Medical examination report interpretation and risk assessment method

Through natural language processing and machine learning algorithms, the multi-source medical test reports are standardized and multi-modal data fusion is solved, and the problem of inconsistent interpretation results with treatment suggestions and insufficient multi-modal data processing capabilities in the existing methods is achieved, and individualization is achieved The interpretation of medical reports and risk assessment have improved the accuracy of diagnosis and targeted treatment recommendations.

CN118942616AInactive Publication Date: 2024-11-12JINHUA SECOND HOSPITAL

Patent Information

Application Number
CN202411164646.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The interpretation of existing medical test reports and risk assessment methods rely on the comparison of multiple test reports, and cannot fully capture the changes in patients' health, ignore the inherent connection between interpretation results and treatment suggestions, and have limited ability to process multimodal data, resulting in inconsistent interpretation results and treatment. The suggestion is not targeted.

Method used

The report format analysis system using natural language processing and machine learning algorithms is used to standardize and hierarchically analyze multi-source medical test reports through adaptive semantic analytical models and multi-dimensional deep learning models, combining time series analysis and contextual context. Fusion of multimodal data, generate personalized interpretation of medical reports, and extract key risk points through intelligent summary generators.

Benefits of technology

It realizes a comprehensive and accurate analysis of the patient's health status, improves the diagnostic accuracy and targeted treatment suggestions, ensures that the interpretation results match the patient's individualized situation, and improves the treatment effect and compliance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942616A_ABST
    Figure CN118942616A_ABST
Patent Text Reader

Abstract

The invention relates to a medical examination report interpretation and risk assessment method. The method comprises the following steps: analyzing a multi-source medical examination report through a report format analysis system integrating natural language processing (NLP) and a machine learning algorithm, and converting original examination data into a standardized structured data set; secondly, based on a hierarchical semantic analysis mechanism, performing hierarchical analysis on the standardized medical examination report information, and dividing data into key information, auxiliary information and background information; the method comprises the following steps: integrating and fusing multi-source data including a blood test result, iconography data and gene sequencing data in a test report by utilizing historical health data, medical record records and a multi-modal data fusion technology of a patient, and generating personalized medical report interpretation through an adaptive enhancement algorithm; based on a context-sensitive intelligent interpretation model, clinical history, current symptoms and pathological features of a patient are comprehensively considered, and an interpretation result with clinical correlation is dynamically generated through an integrated semantic network and a deep learning technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a report interpretation and risk assessment method, in particular to a medical test report interpretation and risk assessment method. Background Art

[0002] Although the current methods for interpreting and risk assessment of medical test reports have made some progress, there are still some shortcomings and drawbacks. By comparing the content of Chinese invention patent 202410406851.X, these shortcomings can be seen more clearly. Patent 202410406851.X proposes an intelligent interpretation and treatment recommendation method based on test reports, which relies on multiple test reports and generates interpretation results and treatment recommendations through a large language model. This method makes up for the shortcomings of traditional medical report interpretation to a certain extent and provides a more intelligent and customized interpretation method. However, in practical applications, existing methods still face many challenges. First, the existing methods for interpreting and risk assessment of medical test reports rely too much on the comparison of multiple test reports. Although this method can identify abnormal items by comparing test results at different time points, in some cases, this reliance leads to deviations in the interpretation results. Especially when the patient's health status changes rapidly or certain diseases develop rapidly, relying only on recent test reports for comparison cannot fully capture the patient's health changes. Secondly, when interpreting abnormal items and generating treatment recommendations, existing methods regard these two tasks as independent processes, which to a certain extent ignores the inherent connection between them. In actual clinical applications, the interpretation of test reports and treatment recommendations are usually highly correlated. Abnormal test results often directly affect the formulation of treatment plans, and the effectiveness of treatment plans in turn has an important impact on future test results. Therefore, existing methods lack modeling of the interactive relationship between these two tasks, resulting in inconsistencies in interpretation results and treatment recommendations, which in turn affects the accuracy and effectiveness of clinical decision-making.

[0003] In addition, existing methods have limited capabilities in processing complex multimodal data. Modern medical reports often include data in multiple forms, including imaging data, gene sequencing data, and laboratory test results. Data of different modalities contain complementary key information, while existing methods mainly focus on the processing of text data, and lack the ability to integrate and analyze multimodal data, resulting in an incomplete understanding of the patient's overall health status. The lack of integrated analysis of multimodal data not only limits the diagnostic accuracy of the system, but also misses some implicit clinical associations, which is particularly unfavorable for the interpretation of complex cases. In addition, existing methods still have deficiencies in generating personalized medical advice. Although patent 202410406851.X mentions that interpretations and suggestions can be generated according to customized requirements under different needs, in actual operation, due to the complexity of individual differences, existing models often fail to fully capture the patient's specific health background and needs when processing individualized information. This deficiency is particularly evident in some special cases or chronic disease management, resulting in the lack of targetedness of the generated advice and the inability to fully consider the individual situation of the patient, which in turn affects the treatment effect and patient compliance. Summary of the invention

[0004] The purpose of the present invention is to provide a method for interpreting and assessing risks of medical test reports, thereby solving some of the drawbacks and deficiencies pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0006] S1. First, we parse multi-source medical test reports through a report format parsing system that integrates natural language processing (NLP) and machine learning algorithms. We use an adaptive semantic parsing model to identify and uniformly process medical terms, abbreviations, and specific expressions from different hospitals and laboratories, and convert the original test data into a standardized structured data set.

[0007] S2. Then, based on the hierarchical semantic analysis mechanism, the standardized medical test report information is analyzed hierarchically, and the data is divided into key information, auxiliary information and background information;

[0008] S2.1. Use a multi-dimensional deep learning model to map information at all levels to a standard medical concept library, especially by combining time series analysis and context, dynamically tracking and analyzing changes in the patient's health status over time, and generating an interpretation framework associated with the patient's clinical background;

[0009] S3. Utilize the patient's historical health data, medical records and multimodal data fusion technology to integrate and fuse the multi-source data in the test report, including blood test results, imaging data, and gene sequencing data, and generate personalized medical report interpretation through adaptive enhancement algorithms;

[0010] S4, a context-sensitive intelligent interpretation model that comprehensively considers the patient's clinical history, current symptoms and pathological characteristics, and dynamically generates clinically relevant interpretation results by integrating semantic networks and deep learning technologies;

[0011] S4.1, and use the intelligent summary generator to extract and highlight the key risk points in the report, and further optimize the interpretation model through real-time feedback mechanism and continuous learning algorithm.

