Multi-mode biochemical test result abnormal mode intelligent early warning system and method thereof

Abstract

The invention relates to the technical field of test result abnormal mode intelligent early warning systems, in particular to a multi-mode biochemical test result abnormal mode intelligent early warning system and method, and the system comprises a dynamic threshold module, a multivariable mode module, a time evolution module and a self-adaption module. Through fusion of a dynamic threshold technology, multivariable pattern analysis, time sequence modeling and an interpretability mechanism, comprehensive, dynamic and intelligent analysis of biochemical test results is realized. According to the method, the accuracy and timeliness of anomaly detection are remarkably improved, and visual and credible decision support is provided for doctors. The innovative solution is expected to play an important role in the aspects of improving medical diagnosis efficiency, reducing missed diagnosis and misdiagnosis rates, promoting precise medical treatment and the like, and substantive improvement is brought to health management of patients.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent early warning systems for abnormal patterns of test results, and in particular to an intelligent early warning system for abnormal patterns of multimodal biochemical test results and a method thereof. Background Art

[0002] In the field of modern medical diagnosis, accurate interpretation of biochemical test results and timely warning have always been an important topic of concern. With the continuous advancement of medical technology and the increasing richness of detection methods, doctors have to face a huge amount of biochemical test data every day. These data are not only huge in quantity, but also involve multiple dimensions, including but not limited to blood routine, liver function, kidney function and other indicators. How to quickly and accurately identify potential health risks from these complex data has become a major challenge facing the current medical community.

[0003] Traditional analysis methods for biochemical test results mainly rely on fixed reference ranges and doctors' experience and judgment. Although this method is intuitive and simple, it has many limitations. First, fixed reference ranges are difficult to adapt to differences between individuals, which may lead to a large number of false positive or false negative results. Secondly, isolated analysis of a single indicator cannot capture the complex interactions between indicators and may ignore some potential health risks. Furthermore, this method is difficult to achieve early warning, and problems can often only be discovered when indicators are obviously abnormal, missing the best time for intervention.

[0004] In recent years, with the development of artificial intelligence technology, some researchers have begun to try to apply machine learning methods to the analysis of biochemical test results. Although these methods have improved the accuracy of anomaly detection to a certain extent, there are still some key problems. First, most existing machine learning models are black-box and lack interpretability, making it difficult for doctors to understand and trust their predictions. Secondly, these models often only focus on static data features and ignore the dynamic characteristics of biochemical indicators over time. Furthermore, most existing methods are limited to single-modality data analysis and cannot effectively integrate multi-source, heterogeneous medical data.

[0005] In addition, existing technologies also face problems such as insufficient data preprocessing, fixed anomaly detection thresholds, and inability to handle long-term dependencies. These problems seriously restrict the further development and clinical application of biochemical test result anomaly detection technology.

[0006] In view of the many deficiencies of the above-mentioned prior art, there is an urgent need for an intelligent system that can comprehensively consider multimodal data, capture time series characteristics, have good interpretability, and can achieve early warning. The present invention is proposed to address this urgent need. Summary of the invention

[0007] The purpose of the present invention is to provide an intelligent early warning system and method for abnormal patterns of multimodal biochemical test results, aiming to solve key technical problems existing in the prior art, such as insufficient data utilization, inaccurate anomaly detection, untimely early warning, and poor interpretability.

[0008] The present invention provides an intelligent early warning system for abnormal patterns of multimodal biochemical test results, comprising a dynamic threshold module, a multivariate pattern module, a time evolution module and an adaptive module. The dynamic threshold module constructs and trains an initial exponentially weighted moving average model according to historical test results, and calculates the upper prediction reference line, the lower prediction reference line and the interval width of the test results by using the real-time changes of the current test results and the historical test results; the multivariate pattern module collects sample test data and test result data from historical samples to construct a tensor data set; and uses a density clustering algorithm to obtain abnormal samples, and divides key clusters according to the density of abnormal samples; finally, a Gaussian mixture model is used to fuse abnormal samples. Samples are collected and hidden normal points are extracted from them; the time evolution module constructs the space-time tensor and inputs it into the time series convolutional neural network, extending the univariate anomaly analysis to the time evolution anomaly analysis; in the back propagation of the time series convolutional neural network, the chain rule is used to solve the gradient of the space-time tensor in each time window, the data constructed by the space-time tensor is gradient updated and the anomaly of the time window is obtained; the adaptive interpretation module uses the hierarchical analysis method to determine the correlation between each anomaly and the corresponding weight size of the time series convolutional neural network input; the time series convolutional neural network is decomposed by the interpretable causal reasoning network to generate an explanatory report.

[0009] Specifically, the exponentially weighted moving average model in the dynamic threshold module calculates the real-time prediction reference line based on the historical data of the inspection quantity and the current inspection result, including: the direction of the reference line, the deviation from the prediction reference line, and the width of the upper and lower intervals of the reference line.

[0010] Specifically, in the multivariate pattern module, abnormal patterns of multiple patterns are hidden; specifically: in tensor data with one or more key clusters, based on density clustering, the density within the cluster is calculated based on the clustering; in the cluster with the highest density, the Gaussian mixture model is used to merge the current data with other clusters with higher density, and the merged tensor data is fitted with the current tensor data to obtain normal samples; and abnormal samples and their density are obtained on the fitted tensor data.

[0011] Specifically, sample test data is collected from historical samples to form a spatiotemporal tensor A∈R n×p×k , b∈R n×q×k , construct the corresponding space-time tensor, whose structure is: A=(......A ji ...), where A ij =(......Aijk ...), (1≤n≤N), N is the total number of detection items in the current time window; b=(......b ij ...), where b ij =(......b iik ...), (1≤s≤SN), SN is the total number of samples in the current time window; density clustering is used to divide the data in the above sampled spatiotemporal tensor into multiple categories; if there are k key clusters, the expectation-maximization cycle of the variable Bayesian expectation-maximization algorithm is used, combined with normal data samples to obtain the final Gaussian mixture model: The probability density function parameter of the Gaussian distribution of each cluster is (μ i ,∑ i ) i μ i is the mean of the ith cluster, ∑ i is the covariance matrix of its ith cluster; then the density value within any cluster is ρ i =N i / V i ; Obtain hidden points through density calculation, obtain abnormal samples with higher density values, combine with the fitted tensor sample data, and obtain key points with higher density through Gaussian mixture model; the formula is: y=G+a+b, where G∈{0,1}, its application: when G=0, it means that it is obtained by fitting the current Gaussian mixture model; when G=1, it means that the current data comes from the fitting points obtained by the high-density cluster.

