi-enhanced feature vector e k .

[0116] As an implementation method, the event recognition result of the health monitoring data to be analyzed is determined by debugging a converged event monitoring model in advance; the model debugging error of the event monitoring model is obtained by the following steps: Step S101: determining an event recognition result of the first health monitoring data sample based on the first health monitoring data sample, event recognition results of a plurality of preceding monitoring data corresponding to the first health monitoring data sample, and an event monitoring model; Step S102: extracting a first recognition result and a second recognition result from the event recognition result of the first health monitoring data sample; Step S103: obtaining a first debugging error based on the first recognition result and the event priori mark corresponding to the first health monitoring data sample; Step S104: obtaining a second debugging error based on the second recognition result and the event priori mark; Step S105: determining a model debugging error of the event monitoring model based on the first debugging error and the second debugging error; The monitoring time point corresponding to the first recognition result is after the monitoring time point corresponding to the second recognition result.

[0117] In step S101, when calculating the model debugging error, the computer system first needs to determine the event recognition result of the first health monitoring data sample based on the given first health monitoring data sample, the event recognition results of multiple previous monitoring data corresponding to the sample, and the event monitoring model.

[0118] The first health monitoring data sample is a part of the health monitoring data used for model debugging. It contains various health-related feature information, such as a person's heart rate, blood pressure, body temperature and other physiological indicators in a specific time period. The event recognition results of multiple previous monitoring data provide relevant information about the individual's health events in the previous time period, which is important for understanding the dynamic changes in the current health status. The event monitoring model is a model that has been initially constructed or has undergone a certain amount of training, which can perform event recognition based on the input data.

[0119] The computer system inputs the first health monitoring data sample and the corresponding event recognition results of multiple preceding monitoring data into the event monitoring model. For example, assuming that the first health monitoring data sample is the health monitoring data of a patient on that day, including information such as a heart rate of 80 beats / minute and a blood pressure of 120 / 80 mmHg, the corresponding event recognition results of multiple preceding monitoring data show that the patient had a short-term abnormal increase in heart rate in the previous two days. After receiving these data, the event monitoring model calculates and analyzes them through its internal algorithms and parameters, and outputs the event recognition results of the first health monitoring data sample. For example, the model may output that the patient's current health status is "normal", or output a vector containing different event categories and their corresponding probabilities, such as [0.8, 0.1, 0.1], which respectively represent the probabilities of "normal", "mild abnormality" and "severe abnormality".

[0120] The specific implementation method can be a forward propagation algorithm based on a neural network. The event monitoring model is usually composed of multiple neural network layers. The data enters from the input layer, undergoes a series of calculations and transformations in the hidden layer, and finally obtains the event recognition result in the output layer. Assume that the event monitoring model is a multi-layer perceptron (MLP), the input layer receives the feature vector X composed of the first health monitoring data sample and the event recognition result of the previous monitoring data, the weight matrix of the hidden layer is W, the bias vector is b, and the activation function is, for example, the ReLU function.

[0121] In step S102, after obtaining the event recognition result of the first health monitoring data sample, the computer system needs to extract the first recognition result and the second recognition result. The two recognition results are divided from the event recognition result based on the time sequence or specific rules, and they represent the event recognition information in different time periods or different feature dimensions.

[0122] For example, assuming that the event recognition result is a sequence data containing multiple time steps, the first recognition result may be the event recognition information of an earlier time step in the sequence, and the second recognition result may be the event recognition information of a later time step. Specifically, if the event recognition result is a sequence of length 10, the event recognition information of the first 5 time steps is extracted as the first recognition result, and the event recognition information of the last 5 time steps is extracted as the second recognition result. Alternatively, based on the type or importance of the event, the event recognition information corresponding to certain key indicators is classified as the first recognition result, and the event recognition information corresponding to other indicators is classified as the second recognition result.

[0123] In terms of implementation, the extraction process can be implemented through indexing or slicing operations. If the event recognition result is stored in the form of an array or matrix in a computer system, the first recognition result and the second recognition result can be extracted from the event recognition result by specifying the corresponding index range or condition. For example, in the Python programming language, array slicing operations can be used to implement this.