[0012] Furthermore, the hierarchical semantic analysis mechanism comprises the following steps:

[0013] S1. First, we build a hierarchical semantic model. By designing a multi-layer semantic analysis model, we can perform hierarchical analysis on the information in the medical test report. And based on the semantic network or hierarchical attention mechanism of the graph structure, we can divide the semantic information at different levels into different semantic layers. The hierarchical analysis defines the higher-level function S Layer (x), where x represents the input semantic data; the function form is:

[0014]

[0015] Among them, φ(x) is the activation function of semantic information, which enhances or suppresses the contribution of specific semantics, γ is a parameter for adjusting the sensitivity of semantic information, and ψ(x) represents the excitation function in the semantic network, which is used to dynamically adjust the response intensity of input data;

[0016] S2. Then, we use the medical knowledge graph and the customized medical concept library to build a semantic network. In the interpretation process, the nodes in the semantic network represent medical concepts, the edges represent the relationship between concepts, and the weights are dynamically adjusted according to the importance of the concepts. Then, we define the calculus formula W Dynamic (t) is used to describe the change of weight over time, where t represents time, and the formula is as follows:

[0017]

[0018] Among them, η is the time scaling parameter that controls the speed of change of semantic weights, ξ(t) represents the time dependency function of semantic relations, which is used to describe the changes in the relations between different concepts in the semantic network over time;

[0019] S3. Then build a hierarchical classifier and use the multi-task learning model Multi-taskLearningModel to process the standardized and semantically annotated report data; divide the input data into key information, auxiliary information and background information, and output the hierarchical semantic information; introduce the differential equation C Layer (x):

[0020]

[0021] Among them, k is the differential order, which is used to control the degree of data refinement during the classification process, ζ(x) represents the feature weight function of the input data, λ is the adjustment factor, which adjusts the flexibility of the classification boundary, and θ(x) is the activation function of the classifier, which affects the nonlinear response of the classification result;

[0022] S4. Finally, based on the classifier, an adaptive adjustment mechanism is used to fine-tune the classified information according to the patient's specific conditions, including medical history, age, and gender; describe the adaptive adjustment process and introduce the integral equation A Adaptive (x,p), where p represents the patient characteristic parameter, the formula is as follows:

[0023]

[0024] Among them, κ(x,p) is the adjustment function used to describe the interaction between the input data x and the patient characteristics p, μ is the adjustment sensitivity coefficient to control the adaptability of the classification results, and σ(p) represents the nonlinear mapping function of the patient characteristics.

[0025] Furthermore, the generation process of the interpretation framework includes:

[0026] S1. First, the high-dimensional vector generated by the multi-dimensional deep learning model is mapped to the standard medical concept library. The mapping process is implemented by a multi-layer perceptron MLP, where the mapping function is defined as M Concept (x), where x represents the high-dimensional vector output by the model, and the mapping function is as follows:

[0027]

[0028] Among them, α i is the mapping weight, is the activation function, θ(x) and ω(x) are the adaptive function for adjusting the output characteristics of the model and the function for controlling the degree of nonlinearity of the mapping, respectively;

[0029] S2. Secondly, in the process of dynamically tracking the health status of patients, a time series model is constructed to capture the pattern of changes in patient test data over time through a long short-term memory network LSTM or a variational autoencoder VA, and predict future trends; the time series analysis is performed through the calculus formula T Sequence (t) is realized, where t represents the time point in the time series:

[0030]

[0031] Among them, κ(t) represents the time-dependent function, λ(t) and η(τ) represent the key patterns and trends captured in the time series, respectively, while σ and ξ(t,τ) are used to adjust the model's ability to respond to short-term fluctuations and long-term trends;

[0032] S3. Finally, by combining time series data with the patient's historical health data and medical records, the multi-head attention mechanism is used to capture important information in the context; the context integration formula C is defined Context (t,h), where t represents the time point in the time series and h represents the patient's historical data:

[0033]

[0034] Among them, μ(t,h) is the context-sensitive coefficient, which controls the degree of dependence of the adjustment model on time series and historical data, ν(t) and ζ(h) represent the feature weights of the current time point and historical data, respectively, and χ(t,h) controls the nonlinear interaction between context information and time series.

[0035] Furthermore, the personalized medical report interpretation and generation method comprises the following steps:

[0036] S1. First, we use the embedding layer in deep learning to convert different types of medical data into a unified high-dimensional vector representation. The imaging data is characterized by the convolutional neural network CNN, the gene sequencing data is characterized by the fully connected network or convolutional network, and the blood test results are represented by the standardized numerical vector. Define the data embedding function E Embed (d), where d represents the data input of different modes, and the embedding function is as follows:

[0037]

[0038] Among them, λ i is the modal weight coefficient, σ(·) is the activation function, and η(d) represents the nonlinear feature extraction function of the modal data;

[0039] S2. Next, we use adaptive fusion technology to fuse multimodal data together. We use multimodal variational autoencoders to combine the embedding vectors of different modalities, and dynamically adjust the weights of each modality through the attention mechanism. The data fusion function is defined as F Fusion (c), where c represents the embedding vector of multiple modalities, and the fusion function is as follows:

[0040]

[0041] Among them, ν(x) is the modal fusion coefficient, γ(x) and ω(θ) represent the modal feature weight and fusion adjustment parameter respectively, and ζ is the adjustment factor for controlling nonlinear interaction;

[0042] S3. Finally, the fused multimodal data is adaptively enhanced using the patient's historical health data and medical records. The adaptive enhancement algorithm weights and adjusts the data according to the patient's personal history and specific health status. A personalized enhancement function is defined, where y represents the fused multimodal data. The personalized enhancement function is as follows:

[0043]

[0044] Among them, ρ j is the personalized weighting coefficient, ξ j (y) is the enhancement function, τ is the parameter for adjusting the sensitivity of personalized enhancement, φ j (y) represents the influencing factor of the patient's specific health condition.

[0045] Furthermore, the generation of the interpretation results of the clinical relevance includes:

[0046] S1. First, a multi-layer deep learning model is designed to further process the patient's contextual data based on the semantic network; the multi-layer deep learning model consists of multiple sub-models, including a convolutional neural network CNN, a long short-term memory network LSTM and an attention mechanism, which are used to extract key features from patient data and perform contextual correlation analysis; the integration function D of the deep learning model Deep (x) is defined as follows, where x represents the multimodal input data:

[0047]

[0048] Among them, θ j is the layer weight of the deep model, σ(·) is the activation function, and γ j (x) and ω(x) represent feature extraction and context relevance functions, respectively;

[0049] S2. Then develop a context-sensitive interpretation generator to dynamically adjust the model's interpretation strategy based on the patient's current clinical context; the interpretation generation function G Context (y) is defined as follows, where y represents the processed context data:

[0050]

[0051] Among them, ζ(y) is the interpretation adjustment coefficient, ρ(y) represents the context feature weight function, ξ(τ) and ψ(y,τ) are time- and context-related adjustment factors.

[0052] Furthermore, the intelligent summary generator uses natural language processing (NLP) technology to understand the semantics of medical reports and convert complex medical terms, diagnostic information and test results into structured data; analyzes structured data based on the attention mechanism to identify key risk points related to the patient's health status;

[0053] The alert report is a concise risk summary that is dynamically adjusted through a sequence-to-sequence Seq2Seq model to make it highly relevant to the patient's clinical context.

