Intelligent tumor marker data processing system

Abstract

The invention provides an intelligent tumor marker data processing system. The system comprises a data extraction module, a data fusion module and a marker analysis module which are connected in sequence, the data extraction module is used for collecting basic data related to tumor markers, performing knowledge extraction from the basic data, extracting structured information including entities, attributes and relationships among the entities, and forming a preliminary knowledge base; the data fusion module is used for fusing knowledge of different sources and different formats in the preliminary knowledge base, performing knowledge fusion by adopting a redundancy resolution and conflict resolution algorithm, and establishing a non-redundancy knowledge graph; the marker analysis module is used for establishing a tumor marker analysis model, evaluating and predicting the tumor marker level of a patient by using a non-redundant knowledge graph, and judging an abnormal value; the system can realize accurate analysis and evaluation of tumor marker data, and provides reliable data support for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data processing, and in particular to a tumor marker data intelligent processing system. Background Art

[0002] Tumor marker detection is widely used in clinical practice as an important auxiliary diagnostic method. However, there are still many challenges in the processing and analysis of tumor marker data. For example, patent CN114579866A discloses a tumor marker analysis method based on machine learning. Although this method can analyze standardized data from a single source, it cannot effectively process multi-source heterogeneous tumor marker data and lacks a systematic solution to data redundancy and conflict.

[0003] Existing tumor marker data processing systems generally have problems such as insufficient knowledge expression capabilities and low data integration. For example, patent CN113899517A describes a tumor marker test result analysis system, which mainly relies on statistical methods for data analysis and fails to make full use of medical knowledge and clinical experience. In addition, its threshold determination method is too static and cannot be dynamically adjusted according to specific case characteristics, resulting in limited accuracy and reliability of the analysis results.

[0004] With the development of artificial intelligence and knowledge graph technology, it has become possible to combine deep learning with knowledge reasoning to build an intelligent tumor marker data processing system. However, how to effectively integrate multi-source heterogeneous data, eliminate data redundancy and conflict, and how to achieve knowledge-driven dynamic analysis are still technical problems that need to be solved urgently. Existing technologies have not yet provided a systematic solution that can simultaneously take into account the comprehensiveness of data processing, the richness of knowledge expression, and the accuracy of analysis results. Summary of the invention

[0005] The present invention provides a tumor marker data intelligent processing system, which uses deep learning and knowledge graph technology to realize intelligent processing, fusion and analysis of multi-source heterogeneous tumor marker data. The specific scheme is as follows:

[0006] A tumor marker data intelligent processing system comprises a data extraction module, a data fusion module and a marker analysis module which are connected in sequence.

[0007] Furthermore, the data extraction module is used to collect basic data related to tumor markers, extract knowledge from the basic data, extract structured information including entities, attributes and relationships between entities, and form a preliminary knowledge base.

[0008] The data extraction module includes: an entity recognition unit, an attribute extraction unit, a relationship mining unit and a knowledge base establishment unit;

[0009] The entity recognition unit is used to identify entities related to tumor markers from text and unstructured data;

[0010] The attribute extraction unit is used to extract and quantify the attributes associated with each entity by adopting an attribute value embedding method;

[0011] The relationship mining unit is used to extract the relationship between entities from the data through the RNA algorithm to form entity-relationship triples;

[0012] The knowledge base establishment unit is used to use a graph database to store the identified entities, attributes and relationships to form a preliminary knowledge base.

[0013] Furthermore, the data fusion module is used to fuse the knowledge from different sources and in different formats in the preliminary knowledge base, use redundancy elimination and conflict resolution algorithms to fuse the knowledge, and establish a knowledge graph without redundancy.

[0014] The data fusion module includes: a standardization unit, a redundancy elimination unit and a conflict resolution unit;

[0015] The standardization unit is used to standardize data from different sources and formats;

[0016] For quantitative tumor marker data, a modified Min-Max normalization method was used: Among them, x is the value before standardization, x ′ is the standardized value, x max and x min are the upper and lower limits of the clinical reference value range of the marker to be standardized, and U and L are the upper and lower limits of the target range of the data after standardization.

