Intelligent interpretation management platform for medical examination reports

Abstract

The invention relates to the technical field of medical data analysis, and discloses an intelligent interpretation management platform for a medical examination report. The platform comprises a data acquisition module used for acquiring multi-source heterogeneous medical examination report data and generating a medical index data set; the feature extraction module is used for constructing a deep feature fusion network based on a multi-head attention mechanism and extracting cross-modal medical features to generate a joint feature vector; the correlation analysis module is used for carrying out dynamic anomaly scoring by adopting a space-time diagram convolutional network and constructing an anomaly propagation map; the dynamic knowledge graph construction module is used for integrating the medical ontology library and the diagnosis and treatment guide based on an incremental knowledge graph embedded learning algorithm; and the personalized interpretation generation module is used for generating a multi-level interpretation report through the conditional variation auto-encoder. In addition, the platform anonymizes patient data based on a differential privacy mechanism. The platform can efficiently process multi-source data, accurately mine index relations, provide personalized interpretation and guarantee data security.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and particularly to an intelligent interpretation and management platform for medical test reports. Background Art

[0002] In the modern medical system, medical test reports are crucial as the key basis for disease diagnosis and treatment, and the accuracy and efficiency of their interpretation are of vital importance. However, there are many drawbacks in the traditional way of interpreting medical test reports.

[0003] Medical test data presents the characteristics of multi-source heterogeneity. Different hospitals and different departments use different test equipment and systems, and the generated data formats vary widely, including structured numerical data, semi-structured text descriptions, and unstructured image data, etc. This makes it extremely difficult to integrate and analyze the data, and it is difficult for medical staff to quickly and comprehensively obtain and understand the patient's test information, which is likely to lead to information omission or misinterpretation, thus affecting the accuracy and timeliness of diagnosis.

[0004] Existing interpretation methods lack in-depth exploration of the potential pathological relationships between indicators. The abnormality of a single indicator may not be sufficient to clarify the diagnosis of a disease, and the associated changes between multiple indicators often contain important pathological information. However, most current interpretations only stay at the analysis of a single indicator, and cannot effectively reveal the internal connections between these indicators, making it difficult to assist doctors in grasping the overall trend of the patient's condition and formulating precise treatment plans.

[0005] With the continuous update of medical knowledge and the rapid development of diagnosis and treatment technologies, the items and knowledge involved in medical tests are becoming increasingly complex. Medical staff need to spend a lot of time and energy learning and mastering the latest medical knowledge to accurately interpret test reports. However, due to limited time and energy, they may not be able to keep up with the pace of knowledge update in a timely manner, resulting in the problem of knowledge lag during the interpretation process and affecting the diagnostic quality.

[0006] In the case of the increasing number of patients, the efficiency of manually interpreting test reports is low and difficult to meet the clinical needs. Especially in large hospitals, the number of test reports generated every day is huge. Manual interpretation not only consumes labor costs, but also is prone to fatigue and errors. Moreover, there are differences in the experience and professional levels of different medical staff, and it is difficult to unify the interpretation standards for test reports, further affecting the quality and consistency of medical services.

[0007] The current interpretation mode lacks personalization. The physical conditions, medical histories, genetic factors, etc. of each patient are different, and the interpretation of test results and treatment suggestions should be targeted. However, traditional methods often use unified standards for interpretation, unable to fully consider the individual differences of patients, and it is difficult to provide precise personalized medical services. In addition, in terms of medical data security and privacy protection, there are significant risks in existing management systems. The test data of patients contains a large amount of sensitive information. If it is leaked or misused, it will cause serious damage to the rights and interests of patients. However, the information systems of many hospitals have vulnerabilities in data encryption, access control, etc., and it is difficult to effectively guarantee the security of patient data. Summary of the Invention

[0008] An object of the present invention is to provide an intelligent interpretation management platform for medical test reports to solve the problems mentioned in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions: An intelligent interpretation management platform for medical test reports, the platform includes:

[0010] Data acquisition module: used to obtain multi-source heterogeneous medical test report data and generate a medical index data set;

[0011] Feature extraction module: Based on the multi-head attention mechanism, a deep feature fusion network is constructed to extract cross-modal medical features from the medical index data set and generate a joint feature vector, and the joint feature vector associates the potential pathological relationships between indicators;

[0012] Association analysis module: Adopt a spatio-temporal graph convolutional network to perform dynamic anomaly scoring on medical indicators, construct an anomaly propagation map by combining patient time-series data, and identify key anomaly nodes and associated paths;

[0013] Dynamic knowledge graph construction module: Integrate the medical ontology library and treatment guidelines based on the incremental knowledge graph embedding learning algorithm, and update the triple relationship of disease-index-treatment plan in real time;

[0014] Personalized interpretation generation module: Map the joint feature vector to the natural language space through a conditional variational autoencoder, and generate a multi-level interpretation report in combination with the patient's test report, including indicator explanation, risk assessment and clinical suggestions.