[0054] Furthermore, the data construction method of the intelligent summary generator includes the following steps:

[0055] S1. First, in the process of semantic understanding and data structuring, complex medical terms, abbreviations and professional vocabulary are mapped to standard medical concepts, including ICD codes and SNOMED terms, through medical terminology standardization technology, and unstructured text data is converted into structured medical data; this is achieved through the following mapping function, mapping function S Norm (t), where t represents the input unstructured text data, and the formula is as follows:

[0056]

[0057] Among them, α i is the mapping weight, φ(t) represents the text parsing function, λ i and γ are adjustment parameters, which are used to control the nonlinearity of the mapping;

[0058] S2. Then, the structured information extracted from JSON and XML is stored in a standardized format to form a unified data set containing medical records, test results, diagnostic information and other elements; in the process of identifying key risk points, the attention mechanism is used to analyze the structured data to identify the key risk points closely related to the patient's health status; the attention weight function A is used Attention (x), where x represents the structured data input, the formula is as follows:

[0059]

[0060] Among them, β(x) is the data importance function, ψ(x) represents the feature extraction function, θ(τ) and ζ(x,τ) are time-related adjustment factors, which are further processed through multi-level analysis to identify predicted health risk points from the perspectives of symptoms, diagnosis, medical history, and test results.

[0061] Furthermore, the data construction method of the prompt report includes the following steps:

[0062] S1. First, a large amount of medical report data is collected, including diagnosis information, test results, treatment records, and corresponding patient clinical background information including medical history, age, gender, and previous diseases, and the data is used to train a Seq2Seq model; the Seq2Seq model consists of an encoder and a decoder, wherein the encoder encodes the input medical report and patient clinical background information into a high-dimensional vector representation, and the decoder generates a concise risk summary based on the encoded vector;

[0063] During the training process, the model parameters are optimized by maximum likelihood estimation, and the loss function L is defined MLE (θ) is used to optimize the objective, where θ represents the model parameters, and the formula is as follows:

[0064]

[0065] Among them, x i Represents the input medical report data, y i is the target risk summary, f θ (x i ,y i ) represents the score function generated by the model, which maximizes the log-likelihood to control the generated summary to capture the key information in the report and reflect the individualized health risks of the patient;

[0066] S2. When generating risk summaries, the Seq2Seq model uses a context-sensitive generation mechanism to dynamically adjust the summary content according to the patient's specific clinical background; when generating each word or sentence, the decoder refers to the patient's historical data and current health status to control the summary content to be highly relevant to the patient's individual situation; define the context-sensitive generation function G Context (s,h), where s represents the generated sequence and h represents the patient's historical information. The formula is as follows:

[0067]

[0068] Among them, α(s,h) is the context relevance weight, φ(s) represents the characteristic function of the generated sequence, λ(h) and η(h,τ) are adjustment factors related to historical information, which control the generated summary content to dynamically adapt to the individual background of the patient;

[0069] S3. Finally, the model’s ability to capture key information is enhanced by integrating an attention mechanism into the Seq2Seq model; this allows the decoder to focus on the important parts of the input when generating summaries, highlighting content that is closely related to the patient’s current health risks; and using the attention weight function A Attention (x), where x represents the structured data input, the formula is as follows:

[0070]

[0071] Among them, β k (x) is the weight function of the attention mechanism, ψ k (x) represents the data characteristic function, θ(x) and ζ k (x) is the adjustment parameter that controls the decoder to focus on the key parts of the input and generate a risk summary that is highly correlated with the patient's health status.

[0072] Beneficial effects of the present invention:

[0073] By integrating multimodal data, including text data, imaging data, and gene sequencing data, the patient's health status can be analyzed comprehensively and accurately. This fusion of multimodal data enables the system to not only identify abnormal information in a single data source when interpreting medical reports, but also extract complementary information from different data sources, thereby improving the accuracy and comprehensiveness of the overall diagnosis.

[0074] The use of context-sensitive generation and attention mechanisms significantly enhances the system's sensitivity to individualized information and its ability to capture key information. During the interpretation generation process, the system can dynamically adjust the interpretation content to ensure that the generated report is highly consistent with the patient's individual health background. This not only improves the accuracy of the report, but also generates more personalized health risk assessments and treatment recommendations, making medical recommendations more targeted and helping to improve patient treatment outcomes and compliance. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 The present invention is a flow chart of the medical test report interpretation and risk assessment method.

[0076] Figure 2 This is a module diagram of the hierarchical semantic analysis mechanism of the present invention.

[0077] Figure 3 The flowchart for generating the interpretation framework of the present invention.

[0078] Figure 4 The present invention is a flowchart of the personalized medical report interpretation and generation method. DETAILED DESCRIPTION

[0079] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0080] Combined with Figure 1In the medical test report interpretation and risk assessment method, the multi-source medical test reports from different hospitals and laboratories are first parsed through a report format parsing system that integrates natural language processing (NLP) and machine learning algorithms. This system uses an adaptive semantic parsing model to identify and uniformly process various medical terms, abbreviations, and specific expressions, and convert these original, unstructured test data into standardized structured data sets. This standardization process ensures data consistency, allowing test results from different sources to be analyzed and interpreted on the same platform, providing a unified data foundation for subsequent steps.

[0081] Next, based on the hierarchical semantic analysis mechanism, the standardized medical test report information is analyzed hierarchically. Specifically, the data is divided into three categories: key information, auxiliary information, and background information. This hierarchical analysis method can help identify the most important parts of the report and process this information in a hierarchical manner. In this process, a multi-dimensional deep learning model is used, which maps each level of information to a standard medical concept library to ensure that each type of information can be accurately interpreted and applied. In particular, the model combines time series analysis and contextual context, allowing the system to dynamically track changes in the patient's health status over time and generate an interpretation framework that is closely related to the patient's clinical background. This dynamic tracking and analysis mechanism not only takes into account the current test results, but also combines the patient's historical data to predict future health trends.

[0082] Subsequently, the system uses the patient's historical health data, medical records, and multimodal data fusion technology to integrate and fuse the multi-source data in the test report. These data include blood test results, imaging data, and gene sequencing data. Through the adaptive enhancement algorithm, these multimodal data are integrated into a unified and personalized medical report interpretation. The adaptive enhancement algorithm adjusts the data weight and interpretation strategy according to the patient's specific situation to generate a highly personalized interpretation report. This report not only provides an assessment of the current health status, but also provides detailed analysis and suggestions for the patient's future health risks.

[0083] In the process of generating interpretation reports, the system takes into account the patient's clinical history, current symptoms and pathological characteristics based on a context-sensitive intelligent interpretation model. The interpretation model dynamically generates interpretation results with high clinical relevance by integrating semantic networks and deep learning technology. The semantic network ensures the consistency and accuracy of concepts in the interpretation process, while deep learning technology is used to extract and analyze complex medical data to ensure the accuracy and practicality of the interpretation results.

[0084] Finally, to improve the practicality and accuracy of the report, the system uses an intelligent summary generator to extract and prompt key risk points in the report. This step ensures that medical staff and patients can quickly obtain the most important health information. In addition, the system integrates a real-time feedback mechanism and a continuous learning algorithm. By continuously collecting and analyzing user feedback, the system can self-optimize and gradually improve the performance and accuracy of the interpretation model. This ability to continuously learn and optimize enables the system to maintain a high degree of accuracy and relevance when faced with new medical data and patient conditions, and provide personalized medical advice and risk assessments.