[0012] Specifically, the time evolution module uses network decomposition technology to autonomously decompose and explain anomalies; the module includes a time series convolutional neural network decomposition and an interpretable causal reasoning network; the corresponding relationship between the encoding result and the time window anomaly using the time series convolutional neural network model is specifically: y = G + a + b; where y is the anomaly at the current time, G i is the weight of the output of the i-th neuron, i is the current moment and I is the total moment; y is the encoding result, a is the historical abnormal data, and b is the normal sample data.

[0013] Specifically, the explanatory causal reasoning network uses neural network decomposition to split the causal network. During the decomposition process, the decomposition technology is used to continuously optimize the weights and abnormal data to obtain the relationship between the abnormal data and the weights, and the causal relationship is obtained through neural network decomposition. The specific steps of decomposition are: first initialize the weights, the initial abnormal data is the current time window data, calculate the weight of the current data at the next moment, and then calculate the weight of the data at the next moment, the abnormal data and its time series; if the difference between the data output by the current module and the current input data is greater than the set threshold, continue to decompose; if convergence, the anomaly obtained in the decomposition and calculation is the final anomaly, and the corresponding neuron weight is the weight corresponding to the anomaly, and the abnormal data is obtained by multiplying the overall abnormal data and the weight.

[0014] The intelligent early warning method for abnormal patterns of multimodal biochemical test results has the following specific steps: step S1. Collect abnormal test data and construct a sample library containing sample test data and test results; step S2. Retrieve the sample library of step S1, screen abnormal data and convert them into space-time tensors, specifically: convert sample test data and test results into Z vectorization, that is, discretized time and multiple continuous variables, and perform historical sampling to form a space-time tensor; apply the space-time tensor to the encoder of the time series convolutional neural network, divide it into multiple time windows by day, hour and every 10 minutes, and jointly characterize the current time window and its past; step S3. Convert the space-time tensor into an input tensor in the form of N×P, and construct the input of the time series convolutional neural network; step S4. Dynamic tensor based on exponentially weighted moving model The state threshold module calculates the Gaussian distribution of the N×P window, adjusts the prediction reference line to calculate the real-time threshold interval; step S5. Encode the input data corresponding to the time window, and input it into the attention mechanism after calculation to determine the key clusters and whether the current data contains the key clusters; step S6. Calculate the density of the key clusters and obtain the density of abnormal samples; step S7. After obtaining the density of abnormal samples, obtain the density of normal points through Gaussian mixture model fitting; step S8. Combine the spatiotemporal tensor with the output layer of the Gaussian mixture model to obtain tensor data containing potential multivariate patterns and their respective densities; step S9. Obtain key nodes through the encoder of the time series convolutional neural network; use neural network decomposition and causal reasoning network decomposition to obtain the results of causal paths and corresponding weights; generate a causal association report based on the results.

[0015] Specifically, causal reasoning network inference is applied in the encoder of the time series convolutional neural network; the inference of the causal reasoning network uses neural network decomposition and interpretable mechanism decomposition, including: decomposing the causal network and calculating understandable causal relationships; the interpretable mechanism decomposition is specifically as follows: the steps of neural network decomposition are as follows: (a) Initialize the discretization process of weights and abnormal data, and perform time series convolutional neural network encoding on the window at the current moment, and obtain the abnormal data and encoding value corresponding to the current weight in the sample test data; (b) If the difference between the abnormal data at the current moment and its source data is not greater than 1%, the time series convolutional neural network converges, and then execute step (d) to extract the final abnormal data; if the difference is greater than 1%, decomposition is performed; (c) The current solution abnormal data and the reconstructed abnormal data are used as encoder input, and the extracted current weight is used as encoder input. By decomposing the encoder, the abnormal data at the current moment and the latent vector k are obtained, and the abnormal data is updated through the latent vector k. The formula is: k =a k-1 +g(x k ,W k-1 ); (d) If k does not converge, then loop step (c); If k converges, then loop step (e); (e) Where, W k Decompose the weights of the neural network, a k is abnormal data, x k is the hidden vector at the current moment, W k-1 is the weight of the previous moment; (f) set the updated abnormal data to the new initial value and the current updated weight to the weight; (g) return to step (b).

[0016] Specifically, the calculation method of the understandable causal relationship is: the abnormal data obtained by encoding is decomposed by the causal network, and the implicit causal graph after decomposition reveals the direct cause of the current abnormal data; the steps are as follows:

[0017] (a) Obtain and analyze the output implicit causal network diagram to determine the causal direction and the nodes of the final causal network diagram; (b) Decompose the weights representing the causal relationship in the neural network and reversely calculate the abnormal data at the current moment; (c) Obtain the abnormal data output by the encoder at the current moment and compare it with the difference calculated by reverse calculation using the weights; (d) If the difference in the abnormal data is not greater than 1%, delete the node and enter step (g), and the current abnormal data is used as the final weight of the causal network; if the abnormal data is greater than 1%, return to step (b) and update the weight by deleting the current node; (e) In step (a), the implicit causal diagram is applied. When the abnormal data is determined, the key nodes are compared with the node information of the corresponding neural network to be decomposed to obtain the corresponding causal information, and the influencing factors of causality are understood in combination with the weights.

[0018] Specifically, the prediction reference line obtained in step S4 is applied to calculate the real-time abnormal data; the causal relationship is obtained through neural network decomposition, and step S3 is called to obtain the implicit causal network diagram.

[0019] Advantages and beneficial effects of the present invention:

[0020] By integrating dynamic threshold technology, multivariate pattern analysis, time series modeling and interpretability mechanisms, the present invention realizes a comprehensive, dynamic and intelligent analysis of biochemical test results. The implementation of the present invention not only significantly improves the accuracy and timeliness of abnormality detection, but also provides doctors with intuitive and reliable decision support. This innovative solution is expected to play an important role in improving the efficiency of medical diagnosis, reducing the rate of missed diagnosis and misdiagnosis, and promoting precision medicine, bringing substantial improvements to the health management of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is the overall logic block diagram of the system of the present invention.

[0022] Figure 2 It is a logic block diagram of the dynamic threshold module of the present invention.