[0015] Preferably, the multi-dimensional feature extraction module includes:

[0016] Design a hierarchical feature extraction architecture for the multi-head attention mechanism, each attention head focuses on a specific medical indicator category, and introduce self-supervised contrast learning to optimize the feature discriminability;

[0017] Generate the joint feature vector by constraining the semantic consistency between the text description and the numerical indicators through a cross-modal alignment loss function.

[0018] Preferably, the association analysis module includes:

[0019] Use a temporal convolutional network to extract the dynamic change patterns of single indicators, and construct a graph attention network to model the association weights between indicators;

[0020] Detect local abnormal events based on the dynamic threshold sliding window algorithm, and mine the abnormal propagation paths across time slices through the random walk algorithm.

[0021] Preferably, the dynamic knowledge graph construction module includes:

[0022] Adopt the incremental TransR algorithm for low-dimensional embedding of newly added medical entities, and optimize the graph relationship reasoning through the negative sampling strategy;

[0023] Design an entity linking mechanism based on reinforcement learning to automatically correct ambiguous terms and alias mappings to the standard medical ontology.

[0024] Preferably, the personalized interpretation generation module includes:

[0025] Construct a hierarchical decoder architecture, where the first layer generates the index interpretation text, the second layer fuses the risk assessment matrix to output the warning level, and the third layer combines knowledge graph reasoning to generate diagnosis and treatment suggestions;

[0026] Constrain the clinical compliance of the generated text through the adversarial training strategy, and introduce an interpretability masking mechanism to highlight the key decision-making basis.

[0027] Preferably, the self-supervised contrastive learning includes:

[0028] Construct positive sample pairs as cross-modal data augmentation versions of the same patient, and negative sample pairs as similar index combinations of different patients;

[0029] Maximize the embedding similarity of positive sample pairs through the InfoNCE loss function.

[0030] Preferably, the dynamic threshold sliding window algorithm includes:

[0031] Calculate the local mean and variance of the index sequence based on the exponential weighted moving average algorithm, and dynamically adjust the abnormal determination threshold;

[0032] Introduce the isolation forest algorithm to identify high-dimensional outliers within the window, and model the joint probability distribution of abnormal events through the Bayesian network.

[0033] Preferably, the incremental TransR algorithm includes:

[0034] Project the newly added entities into the relation-specific space and update the embedding parameters through the adaptive momentum optimizer;

[0035] Design a relation-aware adversarial training framework to generate adversarial samples to enhance the robustness and generalization ability of the knowledge graph.

[0036] Preferably, the adversarial training strategy includes:

[0037] Construct a discriminator network to distinguish the generated text from the real clinical guideline content, and improve the anti-interference ability of the generator through the gradient reversal layer;

[0038] Adopt the curriculum learning strategy to optimize the generation complexity in stages. In the initial stage, focus on grammar correctness, and in the later stage, strengthen the medical logical coherence.

[0039] Preferably, the platform further includes: anonymize the patient data based on the differential privacy mechanism.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] In terms of data processing, through the data acquisition module, multi-source heterogeneous medical test report data can be obtained and a medical index dataset can be generated, effectively solving the problem of traditional data integration. This enables the originally scattered and diverse-formatted data to be uniformly collected and organized, providing a complete and standardized data basis for subsequent analysis, avoiding information errors caused by data missing or format incompatibility, and greatly improving the usability of the data.

[0042] The feature extraction module constructs a deep feature fusion network based on the multi-head attention mechanism, extracts cross-modal medical features from the medical index dataset, and generates a joint feature vector. This vector associates the potential pathological relationships between the indicators, breaking the limitation of traditional single-index analysis. For example, when analyzing the test data of diabetic patients, not only can the blood glucose index be concerned, but also the potential connections between blood glucose and indicators such as insulin and glycated hemoglobin can be discovered through the joint feature vector, assisting doctors to more comprehensively and deeply understand the disease development mechanism, and thus make more accurate diagnoses.

[0043] The association analysis module uses the spatio-temporal graph convolutional network to perform dynamic anomaly scoring on medical indicators, constructs an anomaly propagation graph by combining the patient's time-series data, and identifies key anomaly nodes and associated paths. This helps doctors timely detect the abnormal changes in the patient's condition and track the propagation trajectory of the anomalies. Taking cardiovascular diseases as an example, through the dynamic monitoring and analysis of multiple indicators such as blood pressure, blood lipids, and electrocardiogram, subtle abnormal signals can be captured in the early stage of the disease, predicting the development trend of the disease, providing a strong basis for early intervention and treatment, and improving the cure rate of the disease and the survival rate of patients.