[0085] Embodiment 1:

[0086] View attached Figure 2 In the embodiment, a multi-layer semantic analysis model is designed to hierarchically analyze the patient's medical test report, and the data in the report is divided into different semantic levels based on the semantic network and hierarchical attention mechanism. The input data x is set to represent the patient's blood sugar level, liver function index, blood lipid content and imaging data, etc. The formula is used:

[0087]

[0088] Among them, φ(x) is the activation function of semantic information, which aims to enhance or suppress the contribution of specific semantics, γ is a parameter for adjusting the sensitivity of semantic information, and its value range is 0.1≤γ≤1.0, and ψ(x) represents the excitation function in the semantic network, which is used to dynamically adjust the response strength of the input data.

[0089] Assume that the patient's blood sugar level (8.0mmol / L) is slightly higher than the normal range, which is the key semantic information that the system pays attention to. Set φ(x) = 2, γ = 0.5, ψ(x) = 1.5x. Then within the integral range of α = 0, β = 10, the system will perform a hierarchical analysis of the blood sugar data and substitute it into the formula for calculation:

[0090]

[0091] Set n = 1 (first differential) and perform an approximate solution to obtain:

[0092] S Layer (8)≈15.6

[0093] This result indicates that the patient's blood glucose level is marked as high-risk semantic information in the system and will be further analyzed in subsequent steps.

[0094] The embodiment further utilizes the medical knowledge graph and the customized medical concept library to establish a semantic network, and first maps the patient's test data to the nodes of the semantic network. Node A is set to represent "diabetes", node B represents "abnormal liver function", node C represents "hyperglycemia", and node D represents "insulin resistance". The edges represent the causal relationship or mutual influence between these concepts. In order to dynamically adjust the weights of these nodes, different weights are assigned according to the patient's specific circumstances (family history, previous health data) and current test results. In order to accurately describe the changes in these weights over time, the following calculus formula is introduced:

[0095]

[0096] Where t represents time, η is the time scaling parameter that controls the speed of change of semantic weights, and its value range is 0.01≤η≤0.1. ξ(t) represents the time-dependent function of the semantic relationship, which is used to describe the changes in the relationship between different concepts in the semantic network over time.

[0097] The time scaling parameter η is set to 0.05, the time t ranges from 0 to 12 months, and ξ(t) is the value of the time-dependent function, reflecting the changes in blood glucose levels and liver function. By solving the integral formula, the system can calculate the dynamic weight of each node. In a certain month, the weight of node C (hyperglycemia) increases because the patient's blood glucose level increases significantly in that month. At this time, the system will give node C a higher weight, thereby paying more attention to medical concepts related to hyperglycemia and adjusting the interpretation of diabetes risk.

[0098] In the specific calculation, ξ(t)=0.8+0.2sin(t) is set, and the system obtains W Dynamic The results of (t) changing with time are as follows:

[0099]

[0100] This result suggests that systemic attention to the risks associated with hyperglycemia has increased over time, particularly as patients' health conditions become more significantly abnormal.

[0101] Then the embodiment uses a multi-task learning model to hierarchically classify the report into key information, auxiliary information and background information, and finally outputs the layered semantic information for accurate risk assessment. Construct a hierarchical classifier and use a multi-task learning model to process the medical report data of patients that have been standardized and semantically annotated. The task of the classifier is to divide the input data (blood sugar, liver function indicators, imaging data) into key information, auxiliary information and background information. Use the formula:

[0102]

[0103] Among them, k is the differential order, ranging from 1 to 3, which is used to control the degree of data refinement; ζ(x) represents the feature weight function of the input data, reflecting the importance of the data; λ is the adjustment factor, ranging from 0.1 to 1.0, which is used to adjust the flexibility of the classification boundary; θ(x) is the activation function of the classifier, which affects the nonlinear response of the classification result.

[0104] The patient's blood sugar level x is set to 8.5mmol / L, and the system inputs this data into the hierarchical classifier. According to the above formula, let k = 2, ζ(x) = 2x, λ = 0.5, θ(x) = 1.2x, and substitute it into the differential equation to calculate:

[0105]

[0106] By derivation and simplification, we get:

[0107] C Layer (8.5)≈0.135

[0108] The results showed that the patient's blood glucose level was marked as very important key information in the classification process and would be focused on in the subsequent risk assessment.

[0109] The final embodiment uses an adaptive adjustment mechanism to fine-tune the classification results according to the patient's specific medical history, age, gender and other individualized information to generate a more accurate medical report interpretation and health risk assessment. The adaptive adjustment mechanism is used to fine-tune the previously classified information according to the patient's medical history, age, and gender. By introducing the integral equation A Adaptive (x,p) implementation:

[0110]

[0111] Where x represents input data (such as blood sugar level, liver function test results, etc.), p represents patient characteristic parameters (such as age, gender, medical history, etc.), κ(x, p) is the adjustment function, which is used to describe the interaction between input data x and patient characteristics p; μ is the adjustment sensitivity coefficient, which ranges from 0.1≤μ≤1.0 and controls the adaptability of classification results to patient characteristics; σ(p) is the nonlinear mapping function of patient characteristics, which reflects the influence of factors such as age and medical history on the adjustment process.

[0112] The system will input the patient's blood glucose level x = 8.5mmol / L, as well as the patient's characteristic parameters p, age 45 years old (mid-life risk parameter), gender female, and family history of diabetes. Set the adjustment function κ(x,p) = 1.5x·p, the sensitivity coefficient μ = 0.8, the nonlinear mapping function σ(p) = 0.7p, and define the integral domain Ω from 0 to 1, then the integral equation A Adaptive (x,p) is calculated as follows:

[0113]

[0114] By calculation, we can approximate:

[0115] A Adaptive (8.5,45)≈574.5

[0116] The results showed that the system made significant adjustments to hyperglycemia data based on the patient's specific circumstances, especially taking into account age, gender and family history, which significantly increased the weight of this information in system interpretation and risk assessment.

[0117] Embodiment 2:

[0118] In this embodiment, a high-dimensional vector representation of the patient's test report is first generated through a multi-dimensional deep learning model (convolutional neural network, long short-term memory network), which contains multiple health indicators about the patient. These vectors are mapped to the standard medical concept library through a multi-layer perceptron (MLP). The mapping function is defined as:

[0119]

[0120] Where x represents the high-dimensional vector output by the model, α i is the mapping weight, ranging from 0.1≤α i ≤1.0, As the activation function, ReLU or Sigmoid is selected. θ(x) and ω(x) are the adaptive function for adjusting the output characteristics of the model and the function for controlling the degree of nonlinearity of the mapping, respectively. ∈ is a parameter for controlling the degree of nonlinearity, and its value range is 0.01≤∈≤0.5.

