Medical diagnosis prediction method and system based on knowledge guidance and credible evidence learning

Abstract

The present application belongs to the technical field of medical data analysis and artificial intelligence auxiliary diagnosis, and provides a medical diagnosis prediction method and system based on knowledge guidance and credible evidence learning. Electronic medical record data of a patient to be tested is acquired; a trained credible diagnosis evolution and prediction model is used to perform diagnosis prediction on the electronic medical record data of the patient to be tested, and a high-risk blind area early warning is triggered when the uncertainty score is higher than a set threshold; the training process of the credible diagnosis evolution and prediction model is as follows: a global representation vector of the patient is generated, and then positive evidence and negative evidence of various diseases are obtained and quantified as Dirichlet distribution, and prediction probability and prediction uncertainty score of various diseases are calculated; a binary evidence loss function is used to calculate prediction error, and model parameters are optimized. It can not only provide high-precision diagnosis prediction results, but also accurately quantify cognitive uncertainty.

Description

Technical Field

[0001] This invention belongs to the field of medical data analysis and artificial intelligence-assisted diagnosis technology, and in particular relates to a medical diagnostic prediction method and system based on knowledge-guided and credible evidence learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Electronic Health Records (EHR) systems are widely used in medical institutions, accumulating a massive amount of patient lifecycle diagnosis and treatment data. Mining the temporal patterns in these data is of great significance for early disease screening, risk warning, and assisting clinical decision-making. Existing deep learning-based diagnostic prediction methods still have the following obvious limitations when facing real clinical environments: (1) Existing Graph Neural Networks (GNNs) mainly rely on purely data-driven statistical co-occurrence information when constructing disease associations. Real electronic health record data usually exhibits extremely long-tailed distribution characteristics and is accompanied by high sparsity. Due to the small sample size, many rare diseases or atypical complications are difficult to form significant co-occurrence relationships at the statistical level, causing long-tailed nodes to become "isolated points" in the graph structure and failing to obtain effective feature aggregation; at the same time, high-frequency common diseases are prone to generating a large number of random noise connections without clinical significance. (2) Existing deep learning diagnostic prediction models generally lack the ability to quantify cognitive uncertainty and have serious "overconfidence" risks. Traditional models mostly use functions such as Softmax for point estimation and prediction. Due to their forced normalization characteristics, even when faced with abnormal and difficult medical records with missing information, severe noise interference, or exceeding their cognitive boundaries, the models will still output extremely high confidence probabilities. This defect of not being able to identify its own "cognitive blind spots" makes it impossible for the system to trigger safety rejection and manual intervention warnings in a timely manner, and thus cannot be applied to real medical scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a medical diagnostic prediction method and system based on knowledge-guided and credible evidence learning, which not only provides high-precision diagnostic prediction results but also accurately quantifies cognitive uncertainty.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a medical diagnostic prediction method based on knowledge-guided and credible evidence learning.

[0006] In one or more embodiments, a medical diagnostic prediction method based on knowledge-guided and credible evidence learning is provided, including: Obtain the electronic medical record data of the patients to be tested; The trained reliable diagnostic evolution and prediction model is used to make diagnostic predictions on the electronic medical record data of the patient to be tested. When the uncertainty score is higher than a set threshold, a high-risk blind zone warning is triggered. The training process of the reliable diagnostic evolution and prediction model is as follows: Obtain the patient's historical diagnosis sequence and continuous clinical monitoring indicator matrix from historical electronic medical record data; Disease statistical co-occurrence information is extracted from the historical diagnostic sequences to construct a statistical co-occurrence graph. Based on the diagnosed diseases during medical visits, an asymmetric gating residual mechanism is used to adaptively and weightedly fuse the pre-constructed prior knowledge graph with the statistical co-occurrence graph to generate a knowledge-data dual-driven diagnostic representation vector. Clinical monitoring indicator features are extracted from the clinical monitoring indicator matrix, their semantic alignment with the diagnostic representation vector is calculated, and an intensity-aware gating mechanism is introduced to generate a feature-disease similarity representation vector. The temporal pattern of the disease is determined based on the diagnosis records of two consecutive patient visits. The self-attention mechanism and the cross-attention mechanism are fused with the feature-disease similarity representation vector to generate the patient's global representation vector. Based on the global representation vector, positive and negative evidence for various diseases are obtained and their parameters are parameterized into a Dirichlet distribution. The prediction probability and prediction uncertainty score of various diseases are calculated. The prediction error is calculated using the binary evidence loss function, and the model parameters are then optimized.

[0007] As one implementation method, in the process of constructing the statistical co-occurrence graph, the co-occurrence frequency of disease pairs is statistically analyzed based on the historical diagnostic sequences of all patients in the training set, and low-frequency combinations are filtered out through preset hyperparameters to construct the statistical co-occurrence graph and its adjacency matrix.

[0008] As one implementation method, the process of adaptively weighting and fusing a pre-constructed prior knowledge graph with a statistical co-occurrence graph using an asymmetric gated residual mechanism to generate a knowledge-data dual-driven diagnostic representation vector is as follows: Based on prior knowledge graphs and statistical co-occurrence graphs, diagnostic features of confirmed diseases under knowledge-driven and statistical-driven approaches are extracted respectively, and then concatenated along the feature dimension to obtain concatenated features. Based on the splicing features, the feature correction amount and the fusion gating weight are calculated separately through bilinear branches; After multiplying the feature correction amount element-wise with the fusion gating weight, it is then fused with the diagnostic features of the confirmed disease under statistical drive through residual connection to obtain a knowledge-data dual-driven diagnostic representation vector.