[0044] The dynamic knowledge graph construction module integrates the medical ontology library and clinical practice guidelines based on the incremental knowledge graph embedding learning algorithm, and updates the triple relationship of disease-index-treatment plan in real time. This enables the platform to promptly reflect the latest research findings and clinical experiences in the medical field, providing the most cutting-edge diagnosis and treatment knowledge support for doctors. Regardless of how rapidly medical knowledge is updated, the platform can automatically adjust and improve the knowledge graph, ensuring that the information obtained by doctors when interpreting test reports is always accurate and up-to-date, and avoiding misdiagnosis and mistreatment caused by knowledge lag.

[0045] The personalized interpretation generation module maps the joint feature vector to the natural language space through a conditional variational autoencoder, and generates a multi-level interpretation report in combination with the patient's test report, including index explanation, risk assessment, and clinical suggestions. This personalized interpretation method fully considers the individual differences of patients, providing customized medical services for patients. For example, for patients with the same disease but different ages, physical conditions, and medical histories, the platform can generate targeted risk assessments and treatment suggestions based on their respective test data, improving the accuracy and effectiveness of treatment, while enhancing patients' understanding of their own conditions and their enthusiasm for cooperating with treatment.

[0046] In terms of data security, the patient data is anonymized based on the differential privacy mechanism, effectively protecting the privacy of patients. Even if there are security vulnerabilities during data storage, transmission, or sharing, it is difficult for attackers to obtain patients' personal sensitive information from the anonymized data, providing a reliable guarantee for patients' data security and enhancing patients' trust in the medical system.

[0047] In addition, the intelligent design of the entire platform greatly improves the interpretation efficiency of medical test reports. Compared with the traditional manual interpretation method, it can process a large amount of test data in a short time, reducing the workload of medical staff and improving the efficiency and quality of medical services. At the same time, since the platform uses a unified algorithm and standard for interpretation, it avoids the problem of inconsistent interpretations caused by individual differences among medical staff, ensuring the standardization and consistency of medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the working principle diagram of the intelligent interpretation management platform for the medical test report described in the present invention;

[0049] Figure 2 It is the working flow chart of the multi-dimensional feature extraction module;

[0050] Figure 3 It is the working principle diagram of self-supervised contrast learning;

[0051] Figure 4 It is the working principle diagram of patient data anonymization processing. DETAILED DESCRIPTION OF THE INVENTION

[0052] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0053] Please refer to Figures 1-4 , the present invention provides an intelligent interpretation management platform for medical test reports, and its overall implementation solution is as follows:

[0054] Data acquisition module: This module is responsible for obtaining multi-source heterogeneous medical test report data. These data have a wide range of sources and may include data generated by test equipment in different departments of the hospital, different test systems, etc. The data formats are also diverse, such as structured numerical data, semi-structured text descriptions, and unstructured image data, etc. Through specialized data acquisition interfaces and adaptation programs, these data are collected and integrated, and finally a medical index data set is generated to provide basic data support for subsequent analysis and processing.

[0055] Feature extraction module: Build a deep feature fusion network based on the multi-head attention mechanism. This network designs a hierarchical feature extraction architecture, where each attention head focuses on a specific medical index category. For example, some attention heads focus on blood test indicators, and some focus on biochemical test indicators, etc. At the same time, self-supervised contrast learning is introduced to optimize the feature discriminability, making the extracted features more distinguishable. By using a cross-modal alignment loss function to constrain the semantic consistency between the text description and the numerical indicators, cross-modal medical features are extracted from the medical index data set, and a joint feature vector is generated, which associates the potential pathological relationships between the indicators.

[0056] Association analysis module: Use a spatio-temporal graph convolutional network to perform dynamic anomaly scoring on medical indicators. Specifically, a temporal convolutional network is used to extract the dynamic change patterns of single indicators, and a graph attention network is constructed to model the association weights between indicators. Based on the dynamic threshold sliding window algorithm, local abnormal events are detected, and the abnormal propagation paths across time slices are mined through the random walk algorithm. Combine the patient's temporal data to construct an abnormal propagation map, thereby identifying key abnormal nodes and associated paths to help doctors better understand the development and association of the patient's condition.

[0057] Dynamic Knowledge Graph Construction Module: Integrate the medical ontology library and clinical practice guidelines based on the incremental knowledge graph embedding learning algorithm. Use the incremental TransR algorithm to perform low-dimensional embedding on newly added medical entities, and optimize the graph relationship reasoning through the negative sampling strategy. Design an entity linking mechanism based on reinforcement learning to automatically correct ambiguous terms and alias mappings to the standard medical ontology, and update the triple relationship of disease-index-treatment plan in real time, providing rich knowledge support for interpretation and diagnosis.