[0121] Assume that the patient's blood sugar level, high-density lipoprotein (HDL) and liver function (ALT) data are encoded as a high-dimensional vector x, and the system maps these data to a standard medical concept library through a multi-layer perceptron. Set α1 = 0.8, β1=0.6,θ(x)=1.5x,∈=0.1,ω(x)=2x,and m=1 (first differential), the system calculates:

[0122]

[0123] Assuming x = 8.5 (corresponding to the patient's blood sugar level), we can calculate:

[0124] M Concept (8.5)≈0.8·ReLU(0.672)=0.5376

[0125] The results show that the patient's hyperglycemia data was successfully mapped to the relevant node (diabetes risk) in the standard medical concept library, and the weight was set to 0.5376.

[0126] View attached Figure 3 The embodiment further utilizes the time series model to dynamically track the health status of the patient. The long short-term memory network (LSTM) is used to capture the pattern of patient test data changing over time, and the calculus formula for time series analysis is introduced:

[0127]

[0128] κ(t) represents the time-dependent function, with a value range of 0.01≤κ(t)≤1.0, and is used to describe patterns that change over time; λ(t) and η(τ) represent the key patterns and trends captured in the time series, respectively; σ is a parameter for controlling short-term fluctuations, with a value range of 0.1≤σ≤0.5; ξ(t,τ) is used to adjust the model’s ability to respond to short-term fluctuations and long-term trends.

[0129] The change data of the patient's blood sugar level in the past 12 months is set as x(t), where t is the time point (in months). The system captures the change pattern of blood sugar level through the LSTM model and uses the following parameters for calculation: κ(t) = 0.05·t, λ(t) = 0.8·t, η(τ) = 1.2·sin(τ), σ = 0.3, ξ(t,τ) = 0.6·t·cos(τ). By substituting these parameters into the time series analysis formula, the system can calculate the future trend of the patient's blood sugar level.

[0130]

[0131] Through calculation, we get the following results:

[0132] T Sequence (t)≈0.35·t 2 -0.05·t+1.0

[0133] This result indicates that patients' blood sugar levels will continue to rise in the coming months, especially based on the current trend, predicting a new peak in blood sugar by month 18.

[0134] The final example combines time series data with the patient's historical health data and medical records, and uses a multi-head attention mechanism to capture important information in the context to generate a more accurate health risk assessment report. Define the context integration formula:

[0135]

[0136] Where t represents the time point in the time series, h represents the patient's historical data, μ(t,h) is the context sensitivity coefficient, and its value range is generally 0.01≤μ(t,h)≤1.0, which is used to control the model's dependence on the time series and historical data; ν(t) and ζ(h) represent the feature weights of the current time point and historical data, respectively, and χ(t,h) is a parameter that controls the nonlinear interaction between context information and time series, and its value range is 0.1≤χ(t,h)≤0.5.

[0137] Assume that the patient's time series data t represents the blood sugar level in the past 12 months, and the historical data h includes the patient's weight change and family history of diabetes in the past 5 years. Assume that μ(t,h)=0.5t+0.2h, ν(t)=0.7·t, ζ(h)=1.1·h, χ(t,h)=0.3·t·h, and the integration domain Ω is from 0 to 1, then the context integration formula C Context (t,h) is calculated as follows:

[0138]

[0139] By calculation, setting t = 12 months, h = 45 months (considering weight change data), and setting ρ = 0.4, we get:

[0140] C Context (12,45)≈2.16

[0141] The results showed that after combining the patient's time series data with historical data, the system identified the importance of weight gain and family history of diabetes in current blood sugar trends. Therefore, the system incorporated this information into the final health risk assessment and gave a more accurate prediction.

[0142] Embodiment 3:

[0143] Combined with Figure 4 In the embodiment, a personalized medical report interpretation is generated. First, different types of medical data are converted into a unified high-dimensional vector representation, using a data embedding function:

[0144]

[0145] λ i is the modal weight coefficient, ranging from 0.1≤λ i≤1.0; σ(·) is the activation function, including ReLU or Sigmoid; η(d) represents the nonlinear feature extraction function of the modal data, which is used to capture the key features in each data type.

[0146] After the patient's abdominal ultrasound imaging data d1 is input into the system, the features are extracted through the convolutional neural network (CNN), and the feature data extracted by the system is expressed as η(d1). Set λ1=0.7, σ(x)=ReLU(x), α1=0.9, and the differential order k=2, and substitute them into the formula to calculate:

[0147]

[0148] Assuming that the feature vector η(d1) extracted by the system from the ultrasound image is 5, the calculation result is:

[0149] E Embed (d1)≈0.7·ReLU(0.9·10)=6.3

[0150] Similarly, the system processes the patient’s gene sequencing data d2 and blood test data d3 to obtain high-dimensional vector representations E Embed (d2) and E Embed (d3). These embedded vectors will be input into the subsequent analysis module of the system as unified high-dimensional data to generate the final personalized medical report interpretation.

[0151] Then, the embodiment uses adaptive fusion technology to fuse these multimodal data together, and dynamically adjusts the weight of each modality through the attention mechanism to generate a comprehensive health risk assessment. The patient's medical data has been processed and embedded into a high-dimensional vector, and the abdominal ultrasound image data is represented as a vector E Embed (d1), gene sequencing data is E Embed (d2), blood test data is E Embed (d3). Adaptive fusion technology is used to fuse the multimodal data of patients. This is achieved through a multimodal variational autoencoder (MVAE), which can combine the embedding vectors of different modalities and dynamically adjust the weights of each modality through an attention mechanism. The fusion function is:

[0152]

[0153] Where x represents the embedding vector of multiple modalities, ν(x) is the modal fusion coefficient, which ranges from 0.1≤ν(x)≤1.0 and is used to adjust the influence of each modality in the overall fusion; γ(x) and ω(θ) represent the modal feature weights and fusion adjustment parameters, respectively; ζ is the adjustment factor for controlling nonlinear interaction, which ranges from 0.01≤ζ≤0.5.

[0154] Set the image data vector E Embed (d1) = 6.3, gene data vector E Embed (d2) = 5.7, and the blood test data vector E Embed (d3) = 7.2. The system inputs these embedding vectors into the adaptive fusion module for fusion.

[0155] Set the modal fusion coefficient ν(x) = 0.8, the modal feature weight γ(x) = 1.2·x, the fusion adjustment parameter ω(θ) = 0.9·θ, and the nonlinear interaction adjustment factor ζ = 0.3. The integral range of adaptive fusion θ1 to θ2 is set to 0 to 1. The data fusion function is calculated as follows:

[0156]

[0157] Set x = [6.3, 5.7, 7.2] to represent the embedding vectors of the three modes, and the system calculates and fuses the results. Taking x1 = 6.3 as an example, the calculation results are as follows:

[0158] F Fusion (6.3)≈0.8·1.44≈1.152

[0159] Similarly, the system calculates F Fusion (5.7)≈1.008 and F Fusion (7.2)≈1.296, and the final fusion result will be used as the input for interpreting the patient’s personalized medical report.

[0160] The last example uses the patient's historical health data and medical records to adaptively enhance the fused multimodal data to generate a more personalized medical report interpretation. The personalized enhancement function is used:

[0161]

[0162] Where y represents the fused multimodal data, ρ j is the personalized weighting coefficient, and its value range is generally 0.1≤ρ j ≤1.0, used to adjust the weight of each data feature according to the patient's individualized information; ξ j(y) is the enhancement function, which indicates the enhancement strength of a specific data feature; τ is the parameter for adjusting the sensitivity of personalized enhancement, and its value range is 0.1≤τ≤1.0; φ j (y) represents the influencing factor of the patient's specific health status, reflecting the impact of specific historical health data on the current health assessment.