[0009] As one implementation method, the process of generating feature-disease similarity representation vectors is as follows: Cosine similarity is used to calculate the semantic alignment between clinical monitoring indicator features and diagnostic representation vectors, resulting in a preliminary feature-disease similarity matrix; A deviation feature vector of the abnormality degree of clinical monitoring indicators is constructed, and then the gating weight is dynamically calculated based on the intensity-sensing gating mechanism. Based on the preliminary feature-disease similarity matrix and gating weights, the final feature-disease similarity matrix is ​​obtained; Align the final feature-disease similarity matrix with the disease embeddings normalized in the same space to obtain the feature-disease similarity representation vector.

[0010] As one implementation method, the ICD-9 medical ontology hierarchy is introduced to construct a priori knowledge graph. The process is as follows: Extract the disease hierarchy tree structure from the ICD-9 medical ontology hierarchy, and construct a node set and edge set containing all disease codes; if two diseases have the same parent node in the ICD-9 medical ontology hierarchy, or have a direct superior-subordinate relationship, then establish a bidirectional edge connection in the graph. Based on the node set and edge set containing all disease codes, a medical ontology knowledge graph is constructed. The topology of the medical ontology knowledge graph is represented by an adjacency matrix, which serves as a prior knowledge constraint to guide the model in medical logical reasoning and provides semantic anchors for isolated long-tail rare disease nodes.

[0011] As one implementation method, the temporal patterns of the disease include three temporal patterns: persistent disease, evolving disease, and sporadic disease.

[0012] As one implementation method, a two-stage hybrid training strategy of cross-entropy warm-up and evidence loss fine-tuning is adopted during the training of the credible diagnostic evolution and prediction model: in the early stage of training, the standard binary cross-entropy loss function is used to guide the credible diagnostic evolution and prediction model to quickly learn the basic feature representation; after the performance of the credible diagnostic evolution and prediction model tends to stabilize, the binary evidence loss function is switched to fine-tuning training, and the backpropagation algorithm is used to update the network weight parameters until the model converges.

[0013] A second aspect of the present invention provides a medical diagnostic prediction system based on knowledge-guided and credible evidence learning.

[0014] In one or more embodiments, a medical diagnostic prediction system based on knowledge-guided and credible evidence learning includes: The data acquisition module is used to acquire the electronic medical record data of the patient to be tested. The diagnostic prediction module is used to make diagnostic predictions on the electronic medical record data of the patient under test using a trained reliable diagnostic evolution and prediction model. When the uncertainty score is higher than a set threshold, a high-risk blind zone warning is triggered. The training process of the reliable diagnostic evolution and prediction model is as follows: Obtain the patient's historical diagnosis sequence and continuous clinical monitoring indicator matrix from historical electronic medical record data; Disease statistical co-occurrence information is extracted from the historical diagnostic sequences to construct a statistical co-occurrence graph. Based on the diagnosed diseases during medical visits, an asymmetric gating residual mechanism is used to adaptively and weightedly fuse the pre-constructed prior knowledge graph with the statistical co-occurrence graph to generate a knowledge-data dual-driven diagnostic representation vector. Clinical monitoring indicator features are extracted from the clinical monitoring indicator matrix, their semantic alignment with the diagnostic representation vector is calculated, and an intensity-aware gating mechanism is introduced to generate a feature-disease similarity representation vector. The temporal pattern of the disease is determined based on the diagnosis records of two consecutive patient visits. The self-attention mechanism and the cross-attention mechanism are fused with the feature-disease similarity representation vector to generate the patient's global representation vector. Based on the global representation vector, positive and negative evidence for various diseases are obtained and their parameters are parameterized into a Dirichlet distribution. The prediction probability and prediction uncertainty score of various diseases are calculated. The prediction error is calculated using the binary evidence loss function, and the model parameters are then optimized.

[0015] A third aspect of the present invention provides a computer-readable storage medium.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described above.

[0017] A fourth aspect of the present invention provides an electronic device.

[0018] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described above.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention extracts statistical co-occurrence information of diseases from historical diagnostic sequences, constructs a statistical co-occurrence graph, and uses an asymmetric gating residual mechanism to adaptively weight and fuse a pre-constructed prior knowledge graph with the statistical co-occurrence graph, generating a knowledge-data dual-driven diagnostic representation vector. This provides an effective semantic anchor for rare diseases and complex complications in long-tail distributions, successfully overcoming the problems of graph construction bias and hindered graph feature transmission caused by sample sparsity in pure data-driven models. Combining the semantic alignment of clinical monitoring indicator features with the diagnostic representation vector, an intensity-aware gating mechanism is introduced to generate a feature-disease similarity representation vector and fuse it with the temporal pattern of the disease to generate a global representation vector for the patient. This further identifies positive and negative evidence for various diseases and parameterizes it into a Dirichlet distribution, calculating the predicted probability and prediction uncertainty score for each disease. This not only provides high-precision diagnostic prediction results but also accurately quantifies cognitive uncertainty, effectively outputting high uncertainty scores and triggering a safe rejection mechanism, effectively eliminating the "overconfidence" risk of large medical models and significantly enhancing the safety and reliability of clinical decision-making. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a flowchart of a medical diagnostic prediction method based on knowledge guidance and credible evidence learning according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the training process of the reliable diagnostic evolution and prediction model according to an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, 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 pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Terminology Explanation: The ICD-9 medical ontology uses a three-digit core structure (chapter / category → subcategory → detail). The clinical version (ICD-9-CM) expands to 21 chapters of diseases + 3 volumes of operations. The coding logic is 1 digit for the major category → 2 digits for the main group → 3 digits for the specific disease / operation, with supplementary subdivisions after the decimal point.