[0058] Personalized Interpretation Generation Module: Map the joint feature vector to the natural language space through a conditional variational autoencoder, and generate a multi-level interpretation report in combination with the patient's test report. Construct a hierarchical decoder architecture. The first layer generates the index explanation text, the second layer fuses the risk assessment matrix to output the warning level, and the third layer generates the diagnosis and treatment suggestions in combination with the knowledge graph reasoning. Constrain the clinical compliance of the generated text through the adversarial training strategy, and introduce an interpretability masking mechanism to highlight the key decision-making basis, making the interpretation report more practical and interpretable.

[0059] The following further illustrates the implementation of the present invention in conjunction with Embodiments 1 to 5.

[0060] Embodiment 1:

[0061] This embodiment elaborates in detail the hierarchical feature extraction architecture of the multi-head attention mechanism in the feature extraction module and the specific implementation method of self-supervised contrast learning. Specifically, it includes:

[0062] Construct a hierarchical feature extraction architecture of the multi-head attention mechanism. The medical index dataset is initially divided according to different categories. For example, according to the type of test items, it is divided into blood routine indexes, urine routine indexes, liver function indexes, kidney function indexes, etc. For each category, set an attention head. Taking the blood routine indexes as an example, this attention head focuses on the feature extraction of related indexes such as red blood cell count, white blood cell count, and platelet count.

[0063] In each attention head, the input index data is processed through a series of neural network layers. For numerical blood routine indexes, first perform standardization processing to make them have a unified scale. Then, through a multi-layer perceptron (MLP) for feature mapping, map the original index data to a high-dimensional feature space. For text descriptions related to blood routine, such as "the red blood cell morphology is normal", etc., use the word embedding method in natural language processing technology to convert the text into a vector representation, and then extract the text features through a convolutional neural network (CNN).

[0064] Introduce self-supervised contrastive learning to optimize feature discriminability. Construct positive sample pairs as cross-modal data augmentation versions of the same patient. Taking the blood routine data of a patient as an example, perform data augmentation operations such as adding random noise to its numerical indicators, and at the same time perform operations such as synonym replacement on its text description to obtain the augmented numerical data and text data, which constitute positive sample pairs. Negative sample pairs are similar indicator combinations of different patients. For example, select samples with similar red blood cell counts from the blood routine data of different patients, and combine their numerical data and text data into negative sample pairs.

[0065] Maximize the embedding similarity of positive sample pairs through the InfoNCE loss function. The formula of the InfoNCE loss function is:

[0066]

[0067] where N is the number of samples, M is the number of negative samples, z i is the feature vector of the i-th sample, is the positive sample feature vector of the i-th sample, is the j-th negative sample feature vector of the i-th sample, and sim(a, b) represents the similarity between vector a and vector b, usually calculated using cosine similarity. In actual calculation, by minimizing this loss function, the distance between cross-modal data of the same patient in the feature space is closer, and the distance between similar indicator data of different patients is farther, thereby optimizing the discriminability of features.

[0068] Constrain the semantic consistency between text descriptions and numerical indicators through a cross-modal alignment loss function. This loss function calculates the difference between the text feature vector and the numerical feature vector, and backpropagates to adjust the network parameters to make the two more consistent semantically. For example, when the numerical indicator shows that the red blood cell count is low, the relevant content about red blood cells in the text description should also reflect similar abnormal information, and in this way, a more accurate joint feature vector is generated.

[0069] Example 2:

[0070] This example details the specific implementation process of using a temporal convolutional network to extract the dynamic change patterns of single indicators, constructing a graph attention network to model the association weights between indicators, and a dynamic threshold sliding window algorithm in the association analysis module. Specifically include:

[0071] Use a temporal convolutional network to extract the dynamic change patterns of single indicators. Taking the blood glucose indicator as an example, collect the blood glucose measurement values of patients over a period of time to form time series data. Input this data into the temporal convolutional network. The convolutional layer in the network slides in the time dimension through convolutional kernels of different sizes to capture the change characteristics of blood glucose values at different time scales. For example, smaller convolutional kernels can capture the short-term fluctuations in blood glucose, while larger convolutional kernels can obtain the trend changes in blood glucose over a longer period. Through the combination of multiple convolutional layers and pooling layers, a feature vector that can represent the dynamic change pattern of blood glucose is extracted.

[0072] Construct a graph attention network to model the correlation weights between indicators. Consider medical indicators as nodes in a graph and the correlation relationships between indicators as edges. For each node (indicator), calculate the correlation weights with other nodes through the attention mechanism based on its feature vector and the feature vectors of neighboring nodes. For example, there is a close correlation between the blood glucose indicator and the insulin indicator. The graph attention network will calculate a relatively high correlation weight between the two based on their respective features and changes in the time series. In the specific calculation process, transform and calculate the node features through a specific neural network layer to obtain attention scores, and then normalize them through the softmax function to obtain the final correlation weights.