[0124] In step S103, after extracting the first recognition result, the computer system compares it with the event priori mark corresponding to the first health monitoring data sample, thereby obtaining a first debugging error. The event priori mark is a known and accurate event category label, which represents the actual event situation of the first health monitoring data sample.

[0125] The first debugging error reflects the difference between the first recognition result and the actual situation. A common method for calculating the first debugging error is to use a loss function. A loss function is a function that measures the difference between the model's prediction result and the actual label. Different tasks and data types can choose different loss functions. For example, for classification tasks, the commonly used loss function is the cross entropy loss function.

[0126] Assume that the first recognition result is a vector containing the probabilities of different event categories ,in Represents the probability of belonging to the i-th event category. The event prior label is a one-hot encoded vector ,in Represents the real event category. When i is the real category, =1, otherwise = 0. The calculation formula of the cross entropy loss function is: .

[0127] For example, suppose the first recognition result P=[0.7, 0.2, 0.1], which means the probability of the first type of event is 0.7, the probability of the second type of event is 0.2, and the probability of the third type of event is 0.1. The event prior label Y=[1, 0, 0], which means the real event category is the first type. Then the first debugging error .

[0128] In step S104, after obtaining the second recognition result, the computer system also needs to compare it with the event prior mark to obtain a second debugging error. The calculation method of the second debugging error is similar to that of the first debugging error, but because the second recognition result may have different characteristics or be related to the first recognition result, more factors may be considered in the calculation process.

[0129] For example, in some cases, the second recognition result may depend on the first recognition result or have a temporal relationship with the first recognition result. Therefore, when calculating the second debugging error, the prior information comparison result may be introduced. The prior information comparison result is determined based on the comparison result between the first recognition result and the event prior mark.

[0130] Assume that the second recognition result is also a vector Q=[q1, q2,…, q n ], the prior information comparison result can be expressed as a scalar value r, which reflects the degree of difference between the first recognition result and the event prior label. For example, the prior information comparison result can be obtained by calculating a certain distance metric (such as Euclidean distance) between the first recognition result and the event prior label.

[0131] One method of determining the second debugging error based on the prior information comparison result, the second recognition result, and the event prior mark is: first, determine the error corresponding to each event prediction result according to the prior information comparison result r, the second recognition result Q, and the event prior mark Y. Assume that the error corresponding to the i-th event prediction result is , can be obtained by the formula Then, based on the mean of the errors corresponding to all event prediction results, the second debugging error is determined , the calculation formula is: .

[0132] For example, suppose the second recognition result Q=[0.6, 0.3, 0.1], the event prior label Y=[1, 0, 0], and the prior information comparison result r=0.5. Then the errors corresponding to the prediction results of each event are: . Second debugging error .

[0133] In step S105, after obtaining the first debugging error and the second debugging error, the computer system combines them to determine the model debugging error of the event monitoring model. The model debugging error is a comprehensive evaluation indicator of the performance of the entire model, which reflects the overall error of the model when processing the first health monitoring data sample.

[0134] A feasible method for determining the model debugging error is to perform a weighted summation of the first debugging error and the second debugging error.

[0135] As an implementation manner, step S104, obtaining a second debugging error based on the second recognition result and the event priori mark, includes: Step S1041: determining a second debugging error based on the prior information comparison result, the second recognition result, and the event prior mark; the prior information comparison result is determined based on the comparison result between the first recognition result and the event prior mark; The second recognition result includes one or more event prediction results; and the second debugging error is obtained by the following steps: Step S1041a: determining the error corresponding to each event prediction result based on the prior information comparison result, the second recognition result, and the event prior label; Step S1041b: Determine a second debugging error based on the mean of errors corresponding to one or more event prediction results.

[0136] In step S1041, when executing this step, the computer system comprehensively considers the three key factors of the prior information comparison result, the second recognition result and the event prior mark to determine the second debugging error. The prior information comparison result is obtained based on the comparison between the first recognition result and the event prior mark. It reflects the degree of difference between the first recognition result and the actual situation, and provides important reference information for the subsequent calculation of the second debugging error. The second recognition result is extracted from the event recognition result of the first health monitoring data sample, which represents the model's recognition of events under specific conditions or time periods. The event prior mark is a known and accurate event category label, which is a standard for measuring the accuracy of model prediction.