[0163] It is assumed that the patient's fusion data y includes image data, gene marker data, and blood test results, wherein the fusion vector of the image data is y1=6.3, the fusion vector of the gene data is y2=5.7, and the fusion vector of the blood test results is y3=7.2.

[0164] Set ρ1 = 0.7, ρ2 = 0.8, ρ3 = 0.9, and enhance the function ξ j (y) = 1.2y j , sensitivity parameter τ = 0.5, health status influence factor φ j (y) = 1.5y j For the image data y1, the calculation of the personalized enhancement function is as follows:

[0165]

[0166] The following results are obtained through calculation:

[0167] G Enhance (y1)≈0.7·7.56≈5.292

[0168] Similarly, the system calculates the genetic data and blood test results to obtain the enhanced fusion data result G Enhance (y2)≈4.824 and G Enhance (y3)≈7.776.

[0169] The enhanced data will be used to generate the patient's final health risk assessment report.

[0170] Embodiment 4:

[0171] In the embodiment, the interpretation results of clinical relevance are generated, and a multi-layer deep learning model is used to further process the patient's contextual data based on the semantic network. The multi-layer deep learning model is designed, which consists of multiple sub-models, including convolutional neural network (CNN), long short-term memory network (LSTM) and attention mechanism. CNN is used to process and extract spatial features in imaging data, LSTM is used to capture the dependencies of time series data, and attention mechanism is used to identify key features in the data and dynamically adjust the weights according to the context. The system defines the integration function D of the deep learning model Deep (x), the function is in the following form:

[0172]

[0173] Where x represents the multimodal input data, θ j is the layer weight of the deep model, and its value range is generally 0.1≤θ j ≤1.0; σ(·) is the activation function, usually ReLU or Sigmoid; γ j (x) and ω(x) represent the feature extraction function and context correlation function, respectively, which are used to perform in-depth analysis of the data.

[0174] Assume that the patient's multimodal data x includes image data x1 = 6.3, gene data x2 = 5.7, and blood test data x3 = 7.2. The system uses CNN to extract spatial features from image data, LSTM to capture time series features of gene data, and attention mechanism to identify key indicators in blood test data.

[0175] Set layer weights θ1 = 0.8, θ2 = 0.9, θ3 = 0.7, activation function σ(x) = ReLU(x), feature extraction function γ j (x) = 1.2x, context correlation function ω(x) = 0.9x, and differential order k = 1, substitute into the formula and calculate:

[0176]

[0177] The calculation results are:

[0178] D Deep (x)≈0.8·43.092+0.9·34.722+0.7·56.376=34.4736+31.2498+39.4632

[0179] =105.1866

[0180] The results showed that the system comprehensively considered the patient's imaging characteristics, genetic characteristics, and blood test characteristics, and performed contextual correlation analysis through a deep learning model.

[0181] The embodiment further develops a context-sensitive interpretation generator that dynamically adjusts the interpretation strategy according to the patient's current clinical context, thereby generating a highly personalized medical report interpretation. The interpretation generation function is:

[0182]

[0183] Where y represents the processed context data, ζ(y) is the interpretation adjustment coefficient, which ranges from 0.1≤ζ(y)≤1.0 and is used to dynamically adjust the flexibility of interpretation; ρ(y) represents the context feature weight function, which reflects the importance of certain specific context features in the interpretation process; ξ(τ) and ψ(y,τ) are time- and context-related adjustment factors, and χ is the adjustment coefficient that controls the impact of context information on the interpretation result, which ranges from 0.1≤χ≤0.5.

[0184] The patient’s contextual data y is assumed to include recent symptom reports (fatigue, weight gain), the latest blood sugar and lipid test results, and their family medical history. The system uses this data to dynamically adjust the output of the interpretation generator.

[0185] Assuming the interpretation adjustment coefficient ζ(y) = 0.8·y, the context feature weight function ρ(y) = 1.2·y, the time adjustment factor ξ(τ) = 0.5·τ, the context adjustment factor ψ(y,τ) = 0.3·y·τ, and the integration range τ1 to τ2 is 0 to 12 months, the interpretation generation function is calculated as follows:

[0186]

[0187] Assuming the patient's context data y = 7.5 (comprehensively considering symptoms and test results), the calculation result is:

[0188]

[0189] Through integration and further calculation, the system obtains the following results:

[0190] G Context (7.5)≈45

[0191] The results showed that the system specifically emphasized the importance of recent symptoms and test results in clinical context when interpreting patients' health data.

[0192] Embodiment 5:

[0193] In the embodiment, the key information in the report is extracted by the intelligent summary generator, and the medical terms are first mapped to the standard medical concept library, including ICD codes and SNOMED terms, so as to convert the unstructured text data into structured medical data. The mapping function is used:

[0194]

[0195] Where t represents the input unstructured text data, α i is the mapping weight, the value range is 0.1≤α i≤1.0; φ(t) represents the text parsing function, which is used to parse the input unstructured text data; λ i and γ are adjustment parameters, λ i The value range is 0.1≤λ i ≤1.0, and the value range of γ is 0.1≤γ≤0.5, which is used to control the nonlinearity of the mapping.

[0196] Set the mapping weight α1 = 0.7, the text parsing function φ(t) = 1.2·t, the adjustment parameters λ1 = 0.5, γ = 0.3, and the differential order k = 2, then the mapping function is calculated as follows:

[0197]

[0198] Assuming t = 3.5 corresponds to a complex term in the text, the calculation result is:

[0199]

[0200] Through this mapping function, the complex term “diabetes risk” in the report is mapped to standardized structured data items, corresponding ICD codes or SNOMED terms, making subsequent analysis and intelligent summary generation more efficient and accurate.

[0201] The embodiment further processes this data, stores it in a standardized format, and uses an attention mechanism to analyze the structured data to identify key risk points that are closely related to the patient's health status. The attention weight function is:

[0202]

[0203] Where x represents the structured data input, β(x) is the data importance function, and its value range is 0.1≤β(x)≤1.0; ψ(x) represents the feature extraction function, which is used to extract key features from the input data; θ(τ) and ζ(x,τ) are time-related adjustment factors, and ρ is the adjustment coefficient for controlling the time interaction effect, and its value range is 0.1≤ρ≤0.5.

[0204] Assume that the system has extracted the patient's medical record data x1, blood sugar test results x2, and diagnostic information x3 as structured data input. Input these data into the attention weight function to identify key risk points. Set the data importance function β(x) = 0.8·x, the feature extraction function ψ(x) = 1.2·x, the time adjustment factor θ(τ) = 0.9·τ, the context adjustment factor ζ(x,τ) = 0.7·x·τ, the integral range τ1 to τ2 is 0 to 12 months, and the attention weight function is calculated as follows:

[0205]

[0206] Assume the patient's medical record data x1 = 7.5, and the calculation result is:

[0207]

[0208] Through integration and further calculation, the following results are obtained:

[0209] A Attention (7.5)≈57.6

[0210] This result shows that when analyzing the patient's medical record data, the system identified the high importance of this part of the data in the current health risk assessment, and therefore paid special attention to this when generating the final report.