[0073] Detect local abnormal events based on the dynamic threshold sliding window algorithm. First, calculate the local mean and variance of the indicator sequence based on the exponentially weighted moving average algorithm, and dynamically adjust the abnormal determination threshold. The formula for the exponentially weighted moving average algorithm is:

[0074] EMA t = αX t +(1 - α)EMA t-1

[0075] where EMA t is the exponentially weighted moving average value at the t-th moment, X t is the indicator value at the t-th moment, α is the weight coefficient, and its value range is between 0 and 1. It determines the relative importance of the current value and the historical value in the calculation. By continuously updating EMA t , the local mean of the indicator sequence can be obtained. The calculation of variance is similar and is dynamically updated according to the idea of the exponentially weighted moving average. According to the calculated mean and variance, dynamically adjust the abnormal determination threshold. When the indicator value exceeds the threshold range, it is considered that there may be local abnormal events.

[0076] The isolation forest algorithm is introduced to identify high-dimensional outliers within the window. The isolation forest algorithm is based on the principle that in the data space, outliers are those data points in low-density regions, and they are more "isolated". For the index data within the sliding window, it is regarded as high-dimensional data points, and multiple isolation trees are constructed to calculate the outlier score for each data point. The higher the outlier score, the more likely the data point is an outlier.

[0077] The joint probability distribution of abnormal events is modeled through a Bayesian network. A Bayesian network is a directed acyclic graph, where nodes represent random variables and edges represent the dependence relationships between variables. The detected abnormal events are used as nodes, and a Bayesian network is constructed based on their causal relationships and statistical correlations. For example, an abnormal increase in blood sugar may be related to factors such as insufficient insulin secretion and excessive dietary intake. These relationships can be represented in the Bayesian network, and the joint probability distribution of abnormal events can be calculated through known probability information, so as to more comprehensively understand the possibility and mutual influence of abnormal events occurring.

[0078] Based on the results obtained from the above steps, the random walk algorithm is used to mine the abnormal propagation paths across time slices. The random walk algorithm starts from an initial abnormal node and randomly selects the next node to move according to the association weights between nodes, and records the path during the movement. After multiple random walks, the obtained paths are analyzed to find the frequently occurring paths, which are the possible abnormal propagation paths and help to discover the connections between different index abnormalities in the disease development process.

[0079] Example 3:

[0080] In this example, the specific implementation of the incremental TransR algorithm in the dynamic knowledge graph construction module for low-dimensional embedding of newly added medical entities and the design of an entity linking mechanism based on reinforcement learning are introduced in detail. Specifically, it includes:

[0081] The incremental TransR algorithm is used to perform low-dimensional embedding on newly added medical entities. Suppose a new disease entity, "rare genetic liver disease", appears in the medical knowledge graph. First, the newly added entity is projected into the relation-specific space. In the medical knowledge graph, there are various relationships, such as the "disease - symptom" relationship, the "disease - treatment plan" relationship, etc. For the newly added entity "rare genetic liver disease", according to its different relationships with other entities, it is projected into the corresponding relation-specific space respectively. For example, in the "disease - symptom" relation space, based on the known symptom information of this disease, it is projected into this space through a specific projection matrix to obtain the low-dimensional embedding representation under this relation.

[0082] Update the embedding parameters through the Adaptive Moment Estimation (Adam) optimizer. The Adam optimizer adjusts the learning rate based on the gradient history of the parameters, making the parameter update more efficient and stable. During the update process, according to the relationship information in the training data, calculate the gradient with respect to the embedding parameters, and then use the Adam optimizer to adjust the embedding parameters according to its update rule. The formula is:

[0083] m t = β1 * m t-1 + (1 - β1) * g t

[0084]

[0085] where m t and v t are the first and second moment estimates of the gradient respectively, g t is the gradient at the current moment, β1 and β2 are decay coefficients, usually set to 0.9 and 0.999 respectively, η is the learning rate, is a small constant used to prevent the denominator from being zero, and θ t is the parameter value at the current moment. Through continuous iterative updates, the embedding parameters of the newly added entities can more accurately reflect their relationships in the knowledge graph.

[0086] Design a relationship-aware adversarial training framework to generate adversarial samples to enhance the robustness and generalization ability of the knowledge graph. Construct a Generative Adversarial Network (GAN), where the generator is responsible for generating adversarial samples, and the discriminator is used to distinguish real samples from adversarial samples. In the context of a medical knowledge graph, the generator generates entity-relationship combinations that seem reasonable but are actually adversarial samples based on the existing entity and relationship information. For example, generate a false "rare genetic liver disease - wrong symptom" relationship pair as an adversarial sample. The discriminator learns from real disease-symptom relationship pairs to determine whether the input sample is real or an adversarial sample. During the training process, the generator and the discriminator compete with each other. The generator continuously optimizes to generate more difficult-to-recognize adversarial samples for the discriminator, while the discriminator continuously improves its recognition ability. Through this adversarial training, the knowledge graph has stronger robustness and generalization ability in the face of various interferences and unknown situations.