[0017] For qualitative or semi-quantitative tumor marker data, piecewise function mapping is used for normalization:

[0018]

[0019] Among them, x ref These are clinically typical values ​​for tumor marker data.

[0020] The redundancy elimination unit is used to eliminate redundant data; the conflict resolution unit is used to use game theory models and Bayesian networks to resolve conflicts and ensure the consistency of the knowledge graph.

[0021] Furthermore, the marker analysis module is used to establish a tumor marker analysis model, using the non-redundant knowledge graph to evaluate and predict the patient's tumor marker level, and to determine outliers.

[0022] Furthermore, the redundancy elimination unit includes: a linear transformation subunit and a conditional screening subunit;

[0023] The linear transformation subunit is used to implement the following linear transformation:

[0024] Input layer to hidden layer: h = ReLU(W1x′+b1);

[0025] Hidden layer to output layer: y MLP =ReLU(W2h+b2);

[0026] Among them, W1 is the weight matrix from the input layer to the hidden layer, which is used to weight the importance of different marker features. Each element in the matrix represents the influence of the corresponding marker feature on the hidden layer neurons; b1 is the bias vector of the hidden layer, which is used to adjust the activation threshold of the hidden layer neurons to ensure that the model can adapt to different ranges of marker values; h is the output vector of the hidden layer, which represents the high-level feature representation extracted from the original marker data; W2 is the weight matrix from the hidden layer to the output layer, which is used to map the extracted high-level features to the final marker evaluation results; b2 is the bias vector of the output layer, which is used to fine-tune the final output marker evaluation value; y MLP is the output vector of the output layer, which represents the feature vector of the landmark after being processed by the neural network;

[0027] The conditional screening subunit is used to realize conditional probability distribution: Among them, P(y|y MLP ) is the conditional probability of a specific combination of marker states given the output of the MLP; y is the output sequence of marker states, representing the combined state of multiple related markers; M(y MLP ) is a normalization factor used to ensure that the sum of the probabilities of all possible states is 1; λ k is the weight coefficient of the kth feature function, reflecting the importance of this feature in the judgment of the marker state; f k (y,y MLP ) is the kth characteristic function, which is used to capture the local dependencies in the marker state sequence;

[0028] By maximizing the conditional probability P(y|y MLP ) to obtain the output sequence y of the optimal marker state * :

[0029]

[0030] Further, the entity recognition unit includes a named entity recognition model based on BiLSTM-CRF, and the named entity recognition model includes: an input layer, a BiLSTM layer and a CRF layer;

[0031] The input layer is used to receive text sequence input;

[0032] The BiLSTM layer is used to extract contextual features of the text;

[0033] The CRF layer is used to label the feature sequence output by the BiLSTM layer and identify tumor marker related entities;

[0034] Among them, the forward propagation and back propagation of the BiLSTM layer are:

[0035]

[0036] is the forward hidden layer state at time t, representing the text information encoding from left to right, is the reverse hidden layer state at time t, indicating the text information encoding from right to left; W f is the weight matrix of forward propagation, which is used to control the transmission strength of forward information; W b is the weight matrix of back propagation, which is used to control the transmission strength of reverse information; b f is the bias vector of the forward propagation, which is used to adjust the reference value of the forward hidden layer; b b is the bias vector of the back propagation, which is used to adjust the baseline value of the reverse hidden layer; x t is the input vector at time t, which represents the feature representation of the currently processed word; σ is the activation function, which uses the sigmoid or tanh function to introduce nonlinear transformation.

[0037] Further, the marker analysis module includes: a knowledge reasoning unit and a marker evaluation unit;

[0038] The knowledge reasoning unit is used to perform reasoning based on the non-redundant knowledge graph to calculate the patient marker level:

[0039] Among them, M d is the comprehensive marker level, w i is the weight of the i-th relationship in the knowledge graph, R i is the contribution value of the i-th relationship, v j is the weight of the jth attribute, A j is the normalized value of the jth attribute;

[0040] The marker evaluation unit is used to establish a tumor marker analysis model by combining clinical experience and statistical laws in the knowledge graph:

[0041] E(t)=f(M d ,G,H)+∈, where E(t) is the marker evaluation value at time t, f(Md ,G,H) is a nonlinear evaluation function, G is the group feature provided by the knowledge graph, H is the historical data feature, and ∈ is the random error term.