[0137] For example, suppose that in a health monitoring scenario, the first recognition result indicates that the health status of a patient in a certain time period is predicted to be "mildly abnormal", while the event prior mark shows that the patient's actual health status in this time period is "normal". Through a certain comparison method (such as calculating the difference between the predicted category and the actual category), the computer system obtains a prior information comparison result of 0.3 (the 0.3 here is only an example, and the value and calculation method of the comparison result in actual application will depend on the specific situation). The second recognition result shows that in another subsequent time period, the probability of the patient's health status being predicted to be "normal" is 0.8, the probability of "mildly abnormal" is 0.1, and the probability of "severely abnormal" is 0.1, that is, the second recognition result can be represented as a vector Q=[0.8, 0.1, 0.1]. The event prior mark indicates that the patient's actual health status in this subsequent time period is "normal", which is represented by one-hot encoding as a vector Y=[1, 0, 0].

[0138] In step S1041a, the corresponding error is calculated for each event prediction result based on the prior information comparison result, the second recognition result and the event prior label. The key to this step is how to comprehensively use these three factors to accurately measure the deviation between each event prediction result and the actual situation.

[0139] One way to achieve this is to define an error calculation function that takes the prior information comparison result, the predicted probability of each event in the second recognition result, and the true category information in the event prior label as input, and outputs the error corresponding to each event prediction result. For example, the following formula can be used to calculate the error corresponding to the i-th event prediction result : ; Among them, r represents the comparison result of prior information, q i is the predicted probability of the i-th event category in the second recognition result, y i is the true label of the i-th event category in the event prior label (1 means it belongs to this category, 0 means it does not belong to this category).

[0140] Taking the previous example, for the event category "normal" (i=1), r=0.3, q1=0.8, y1=1, the error corresponding to the prediction result of the event is: ; For the event category of "mild anomaly" (i=2), q2=0.1, y2=0, the error is: ; For the event category "severe anomaly" (i=3), q3=0.1, y3=0, the error is: ; In this way, the computer system can calculate the error corresponding to each event prediction result separately, thereby analyzing the prediction accuracy of the model on different event categories in more detail.

[0141] In step S1041b, after obtaining the errors corresponding to the prediction results of each event, the computer system determines the second debugging error by calculating the mean of these errors. The purpose of this step is to comprehensively consider the errors of all event prediction results and obtain an indicator that can reflect the overall prediction accuracy.

[0142] The formula for calculating the mean is: ; Where L2 represents the second debugging error, n is the number of event categories, and e i is the error corresponding to the prediction result of the i-th event.

[0143] Continuing with the previous example, the number of event categories n=3, the errors corresponding to the prediction results of each event are e1=0.012, e2=0.003, and e3=0.003. Then the second debugging error is: ; By calculating the mean, the computer system can obtain an indicator that comprehensively reflects the degree of difference between the model's second recognition result and the actual situation, namely the second debugging error. This indicator is of great significance for evaluating the performance of the model and guiding the optimization of the model. For example, if the second debugging error is large, it means that the model's prediction accuracy on the second recognition result is low, and the model's parameters, structure or training data may need to be adjusted and improved.

[0144] As an implementation method, the event recognition result of the health monitoring data to be analyzed is determined by debugging a converged event monitoring model in advance; the event monitoring model is debugged by the following steps: Step S1: Debugging the initialized machine learning model through one or more second health monitoring data samples and event recognition priori marks of each second health monitoring data sample to obtain a model to be debugged; Step S2: Debugging the restored mapping network in the to-be-debugged model through one or more health monitoring data samples and a priori marks of dynamic event recognition of each health monitoring data sample to obtain an event monitoring model.

[0145] In step S1, the second health monitoring data sample and its corresponding event recognition prior marker are used to adjust and optimize the parameters of the initialized machine learning model so that it can more accurately perform event recognition on the health monitoring data, thereby obtaining a simulated debugging model.