[0026] according to Figure 1 The medical diagnostic prediction method based on knowledge guidance and credible evidence learning in this embodiment may include the following steps S101~S102.

[0027] The specific implementation process of steps S101 to S102 is as follows: Step S101: Obtain the electronic medical record data of the patient to be tested.

[0028] All electronic medical record data here has been obtained with the patient's consent and through legal means.

[0029] Step S102: Use the trained reliable diagnostic evolution and prediction model to perform diagnostic prediction on the electronic medical record data of the patient to be tested. When the uncertainty score is higher than the set threshold, a high-risk blind zone warning is triggered.

[0030] according to Figure 2 The training process of the reliable diagnostic evolution and prediction model is as follows: Step S1021: Obtain the patient's historical diagnosis sequence and continuous clinical monitoring indicator matrix from historical electronic medical record data.

[0031] Electronic medical record data includes discrete historical diagnostic records and continuous clinical monitoring indicators.

[0032] The acquired electronic medical record data undergoes data preprocessing, including data cleaning, missing data completion, data definition, and storage.

[0033] Specifically, this application example uses historical medical data obtained from an intensive care unit (ICU) database in a specific city or region. Patient data with two or more hospitalization records are selected, and the dataset is divided into training, validation, and experimental samples according to a preset ratio. Furthermore, based on the ICD-9-CM standard, multi-heat vectors of the patients' disease diagnosis codes are extracted, along with a continuous clinical feature matrix covering vital signs and laboratory test results.

[0034] Step S1022: Extract disease statistical co-occurrence information from the historical diagnostic sequence and construct a statistical co-occurrence graph; based on the diagnosed disease during the consultation, use an asymmetric gating residual mechanism to adaptively weight and fuse the pre-constructed prior knowledge graph with the statistical co-occurrence graph to generate a knowledge-data dual-driven diagnostic representation vector.

[0035] For the acquired medical data samples, the data is first standardized and processed to select features that influence medical treatment factors, eliminate feature redundancy, and extract features that are highly correlated with hospitalization behavior.

[0036] Specifically, the acquired continuous clinical monitoring indicators and discrete diagnostic codes are cleaned, standardized, and vectorized to eliminate scale differences in heterogeneous features and maintain the continuity of time-series data. Furthermore, addressing the limitations of purely data-driven methods in medical scenarios, this approach combines local and global disease statistical patterns with authoritative medical ontology knowledge to eliminate random noise connections in the graph structure and compensate for semantic deficiencies caused by rare disease samples.

[0037] For the acquired electronic medical record data samples, in real clinical settings, the small number of rare disease samples makes it difficult to form significant co-occurrence connections at the statistical level. Therefore, when extracting features and constructing disease associations, authoritative medical domain knowledge (ICD-9) is used to guide the construction of statistical co-occurrence graphs. Semantic topological supplementation is performed on long-tail nodes that influence the evolution of patients' diseases to address the graph construction bias caused by extremely sparse data.

[0038] Define the total dimension of all disease diagnosis codes in the electronic medical record dataset as follows: .set up This is a patient dataset constructed from an electronic medical record system. For any patient... In its first Second visit ( In this context, the patient's diagnostic results are derived from multiple heat vectors. Indicates. If The If an element is 1, it means that the patient is in the first... During the second medical visit, the patient was diagnosed with the first... Disease Conversely, it indicates that the patient does not have the disease. Indicates the patient was in the first A continuous clinical feature matrix during each visit, including This refers to the length of the sampling sequence during this hospitalization. These are the characteristic dimensions of predefined standardized examination indicators. For patient u, the complete medical event sequence fully covers the entire diagnosis and treatment trajectory within the patient's observation period, consisting of a series of chronologically arranged medical records, denoted as... .in This represents the patient's total number of outpatient visits. The data were standardized by using a combination of forward imputation and mean imputation to fill missing values, and population quantile imputation to fill single test indicators. Outliers were truncated based on clinical reference thresholds and Z-score standardized based on statistical distribution.

[0039] In one or more embodiments, during the construction of the statistical co-occurrence graph, the co-occurrence frequency of disease pairs is statistically analyzed based on the historical diagnostic sequences of all patients in the training set, and low-frequency combinations are filtered out by preset hyperparameters to construct the statistical co-occurrence graph and its adjacency matrix.

[0040] Disease co-occurrence information was extracted from historical diagnostic sequences, and a statistical co-occurrence graph was constructed. Based on the medical records of all patients in the training set, disease pairs were statistically analyzed. co-occurrence frequency And through hyperparameters Filter low-frequency combinations and construct a statistical co-occurrence graph. and its adjacency matrix The calculation formula is as follows: ; in For qualified neighbors with diseases and illnesses The sum of the collinear frequencies that co-occur.

[0041] The process of introducing the ICD-9 medical ontology hierarchy to construct a prior knowledge graph is as follows: Extract the disease hierarchy tree structure from the ICD-9 medical ontology hierarchy and construct a structure containing all A set of nodes for a disease code and edge set If two diseases share the same parent node in the ICD-9 medical ontology hierarchy, or have a direct hierarchical relationship, then a bidirectional edge connection is established in the graph. This constructs a medical ontology knowledge graph. Its topology is composed of an adjacency matrix. This adjacency matrix serves as a priori knowledge constraint guiding the model's medical logical reasoning, providing semantic anchors for isolated long-tail rare disease nodes.