[0042] Furthermore, the marker analysis module also includes:

[0043] Dynamic threshold unit, used to dynamically adjust the outlier determination threshold based on the knowledge graph:

[0044] T=α×T b +(1-α)×T d ; Where T is the dynamic threshold, T b is the benchmark threshold in the knowledge graph, T d is a data-driven threshold;

[0045] The calculation method of the balance factor α is:

[0046] Among them, n ref is the number of standard cases referenced in the knowledge graph, n sim is the number of historical cases in the knowledge graph that are similar to the current case.

[0047] Furthermore, the data-driven threshold T d Calculate by the following steps:

[0048] Step S1, using cosine similarity to calculate the similarity between the current case and the historical case, and selecting cases with similarity greater than 0.8 as the similar case set S;

[0049] Step S2: cluster the marker values ​​in the similar case set S, T d Take the mean of the largest cluster.

[0050] Furthermore, the system also includes:

[0051] The data storage module is used to store the patient's historical test data, clinical standard data and intermediate data generated during the system operation.

[0052] Furthermore, the system also includes:

[0053] The user interface module is used to receive the detection data and system parameters input by the user, and to display the analysis results and abnormal warning information.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention realizes intelligent processing of the entire process from data extraction, fusion to analysis by establishing a multi-level data processing architecture. The improved Min-Max standardization method and piecewise function mapping realize unified standardized processing of different types of marker data. The conflict resolution mechanism based on game theory and Bayesian network ensures the consistency of the knowledge graph. The dynamic threshold adjustment mechanism improves the accuracy of outlier judgment, and ultimately realizes accurate analysis and evaluation of tumor marker data, providing reliable data support for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The figure is a schematic diagram of the composition of a tumor marker data intelligent processing system of the present invention. DETAILED DESCRIPTION

[0057] To make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Based on the technical solution disclosed below, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] like Figure 1 FIG. 1 is a schematic diagram of a tumor marker data intelligent processing system of the present invention, wherein the system comprises a data extraction module, a data fusion module and a marker analysis module connected in sequence.

[0059] The data extraction module is used to collect basic data related to tumor markers, extract knowledge from the basic data, extract structured information including entities, attributes and relationships between entities, and form a preliminary knowledge base.

[0060] The sources of basic data may include, but are not limited to: various types of text data such as hospital test reports, scientific research literature, clinical guidelines, and numerical data directly output by test instruments.

[0061] The data extraction module includes: an entity recognition unit, an attribute extraction unit, a relationship mining unit and a knowledge base establishment unit; each unit can be independently upgraded and optimized to improve the overall performance of the system; for example, the entity recognition unit can select different deep learning models according to specific application scenarios, such as BERT, XLNet and other pre-trained models.

[0062] The entity recognition unit is used to identify entities related to tumor markers from text and unstructured data; entity types include but are not limited to: marker name (such as AFP, CEA, etc.), detection method, reference value range, unit, etc. In addition, the unit can also identify clinical information such as disease name and symptom description related to the marker.

[0063] The attribute extraction unit is used to extract and quantify the attributes associated with each entity by adopting the attribute value embedding method; in addition to conventional numerical attributes, this unit can also process text description attributes, such as qualitative descriptions such as "significantly increased" and "mildly abnormal", and convert these descriptions into standardized numerical representations through a deep learning model.

[0064] The relationship mining unit is used to extract the relationship between entities from the data through the RNA algorithm to form entity-relationship triples; in addition to the RNA algorithm, this unit can also integrate other relationship extraction algorithms, such as a relationship extraction method based on remote supervision, or a neural network model based on an attention mechanism, to improve the accuracy and coverage of relationship extraction.

[0065] The knowledge base establishment unit is used to use a graph database to store the identified entities, attributes and relationships to form a preliminary knowledge base.

[0066] This system uses Neo4j graph database, which has high query efficiency and is easy to expand. The nodes in the graph database are used to store entities and their attributes, and the edges are used to represent the relationship between entities, supporting complex graph traversal and path query operations.