[0087] Design an entity linking mechanism based on reinforcement learning to automatically correct the mapping of ambiguous terms and aliases to the standard medical ontology. Take the common alias "hepatitis B" as an example. In medical texts, different expressions such as "hepatitis B" and "hepatitis B virus infection" may appear. Consider the entity linking problem as a reinforcement learning task. The state of the agent is the current text segment being processed and the existing knowledge graph information, and the action is to select a possible standard medical ontology concept for mapping. For example, when encountering "hepatitis B", the action can be to select the standard concept "hepatitis B". The reward function is designed based on the correctness of the mapping result and its contribution to the integrity of the knowledge graph. If "hepatitis B" is successfully mapped to "hepatitis B" and the relevant relationships in the knowledge graph become more complete and accurate, a positive reward is given; otherwise, a negative reward is given. By continuously interacting with the environment, the agent learns the optimal mapping strategy according to the reward feedback, thereby automatically correcting the mapping of ambiguous terms and aliases to the standard medical ontology and ensuring the consistency and accuracy of the terms in the knowledge graph.

[0088] Example 4:

[0089] This example elaborates in detail the specific implementation methods of constructing a hierarchical decoder architecture to generate a multi-level interpretation report and the adversarial training strategy in the personalized interpretation generation module. Specifically, it includes:

[0090] Construct a hierarchical decoder architecture to generate a multi-level interpretation report. First, the first layer generates an index explanation text. Taking the blood glucose index as an example, when the blood glucose value is higher than the normal range, the first layer of the decoder generates an explanation text about the elevated blood glucose based on the pre-trained language model and medical knowledge, such as "Your blood glucose value is higher than the normal range, which may mean that there is a problem with your body's regulation of blood glucose. Common causes include insufficient insulin secretion and excessive sugar intake in recent diet." During the generation process, the model combines the specific value of the blood glucose index and the degree of deviation from the normal range to generate a more targeted explanation.

[0091] The second layer fuses the risk assessment matrix to output the warning level. The risk assessment matrix is established based on a large amount of clinical data and medical research. It considers multiple factors, such as the degree of abnormality of the index, the duration of the abnormality, and the patient's underlying diseases. For the case of elevated blood glucose, if the blood glucose value is only slightly higher than the normal range and the patient has no other underlying diseases, the risk level given by the risk assessment matrix may be "low risk"; if the blood glucose value is significantly higher than the normal range and the patient already has underlying diseases such as diabetes, the risk level may be "high risk". The second layer of the decoder outputs the corresponding warning level based on this information, such as "According to your blood glucose index and related conditions, the current risk level is high risk. Please pay close attention."

[0092] The third layer combines a knowledge graph for reasoning to generate diagnosis and treatment suggestions. The knowledge graph contains rich triple relationships of disease-index-treatment plan. When it is determined that blood sugar is elevated and the risk level is high, through knowledge graph reasoning, treatment plans related to high blood sugar are found. For example, it may include dietary adjustment suggestions such as "reduce the intake of high-sugar and high-fat foods and increase the proportion of vegetables and coarse grains"; exercise suggestions such as "perform at least 150 minutes of moderate-intensity aerobic exercise per week, such as brisk walking, jogging, etc."; drug treatment suggestions. If the patient's blood sugar is too high, it may be recommended to use insulin or oral hypoglycemic drugs, etc., and specific dosage and frequency of medication are given.

[0093] The clinical compliance of the generated text is constrained through an adversarial training strategy, and an interpretability masking mechanism is introduced to highlight the key decision-making basis. A discriminator network is constructed to distinguish the generated text from the content of real clinical guidelines. The discriminator network is a neural network-based classifier that learns the text features and patterns in real clinical guidelines. The text such as the diagnosis and treatment suggestions generated by the generator is used as the input of the discriminator, and the discriminator judges whether these texts conform to the norms of real clinical guidelines. The anti-interference ability of the generator is improved through a gradient reversal layer. The gradient reversal layer changes the direction of the gradient during training, enabling the generator to continuously optimize and generate text that is more compliant with clinical requirements when facing the judgment of the discriminator.