[0042] The process of adaptively weighting and fusing a pre-constructed prior knowledge graph and a statistical co-occurrence graph using an asymmetric gated residual mechanism to generate a knowledge-data dual-driven diagnostic representation vector is as follows: Step a1: Based on the prior knowledge graph and the statistical co-occurrence graph, extract the diagnostic features of the confirmed disease under the knowledge-driven approach and the diagnostic features under the statistical approach, respectively, and then concatenate them in the feature dimension to obtain the concatenated features; For the patient's first The second visit, based on its current diagnostic set and multiple heat vectors Dynamically decompose the subgraph into three types of adjacency matrices: including the first type... Local diagnostic map adjacency matrix of all diagnoses during this visit ; This represents a global diagnostic-neighborhood graph adjacency matrix that associates diagnosed diseases with adjacent, untreated diseases. ; and the global neighborhood graph adjacency matrix representing the association of untreated disease pairs. To avoid the computational burden of frequently accessing high-dimensional tensors during deep model training, an implicit parameter-sharing representation mechanism is adopted to map the aforementioned dynamic subgraph structure into a global statistical co-occurrence graph adjacency matrix. A linear combination of . Definition To record the indicator vectors of all neighboring nodes at time t, let... and Let be the initial embedding matrices of the disease at the time of diagnosis and when it is a neighbor, respectively. Then, under the statistical co-occurrence channel, the aggregated context of the diagnosed disease is... Aggregate context with neighboring nodes The calculation is as follows: ; ; in This indicates an element-wise multiplication operation.

[0043] Introducing medical ontology knowledge channels for parallel utilization of diagnosed diseases. and Perform feature aggregation. Diagnostic features of confirmed diseases under statistical driving principles. Knowledge-driven diagnostic features The calculation is as follows: ; ; in, To make the above The statistical co-occurrence graph adjacency matrix in the formula Replace with ontology knowledge adjacency matrix The resulting knowledge aggregation context; and These are the trainable weight transformation matrices for the statistical graph and knowledge graph channels, respectively. For neighboring disease nodes that were not diagnosed in this visit but have a correlation, to maintain the true statistical distribution of their local neighborhood semantics, the model relies solely on the data-driven statistical co-occurrence graph for feature extraction, and the final neighbor node embeddings... The calculation is as follows: ; Step a2: Based on the spliced ​​features, calculate the feature correction amount and fusion gate weights respectively through bilinear branches; An asymmetric gated residual concatenation mechanism is introduced for diagnosed diseases. Statistical features and knowledge features are concatenated along the feature dimension to obtain... The feature correction is calculated using bilinear branches. With fusion gating weights : ; ; in, and is a learnable parameter matrix.

[0044] Step a3: Multiply the feature correction amount and the fusion gating weight element by element, and then fuse it with the diagnostic features of the confirmed disease under statistical drive through residual connection to obtain the knowledge-data dual-driven diagnostic representation vector.

[0045] Knowledge-data dual-driven diagnostic representation vector The result is obtained through residual connection: ; This mechanism enables the model to adaptively control the degree of prior knowledge injection, preserving statistical features for simple or high-frequency samples, while using prior medical knowledge to correct complex or sparse samples, ultimately outputting a knowledge- and data-driven diagnostic representation vector. .

[0046] This embodiment dynamically decomposes the statistical co-occurrence subgraph at each time step to capture the local pathological dependencies of the current medical visit, and performs feature aggregation by utilizing the statistical co-occurrence channel and the ontology knowledge channel in parallel. To address the problem of irrelevant knowledge noise diffusion that can easily occur when introducing medical prior knowledge, an asymmetric aggregation and gating strategy is implemented for diagnosed diseases and undiagnosed neighboring diseases.

[0047] Step S1023: Extract clinical monitoring indicator features from the clinical monitoring indicator matrix, calculate their semantic alignment with the diagnostic representation vector, and introduce an intensity-aware gating mechanism to generate a feature-disease similarity representation vector.

[0048] A continuous clinical monitoring index matrix is ​​subjected to time-dimensional pooling and linear mapping, and a feature-disease similarity representation vector is generated using cosine similarity and intensity-aware gating mechanisms. By introducing cosine similarity measurement and feature intensity-aware gating mechanisms, the model can adaptively adjust the contribution ratio between the global disease contribution pattern and the local pathological mechanism, thereby effectively eliminating the distribution bias between different features and providing an accurate contextual representation for capturing the subsequent dynamic evolution trajectory of the disease.

[0049] Specifically, the process of generating the feature-disease similarity representation vector is as follows: Step b1: Calculate the semantic alignment between clinical monitoring indicator features and diagnostic representation vectors using cosine similarity to obtain a preliminary feature-disease similarity matrix; For patients in the first Continuous clinical feature matrix during each visit First, mean pooling over time is used to eliminate differences in sampling intervals, extracting the global trend of indicators within the medical visit cycle, and then mapping it to the disease semantic space through linear transformation. Features after mapping each indicator to the disease semantic space ( The dimension for disease embedding is calculated as follows: ; in, Indicates the first The first time step Individual indicator values, and These are learnable parameters.