[0023] Figure 3 It is a logic block diagram of the multivariable mode module of the present invention.

[0024] Figure 4 It is a logic block diagram of the time evolution module of the present invention.

[0025] Figure 5 It is a logic block diagram of the adaptive interpretation module of the present invention. DETAILED DESCRIPTION

[0026] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0027] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0028] See also Figure 1-5 The present invention discloses an intelligent early warning system and method for abnormal patterns of multimodal biochemical test results. The following are specific embodiments of the present invention:

[0029] The intelligent early warning system for abnormal patterns of multimodal biochemical test results includes a dynamic threshold module 1, a multivariate pattern module 2, a time evolution module 3 and an adaptive module 4, specifically:

[0030] Dynamic threshold module 1, based on historical test results, builds and trains the initial exponentially weighted moving average model, and uses the real-time changes of current test results and historical test results to calculate the upper prediction reference line, lower prediction reference line and interval width of the test results;

[0031] Multivariate model module 2 collects sample test data and test result data from historical samples to construct a tensor data set; uses density clustering algorithm to obtain abnormal samples, and divides key clusters according to the density of abnormal samples; finally, uses Gaussian mixture model to fuse abnormal samples and extract hidden normal points from them;

[0032] Time evolution module 3, inputs the space-time tensor construction into the time series convolutional neural network, and expands the single variable anomaly analysis to the time evolution anomaly analysis; in the back propagation of the time series convolutional neural network, the chain rule is used to solve the gradient of the space-time tensor construction in each time window, and the data of the space-time tensor construction is updated with gradients to obtain the anomalies of the time window;

[0033] The adaptive interpretation module 4 uses the hierarchical analysis method to determine the correlation between each anomaly and the corresponding weight size of the time series convolutional neural network input; uses the interpretable causal reasoning network to decompose the time series convolutional neural network and generate an explanatory report.

[0034] In one embodiment of the present invention, the dynamic threshold module 1 adopts an exponentially weighted moving average model. The calculation formula of the model is:

[0035] S t =α*X t +(1-α)*S t-1 , where S t is the exponentially weighted moving average at time t, X t is the actual value at time t, α is the smoothing coefficient 0<α<1, S t -1 is the exponentially weighted moving average at time t-1.

[0036] Preferably, the value range of α is between 0.1 and 0.3. This is because in the analysis of biochemical test results, it is necessary to maintain sufficient sensitivity to new data while avoiding interference caused by short-term fluctuations. For example, in blood sugar testing, α of 0.2 can well balance these two requirements.

[0037] Next, the system calculates the upper and lower forecast reference lines:

[0038] Upper prediction reference line U t=S t +k*σ t

[0039] Lower prediction reference line L t =S t -k*σ t

[0040] Interval width W t =U t -L t =2*k*σ t

[0041] Where k is the control parameter, usually 2 or 3; σ t is the standard deviation at time t, which can be calculated by the exponentially weighted moving standard deviation:

[0042]

[0043] β is the smoothing coefficient (0<β<1).

[0044] One advantage of the present invention is that the dynamic threshold module 1 can adjust the prediction reference line in real time, making the system more sensitive to abnormalities. For example, in liver function testing, if the value of ALT enzyme suddenly increases, the traditional method may need to wait for multiple tests to confirm the abnormality, while the present system can quickly adjust the prediction reference line and issue a warning in time.

[0045] The exponentially weighted moving average model in the dynamic threshold module 1 calculates the real-time prediction reference line based on the historical data of the inspection volume and the current inspection results, including: the direction of the reference line, the deviation from the prediction reference line, and the width of the upper and lower intervals of the reference line. The function of the dynamic threshold module 1 is further refined. First, the system calculates the direction of the reference line. The reference line direction D t Calculated by the following formula:

[0046] D t =sign(S t -S t-1 )

[0047] Deviation from the forecast reference line δ t The calculation is as follows:

[0048] δ t =X t -S t

[0049] Finally, calculate the width of the upper and lower intervals of the reference line:

[0050] W t =U t -L t =2*k*σ t .

[0051] These calculation results provide the system with rich information, including not only whether the current value is abnormal, but also the degree and trend of the abnormality. For example, in a thyroid function test, if the TSH level continues to rise but has not yet exceeded the traditional reference range, the system can issue an early warning based on the direction and degree of deviation from the reference line.

[0052] In the multivariate pattern module 2, the abnormal patterns of multiple patterns are hidden; specifically: in the tensor data with one or more key clusters, based on density clustering, the density within the cluster is calculated on the basis of clustering; in the cluster with the highest density, the Gaussian mixture model is used to merge the current data with other clusters with higher density, and the merged tensor data is fitted with the current tensor data to obtain normal samples; and abnormal samples and their density are obtained on the fitted tensor data.

[0053] The innovation of Multivariate Pattern Module 2 is that it can handle the complex relationship between multiple biochemical indicators. For example, in renal function tests, creatinine and urea nitrogen often need to be considered together. This module first uses the density clustering algorithm DBSCAN to identify abnormal samples. The advantage of DBSCAN is that it does not require the number of clusters to be specified in advance, but automatically determines the number and shape of clusters based on the distribution of the data.

[0054] In DBSCAN, two key parameters need to be set: ε (neighborhood radius) and MinPts (minimum number of points). In the biochemical test scenario, ε can be set to 1.5 times the standard deviation of each indicator, and MinPts can be set to 1% of the total number of samples. Such parameter settings can effectively identify abnormal samples while avoiding misjudging normal fluctuations as abnormalities. Next, the system calculates the intra-cluster density ρ i :

[0055]

[0056] N i is the number of points in cluster i, V i is the volume of cluster i. In high-dimensional space, V i It can be approximated by the Monte Carlo method.

[0057] The system then uses a Gaussian mixture model (GMM) for further refinement. The probability density function of the GMM is:

[0058]

[0059] Among them, π k is the weight of the kth Gaussian distribution, N(x|μ k ,∑ k ) is the mean value μ k , the covariance matrix is ​​∑k Gaussian distribution.

[0060] In this way, the system can effectively identify complex abnormal patterns. For example, in diabetes screening, even if fasting blood sugar and glycosylated hemoglobin are within the normal range, if both are close to the upper limit, the system can also identify this potential abnormal state.