[0146] The second health monitoring data sample is a set of health monitoring data used for model debugging. These data contain rich health-related information, such as physiological indicator data such as heart rate, blood pressure, body temperature, etc. of different individuals in a specific time period, as well as possible living habits, environmental factors and other information. Each second health monitoring data sample has a corresponding event recognition prior tag, which is a known and accurate event category label used to represent the real health event situation corresponding to the sample. For example, in a health monitoring scenario, event recognition prior tags may include categories such as "normal", "mild abnormality", and "severe abnormality".

[0147] An initialized machine learning model is a model that has not been fully trained and optimized. It has a certain structure and initial parameters. Viable machine learning models include neural networks, support vector machines, decision trees, etc. Taking a neural network as an example, it usually consists of an input layer, a hidden layer, and an output layer, and each layer is connected by weights. In the initialization stage, these weights are usually randomly initialized.

[0148] When debugging and initializing a machine learning model, a computer system can use some optimization algorithms to adjust the model's parameters so that the model's prediction results are as close as possible to the event recognition prior label. A feasible optimization algorithm is the gradient descent method. The basic idea of ​​the gradient descent method is to calculate the error between the model's prediction results and the true label (usually measured by a loss function), and then update the parameters according to the gradient direction of the error to the model parameters so that the error gradually decreases.

[0149] Specifically, assume that the initialized machine learning model is a simple neural network, the input is the second health monitoring data sample X, and the output is the predicted event recognition result , event recognition priors are marked as The loss function can be selected as the cross entropy loss function, and its calculation formula is: ; Where N is the number of second health monitoring data samples, C is the number of event categories, is the true label of the i-th sample belonging to the j-th category (1 means it belongs to the category, 0 means it does not belong to the category), is the probability that the model predicts that the i-th sample belongs to the j-th category.

[0150] In each iteration, the computer system first inputs the second health monitoring data sample into the initialized machine learning model to obtain the prediction result . Then, the current error L is calculated based on the loss function. Next, the gradient of the loss function to the model parameters is calculated through the back propagation algorithm. The back propagation algorithm is an effective method for calculating gradients. It starts from the output layer and calculates the gradient of the parameters of each layer in turn. Finally, the parameters of the model are updated according to the gradient and the learning rate (a pre-set hyperparameter used to control the step size of the parameter update). The formula for parameter update is: ; in, represents the parameters of the model, Represents the gradient of the loss function with respect to the parameters.

[0151] For example, suppose there are 100 second health monitoring data samples, each sample contains 5 physiological indicators (such as heart rate, blood pressure, body temperature, exercise steps, sleep time), and there are 3 event categories ("normal", "mild abnormality", "severe abnormality"). The initialized machine learning model is a neural network with one hidden layer, and the hidden layer contains 10 neurons. In the first iteration, the computer system inputs these 100 samples into the model to obtain the prediction results. Assume that after calculation, the current cross entropy loss function value L=0.8. Then, the gradient of each parameter is calculated by the back propagation algorithm. Assuming that the learning rate is 0.01, the model parameters such as weights and biases are updated according to the parameter update formula. After many iterations, the error of the model gradually decreases. When a certain stop condition is reached (such as the error is less than a certain threshold or the number of iterations reaches the upper limit), the computer system obtains a model that has been preliminarily debugged, that is, the quasi-debugged model.

[0152] In step S2, after obtaining the simulated debugging model, the computer system further debugs the restoration mapping network therein to improve the model's ability to recognize dynamic events in health monitoring data, and finally obtains an event monitoring model.

[0153] The health monitoring data samples also contain rich health-related information, similar to the second health monitoring data samples, but here they focus more on the dynamic event characteristics in the data. The dynamic event recognition prior label is the known and accurate label information of the dynamic events in these health monitoring data samples, which reflects the real situation of the dynamic events, such as the type, occurrence time, duration and other information of the dynamic events.

[0154] The restoration mapping network plays a key role in the event monitoring model. It is responsible for restoring the feature vector after encoding and other processing back to the original data space to obtain more interpretable event recognition results. When debugging the restoration mapping network, the computer system will also use some optimization algorithms to adjust the parameters of the network so that the restoration mapping results match the dynamic event recognition prior markers as much as possible.