[0050] Cosine similarity is used to measure the semantic relevance between continuous clinical indicators and disease nodes by the alignment of vector directions in the vector space. The fused diagnostic representation vector is then calculated. The Middle Disease characteristics With the Individual indicators mapping characteristics cosine similarity : ; in, .

[0051] Step b2: Construct a deviation feature vector of the abnormality degree of the clinical monitoring indicators, and then dynamically calculate the gating weight based on the intensity-sensing gating mechanism; To quantify the degree of abnormality of clinical monitoring indicators, a deviation feature vector of the indicators is further constructed. : ; in, and These represent the industry standard reference mean and standard deviation, respectively, based on large-scale real-world clinical data statistics. It is a very small constant used to prevent division by zero. Indicates the first During the second visit, the total number of sampling time steps for the patient's clinical test indicators was... Indicates the first During the second visit, the first The vector of clinical test indicators corresponding to each sampling time.

[0052] Considering that abnormalities in a specific clinical indicator are often common symptoms of multiple diseases, a feature intensity-aware gating mechanism is introduced to adaptively quantify the different contributions of abnormal indicators to different diseases. This is achieved by utilizing deviation feature vectors. Dynamically calculate gate weights : ; in, For the Sigmoid function, and These are the gating parameters.

[0053] Step b3: Based on the preliminary feature-disease similarity matrix and gating weights, obtain the final feature-disease similarity matrix; The gating weights are applied to the initial similarity matrix to obtain the final feature-disease similarity matrix. : ; in, This indicates element-wise multiplication. Represents the tensor product. It is an all-1 vector used to implement the broadcast calculation of gating weights on the disease dimension.

[0054] Step b4: Align the final feature-disease similarity matrix with the disease embeddings normalized in the same space to obtain the feature-disease similarity representation vector.

[0055] To eliminate distributional biases between different features and achieve dynamic feature fusion, the fusion process of the similarity matrix is ​​aligned with the disease embeddings normalized in the same space, resulting in a final feature representation that incorporates clinical monitoring indicators. : ; in, and It is a learnable weight vector. Presentation layer normalization operation, This is the weight transformation matrix for the similarity space. This feature represents... It contains pathological state information driven by abnormal clinical indicators, which will serve as context input for subsequent time-series dynamic evolution.

[0056] Step S1024: Determine the temporal pattern of the disease based on the diagnosis records of two consecutive visits of the patient, and use the self-attention mechanism and cross-attention mechanism to fuse with the feature-disease similarity representation vector to generate the patient's global representation vector.

[0057] Based on adjacent medical visit statuses, disease nodes are divided into three temporal patterns: persistent, evolving, and sporadic. Attention mechanisms and matrix-gated cyclic units are used to perform time-aware dynamic diagnostic evolution modeling.

[0058] Disease development typically exhibits different dynamic patterns: some persist long-term as chronic diseases, some evolve and are diagnosed as potential risks, or some experience acute onset due to external factors. Therefore, this step overcomes this limitation by implementing differentiated feature fusion and state transition calculations for diseases with different temporal roles. Based on the aforementioned differentiated evolutionary characteristics and the acquired historical visit sequences and feature-disease similarity representations, the model utilizes self-attention and cross-attention mechanisms to extract the historical temporal context and feature-driven context of newly diagnosed diseases, respectively. Simultaneously, an optimized M-GRU is used to specifically process matrix-style medical data, enabling dynamic evolutionary modeling of persistent diseases across visits. Finally, the evolutionary states of different patterns are aggregated to generate a global representation vector for a single patient visit.

[0059] The general generative process for time-aware dynamic diagnostic evolution modeling is as follows: Step c1: Based on the states of two consecutive medical visits, divide the disease node into three non-overlapping temporal pattern subsets. Given the first... Second visit ( Diagnostic multi-thermal vector Define persistent disease indicator vector Evolutionary disease indicator vector with incidental disease indicator vector : ; ; ; in, , , These represent the element-wise logical AND, logical OR, and logical NOT operations of the multi-hot vector, respectively. This is the neighbor node indicator vector at the (t-1)th visit.

[0060] Step c2: For both evolving and sporadic diseases, a hierarchical pathological feature fusion mechanism is employed. First, a self-attention mechanism is used to learn the temporal context features based on the disease's historical embedding. and : ; ; in, Embed the neighbor nodes from the previous time step. The coding matrix for unrelated diseases; This represents the self-attention mechanism function. Subsequently, a cross-attention mechanism is used to integrate the features-disease similarity representation. Contextual information is used to capture evolutionary patterns driven by abnormal clinical indicators, resulting in feature-driven contextual information. and : ; ; in, Represents the function for the cross-attention mechanism; Finally, a gating mechanism is used to fuse the hidden states, resulting in an evolutionary representation of the emerging disease. and : ; ; in, It is the hyperbolic tangent function; and This is a learnable weight vector used to control the fusion ratio.

[0061] Step c3: For persistent diseases The cross-visit state transition is achieved through the improved gated recurrent mechanism M-GRU. Among them, the improved gated recurrent mechanism M-GRU retains the core dual-gate structure of the original GRU: (1) Reset gate: calculated by the sigmoid activation function, it controls the proportion of information retained in the previous hidden state and determines the degree of forgetting of historical information. (2) Update gate: also output by the sigmoid activation function, it controls the proportion of previous memory passed to the current time, taking into account the long-term information retention. The calculation logic of the basic gate follows the mature paradigm of the original GRU, ensuring that the number of model parameters will not increase significantly and maintaining training efficiency. A memory filter is embedded inside the unit of the original GRU, introducing the long memory modeling idea of ​​dividing the time series: the memory filter can filter and retain the long-distance dependency information of the sequence, making up for the defect that the original GRU does not have a long memory effect, and allowing the model to capture a larger temporal correlation. The new module and the original candidate hidden state calculation path are integrated, which will not introduce too many redundant parameters, and still maintains the computational efficiency advantage compared with LSTM.