[0067] The data fusion module is used to fuse the knowledge from different sources and in different formats in the preliminary knowledge base, use redundancy elimination and conflict resolution algorithms to fuse the knowledge, and establish a knowledge graph without redundancy.

[0068] In practical applications, data from different sources may be duplicate, contradictory or incomplete, so systematic data fusion processing is required; the data fusion module includes: a standardization unit, a redundancy elimination unit and a conflict resolution unit; when data conflicts are found, the system will record the causes and solutions of the conflicts to facilitate subsequent optimization and improvement.

[0069] The standardization unit is used to standardize data from different sources and formats;

[0070] For quantitative tumor marker data, a modified Min-Max normalization method was used: Among them, x is the value before standardization, x ′ is the standardized value, x max and x min are the upper and lower limits of the clinical reference value range of the marker to be standardized, and U and L are the upper and lower limits of the target range of the data after standardization.

[0071] For qualitative or semi-quantitative tumor marker data, piecewise function mapping is used for normalization:

[0072]

[0073] Among them, x ref These are clinically typical values ​​for tumor marker data.

[0074] Regarding the specific implementation of the standardization unit: For quantitative tumor marker data, the improved Min-Max standardization method is used for processing. Taking alpha-fetoprotein (AFP) as an example, its clinical reference range is 0-7ng / mL. If the standardized range is set to [0,1], then U=1, L=0. When the detection value x=15ng / mL, the standardized value can be calculated by the formula; for the standardization of qualitative or semi-quantitative data, piecewise function mapping is used. For example, for the qualitative description of prostate-specific antigen (PSA), descriptions such as "normal", "mildly elevated", and "significantly elevated" can be mapped to different numerical segments within the standardized interval, ensuring the unified expression of qualitative descriptions and quantitative data.

[0075] The redundancy elimination unit is used to eliminate redundant data; the conflict resolution unit is used to use game theory models and Bayesian networks to resolve conflicts and ensure the consistency of the knowledge graph.

[0076] The marker analysis module is used to establish a tumor marker analysis model, and utilize the non-redundant knowledge graph to evaluate and predict the patient's tumor marker level and perform outlier determination.

[0077] The redundancy elimination unit includes: a linear transformation subunit and a conditional screening subunit; the redundancy elimination unit uses a combination of linear transformation and conditional screening to process data redundancy. The linear transformation subunit uses a multi-layer perceptron (MLP) to achieve feature dimension reduction and information compression, while the conditional screening subunit uses a probability model to screen the most representative data points. Each feature function f k It can be customized according to specific application scenarios. For example, time series correlation features, numerical interval features, etc. can be designed.

[0078] The linear transformation subunit is used to implement the following linear transformation:

[0079] Input layer to hidden layer: h = ReLU(W1x′+b1);

[0080] Hidden layer to output layer: y MLP =ReLU(W2h+b2);

[0081] Among them, W1 is the weight matrix from the input layer to the hidden layer, which is used to weight the importance of different marker features. Each element in the matrix represents the influence of the corresponding marker feature on the hidden layer neurons; b1 is the bias vector of the hidden layer, which is used to adjust the activation threshold of the hidden layer neurons to ensure that the model can adapt to different ranges of marker values; h is the output vector of the hidden layer, which represents the high-level feature representation extracted from the original marker data; W2 is the weight matrix from the hidden layer to the output layer, which is used to map the extracted high-level features to the final marker evaluation results; b2 is the bias vector of the output layer, which is used to fine-tune the final output marker evaluation value; y MLP is the output vector of the output layer, which represents the feature vector of the landmark after being processed by the neural network;

[0082] The conditional screening subunit is used to realize conditional probability distribution: Among them, P(y|y MLP ) is the conditional probability of a specific combination of marker states given the output of the MLP; y is the output sequence of marker states, representing the combined state of multiple related markers; M(y MLP ) is a normalization factor used to ensure that the sum of the probabilities of all possible states is 1; λ k is the weight coefficient of the kth feature function, reflecting the importance of this feature in the judgment of the marker state; f k (y,y MLP ) is the kth characteristic function, which is used to capture the local dependencies in the marker state sequence;

[0083] By maximizing the conditional probability P(y|y MLP ) to obtain the output sequence y of the optimal marker state * :