[0155] A feasible debugging method is fine-tuning based on the back-propagation algorithm. The computer system inputs the health monitoring data samples into the model to be debugged, and obtains the output result of the restored mapping network through forward propagation. , which is then combined with the dynamic event recognition prior Compare and calculate the error between the two. The error can be calculated in many ways, such as the mean square error (MSE) loss function, which is calculated as: ; Where N is the number of health monitoring data samples, is the prediction result of the restoration mapping network of the i-th sample, is the dynamic event recognition prior label of the i-th sample.

[0156] Next, the computer system calculates the gradient of the error to the parameters of the restored mapping network through the back propagation algorithm. During the back propagation process, the error propagates backward from the output layer, and the gradient of the parameters of each layer is calculated in turn. Then, the parameters of the restored mapping network are updated according to the gradient and the learning rate. The formula for parameter update is similar to that in step S1: ; For example, suppose there are 50 health monitoring data samples, each of which contains relevant features of dynamic events (such as changes in heart rate in different time periods, fluctuations in blood pressure, etc.). The dynamic event recognition prior label describes the specific information of the dynamic event in each sample. The restoration mapping network in the simulated debugging model receives these sample data and obtains the prediction result through forward propagation. . Assuming that the calculated mean square error MSE=0.5, after calculating the gradient of each parameter through the back propagation algorithm, the parameters of the restoration mapping network are updated according to the learning rate of 0.005. After multiple iterations, the error of the restoration mapping network gradually decreases, and the ability to recognize dynamic events continues to improve. When the stop condition is met, the computer system completes the debugging of the restoration mapping network and obtains the final event monitoring model.

[0157] As an implementation manner, the one or more events include at least one event to be continuously monitored; the dynamic event identification feature vector includes the identification feature vector corresponding to each event to be continuously monitored under multiple preceding monitoring data.

[0158] When a computer system processes health monitoring data, one or more events that need to be paid attention to include at least one event to be continuously monitored. Events to be continuously monitored refer to those events that need to be tracked dynamically, that is, health-related events whose changes need to be continuously monitored over a period of time. The dynamic event identification feature vector is a vector representation of event features in health monitoring data. It contains the identification feature vectors corresponding to each event to be continuously monitored under multiple previous monitoring data. This means that the dynamic event identification feature vector can comprehensively reflect the characteristic information of the event to be continuously monitored at different time points, thereby providing the computer system with a more comprehensive and accurate basis for event analysis.

[0159] In order to obtain the dynamic event recognition feature vector, the computer system can adopt a variety of implementation methods. One feasible method is the feature extraction algorithm, which can automatically extract the features related to the event to be continuously monitored from the original health monitoring data. For example, for heart rate data, the recognition feature vector can be constructed by calculating the statistical features such as the mean, variance, maximum value, minimum value of the heart rate, and the dynamic features such as the slope and frequency of the heart rate change. Another implementation method is data preprocessing and normalization. Since different health monitoring data may have different value ranges and distributions, in order to make each feature comparable in the dynamic event recognition feature vector, the computer system can preprocess and normalize the data. For example, normalizing the heart rate data to the interval [0, 1] can eliminate the dimensional difference of the data and improve the effectiveness of the feature vector. By including the corresponding recognition feature vectors of each event to be continuously monitored under multiple previous monitoring data, the dynamic event recognition feature vector can more comprehensively reflect the dynamic change process of the event. The computer system can use this rich information to more accurately analyze and predict the development trend of health conditions and provide strong support for timely intervention measures. For example, based on the characteristic changes of the "abnormal heart rate" event under multiple previous monitoring data, the computer system can determine whether the individual's abnormal heart rate has a tendency to worsen, thereby issuing a health warning in advance and reminding the individual to seek medical treatment in time or adjust his or her lifestyle habits.