[0061] Collect sample test data from historical samples to form a space-time tensor A∈R n×p×k ,b∈R n×q×k , construct the corresponding space-time tensor, whose structure is: A=(......A ji ...), where A ij =(......A ijk ...), (1≤n≤N), N is the total number of detection items in the current time window; b=(......b ij …), where b ij =(......b iik ...), (1≤s≤SN), SN is the total number of samples in the current time window; density clustering is used to divide the data in the above sampled spatiotemporal tensor into multiple categories; if there are k key clusters, the expectation-maximization cycle of the variable Bayesian expectation-maximization algorithm is used, combined with normal data samples to obtain the final Gaussian mixture model: The probability density function parameter of the Gaussian distribution of each cluster is (μ i ,∑ i ) ι μ i is the mean of the i-th cluster, ∑i is the covariance matrix of the i-th cluster; then the density value within any cluster is ρ i =N i / V i : Obtain hidden points through density calculation, obtain abnormal samples with higher density values, combine the fitted tensor sample data, and obtain key points with higher density through the Gaussian mixture model; the formula is: y=G+a+b, where G∈{0,1}, and its application: when G=0, it means that it is obtained by fitting the current Gaussian mixture model; when G=1, it means that the current data comes from the fitting points obtained by the high-density cluster.

[0062] The use of spatiotemporal tensors enables the system to simultaneously consider the changes in multiple biochemical indicators in the time dimension. For example, when monitoring liver function, we should not only pay attention to the absolute values ​​of indicators such as ALT and AST, but also pay attention to their changing trends over time.

[0063] In one embodiment of the present invention, the time window can be set to 30 days, which can capture short-term fluctuations without being affected by too much noise. During these 30 days, the system collects data once a day, so k = 30. For a general biochemical full set of tests, there may be 20 test items, so N = 20.

[0064] The application of the density clustering algorithm DBSCAN here is similar to that described above. Then, the system uses the variable Bayesian expectation maximization algorithm (VBEM) to optimize the Gaussian mixture model. The advantage of VBEM is that it can automatically determine the optimal number of clusters, avoiding the risk of overfitting.

[0065] Finally, the system obtains key points through the formula y=G+a+b. Here G is a selection matrix. When G=0, y comes from the fitting result of the current Gaussian mixture model; when G=1, y comes from the fitting points of the high-density cluster. This method can effectively identify potential abnormal points, even if these points appear normal in terms of a single indicator.

[0066] The time evolution module 3 uses network decomposition technology to autonomously decompose and explain the anomaly. This module includes the time series convolutional neural network decomposition and the interpretable causal reasoning network. The corresponding relationship between the encoding result and the time window anomaly using the time series convolutional neural network model is: y = Ga + b; where y is the anomaly at the current time, G i is the weight of the output of the i-th neuron, i is the current moment, and I is the total moment; y is the encoding result, a is the historical abnormal data, and b is the normal sample data.

[0067] The core innovation of the Time Evolution Module 3 is that it can capture the patterns of changes in biochemical indicators over time. This is particularly important in the monitoring of many chronic diseases. For example, in the progression monitoring of chronic kidney disease, a slow rise in creatinine levels may be more clinically significant than a single high value.

[0068] This module uses a temporal convolutional neural network (TCN) to simulate time series data. A key advantage of TCN is its ability to handle long-term dependencies, which is very important in biochemical test data analysis because changes in certain indicators may take weeks or even months to appear.

[0069] The convolution operation of TCN is defined as follows:

[0070]

[0071] Where x is the input sequence, f is the convolution kernel, and k is the convolution kernel size. In this system, k=7 can be set to capture the change pattern within a week.

[0072] Network decomposition techniques make it possible to understand how TCN makes its predictions. This interpretability is particularly important in the medical field because it can help doctors understand the system's decision-making process and make more accurate diagnosis and treatment decisions.

[0073] The explanatory causal reasoning network uses neural network decomposition to split the causal network. During the decomposition process, the decomposition technology is used to continuously optimize the weights and abnormal data, obtain the relationship between the abnormal data and the weights, and obtain the causal relationship through neural network decomposition; the specific steps of decomposition are: first initialize the weights, the initial abnormal data is the current time window data, calculate the weight of the current data at the next moment, and then calculate the weight of the data at the next moment, the abnormal data and its time series; if the difference between the data output by the current module and the current input data is greater than the set threshold, continue to decompose; if convergence, the anomaly obtained in the decomposition and calculation is the final anomaly, and the corresponding neuron weight is the weight corresponding to the anomaly, and the abnormal data is obtained by multiplying the overall abnormal data and the weight.

[0074] In one embodiment of the present invention, the explanatory causal reasoning network uses an innovative neural network decomposition technology. First, the system initializes the weights, which are usually initialized with random decimals ranging from -0.5 to 0.5. Then, the system uses the data of the current time window as the initial abnormal data. Preferably, the time window can be set to 30 days, so that a sufficiently long time series can be captured without introducing too much noise.

[0075] Secondly, the system calculates the weight of the current data at the next moment. This step uses the gradient descent method, and the specific formula is as follows:

[0076]

[0077] Where η is the learning rate, which is usually set between 0.001 and 0.01. L is the loss function, which can be the mean square error (MSE) or cross entropy loss.

[0078] Next, the system calculates the weight, abnormal data and its time series at the next moment. The update formula for abnormal data is:

[0079] a t =a t-1 +G t *(x t -a t-1 )

[0080] Among them, G t is the weight at time t, x t is the input data at time t.

[0081] Next, the system checks the difference between the data output by the current module and the input data. If the difference is greater than the set threshold (for example, it can be set to 0.05 or 5%), the decomposition continues. Otherwise, the system considers that it has converged and takes the obtained anomaly as the final anomaly, and the corresponding neuron weight is the weight corresponding to the anomaly.

[0082] Finally, the system multiplies the total abnormal data and the weight to obtain the final abnormal data. The advantage of this method is that it can provide the source and impact of the abnormality, allowing doctors to better understand the nature and possible causes of the abnormality.