[0084]

[0085] The entity recognition unit includes a named entity recognition model based on BiLSTM-CRF, and the named entity recognition model includes: an input layer, a BiLSTM layer and a CRF layer;

[0086] The input layer is used to receive text sequence input;

[0087] The BiLSTM layer is used to extract contextual features of the text;

[0088] The CRF layer is used to label the feature sequence output by the BiLSTM layer and identify tumor marker related entities;

[0089] Among them, the forward propagation and back propagation of the BiLSTM layer are:

[0090]

[0091] is the forward hidden layer state at time t, representing the text information encoding from left to right, is the reverse hidden layer state at time t, indicating the text information encoding from right to left; W f is the weight matrix of forward propagation, which is used to control the transmission strength of forward information; W b is the weight matrix of back propagation, which is used to control the transmission strength of reverse information; b f is the bias vector of the forward propagation, which is used to adjust the reference value of the forward hidden layer; b b is the bias vector of the back propagation, which is used to adjust the baseline value of the reverse hidden layer; x t is the input vector at time t, which represents the feature representation of the currently processed word; σ is the activation function, which uses the sigmoid or tanh function to introduce nonlinear transformation.

[0092] In the implementation of the BiLSTM-CRF named entity recognition model, the input layer uses word embedding technology to convert the input text into a dense vector representation. In practice, a pre-trained medical domain word vector model (such as Medical-BioBERT) can be used to initialize the word embedding layer to improve the feature extraction capability of the model.

[0093] The BiLSTM layer extracts contextual features of the text through bidirectional information flow. For example, for the sentence "serum AFP levels increased significantly", the forward LSTM can capture the limiting relationship between "serum" and "AFP", while the reverse LSTM can capture the state description of "increase" to "AFP". The initial values ​​of the weight matrices W_f and W_b can be determined by the Xavier initialization method.

[0094] The marker analysis module includes: a knowledge reasoning unit and a marker evaluation unit;

[0095] The knowledge reasoning unit is used to perform reasoning based on the non-redundant knowledge graph to calculate the patient marker level:

[0096] Among them, M d is the comprehensive marker level, w i is the weight of the i-th relationship in the knowledge graph, R i is the contribution value of the i-th relationship, v j is the weight of the jth attribute, A j is the normalized value of the jth attribute;

[0097] The marker evaluation unit is used to establish a tumor marker analysis model by combining clinical experience and statistical laws in the knowledge graph:

[0098] E(t)=f(M d ,G,H)+∈, where E(t) is the marker evaluation value at time t, f(M d ,G,H) is a nonlinear evaluation function, G is the group feature provided by the knowledge graph, H is the historical data feature, and ∈ is the random error term.

[0099] In the marker evaluation model, the nonlinear evaluation function f(M d ,G,H) can choose from a variety of implementation methods:

[0100] Based on deep neural network: build a multi-layer perceptron, the input layer receives three types of features M_d, G, and H, and combines and transforms features through multiple hidden layers.

[0101] Based on ensemble learning: Comprehensively use multiple base learners such as random forest and XGBoost to improve the generalization ability of the model.

[0102] Based on Gaussian process: kernel function is used to capture the nonlinear relationship between features and obtain the uncertainty estimate of the prediction.

[0103] The marker analysis module also includes:

[0104] Dynamic threshold unit, used to dynamically adjust the outlier determination threshold based on the knowledge graph:

[0105] T=α×T b +(1-α)×T d ; Where T is the dynamic threshold, T b is the benchmark threshold in the knowledge graph, T d It is a data-driven threshold; in the dynamic threshold adjustment mechanism, the calculation of cosine similarity takes into account multiple dimensional features, including: demographic characteristics (age, gender, etc.), clinical characteristics (disease stage, treatment plan, etc.) and marker characteristics (historical fluctuation trend, etc.). The similarity threshold of 0.8 can be adjusted according to actual application needs.

[0106] The calculation method of the balance factor α is:

[0107] Among them, n ref is the number of standard cases referenced in the knowledge graph, n sim is the number of historical cases in the knowledge graph that are similar to the current case.