[0211] Embodiment 6:

[0212] In this embodiment, medical report data and corresponding patient clinical background information, including various diagnostic information, test results and treatment records, are first collected, and these data are used to train the Seq2Seq model. In the Seq2Seq model, the encoder encodes the input medical report and the patient's clinical background information into a high-dimensional vector representation. The patient's blood sugar level, medical history and current symptoms are encoded into a comprehensive vector. The decoder generates a risk summary based on the encoded vector. During the training process, the system optimizes the model parameters through maximum likelihood estimation (Maximum Likelihood Estimation, MLE). The loss function is as follows:

[0213]

[0214] where θ represents the model parameters, x i Represents the input medical report data, y i is the target risk summary, f θ (x i ,y i ) represents the score function generated by the model. By maximizing the log-likelihood, the system ensures that the generated summary accurately captures the key information in the report and reflects the patient's individualized health risks.

[0215] The system is assumed to have collected 1,000 medical reports of similar patients for training the Seq2Seq model. During the training process, the system will convert each report data into i The goal is to generate a risk summary y that is consistent with the actual diagnosis and physician recommendations. i Suppose the main risk in a particular report x1 is diabetes, and the target summary y1 is "The patient has a high risk of diabetes, and it is recommended to control diet and test blood sugar regularly." The system calculates the score function f θ The result obtained by (x1,y1) is:

[0216] f θ (x1,y1)=0.85

[0217] Other abstracts scored lower:

[0218] f θ (x1,y2)=0.60,f θ (x1,y3)=0.55

[0219] Then, the log-likelihood calculated by the system will be biased towards selecting y1 as the final output summary.

[0220]

[0221] The system adjusts the model parameter θ during continuous training so that each generated summary more accurately reflects the important information in the report and gradually reduces the value of the loss function.

[0222] The next step of the embodiment is to introduce a context-sensitive generation mechanism. When generating the risk summary, the decoder of the Seq2Seq model adopts a context-sensitive generation mechanism. When generating each word or sentence, the decoder will refer to the patient's historical data and current health status to ensure that the generated content is highly relevant to the individual situation. The formula of the generation function is as follows:

[0223]

[0224] Where s represents the generated sequence, h is the patient's historical information, α(s,h) is the context relevance weight, ranging from 0.1≤α(s,h)≤1.0, which is used to adjust the dependence of the generated content on the context; φ(s) represents the characteristic function of the generated sequence, which controls the characteristics of the generated sentences or words; λ(h) and η(h,τ) are adjustment factors related to historical information, and γ is the adjustment coefficient, ranging from 0.01≤γ≤0.1, which controls how the generated summary content dynamically adapts to the patient's individual background.

[0225] Set the context relevance weight α(s,h) = 0.8·s·h, generate the characteristic function of the sequence φ(s) = 1.2·s, the historical information adjustment factor λ(h) = 1.1·h, η(h,τ) = 0.5·h·τ, the integration range τ1 to τ2 is 0 to 12 months, and calculate the context-sensitive generating function as follows:

[0226]

[0227] Assuming the patient's generation sequence s = 3.5 and historical information h = 7.5, the calculation results are:

[0228]

[0229] Through integration and further calculation, the system obtains the following results:

[0230] G Context (3.5,7.5)≈199.5

[0231] The results showed that the system took into account the patient's historical medical history and current health status when generating the patient's risk summary, so the generated summary content will emphasize diabetes risks, necessary preventive measures, and future health monitoring recommendations.

[0232] The embodiment further enhances the model's ability to capture key information by integrating an attention mechanism into the Seq2Seq model, enabling the decoder to focus on the most important parts of the input data when generating a summary, especially content that is closely related to the patient's current health risks.

[0233] In the Seq2Seq model, the attention mechanism is integrated into the decoder, enabling it to focus on the key parts of the input when generating summaries. The attention mechanism dynamically adjusts the decoder's attention to different input features by calculating attention weights, thereby generating a more accurate summary. The attention weight function is used:

[0234]

[0235] Where x represents the structured data input, β k (x) is the weight function of the attention mechanism, and its value range is 0.1≤β k (x)≤1.0, used to adjust the degree of attention to each input feature; ψ k (x) represents the data feature function, which controls the influence of different features; θ(x) and ζ k (x) is the adjustment parameter, ρ is the adjustment coefficient for controlling the response of the attention mechanism, and its value range is 0.1≤ρ≤0.5.

[0236] Suppose the patient's structured data x1 is blood glucose level, x2 is recent weight change, and x3 is family medical history. The system inputs this data into the attention weight function to identify the most critical information and generate a risk summary that highlights this information. Set the attention weight function β1(x) = 0.8·x1, the data feature function ψ1(x) = 1.2·x1, the adjustment parameters θ(x) = 0.9·x, ζ1(x) = 0.7·x, the differential order l = 2, and the attention weight function is calculated as follows:

[0237]

[0238] Assuming the patient's blood sugar level x1 = 8.5, the calculation result is:

[0239]

[0240] The results showed that the system paid special attention to the patient's blood sugar level when generating the summary and highlighted it as the most important part of the report.

[0241] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. Medical test report interpretation and risk assessment method, characterized by The following steps are involved: S1. First, we parse multi-source medical test reports through a report format parsing system that integrates natural language processing (NLP) and machine learning algorithms. We use an adaptive semantic parsing model to identify and uniformly process medical terms, abbreviations, and specific expressions from different hospitals and laboratories, and convert the original test data into a standardized structured data set. S2. Then, based on the hierarchical semantic analysis mechanism, the standardized medical test report information is analyzed hierarchically, and the data is divided into key information, auxiliary information and background information; S2.

1. Use a multi-dimensional deep learning model to map information at all levels to a standard medical concept library, especially by combining time series analysis and context, dynamically tracking and analyzing changes in the patient's health status over time, and generating an interpretation framework associated with the patient's clinical background; S3. Utilize the patient's historical health data, medical records and multimodal data fusion technology to integrate and fuse the multi-source data in the test report, including blood test results, imaging data, and gene sequencing data, and generate personalized medical report interpretation through adaptive enhancement algorithms; S4, a context-sensitive intelligent interpretation model that comprehensively considers the patient's clinical history, current symptoms and pathological characteristics, and dynamically generates clinically relevant interpretation results by integrating semantic networks and deep learning technologies; S4.1, and use the intelligent summary generator to extract and highlight the key risk points in the report, and further optimize the interpretation model through real-time feedback mechanism and continuous learning algorithm.