[0094] A curriculum learning strategy is adopted to optimize the generation complexity in stages. The initial stage focuses on grammar correctness. At this stage, when training the generator, more attention is paid to whether the grammar of the generated text is correct and whether the sentences are smooth. For example, check whether there are problems such as incomplete subject-predicate-object structures and incorrect collocations of parts of speech. As the training progresses, the medical logical coherence is strengthened in the later stage. At this time, it is concerned whether the generated text is reasonable in medical logic, such as whether the treatment suggestions match the disease and symptoms, and whether there are conflicts between different suggestions. Through this staged optimization, the quality of the generated text is improved.

[0095] An interpretability masking mechanism is introduced to highlight the key decision-making basis. In the generated interpretation report, for some key decision-making bases, such as blood sugar values, risk levels, etc., they are highlighted through the interpretability masking mechanism. For example, these key information are highlighted in the text by bolding, changing colors, etc., enabling doctors and patients to more intuitively understand the basis and key points of the interpretation report and enhancing the interpretability of the interpretation report.

[0096] Example 5:

[0097] In the intelligent interpretation management platform of medical test reports, the privacy protection of patient data is of great importance. This example details the specific implementation process of anonymizing patient data based on the differential privacy mechanism.

[0098] The basic principle of differential privacy is to add carefully designed random noise to the original data during data release or use, making it difficult for attackers to infer sensitive information of specific individuals from the processed data. The core lies in controlling the distribution and intensity of the noise to maximize data privacy while ensuring data availability. The privacy budget (represented by the parameter is a key indicator to measure the strength of differential privacy protection. The smaller the value, the higher the degree of privacy protection, but the availability of the data may be correspondingly reduced; conversely, the larger the value, the higher the data availability, but the privacy protection will be weakened.

[0099] First, determine the data scope to be protected, covering the patient's basic information (such as information that can directly identify personal identity, like name, ID number, contact information, home address, etc.), various indicator data in the test report (such as blood routine, biochemical indicators, imaging examination results, etc.), and information such as diagnosis records and treatment history related to the patient. For different types of data, different differential privacy implementation methods are adopted.

[0100] For numerical data, taking the numerical value of the test indicator as an example, determine the amount of noise to be added according to the sensitivity of the data and the privacy budget. For example, the blood glucose value is relatively sensitive data. If the privacy budget is adopted, the Laplace mechanism is used to add noise. The Laplace mechanism formula is

[0101]

[0102] where M(x) is the output result after adding noise, that is, the blood glucose value after privacy protection processing; f(x) is the function value of the original data, that is, the actually measured blood glucose value; Δf is the sensitivity of the function f. For numerical data, the sensitivity usually takes the maximum change amount of the data. For example, in a specific test scenario, the maximum change amount of the blood glucose value is 10 mmol / L; Lap(·) represents the Laplace distribution, and its probability density function is Here Suppose the original blood glucose value is 8 mmol / L. According to the above settings, randomly draw a noise value (suppose it is 3) from the Laplace distribution Lap(20). Then the blood glucose value M(x) after adding noise = 8 + 3 = 11 mmol / L. In this way, the blood glucose value is perturbed to a certain extent, reducing the risk for attackers to accurately identify the patient through the blood glucose value, while retaining the availability of the data in statistical analysis and other aspects.

[0103] For text data, such as the diagnosis description of a patient, a method combining semantic replacement and generalization with the differential privacy mechanism is adopted. For example, for a diagnosis description like "suffering from type 2 diabetes", first perform semantic analysis on the text to determine its key semantic information. Then, according to the pre-set privacy policy and noise addition rules, replace it with a broader and more general expression. At the same time, to meet the requirements of differential privacy, a noise mechanism based on text similarity is introduced. Suppose there is a text dataset containing various diagnosis descriptions, calculate the similarity matrix between each diagnosis description and other descriptions. According to the privacy budget ∈, add noise to the similarity matrix (such as using the exponential mechanism), so that the original diagnosis description has a certain degree of randomness when selecting a replacement expression, but can still ensure the general relevance of semantics. For example, "suffering from type 2 diabetes" may be replaced with "suffering from endocrine system diseases". This replacement not only obscures the specific disease information but also retains the semantic direction of the text to a certain extent, making it difficult for attackers to accurately obtain the patient's true condition from the replaced text.

[0104] For complex datasets containing multiple data types, such as the entire medical test report of a patient, comprehensively apply the above differential privacy techniques for different data types. When processing data, first add noise to numerical data according to the corresponding numerical processing methods, and then perform semantic replacement and generalization processing on text data. At the same time, consider the correlation between different data types to ensure that the overall logical consistency of the data is not damaged during the process of protecting privacy. For example, when processing a medical test report with both blood glucose values and diabetes diagnosis descriptions, ensure that the perturbation of blood glucose values and the replacement of diagnosis descriptions are coordinated with each other, and there will be no contradiction such as normal blood glucose values but a diabetes diagnosis description.