[0108] The data driven threshold T d Calculate by the following steps:

[0109] Step S1, using cosine similarity to calculate the similarity between the current case and the historical case, and selecting cases with similarity greater than 0.8 as the similar case set S;

[0110] Step S2: cluster the marker values ​​in the similar case set S, T d Take the mean of the largest cluster.

[0111] The system also includes: a data storage module for storing the patient's historical test data, clinical standard data and intermediate data generated during the system operation.

[0112] The system also includes: a user interface module, which is used to receive detection data and system parameters input by a user, and to display analysis results and abnormal warning information.

[0113] In the implementation of the data storage module, a distributed storage architecture is adopted:

[0114] Structured data (such as test values, reference ranges, etc.) are stored in relational databases (such as PostgreSQL); knowledge graph data are stored in graph databases (such as Neo4j); unstructured data (such as medical literature, examination reports, etc.) are stored in document databases (such as MongoDB).

[0115] In addition to basic data input and result display functions, the user interface module also provides the following special features: visualization of marker change trends; push of abnormal value warning information; interactive query of knowledge graphs; and explanatory display of analysis results.

[0116] It should be noted that those skilled in the art should understand that various changes and equivalent substitutions may be made to the present invention without departing from the scope of the present invention. In addition, various modifications may be made to the present invention for specific situations or materials without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed, but should include all embodiments falling within the scope of the claims of the present invention.

Claims

1. A tumor marker data intelligent processing system, characterized in that: The system comprises a data extraction module, a data fusion module and a marker analysis module connected in sequence; The data extraction module is used to collect basic data related to tumor markers, extract knowledge from the basic data, extract structured information including entities, attributes and relationships between entities, and form a preliminary knowledge base; The data extraction module includes: an entity recognition unit, an attribute extraction unit, a relationship mining unit and a knowledge base establishment unit; The entity recognition unit is used to identify entities related to tumor markers from text and unstructured data; The attribute extraction unit is used to extract and quantify the attributes associated with each entity by adopting an attribute value embedding method; The relationship mining unit is used to extract the relationship between entities from the data through the RNA algorithm to form entity-relationship triples; The knowledge base building unit is used to use a graph database to store the identified entities, attributes and relationships to form a preliminary knowledge base; The data fusion module is used to fuse knowledge from different sources and in different formats in the preliminary knowledge base, and use redundancy elimination and conflict resolution algorithms to fuse knowledge and establish a knowledge graph without redundancy; The data fusion module includes: a standardization unit, a redundancy elimination unit and a conflict resolution unit; The standardization unit is used to standardize data from different sources and formats; For quantitative tumor marker data, a modified Min-Max normalization method was used: Among them, x is the value before standardization, x′ is the value after standardization, and x max and x min are the upper and lower limits of the range of clinical reference values ​​of the marker to be standardized, respectively, and U and L are the upper and lower limits of the target range of the data after standardization; For qualitative or semi-quantitative tumor marker data, piecewise function mapping is used for normalization: Among them, x ref are clinically typical values ​​of tumor marker data; The redundancy elimination unit is used to eliminate redundant data; The conflict resolution unit is used to resolve conflicts using a game theory model and a Bayesian network to ensure the consistency of the knowledge graph; The marker analysis module is used to establish a tumor marker analysis model, and utilize the non-redundant knowledge graph to evaluate and predict the patient's tumor marker level and perform outlier determination.

2. The tumor marker data intelligent processing system according to claim 1, characterized in that: The redundancy elimination unit includes: a linear transformation subunit and a conditional screening subunit; The linear transformation subunit is used to implement the following linear transformation: Input layer to hidden layer: h = ReLU(W1x′+b1); Hidden layer to output layer: y MLP =ReLU(W2h+b2); Among them, W1 is the weight matrix from the input layer to the hidden layer, which is used to weight the importance of different marker features. Each element in the matrix represents the influence of the corresponding marker feature on the hidden layer neurons; b1 is the bias vector of the hidden layer, which is used to adjust the activation threshold of the hidden layer neurons to ensure that the model can adapt to different ranges of marker values; h is the output vector of the hidden layer, which represents the high-level feature representation extracted from the original marker data; W2 is the weight matrix from the hidden layer to the output layer, which is used to map the extracted high-level features to the final marker evaluation results; b2 is the bias vector of the output layer, which is used to fine-tune the final output marker evaluation value; y MLP is the output vector of the output layer, which represents the feature vector of the landmark after being processed by the neural network; The conditional screening subunit is used to realize conditional probability distribution: Among them, P(y|y MLP ) is the conditional probability of a specific combination of marker states given the output of the MLP; y is the output sequence of marker states, representing the combined state of multiple related markers; M(y MLP ) is a normalization factor used to ensure that the sum of the probabilities of all possible states is 1; λ k is the weight coefficient of the kth feature function, reflecting the importance of this feature in the judgment of the marker state; f k (y,y MLP ) is the kth characteristic function, which is used to capture the local dependencies in the marker state sequence; By maximizing the conditional probability P(y|y MLP ) to obtain the output sequence y of the optimal marker state * :