[0083] Another embodiment of the present invention discloses an intelligent early warning method for abnormal patterns of multimodal biochemical test results, which comprises the following specific steps:

[0084] Step S1. Collect abnormal test data and build a sample library containing sample test data and test results;

[0085] Step S2. Retrieve the sample library of step S1, filter abnormal data and convert them into space-time tensors, specifically: convert sample test data and test results into Z vectorization, that is, discretized time and multiple continuous variables, and perform historical sampling to form space-time tensors; apply the space-time tensor to the encoder of the time series convolutional neural network, divide it into multiple time windows by day, hour and every 10 minutes, and jointly characterize the current time window and its past;

[0086] Step S3. Convert the spatiotemporal tensor into an input tensor in the form of N×P and construct the input of the temporal convolutional neural network;

[0087] Step S4. Based on the dynamic threshold module of the exponentially weighted mobile model, the Gaussian distribution of the N×P window is calculated, and the prediction reference line is adjusted to calculate the real-time threshold interval;

[0088] Step S5. Encode the input data corresponding to the time window, calculate and input it into the attention mechanism to determine the key cluster and whether the current data contains the key cluster;

[0089] Step S6. Calculate the density of key clusters and obtain the density of abnormal samples;

[0090] Step S7. After obtaining the abnormal sample density, obtain the normal point density through Gaussian mixture model fitting;

[0091] Step S8. Combine the spatiotemporal tensor with the Gaussian mixture model output layer to obtain tensor data containing potential multivariate patterns and their respective densities;

[0092] Step S9. Obtain key nodes through the encoder of the temporal convolutional neural network; obtain the results of the causal path and the corresponding weights by using the neural network decomposition and the causal reasoning network decomposition; and generate a causal association report based on the results.

[0093] The multimodal biochemical test result abnormal pattern intelligent early warning method of the present invention includes nine main steps, each of which plays an important role in the entire early warning system. First, in step S1, the system collects abnormal test data and builds a sample library. This step lays the data foundation for subsequent analysis. Preferably, the sample library should contain at least one year of historical data in order to capture seasonal changes and long-term trends.

[0094] Then, in step S2, the system preprocesses and transforms the collected data. Converting sample test data and test results into Z vectors is a key step, which allows indicators of different scales to be compared under the same framework. The division of time windows is also very important. In one embodiment of the present invention, multi-scale time windows can be used: by day, hour, and every 10 minutes. This multi-scale approach allows the system to capture long-term trends and short-term fluctuations at the same time.

[0095] Secondly, in steps S3 to S4, the system starts the core calculation of anomaly detection. The construction of spatiotemporal tensors and the calculation of dynamic thresholds are the key to this stage. The dynamic threshold module uses an exponentially weighted moving average model. The advantage of this model is that it can adaptively adjust the threshold and give higher weight to the most recent data.

[0096] Again, steps S5 to S7 involve the recognition of multivariate patterns. The attention mechanism and Gaussian mixture model are used here. The combination of these two technologies enables the system to effectively identify complex abnormal patterns. For example, in liver function tests, the ratio of ALT and AST may be more clinically significant than the absolute value of a single indicator. This system can automatically learn and recognize this complex relationship.

[0097] Finally, steps S8 and S9 perform the final anomaly detection and interpretation. The use of time series convolutional neural networks enables the system to capture complex patterns in time series data. The application of neural network decomposition and causal inference networks provides interpretability of the results, which is particularly important in the medical field because it can help doctors understand the decision-making process of the system.

[0098] Causal reasoning network inference is applied in the encoder of the temporal convolutional neural network; the inference of the causal reasoning network uses neural network decomposition and interpretable mechanism decomposition, including: decomposing the causal network and calculating understandable causal relationships. The working principle of the temporal convolutional neural network (TCN) encoder is refined. First of all, the advantage of TCN is that it can effectively process long sequence data, which is very important in the analysis of biochemical test results, because changes in some indicators may take a long time to appear.

[0099] Then, the application of causal inference networks enables the system to not only identify anomalies but also infer the possible causes of anomalies. This is achieved through neural network decomposition and interpretability mechanisms. Specifically, neural network decomposition breaks down complex neural networks into simpler sub-networks, each of which is responsible for a specific function.

[0100] Secondly, the explainable mechanism decomposition further analyzes the working principles of these sub-networks. For example, in a renal function test, if the system detects an abnormal increase in creatinine levels, it will not only issue an alarm, but also analyze whether this increase is related to changes in other indicators (such as urea nitrogen), thereby providing doctors with more comprehensive information.

[0101] Finally, the system computes understandable causal relationships. Understandable here means that these relationships can be intuitively understood and explained by medical professionals. For example, the system may output a result such as an 87% correlation between elevated creatinine levels and recent use of a certain medication. This interpretability is critical for medical decision support systems because it increases doctors' trust in the system's output.

[0102] The decomposition of the interpretable mechanism is as follows: The decomposition of the interpretable mechanism is as follows: The steps of neural network decomposition are as follows: (a) Initialize the discretization process of weights and abnormal data, and perform temporal convolutional neural network encoding on the window at the current moment to obtain the abnormal data and encoding value corresponding to the current weight in the sample test data; (b) If the difference between the abnormal data at the current moment and its source data is not greater than 1%, the temporal convolutional neural network converges, and then execute step (d) to extract the final abnormal data; if the difference is greater than 1%, decomposition is performed; (c) The current solution abnormal data and the reconstructed abnormal data are used as encoder inputs, and the extracted current weights are used as encoder inputs. By decomposing the encoder, the abnormal data at the current moment and the latent vector k are obtained, and the abnormal data is updated through the latent vector k. The formula is: k =a k-1 +g(x k ,W k-1 ); (d) If k does not converge, then loop step (c); If k converges, then loop step (e); (e) Where, Wk Decompose the weights of the neural network, a k is abnormal data, x k is the hidden vector at the current moment, W k-1 is the weight of the previous moment; (f) set the updated abnormal data to the new initial value and the current updated weight to the weight; (g) return to step (b).

[0103] First, the system initializes weights and abnormal data. Preferably, the weights can use the Xavier initialization method, which can effectively prevent the gradient from disappearing or exploding. The discretization process of abnormal data can use the equal frequency binning method to divide the continuous data into several discrete intervals.

[0104] Then, the system performs temporal convolutional neural network encoding on the window at the current moment. The purpose of this step is to capture the temporal dependency of the data. In one embodiment of the present invention, a dilated convolution can be used to increase the receptive field, thereby capturing longer-term dependencies.

[0105] Secondly, the system compares the difference between the abnormal data at the current moment and its source data. The 1% threshold set here is based on medical practice experience, because in most biochemical tests, a 1% error is generally considered acceptable. If the difference is not greater than 1%, the network is considered to have converged and the final abnormal data can be directly extracted.