[0062] M-GRU takes the persistent disease features at the current time step, the evolutionary representation of newly emerging diseases, and the hidden state at the previous time step as input, and calculates the matrix-based hidden state at the current time step. : ; in, This is an improved gated loop mechanism function; the improved gated loop mechanism preserves the hidden state of emerging situations for future integration, enabling effective dynamic modeling of matrix-form disease data.

[0063] Step c4: After obtaining the dynamic evolution representation at each time step, apply max pooling to obtain the patient-level representation. : ; in, This is the max pooling function.

[0064] Subsequently, a location-based attention mechanism was used to aggregate the representation sequences of the patient's previous visits. Generate the final patient global representation vector. It is directly used as the input tensor for the subsequent credible evidence prediction module.

[0065] Step S1025: Based on the global representation vector, obtain positive and negative evidence for various diseases and parameterize them into a Dirichlet distribution, and calculate the prediction probability and prediction uncertainty score for various diseases.

[0066] By introducing the theory of evidence deep learning, the point estimation classification output is transformed into a non-negative evidence quantity based on the Dirichlet distribution, thus accurately quantifying cognitive uncertainty.

[0067] The general process for generating reliable predictions and uncertainty quantification is as follows: Obtain the global representation vector output by the temporal dynamic evolution module. By utilizing fully connected layers and the Softplus activation function, for the first... Diseases ( Output the amount of non-negative positive evidence. Compared with negative evidence : ; ; in, This is the activation function.

[0068] For binary classification prediction scenarios involving a single disease, we employ the Beta distribution, a special case of the Dirichlet distribution with two categories, for specific modeling. This model parameterizes the aforementioned non-negative evidence quantity into the parameters of the Beta distribution. and : ; ; Define the total strength of evidence Based on this calculation, the patient is likely to develop the disease at their next visit. Predicted probability of a disease : ; The cognitive uncertainty score is accurately calculated based on the total strength of evidence. : ; When the model encounters complex and difficult cases with severely missing features, the effective evidence extracted approaches zero. The system will output a high uncertainty score and trigger a safety warning and manual intervention mechanism.

[0069] Step S1026: Calculate the prediction error using the binary evidence loss function and optimize the model parameters.

[0070] During the training of the credible diagnostic evolution and prediction model, a two-stage hybrid training strategy of cross-entropy warm-up and evidence loss fine-tuning is adopted: in the early stage of training, the standard binary cross-entropy loss function is used to guide the credible diagnostic evolution and prediction model to quickly learn the basic feature representation; after the performance of the credible diagnostic evolution and prediction model tends to stabilize, the binary evidence loss function is switched to fine-tuning training, and the backpropagation algorithm is used to update the network weight parameters until the model converges.

[0071] ; in, For the first The true label of similar diseases Let be the Digamma function. To eliminate potentially misleading evidence, KL divergence is introduced as a regularization penalty term. Construct the overall binary evidence loss function: .in, To adjust the annealing coefficient for regularization intensity, This is the classification error loss term under the binary evidence deep learning framework.

[0072] Table 1. Accuracy of Medical Diagnostic Predictions;

[0073] The method performs diagnostic prediction and evaluation on the samples to be predicted, outputs the prediction results, and compares them with the actual clinical diagnosis labels of the patients. To comprehensively verify the performance advantages of this method, the experiment selects representative methods covering multiple mainstream technical routes, including basic deep learning, temporal modeling, graph neural networks, and self-supervised representation learning, as comparison baselines. These include: Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), Graph-based Attention Model (GRAM), Collaborative Graph Learning model (CGL), Context-aware Health Event Prediction Framework via TransitionFunctions on Dynamic Disease Graphs (CHET), and Self-supervised Graph Learning Framework with Hyperbolic Embeddings for Temporal Health Event Prediction. Sherbet, a multi-perspective patient representation extractor for disease prediction (MPRE), a diagnostic prediction model based on multi-level dynamic fusion and role-aware network (DFR-Net), and a medical diagnostic prediction model based on knowledge-guided and credible evidence learning (TKDF-Net) proposed in this embodiment.

[0074] Table 1 presents a comparison of the accuracy of the method in this embodiment with other baseline models on the multi-label diagnostic prediction task, based on the test set. The evaluation parameters selected in the table are as follows: Weighted F1 score ( Precision (PPR) is a core metric that comprehensively measures the overall predictive performance of a model. It integrates prediction precision and recall, and is weighted according to the frequency of occurrence of various diseases in real-world data. This allows it to simultaneously consider the predictive effects of high-frequency common diseases and low-frequency rare diseases, objectively reflecting the model's overall generalization ability across the entire disease dataset.

[0075] forward Prediction accuracy ( The confidence level (V): refers to the proportion of the top k predicted diseases in the model's output that match the patient's actual diagnosed disease. This metric measures the accuracy of the model's predictions and corresponds to its ability to reduce misdiagnosis (false positives) in clinical settings.

[0076] The top k predicted recall rates ( ): This refers to the proportion of all diseases actually diagnosed by a patient that are successfully covered by the model's top k predictions. This metric measures the model's ability to cover real diseases, corresponding to its ability to reduce missed diagnoses (false negatives) in clinical settings.