3. The tumor marker data intelligent processing system according to claim 2, characterized in that: The entity recognition unit includes a named entity recognition model based on BiLSTM-CRF, and the named entity recognition model includes: an input layer, a BiLSTM layer and a CRF layer; The input layer is used to receive text sequence input; The BiLSTM layer is used to extract contextual features of the text; The CRF layer is used to label the feature sequence output by the BiLSTM layer and identify tumor marker related entities; Among them, the forward propagation and back propagation of the BiLSTM layer are: is the forward hidden layer state at time t, representing the text information encoding from left to right, is the reverse hidden layer state at time t, indicating the text information encoding from right to left; W f is the weight matrix of forward propagation, which is used to control the transmission strength of forward information; W b is the weight matrix of back propagation, which is used to control the transmission strength of reverse information; b f is the bias vector of the forward propagation, which is used to adjust the reference value of the forward hidden layer; b b is the bias vector of the back propagation, which is used to adjust the baseline value of the reverse hidden layer; x t is the input vector at time t, which represents the feature representation of the currently processed word; σ is the activation function, which uses the sigmoid or tanh function to introduce nonlinear transformation.

4. The tumor marker data intelligent processing system according to claim 3, characterized in that: The marker analysis module includes: a knowledge reasoning unit and a marker evaluation unit; The knowledge reasoning unit is used to perform reasoning based on the non-redundant knowledge graph to calculate the patient marker level: Among them, M d is the comprehensive marker level, w i is the weight of the i-th relationship in the knowledge graph, R i is the contribution value of the i-th relationship, v j is the weight of the jth attribute, A j is the normalized value of the jth attribute; The marker evaluation unit is used to establish a tumor marker analysis model by combining clinical experience and statistical laws in the knowledge graph: E(t)=f(M d ,G,H)+∈, where E(t) is the marker evaluation value at time t, f(M d ,G,H) is a nonlinear evaluation function, G is the group feature provided by the knowledge graph, H is the historical data feature, and ∈ is the random error term.

5. The tumor marker data intelligent processing system according to claim 4, characterized in that: The marker analysis module also includes: Dynamic threshold unit, used to dynamically adjust the outlier determination threshold based on the knowledge graph: T=α×T b +(1-α)×T d ; Where T is the dynamic threshold, T b is the benchmark threshold in the knowledge graph, T d is a data-driven threshold; The calculation method of the balance factor α is: Among them, n ref is the number of standard cases referenced in the knowledge graph, n sim is the number of historical cases in the knowledge graph that are similar to the current case.

6. The tumor marker data intelligent processing system according to claim 5, characterized in that: The data driven threshold T d Calculate by the following steps: Step S1, using cosine similarity to calculate the similarity between the current case and the historical case, and selecting cases with similarity greater than 0.8 as the similar case set S; Step S2: cluster the marker values ​​in the similar case set S, T d Take the mean of the largest cluster.

7. The tumor marker data intelligent processing system based on knowledge graph according to claim 6, characterized in that: The system further comprises: The data storage module is used to store the patient's historical test data, clinical standard data and intermediate data generated during the system operation.

8. The tumor marker data intelligent processing system based on knowledge graph according to claim 7, characterized in that: The system further comprises: The user interface module is used to receive the detection data and system parameters input by the user, and to display the analysis results and abnormal warning information.

Images