[0106] Again, if the difference is greater than 1%, the system performs a further decomposition. This process uses an innovative formula: a k =a k-1 +g(x k ,W k-1 )Here, g is a nonlinear activation function. Usually the ReLU function can be selected because it can effectively solve the gradient disappearance problem. k is the hidden vector at the current moment, which contains all relevant information at the current moment. k-1 is the weight at the previous moment, which is used to calculate the update at the current moment.

[0107] Finally, the system will continue to iterate this process until convergence or the preset maximum number of iterations is reached. This iterative method can gradually refine the anomaly detection results and improve the accuracy of the system.

[0108] The calculation method of understandable causal relationship is as follows: the abnormal data obtained by encoding is decomposed by a causal network, and the implicit causal graph after decomposition reveals the direct cause of the current abnormal data; steps: (a) obtaining and analyzing the output implicit causal network graph, determining the causal direction and the nodes of the final causal network graph; (b) decomposing the weights representing the causal relationship in the neural network, and reversely calculating the abnormal data at the current moment; (c) obtaining the abnormal data output by the encoder at the current moment, and the difference with the reverse calculation using the weight; (d) if the difference in the abnormal data is not greater than 1%, delete the node and enter step (g), and the current abnormal data is used as the final weight of the causal network; if the abnormal data is greater than 1%, return to step (b) and update the weight by deleting the current node; wherein in step (a), the implicit causal graph is applied, and when the abnormal data is determined, it is compared with the node information of the corresponding neural network to be decomposed through the key nodes to obtain the corresponding causal information, and the influencing factors of the causal relationship are understood in combination with the weights.

[0109] First, the system inputs the encoded abnormal data into the causal network for decomposition. The core of this process is to generate an implicit causal graph, which can intuitively show the causes of abnormal data.

[0110] Then, the system analyzes the output implicit causal network diagram to determine the causal direction and the nodes of the final causal network diagram. In this step, the system uses graph theory to analyze the connection relationship between nodes to determine the causal direction. Preferably, a directed acyclic graph (DAG) can be used to represent the causal relationship, because DAG can effectively avoid the problem of circular dependency.

[0111] Secondly, the system decomposes the weights representing causal relationships in the neural network and reversely calculates the abnormal data at the current moment. This step uses the back propagation algorithm, which can effectively calculate the contribution of each weight to the final output. In one embodiment of the present invention, the Layer-wise Relevance Propagation (LRP) method can be used for reverse calculation because LRP can better preserve the semantic information of the original input.

[0112] Again, the system compares the difference between the anomaly data obtained by reverse calculation and the anomaly data directly output by the encoder. If the difference is no greater than 1%, the current causal explanation is considered reliable. The 1% threshold is also used here for the same reason as before.

[0113] Finally, if the difference is greater than 1%, the system updates the weights by deleting the current node and then recalculates. This process is actually pruning, which can effectively reduce the complexity of the model and improve the simplicity and understandability of the explanation.

[0114] One of the innovations of the present invention is the application of implicit causal graphs. When the abnormal data is determined, the system will compare the key nodes with the node information of the neural network to be decomposed to obtain the corresponding causal information. This method can provide a more fine-grained causal explanation. For example, in liver function tests, the system can not only point out that ALT is elevated, but also explain that this increase may be related to specific factors such as eating habits and drug use.

[0115] The prediction reference line obtained in step S4 is applied to calculate the real-time abnormal data; the causal relationship is obtained through neural network decomposition, and step S3 is called to obtain the implicit causal network diagram. First, the system uses the prediction reference line obtained in step S4 to calculate the real-time abnormal data. The advantage of this method is that it can dynamically adjust the definition of abnormality to adapt to the individual differences and time changes of different patients.

[0116] The system then uses neural network decomposition to obtain causal relationships. This step uses the neural network decomposition technique described above to provide the cause and impact of the anomaly. Preferably, Shapley values ​​can be used to quantify the contribution of each input feature to the prediction result, which has a good theoretical basis for explaining the decisions of machine learning models.

[0117] Secondly, the system calls step S3 to obtain an implicit causal network diagram. This diagram not only shows the relationship between anomalies, but also reflects the time sequence of anomalies. For example, in the long-term monitoring of diabetes, the system may find that the abnormality of blood sugar level precedes the abnormality of renal function indicators. This time sequence information is of great significance for the early prevention and intervention of the disease.

[0118] The sample library preprocessing step in step S1-S2 includes: abnormal data processing. First, the system will process the abnormal data, which includes several key steps:

[0119] 1. Missing value processing: use mean, median or interpolation to fill missing values. In one embodiment of the present invention, multiple interpolation can be used to process missing values ​​because this method can take into account the uncertainty of data.

[0120] 2. Outlier detection: Use Z-score or IQR method to detect outliers. Preferably, I isolation forest algorithm can be used to detect outliers because it works well for high-dimensional data and mixed types of features.

[0121] 3. Data normalization: Use Min-Max or Z-score normalization methods. In biochemical test data, the magnitudes of different indicators may vary greatly, so normalization is very important.

[0122] 4. Data smoothing: Use moving average or exponential smoothing methods. This step can effectively reduce the impact of short-term fluctuations and highlight long-term trends.

[0123] Through these preprocessing steps, the system can significantly improve the accuracy and reliability of subsequent analysis. For example, in a blood lipid test, if the total cholesterol value is much higher than the normal range (perhaps due to instrument failure or recording error), the system will mark it as an outlier and handle it accordingly to prevent this erroneous value from interfering with the overall analysis.

[0124] In general, the multimodal biochemical test result abnormal pattern intelligent early warning system and method of the present invention realizes accurate early warning and interpretable analysis of biochemical test results by comprehensively applying a variety of advanced machine learning and statistical techniques. It can not only detect potential health risks in a timely manner, but also provide decision support for doctors, which is expected to significantly improve the efficiency and accuracy of medical diagnosis.

[0125] In the practice of the present invention, a series of experiments were conducted to verify the effectiveness and superiority of the intelligent early warning system for abnormal patterns of multimodal biochemical test results. In order to comprehensively evaluate the performance of the system, an embodiment and two comparative examples were designed, and multiple key indicators were tested and analyzed in detail.