[0105] In addition, the differential privacy mechanism is also applied for protection during data storage and transmission. For patient data stored in the database, add noise for anonymization during writing; during data transmission, perform real-time privacy protection processing on the transmitted data to ensure that data is effectively protected in all aspects. Through the above comprehensive and systematic anonymization processing of patient data based on the differential privacy mechanism, the privacy security of patient data is maximally guaranteed on the premise of not affecting the normal functions of the intelligent interpretation management platform for medical test reports.

[0106] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0107] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent interpretation and management platform for medical test reports, characterized in that, Including: Data acquisition module: used to obtain multi-source heterogeneous medical test report data and generate a medical index dataset; Feature extraction module: constructs a deep feature fusion network based on the multi-head attention mechanism, extracts cross-modal medical features from the medical index dataset and generates a joint feature vector, and the joint feature vector associates the potential pathological relationships between the indicators; Association analysis module: uses a spatio-temporal graph convolutional network to perform dynamic anomaly scoring on medical indicators, constructs an anomaly propagation map by combining patient time-series data, and identifies key anomaly nodes and associated paths; Dynamic knowledge graph construction module: integrates the medical ontology library and the diagnosis and treatment guidelines based on the incremental knowledge graph embedding learning algorithm, and updates the triple relationship of disease-index-treatment plan in real time; Personalized interpretation generation module: maps the joint feature vector to the natural language space through a conditional variational autoencoder, and generates a multi-level interpretation report in combination with the patient's test report, including index interpretation, risk assessment, and clinical suggestions.

2. The intelligent interpretation management platform for medical test report forms according to claim 1, characterized in that, The multi-dimensional feature extraction module includes: Design a hierarchical feature extraction architecture for the multi-head attention mechanism, each attention head focuses on a specific medical index category, and introduce self-supervised contrast learning to optimize the feature discriminability; Generate the joint feature vector by constraining the semantic consistency of the text description and the numerical index through a cross-modal alignment loss function.

3. The intelligent interpretation management platform for medical test report sheets according to claim 1, characterized in that, The association analysis module includes: Use a temporal convolutional network to extract the dynamic change pattern of a single index, and construct a graph attention network to model the association weights between the indexes; Detect local abnormal events based on the dynamic threshold sliding window algorithm, and mine the abnormal propagation path across time slices through the random walk algorithm.

4. The intelligent interpretation management platform for medical test reports according to claim 1, wherein The dynamic knowledge graph construction module includes: Adopt the incremental TransR algorithm to perform low-dimensional embedding on the newly added medical entities, and optimize the graph relationship reasoning through the negative sampling strategy; Design an entity linking mechanism based on reinforcement learning to automatically correct the mapping of ambiguous terms and aliases to the standard medical ontology.

5. The intelligent interpretation management platform for medical test report sheets according to claim 1, characterized in that, The personalized interpretation generation module includes: Construct a hierarchical decoder architecture, the first layer generates index interpretation text, the second layer fuses the risk assessment matrix to output the warning level, and the third layer combines knowledge graph reasoning to generate diagnosis and treatment suggestions; Constrain the clinical compliance of the generated text through an adversarial training strategy, and introduce an interpretability mask mechanism to highlight the key decision-making basis.

6. The intelligent interpretation management platform for medical test report sheets according to claim 2, wherein The self-supervised contrast learning includes: Construct positive sample pairs as cross-modal data augmentation versions of the same patient, and negative sample pairs as similar index combinations of different patients; Maximize the embedding similarity of positive sample pairs through the InfoNCE loss function.

7. The intelligent interpretation management platform for medical test report sheets according to claim 3, characterized in that, The dynamic threshold sliding window algorithm includes: Calculate the local mean and variance of the index sequence based on the exponentially weighted moving average algorithm, and dynamically adjust the anomaly determination threshold; Introduce the isolation forest algorithm to identify high-dimensional outliers within the window, and model the joint probability distribution of abnormal events through a Bayesian network.

8. The intelligent interpretation management platform for medical test reports according to claim 4, characterized in that The incremental TransR algorithm includes: Project the newly added entity into the relationship-specific space, and update the embedding parameters through an adaptive momentum optimizer; Design a relationship-aware adversarial training framework to generate adversarial samples to enhance the robustness and generalization ability of the graph.

9. The intelligent interpretation management platform for medical test reports according to claim 5, characterized in that, The adversarial training strategy includes: Build a discriminator network to distinguish generated text from real clinical guideline content, and improve the anti-interference ability of the generator through a gradient reversal layer; Adopt a curriculum learning strategy to optimize the generation complexity in stages. In the initial stage, focus on grammatical correctness, and in the later stage, strengthen medical logical coherence.

10. The intelligent interpretation management platform for medical test report forms according to claim 1, characterized in that, It also includes: Anonymize patient data based on the differential privacy mechanism.

Images