[0077] Based on the results in Table 1, the prediction accuracy of the proposed method (TKDF-Net) in this embodiment is higher than that of other prediction methods.

[0078] In one or more embodiments, a medical diagnostic prediction system based on knowledge-guided and credible evidence learning is also provided, which can be implemented in software. The medical diagnostic prediction system based on knowledge-guided and credible evidence learning includes the following software modules: The data acquisition module is used to acquire the electronic medical record data of the patient to be tested. The diagnostic prediction module is used to make diagnostic predictions on the electronic medical record data of the patient under test using a trained reliable diagnostic evolution and prediction model. When the uncertainty score is higher than a set threshold, a high-risk blind zone warning is triggered. The training process of the reliable diagnostic evolution and prediction model is as follows: Obtain the patient's historical diagnosis sequence and continuous clinical monitoring indicator matrix from historical electronic medical record data; Disease statistical co-occurrence information is extracted from the historical diagnostic sequences to construct a statistical co-occurrence graph. Based on the diagnosed diseases during medical visits, an asymmetric gating residual mechanism is used to adaptively and weightedly fuse the pre-constructed prior knowledge graph with the statistical co-occurrence graph to generate a knowledge-data dual-driven diagnostic representation vector. Clinical monitoring indicator features are extracted from the clinical monitoring indicator matrix, their semantic alignment with the diagnostic representation vector is calculated, and an intensity-aware gating mechanism is introduced to generate a feature-disease similarity representation vector. The temporal pattern of the disease is determined based on the diagnosis records of two consecutive patient visits. The self-attention mechanism and the cross-attention mechanism are fused with the feature-disease similarity representation vector to generate the patient's global representation vector. Based on the global representation vector, positive and negative evidence for various diseases are obtained and their parameters are parameterized into a Dirichlet distribution. The prediction probability and prediction uncertainty score of various diseases are calculated. The prediction error is calculated using the binary evidence loss function, and the model parameters are then optimized.

[0079] It should be noted that each module in the medical diagnostic prediction system based on knowledge-guided and credible evidence learning in the embodiments of the present invention corresponds one-to-one with each step in the medical diagnostic prediction method based on knowledge-guided and credible evidence learning in the above embodiments, and their specific implementation processes are the same, so they will not be repeated here.

[0080] The structure of the electronic device according to embodiments of the present invention is described in detail below. The electronic device provided in embodiments of the present invention includes: at least one processor, a memory, a user interface, and at least one network interface. The various components in the knowledge-guided and trusted evidence-based medical diagnostic prediction system are coupled together through a bus system. It can be understood that the bus system is used to realize the connection and communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus. The user interface may include a display, keyboard, mouse, trackball, click wheel, buttons, a touchpad, or a touch screen, etc.

[0081] It is understood that the memory can be volatile memory or non-volatile memory, or both. The memory in this embodiment of the invention is capable of storing data to support the operation of the terminal. Examples of this data include any computer programs used to operate on the terminal, such as operating systems and applications. The operating system includes various system programs, such as the framework layer, core library layer, driver layer, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications.

[0082] In some embodiments, the medical diagnostic prediction system based on knowledge-guided and credible evidence learning provided in this invention can be implemented using a combination of hardware and software. For example, the system can be a processor in the form of a hardware decoding processor, programmed to execute the medical diagnostic prediction method based on knowledge-guided and credible evidence learning provided in this invention. For instance, the processor can employ one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0083] As an example, a processor can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where a general-purpose processor can be a microprocessor or any conventional processor, etc.

[0084] As an example of the hardware implementation of the medical diagnostic prediction system based on knowledge-guided and credible evidence learning provided in this embodiment of the invention, the device provided in this embodiment of the invention can be directly executed by a processor in the form of a hardware decoding processor. For example, it can be executed by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components to implement the medical diagnostic prediction method based on knowledge-guided and credible evidence learning provided in this embodiment of the invention.

[0085] The memory in this embodiment of the invention is used to store various types of data to support the operation of a medical diagnostic prediction system based on knowledge-guided and credible evidence learning, or to store data for execution. Figure 1The program code for the method shown. Examples of this data include: any executable instructions for operation on a medical diagnostic prediction system based on knowledge-guided and credible evidence learning, such as executable instructions that can be included in the executable instructions to implement the medical diagnostic prediction method based on knowledge-guided and credible evidence learning of the embodiments of the present invention.

[0086] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including functions for executing... Figure 1 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the apparatus of this application.