[0126] Example 1 adopts the complete solution of the present invention, including a dynamic threshold module, a multivariate pattern module, a time evolution module and an adaptive interpretation module. Comparative Example 1 uses a traditional fixed threshold method for anomaly detection, while Comparative Example 2 uses a single machine learning model (random forest) to predict anomalies.

[0127] Three key metrics were selected to evaluate system performance: precision, recall, and F1 score. These metrics are calculated as follows:

[0128] 1. Accuracy = true positive / (true positive + false positive);

[0129] 2. Recall = true positives / (true positives + false negatives);

[0130] 3. F1 score = 2*(precision*recall) / (precision+recall);

[0131] In addition, the system's interpretability and warning lead time were evaluated. Interpretability was measured by the doctor's satisfaction with the system output explanation, using a Likert scale of 1-5, with 5 indicating very satisfied. Warning lead time recorded the difference between the system warning time and the time it takes for the traditional method to detect the abnormality.

[0132] A six-month test was conducted in a tertiary hospital, involving biochemical test data of 5,000 patients. The test results are shown in the following table:

[0133] index Example 1 Comparative Example 1 Comparative Example 2 Accuracy 0.95 0.82 0.88 Recall 0.93 0.75 0.84 F1 score 0.94 0.78 0.86 Explainability 4.7 3.2 3.8 Early warning lead time (days) 3.5 0.5 1.8

[0134] It can be clearly seen from the data in the table that Example 1 of the present invention is significantly better than the two comparative examples in all indicators. Let's analyze the meaning behind these results in depth.

[0135] First, Example 1 achieved a precision and recall rate of more than 0.9, which means that the system can not only accurately identify abnormal situations, but also capture the vast majority of abnormal events. In contrast, although the fixed threshold method of Comparative Example 1 is simple and intuitive, it cannot adapt to the individual differences of different patients, resulting in a large number of false positive and false negative results. Although the single machine learning model of Comparative Example 2 outperforms the fixed threshold method, its performance is still significantly lower than that of the solution of the present invention due to the lack of time series analysis and multimodal fusion.

[0136] Secondly, the F1 score, as the harmonic mean of precision and recall, more comprehensively reflects the overall performance of the system. The F1 score of Example 1 reaches 0.94, which fully demonstrates the superiority of the present invention in balancing accuracy and coverage.

[0137] Furthermore, the high score of the interpretability index (4.7 points) reflects the high recognition of the system output by doctors. This is particularly important because in the medical field, high accuracy alone is not enough, and the system also needs to be able to provide understandable and trustworthy explanations. The present invention successfully transforms complex machine learning models into a form that doctors can understand and trust through innovative neural network decomposition and causal reasoning networks.

[0138] Finally, the significant advantage of early warning time (3.5 days) highlights the excellent performance of this system in early warning. This means that doctors can discover potential health risks 3.5 days earlier than traditional methods, which buys precious time for timely intervention and treatment. In the treatment of many acute diseases, this 3.5-day advance may be the key to saving lives.

[0139] In summary, the multimodal biochemical test result abnormal pattern intelligent early warning system of the present invention not only performs well in accuracy and comprehensiveness, but also shows significant advantages in interpretability and early warning capabilities. This all-round excellent performance makes the system likely to play an important role in clinical practice, improve the efficiency and quality of medical diagnosis, and ultimately benefit a large number of patients.

[0140] In summary, Example 1 is undoubtedly the best example. It demonstrates the great potential of the present invention in practical applications and opens up a new path for anomaly detection in the financial industry. This method is not only applicable to the financial field, but can also be extended to other fields that require high privacy protection, such as medical care and network security, and has broad application prospects.

[0141] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent early warning system for abnormal patterns of multimodal biochemical test results, characterized in that: It includes a dynamic threshold module, a multivariate mode module, a time evolution module and an adaptive module. The dynamic threshold module constructs and trains an initial exponentially weighted moving average model based on historical test results, and calculates the upper prediction reference line, the lower prediction reference line and the interval width of the test results by using the real-time changes of the current test results and the historical test results. The multivariate mode module collects sample test data and test result data from historical samples to construct a tensor data set; and uses a density clustering algorithm to obtain abnormal samples, and divides key clusters according to the density of abnormal samples; finally, a Gaussian mixture model is used to fuse abnormal samples and extract hidden normal points from them; the time evolution module constructs the space-time tensor and inputs it into the time series convolutional neural network to expand the single variable anomaly analysis to the time evolution anomaly analysis; in the back propagation of the time series convolutional neural network, the chain rule is used to solve the gradient of the space-time tensor construction in each time window, the data constructed by the space-time tensor is gradient updated and the anomaly of the time window is obtained; The adaptive interpretation module uses the hierarchical analysis method to determine the correlation between each anomaly and the corresponding weight size of the time series convolutional neural network input; uses an interpretable causal reasoning network to decompose the time series convolutional neural network and generate an explanatory report.

2. The intelligent early warning system for abnormal patterns of multimodal biochemical test results according to claim 1 is characterized in that: The exponentially weighted moving average model in the dynamic threshold module calculates the real-time prediction reference line based on the historical data of the inspection quantity and the current inspection result, including: the direction of the reference line, the deviation from the prediction reference line, and the width of the upper and lower intervals of the reference line.

3. The intelligent early warning system for abnormal patterns of multimodal biochemical test results according to claim 1 is characterized in that: In the multivariate pattern module, abnormal patterns of multiple patterns are hidden; specifically: in tensor data with one or more key clusters, based on density clustering, the density within the cluster is calculated based on the clustering; in the cluster with the highest density, the Gaussian mixture model is used to merge the current data with other clusters with higher density, and the merged tensor data is fitted with the current tensor data to obtain normal samples; and abnormal samples and their density are obtained on the fitted tensor data.

4. The intelligent early warning system for abnormal patterns of multimodal biochemical test results according to claim 1, characterized in that: Collect sample test data from historical samples to form a space-time tensor A∈R n×p×k ,b∈R n×q×k , construct the corresponding space-time tensor, whose structure is: A =(……A ji …), where A ij =(……A ijk …), (1≤n≤N), N is the total number of detection items in the current time window; b=(……b ij …), where b ij =……b iik …), (1≤s≤SN), SN is the total number of samples in the current time window; density clustering is used to divide the data in the above sampled spatiotemporal tensor into multiple categories; if there are k key clusters, the expectation-maximization cycle of the variable Bayesian expectation-maximization algorithm is used to combine the normal data samples to obtain the final Gaussian mixture model: The probability density function parameter of the Gaussian distribution of each cluster is (μ i ,Σ i ) ι μ i is the mean of its ith cluster, Σ i is the covariance matrix of its ith cluster; then the density value within any cluster is ρ i =N i / V i ; Obtain hidden points through density calculation, obtain abnormal samples with higher density values, combine with the fitted tensor sample data, and obtain key points with higher density through Gaussian mixture model; the formula is: y=G+a+b, where G∈{0,1}, its application: when G=0, it means that it is obtained by fitting the current Gaussian mixture model; when G=1, it means that the current data comes from the fitting points obtained by the high-density cluster.