2. The method for interpreting and assessing medical test reports according to claim 1, characterized in that The hierarchical semantic analysis mechanism comprises the following steps: S1. First, we build a hierarchical semantic model. By designing a multi-layer semantic analysis model, we can perform hierarchical analysis on the information in the medical test report. And based on the semantic network or hierarchical attention mechanism of the graph structure, we can divide the semantic information at different levels into different semantic layers. The hierarchical analysis defines the higher-level function S Layer (x), where x represents the input semantic data; the function form is: Among them, φ(x) is the activation function of semantic information, which enhances or suppresses the contribution of specific semantics, γ is a parameter for adjusting the sensitivity of semantic information, and ψ(x) represents the excitation function in the semantic network, which is used to dynamically adjust the response intensity of input data; S2. Then, a semantic network is established using the medical knowledge graph and a custom medical concept library. During the interpretation process, the nodes in the semantic network represent medical concepts, the edges represent the relationships between concepts, and the weights are dynamically adjusted according to the importance of the concepts. S3. Then build a hierarchical classifier and use the multi-task learning model Multi-taskLearningModel to process the standardized and semantically annotated report data; divide the input data into key information, auxiliary information and background information, and output the layered semantic information; S4. Finally, based on the classifier, an adaptive adjustment mechanism is used to fine-tune the classified information according to the patient's specific conditions, including medical history, age, and gender; describe the adaptive adjustment process and introduce the integral equation Adaptive (x,p), where p represents the patient characteristic parameter, the formula is as follows: Among them, κ(x,p) is the adjustment function used to describe the interaction between the input data x and the patient characteristics p, μ is the adjustment sensitivity coefficient to control the adaptability of the classification results, and σ(p) represents the nonlinear mapping function of the patient characteristics.

3. The method for interpreting and assessing medical test reports according to claim 1, characterized in that The generation process of the interpretation framework includes: S1. First, the high-dimensional vector generated by the multi-dimensional deep learning model is mapped to the standard medical concept library. The mapping process is implemented by a multi-layer perceptron MLP, where the mapping function is defined as M Concept (x), where x represents the high-dimensional vector output by the model, and the mapping function is as follows: Among them, α i is the mapping weight, is the activation function, θ(x) and ω(x) are the adaptive function for adjusting the output characteristics of the model and the function for controlling the degree of nonlinearity of the mapping, respectively; S2. Secondly, in the process of dynamically tracking the health status of patients, a time series model is constructed to capture the pattern of changes in patient test data over time through the long short-term memory network LSTM or variational autoencoder VA, and predict future trends; S3. Finally, by combining time series data with the patient's historical health data and medical records, the multi-head attention mechanism is used to capture important information in the context; the context integration formula C is defined Context (t,h), where t represents the time point in the time series and h represents the patient's historical data: Among them, μ(t,h) is the context-sensitive coefficient, which controls the degree of dependence of the adjustment model on time series and historical data, ν(t) and ζ(h) represent the feature weights of the current time point and historical data, respectively, and χ(t,h) controls the nonlinear interaction between context information and time series.

4. The method for interpreting and assessing medical test reports according to claim 1, characterized in that The personalized medical report interpretation and generation method comprises the following steps: S1. First, we use the embedding layer in deep learning to convert different types of medical data into a unified high-dimensional vector representation. The imaging data is characterized by the convolutional neural network CNN, the gene sequencing data is characterized by the fully connected network or convolutional network, and the blood test results are represented by the standardized numerical vector. Define the data embedding function E Embed (d), where d represents the data input of different modes, and the embedding function is as follows: Among them, λ i is the modal weight coefficient, σ(·) is the activation function, and η(d) represents the nonlinear feature extraction function of the modal data; S2. Next, we use adaptive fusion technology to fuse multimodal data together, use multimodal variational autoencoders to combine the embedding vectors of different modalities, and dynamically adjust the weights of each modality through the attention mechanism; S3. Finally, the fused multimodal data is adaptively enhanced using the patient’s historical health data and medical records. The adaptive enhancement algorithm weights and adjusts the data based on the patient’s personal history and specific health conditions.

5. The method for interpreting medical test reports and assessing risk according to claim 1, characterized in that The generation of clinically relevant interpretation results includes: S1. First, a multi-layer deep learning model is designed to further process the patient's contextual data based on the semantic network; the multi-layer deep learning model consists of multiple sub-models, including a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism, which are used to extract key features from patient data and perform contextual relevance analysis; S2. Then develop a context-sensitive interpretation generator to dynamically adjust the model's interpretation strategy based on the patient's current clinical context.

6. The method for interpreting and assessing medical test reports according to claim 1, characterized in that The intelligent summary generator uses natural language processing (NLP) technology to understand the semantics of medical reports and convert complex medical terms, diagnostic information and test results into structured data; Analyze structured data based on attention mechanism to identify key risk points related to patient health status; The alert report is a concise risk summary that is dynamically adjusted by a sequence-to-sequence Seq2Seq model.

7. The method for interpreting medical test reports and assessing risk according to claim 6, characterized in that The data construction method of the intelligent summary generator includes the following steps: S1. First, in the process of semantic understanding and data structuring, complex medical terms, abbreviations and professional vocabulary are mapped to standard medical concepts, including ICD codes and SNOMED terms, through medical terminology standardization technology, and unstructured text data is converted into structured medical data; this is achieved through the following mapping function, mapping function S Norm (t), where t represents the input unstructured text data, and the formula is as follows: Among them, α i is the mapping weight, φ(t) represents the text parsing function, λ i and γ are adjustment parameters, which are used to control the nonlinearity of the mapping; S2. Then, the structured information including JSON and XML is extracted and stored in a standardized format to form a unified data set containing medical records, test results, diagnostic information and other elements. In the process of identifying key risk points, the attention mechanism is used to analyze the structured data to identify key risk points that are closely related to the patient's health status.

8. The method for interpreting medical test reports and assessing risk according to claim 6, characterized in that The data construction method of the prompt report includes the following steps: S1. First, medical report data is collected, including diagnosis information, test results, treatment records, and corresponding patient clinical background information including medical history, age, gender, and previous diseases, and the data is used to train a Seq2Seq model; the Seq2Seq model consists of an encoder and a decoder, wherein the encoder encodes the input medical report and patient clinical background information into a high-dimensional vector representation, and the decoder generates a concise risk summary based on the encoded vector; S2. When generating risk summaries, the Seq2Seq model uses a context-sensitive generation mechanism to dynamically adjust the summary content based on the patient's specific clinical background; when generating each word or sentence, the decoder refers to the patient's historical data and current health status to control the summary content to be highly relevant to the patient's individual situation; S3. Finally, to enhance the model’s ability to capture key information, an attention mechanism is integrated into the Seq2Seq model; this allows the decoder to focus on important parts of the input when generating summaries, highlighting content that is closely related to the patient’s current health risks.

Citation Information

Patent Citations

  • Intelligent interpretation and treatment suggestion method and device based on inspection report

    CN117995334A

Cited By

  • Postpartum hemorrhage risk assessment system based on medical information processing

    CN119400433A

  • Workload quantification method and device for image report, and storage medium

    CN119539625A

  • Method, device and storage medium for quantifying workload of image report

    CN119539625B

  • Knowledge mining method and system for tumor field

    CN119673479A

  • Physical examination report interpretation tool based on AI general large model

    CN119864120A