[0160] As an implementation mode, the health monitoring data to be analyzed is health monitoring data obtained in a time series set of health monitoring data; the event recognition result includes one or more event prediction results; after obtaining the event recognition result of the health monitoring data to be analyzed, the method further includes: Step S400: extracting event recognition results that meet preset requirements from the event recognition results; Step S500: Iterate the health monitoring data to be analyzed through the next health monitoring data in the health monitoring data sample, iterate the dynamic event recognition feature vector through the recognition feature vector in the event recognition result that meets the preset requirements, and return to execute the process of characterizing and learning the health monitoring data to be analyzed.

[0161] Step S400 selects the part that meets the preset requirements from the obtained event recognition results. The event recognition results contain description information of one or more health monitoring events in the health monitoring data to be analyzed, such as the category confidence corresponding to the event. The preset requirements are some pre-set conditions or rules used to determine which event recognition results meet specific requirements.

[0162] The preset requirements can be set according to the specific application scenario and analysis purpose. For example, in the field of health monitoring, you may want to focus on event recognition results with higher class confidence, because these results are more likely to represent the true health status. Assume that the preset requirement is to extract event recognition results with class confidence greater than 0.8. The computer system traverses all event recognition results, and for each result, checks whether its class confidence meets the preset requirements.

[0163] For example, suppose the event recognition results obtained by the computer system are as follows: Event 1: The category is "mild anomaly" and the category confidence is 0.9; Event 2: The category is "normal" and the category confidence is 0.7; Event 3: The category is "severe anomaly" and the category confidence is 0.85.

[0164] According to the preset requirements, the computer system extracts Event 1 and Event 3 because their category confidences are greater than 0.8, while the category confidence of Event 2 does not meet the preset requirements and will be excluded.

[0165] In order to implement this screening process, the computer system can use simple conditional judgment statements. For each event recognition result Ri, where i represents the i-th event, its category confidence is represented by Ci. The computer system determines whether it meets the preset requirements by comparing Ci with the preset threshold T (in the above example T=0.8). If Ci>T, the event recognition result is extracted to form a set of event recognition results that meet the preset requirements.

[0166] In step S500, the computer system iterates the health monitoring data to be analyzed using the next health monitoring data in the health monitoring data sample. This means incorporating new data into the analysis process to update and expand the monitoring and analysis of health conditions. For example, assuming that the previous health monitoring data to be analyzed is a person's health data for the previous three days, then the next health monitoring data is the health data for the fourth day. The computer system merges or replaces the data for the fourth day with the data for the previous three days to form new health monitoring data to be analyzed.

[0167] At the same time, the computer system iterates the dynamic event recognition feature vector through the recognition feature vectors in the event recognition results that meet the preset requirements. The dynamic event recognition feature vector contains the recognition feature vectors corresponding to each event to be continuously monitored under multiple previous monitoring data. By introducing new recognition feature vectors, the dynamic event recognition feature vector can more comprehensively and accurately reflect the dynamic changes of the event.

[0168] For example, suppose that the previous dynamic event recognition feature vector was obtained based on the data of the first three days, and now there is a recognition feature vector V4 in the event recognition result that meets the preset requirements on the fourth day. The computer system can iteratively update the dynamic event recognition feature vector in a variety of ways. One feasible method is to splice the new recognition feature vector V4 with the previous dynamic event recognition feature vector to form a new dynamic event recognition feature vector. For example, if the previous dynamic event recognition feature vector is D=[V1, V2, V3], the updated dynamic event recognition feature vector is D'=[V1, V2, V3, V4].

[0169] Another method is to use weighted average to iterate and update. Assuming that the weight of the previous dynamic event recognition feature vector D is w1, and the weight of the new recognition feature vector V4 is w2, the updated dynamic event recognition feature vector D' can be calculated by the following formula: ; In this way, the computer system can continuously update the dynamic event identification feature vector based on new data, so that it can better reflect the dynamic change trend of the event.

[0170] After completing the iterative update of the data, the computer system returns to the process of performing the representation learning process on the health monitoring data to be analyzed. This means that the entire analysis process will start again, and the computer system will perform a new round of representation learning process on the updated health monitoring data to be analyzed, obtain new representation learning results, and then perform subsequent steps such as embedding mapping operations and restoring mapping operations to obtain new event recognition results.