[0087] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A medical diagnostic prediction method based on knowledge-guided and credible evidence learning, characterized in that, include: Obtain the electronic medical record data of the patients to be tested; The trained reliable diagnostic evolution and prediction model is used to make diagnostic predictions on the electronic medical record data of the patient to be tested. When the uncertainty score is higher than a set threshold, a high-risk blind zone warning is triggered. The training process of the reliable diagnostic evolution and prediction model is as follows: Obtain the patient's historical diagnosis sequence and continuous clinical monitoring indicator matrix from historical electronic medical record data; Disease statistical co-occurrence information is extracted from the historical diagnostic sequences to construct a statistical co-occurrence graph. Based on the diagnosed diseases during medical visits, an asymmetric gating residual mechanism is used to adaptively and weightedly fuse the pre-constructed prior knowledge graph and the statistical co-occurrence graph to generate a knowledge-data dual-driven diagnostic representation vector. The process is as follows: Based on the prior knowledge graph and the statistical co-occurrence graph, the diagnostic features of the diagnosed diseases under knowledge-driven and statistical-driven approaches are extracted respectively, and concatenated along the feature dimension to obtain the concatenated features. Based on the concatenated features, the feature correction amount and fusion gating weight are calculated through bilinear branches respectively. After multiplying the feature correction amount and the fusion gating weight element-wise, they are fused with the diagnostic features of the diagnosed diseases under statistical-driven approaches through a residual connection method to obtain the knowledge-data dual-driven diagnostic representation vector. Clinical monitoring indicator features are extracted from the clinical monitoring indicator matrix, their semantic alignment with the diagnostic representation vector is calculated, and an intensity-aware gating mechanism is introduced to generate a feature-disease similarity representation vector. The temporal pattern of the disease is determined based on the diagnosis records of two consecutive patient visits. The self-attention mechanism and the cross-attention mechanism are fused with the feature-disease similarity representation vector to generate the patient's global representation vector. Based on the global representation vector, positive and negative evidence for various diseases are obtained and their parameters are parameterized into a Dirichlet distribution. The prediction probability and prediction uncertainty score of various diseases are calculated. The prediction error is calculated using the binary evidence loss function, and the model parameters are then optimized.

2. The medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in claim 1, characterized in that, In the process of constructing the statistical co-occurrence graph, based on the historical diagnostic sequences of all patients in the training set, the co-occurrence frequency of disease pairs is statistically analyzed, and low-frequency combinations are filtered out through preset hyperparameters to construct the statistical co-occurrence graph and its adjacency matrix.

3. The medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in claim 1, characterized in that, The process of generating feature-disease similarity representation vectors is as follows: Cosine similarity is used to calculate the semantic alignment between clinical monitoring indicator features and diagnostic representation vectors, resulting in a preliminary feature-disease similarity matrix; A deviation feature vector of the abnormality degree of clinical monitoring indicators is constructed, and then the gating weight is dynamically calculated based on the intensity-sensing gating mechanism. Based on the preliminary feature-disease similarity matrix and gating weights, the final feature-disease similarity matrix is ​​obtained; Align the final feature-disease similarity matrix with the disease embeddings normalized in the same space to obtain the feature-disease similarity representation vector.

4. The medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in claim 1, characterized in that, The process of introducing the ICD-9 medical ontology hierarchy to construct a prior knowledge graph is as follows: Extract the disease hierarchy tree structure from the ICD-9 medical ontology hierarchy, and construct a node set and edge set containing all disease codes; if two diseases have the same parent node in the ICD-9 medical ontology hierarchy, or have a direct superior-subordinate relationship, then establish a bidirectional edge connection in the graph. Based on the node set and edge set containing all disease codes, a medical ontology knowledge graph is constructed. The topology of the medical ontology knowledge graph is represented by an adjacency matrix, which serves as a prior knowledge constraint to guide the model in medical logical reasoning and provides semantic anchors for isolated long-tail rare disease nodes.

5. The medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in claim 1, characterized in that, The temporal patterns of the diseases include three types: persistent diseases, evolving diseases, and sporadic diseases.

6. The medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in claim 1, characterized in that, During the training of the credible diagnostic evolution and prediction model, a two-stage hybrid training strategy of cross-entropy warm-up and evidence loss fine-tuning is adopted: in the early stage of training, the standard binary cross-entropy loss function is used to guide the credible diagnostic evolution and prediction model to quickly learn the basic feature representation; after the performance of the credible diagnostic evolution and prediction model tends to stabilize, the binary evidence loss function is switched to fine-tuning training, and the backpropagation algorithm is used to update the network weight parameters until the model converges.

7. A medical diagnostic prediction system based on knowledge-guided and credible evidence learning, characterized in that, The medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in any one of claims 1-6 includes: The data acquisition module is used to acquire the electronic medical record data of the patient to be tested. The diagnostic prediction module is used to make diagnostic predictions on the electronic medical record data of the patient under test using a trained reliable diagnostic evolution and prediction model. When the uncertainty score is higher than a set threshold, a high-risk blind zone warning is triggered. The training process of the reliable diagnostic evolution and prediction model is as follows: Obtain the patient's historical diagnosis sequence and continuous clinical monitoring indicator matrix from historical electronic medical record data; Disease statistical co-occurrence information is extracted from the historical diagnostic sequences to construct a statistical co-occurrence graph. Based on the diagnosed diseases during medical visits, an asymmetric gating residual mechanism is used to adaptively and weightedly fuse the pre-constructed prior knowledge graph with the statistical co-occurrence graph to generate a knowledge-data dual-driven diagnostic representation vector. Clinical monitoring indicator features are extracted from the clinical monitoring indicator matrix, their semantic alignment with the diagnostic representation vector is calculated, and an intensity-aware gating mechanism is introduced to generate a feature-disease similarity representation vector. The temporal pattern of the disease is determined based on the diagnosis records of two consecutive patient visits. The self-attention mechanism and the cross-attention mechanism are fused with the feature-disease similarity representation vector to generate the patient's global representation vector. Based on the global representation vector, positive and negative evidence for various diseases are obtained and their parameters are parameterized into a Dirichlet distribution. The prediction probability and prediction uncertainty score of various diseases are calculated. The prediction error is calculated using the binary evidence loss function, and the model parameters are then optimized.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the medical diagnostic prediction method based on knowledge-guided and credible evidence learning as described in any one of claims 1-6.

Images