5. The intelligent early warning system for abnormal patterns of multimodal biochemical test results according to claim 1, characterized in that: The time evolution module uses network decomposition technology to autonomously decompose and explain the anomaly; the module includes a time series convolutional neural network decomposition and an interpretable causal reasoning network; the corresponding relationship between the encoding result and the time window anomaly using the time series convolutional neural network model is: y = G + a + b; where y is the anomaly at the current time, G i is the weight of the output of the i-th neuron, i is the current moment and I is the total moment; y is the encoding result, a is the historical abnormal data, and b is the normal sample data.

6. The intelligent early warning system for abnormal patterns of multimodal biochemical test results according to claim 5, characterized in that: The explanatory causal reasoning network uses neural network decomposition to split the causal network. During the decomposition process, the decomposition technology is used to continuously optimize the weights and abnormal data, obtain the relationship between the abnormal data and the weights, and obtain the causal relationship through neural network decomposition; The specific steps of decomposition are: first initialize the weights, the initial abnormal data is the current time window data, calculate the weight of the current data at the next moment, and then calculate the weight of the data at the next moment, the abnormal data and its time series; if the difference between the data output by the current module and the current input data is greater than the set threshold, continue to decompose; if convergence occurs, the anomaly obtained in the decomposition and calculation is the final anomaly, and the corresponding neuron weight is the weight corresponding to the anomaly, and the abnormal data is obtained by multiplying the overall abnormal data and the weight.

7. An intelligent early warning method for abnormal patterns of multimodal biochemical test results, characterized in that: Specific steps: Step S1. Collect abnormal test data and build a sample library containing sample test data and test results; Step S2. Retrieve the sample library of Step S1, filter abnormal data and convert them into space-time tensors, specifically: convert sample test data and test results into Z vectorization, that is, discretized time and multiple continuous variables, and perform historical sampling to form a space-time tensor; apply the space-time tensor to the encoder of the time series convolutional neural network, divide it into multiple time windows by day, hour and every 10 minutes, and jointly characterize the current time window and its past; Step S3. Convert the space-time tensor into an input tensor in the form of N×P, and build the input of the time series convolutional neural network; Step S4. Based on the dynamic threshold module of the exponentially weighted moving model, calculate the Gaussian distribution of the N×P window, adjust the prediction reference line to calculate the real-time threshold interval; Step S5. Encode the input data corresponding to the time window, calculate and input it into the attention mechanism to determine the key clusters and whether the current data contains the key clusters; Step S6. Calculate the density of the key clusters and obtain the density of abnormal samples; Step S7. After obtaining the density of abnormal samples, obtain the density of normal points through Gaussian mixture model fitting; Step S8. Combine the spatiotemporal tensor with the output layer of the Gaussian mixture model to obtain tensor data containing potential multivariate patterns and their respective densities; Step S9. Obtain key nodes through the encoder of the time series convolutional neural network; Use neural network decomposition and causal reasoning network decomposition to obtain the results of causal paths and corresponding weights; Generate a causal association report based on the results.

8. The method according to claim 7, characterized in that Causal reasoning network inference is applied in the encoder of the time series convolutional neural network; the inference of the causal reasoning network uses neural network decomposition and interpretable mechanism decomposition, including: decomposing the causal network and calculating understandable causal relationships; the interpretable mechanism decomposition is specifically as follows: the steps of neural network decomposition are as follows: (a) Initialize the discretization process of weights and abnormal data, and perform time series convolutional neural network encoding on the window at the current moment, and obtain the abnormal data and encoding value corresponding to the current weight in the sample test data; (b) If the difference between the abnormal data at the current moment and its source data is not greater than 1%, the time series convolutional neural network converges, and then execute step (d) to extract the final abnormal data; if the difference is greater than 1%, decomposition is performed; (c) The current solution abnormal data and the reconstructed abnormal data are used as encoder input, and the extracted current weight is used as encoder input. By decomposing the encoder, the abnormal data at the current moment and the latent vector k are obtained, and the abnormal data is updated through the latent vector k. The formula is: k =a k-1 +g(x k , W k-1 ); (d) If k does not converge, then loop step (c); If k converges, then loop step (e); (e) Where, W k Decompose the weights of the neural network, a k is abnormal data, x k is the hidden vector at the current moment, W k-1 is the weight of the previous moment; (f) set the updated abnormal data to the new initial value and the current updated weight to the weight; (g) return to step (b).

9. According to the method of claim 8, the calculation method of the causal relationship can be understood as follows: the abnormal data obtained by encoding is decomposed by a causal network, and the implicit causal graph after decomposition reveals the direct cause of the current abnormal data; the steps are as follows: (a) Obtain and analyze the output implicit causal network diagram to determine the causal direction and the nodes of the final causal network diagram; (b) Decompose the weights representing the causal relationship in the neural network and reversely calculate the abnormal data at the current moment; (c) Obtain the abnormal data output by the encoder at the current moment and compare it with the difference calculated by reverse calculation using the weights; (d) If the difference in the abnormal data is not greater than 1%, delete the node and enter step (g), and the current abnormal data is used as the final weight of the causal network; if the abnormal data is greater than 1%, return to step (b) and update the weight by deleting the current node; (e) In step (a), the implicit causal diagram is applied. When the abnormal data is determined, the key nodes are compared with the node information of the corresponding neural network to be decomposed to obtain the corresponding causal information, and the influencing factors of causality are understood in combination with the weights.

10. The intelligent early warning method for abnormal patterns of multimodal biochemical test results according to claim 7, characterized in that: The prediction reference line obtained in step S4 is applied to calculate the real-time abnormal data; the causal relationship is obtained through neural network decomposition, and step S3 is called to obtain the implicit causal network diagram.

Images