[0171] Through continuous iterative updates and cyclic analysis, the computer system can track changes in health monitoring data in real time, promptly discover dynamic trends in health conditions, and provide more accurate and effective support for health monitoring and intervention. For example, in long-term health monitoring, as new data is continuously added, the computer system can continuously adjust and optimize event recognition results, promptly discover potential health problems, and provide personalized health advice and intervention measures for individuals.

[0172] In summary, steps S400 and S500 are a closely connected process. By screening event recognition results that meet preset requirements and using new data for iterative updates, the computer system can continuously optimize the analysis and processing of health monitoring data to achieve dynamic monitoring and accurate analysis of health conditions.

[0173] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0174] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A comprehensive health data analysis method based on artificial intelligence, characterized in that: The method comprises: Performing representation learning processing on the health monitoring data to be analyzed to obtain representation learning results; Performing an embedding mapping operation on the representation learning result to obtain a monitoring recognition feature vector; A restoration mapping operation is performed based on the monitoring identification feature vector and the dynamic event identification feature vector corresponding to the health monitoring data to be analyzed, obtaining an event identification result of the health monitoring data to be analyzed; the dynamic event identification feature vector is determined based on the event identification results of multiple preceding monitoring data corresponding to the health monitoring data to be analyzed; the event identification result of the health monitoring data to be analyzed is used to describe one or more health monitoring events in the health monitoring data to be analyzed; Wherein, the plurality of preceding monitoring data include first health monitoring data and second health monitoring data; The dynamic event recognition feature vector is obtained by the following steps: Extracting a first dynamic event recognition result from the event recognition results of the first health monitoring data; The dynamic event recognition feature vector is determined based on the first dynamic event recognition result and the event recognition result of the second health monitoring data.

2. The method according to claim 1, characterized in that The determining the dynamic event recognition feature vector based on the first dynamic event recognition result and the event recognition result of the second health monitoring data includes: Determine a second dynamic event recognition result based on the event monitoring time point distribution vector in the first dynamic event recognition result and the event content feature vector in the event recognition result of the second health monitoring data; Determining the dynamic event recognition feature vector based on the second dynamic event recognition result and the first dynamic event recognition result; Among them, the health monitoring data to be analyzed and multiple preceding monitoring data corresponding to the health monitoring data to be analyzed belong to a monitoring data set of the same monitoring object; the monitoring time point of the health monitoring data to be analyzed is after the monitoring time point of the first health monitoring data; the monitoring time point of the first health monitoring data is after the monitoring time point of the second health monitoring data.

3. The method according to claim 1, characterized in that The performing a restoration mapping operation based on the monitoring identification feature vector and the dynamic event identification feature vector corresponding to the health monitoring data to be analyzed to obtain an event identification result of the health monitoring data to be analyzed includes: Determining a to-be-processed identification feature vector based on the monitoring identification feature vector and the dynamic event identification feature vector; A restoration mapping operation is performed based on the identification feature vector to be processed to obtain an event identification result of the health monitoring data to be analyzed.

4. The method according to claim 1, characterized in that: The event recognition result of the health monitoring data to be analyzed is obtained by determining the converged reduction mapping model debugged in advance; The restoration mapping model includes one or more restoration mapping networks, wherein an internal attention operator in the restoration mapping network is constructed by a first selective attention filter, and the first selective attention filter is used to block information exchange between corresponding identification feature vectors belonging to the same event under different previous monitoring data; The one or more events include at least one event to be continuously monitored; the one or more restoration mapping networks include a first restoration mapping network and a second restoration mapping network; The restoration mapping model also includes a feature vector integration network; The feature vector integration network is configured to integrate the dynamic event recognition data output by the first restoration mapping network to obtain dynamic event recognition integration data; the dynamic event recognition data is determined based on the first restoration mapping network and the dynamic event recognition feature vector; the dynamic event recognition data includes the processed recognition feature vectors obtained by the first restoration mapping network for each of the events to be continuously monitored; the dynamic event recognition integration data includes the integrated recognition feature vectors output by the feature vector integration network for each of the processed recognition feature vectors; The second restoration mapping network is configured to process the dynamic event identification integration data and the event identification data output by the first restoration mapping network; The event identification data is determined based on the first restoration mapping network and the monitoring identification feature vector.

5. The method according to claim 4, characterized in that The feature vector integration network includes a feature simplification subcomponent and a feature enhancement subcomponent; the dynamic event recognition integration data is obtained by the following steps, including: Processing the dynamic event recognition data through the feature reduction subcomponent to obtain a retained feature vector; Processing the dynamic event recognition data by the feature enhancement subcomponent to obtain a pseudo-enhanced feature vector; Determining the dynamic event recognition integrated data based on the retained feature vector and the proposed enhanced feature vector; The retained feature vector is obtained by the following steps: Extracting the time series position of the dynamic event recognition data to obtain a first time series position; Performing internal weight focusing on the first time series position and the dynamic event recognition data to obtain an internal weight focusing result; Performing global information integration on the internal weight focusing result to obtain a global information integration result; Based on the global information integration result, determining a feature vector to be filtered; Based on the proposed filtering feature vector, the dynamic event recognition data is simplified to obtain the retained feature vector; The proposed enhanced feature vector is obtained by the following steps: Extracting the time series position of the dynamic event recognition data to obtain a second time series position; Internal weight focusing is performed on the second time series position and the dynamic event recognition data to obtain the quasi-enhanced feature vector.

6. The method according to claim 1, characterized in that The event recognition result of the health monitoring data to be analyzed is determined by debugging a convergent event monitoring model in advance; the model debugging error of the event monitoring model is obtained by the following steps: Determine an event recognition result of the first health monitoring data sample based on the first health monitoring data sample, the event recognition results of a plurality of preceding monitoring data corresponding to the first health monitoring data sample, and the event monitoring model; Extracting a first recognition result and a second recognition result from the event recognition result of the first health monitoring data sample; Obtaining a first debugging error based on the first recognition result and an event priori label corresponding to the first health monitoring data sample; Obtaining a second debugging error based on the second recognition result and the event priori mark; Determining a model debugging error of the event monitoring model based on the first debugging error and the second debugging error; The monitoring time point corresponding to the first recognition result is after the monitoring time point corresponding to the second recognition result.

7. The method according to claim 6, characterized in that The obtaining a second debugging error based on the second recognition result and the event priori mark includes: Determine the second debugging error based on a priori information comparison result, the second recognition result, and the event a priori mark; the priori information comparison result is determined based on a comparison result between the first recognition result and the event a priori mark; The second recognition result includes one or more event prediction results; and the second debugging error is obtained by the following steps: Determining errors corresponding to the event prediction results based on the prior information comparison result, the second recognition result, and the event prior label; The second debugging error is determined based on a mean value of the errors corresponding to the one or more event prediction results.

8. The method according to claim 1, characterized in that The event recognition result of the health monitoring data to be analyzed is determined by debugging a convergent event monitoring model in advance; the event monitoring model is debugged by the following steps: Debugging the initialized machine learning model through one or more second health monitoring data samples and event identification priori marks of each of the second health monitoring data samples to obtain a model to be debugged; Debugging the restoration mapping network in the to-be-debugged model through one or more health monitoring data samples and a priori marks of dynamic event identification of each of the health monitoring data samples to obtain the event monitoring model; The one or more events include at least one event to be continuously monitored; the dynamic event identification feature vector includes identification feature vectors corresponding to each of the events to be continuously monitored under multiple preceding monitoring data.

9. The method according to claim 8, characterized in that The health monitoring data to be analyzed is health monitoring data obtained in a time series set of health monitoring data; the event recognition result includes one or more event prediction results; After obtaining the event identification result of the health monitoring data to be analyzed, the method further includes: Extracting event recognition results that meet preset requirements from the event recognition results; The health monitoring data to be analyzed is iterated through the next health monitoring data in the health monitoring data sample, the dynamic event recognition feature vector is iterated through the recognition feature vector in the event recognition result that meets the preset requirements, and the process of performing characterization learning processing on the health monitoring data to be analyzed is returned.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images