An anti-tumor treatment adverse reaction risk warning method and system

Abstract

The present invention discloses an anti-tumor treatment adverse reaction risk early warning method and system, which relates to the technical field of medical information processing, and includes: obtaining physiological index data of a patient, and based on the temporal change characteristics of the physiological index data of a deep neural network, obtaining a short-term risk prediction result; obtaining clinical characteristic data of the patient, constructing a susceptibility assessment model, and analyzing the correlation characteristics of the clinical characteristic data through the susceptibility assessment model to obtain a baseline risk assessment result; calculating a comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result, and when the comprehensive risk score exceeds the early warning threshold corresponding to the adverse reaction type, triggering an early warning signal. By combining the immediate physiological index changes and the patient's clinical characteristics, the present invention constructs a two-level risk assessment system, solves the technical problems of insufficient consideration of patient individual differences and poor early warning timeliness in the existing early warning methods, and improves the accuracy and timeliness of anti-tumor treatment adverse reaction early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information processing, and specifically provides an anti-tumor treatment adverse reaction risk warning method and system. Background Art

[0002] In the field of anti-tumor treatment, the monitoring and risk warning of patients' adverse reactions have always been the core topics of clinical medicine. Traditional adverse reaction monitoring methods mainly rely on the threshold judgment of single physiological indicators and empirical evaluation. Although this method is simple to operate, it is difficult to comprehensively reflect the collaborative change characteristics of the human multi-organ system. In recent years, with the in-depth application of machine learning technology in the medical field, risk warning models based on multi-source data fusion have gradually become a research hotspot. These models integrate multi-dimensional information such as clinical feature data, laboratory test results, and treatment parameters to construct a risk assessment framework. However, existing risk warning models often use simple feature combinations or linear models for risk assessment, lacking in-depth analysis of the complex association patterns between clinical features and being difficult to accurately capture the interactive effects between different organ systems.

[0003] The current anti-tumor treatment adverse reaction risk warning system has the following technical problems: First, most existing monitoring methods use single-index threshold judgment, ignoring the collaborative change characteristics of multiple indicators within the same organ system, resulting in insufficient sensitivity and specificity of the warning; Second, existing risk assessment models do not analyze the relevance of clinical features deeply enough, failing to fully consider the interaction and temporal dependence relationships between different features, affecting the accuracy of risk assessment; Third, traditional warning mechanisms lack quantitative analysis of the risk propagation characteristics of multi-organ systems and cannot effectively predict potential chain reactions; Finally, there is generally a contradiction between response speed and accuracy in existing technologies, making it difficult to achieve rapid response while ensuring warning accuracy. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed.

[0005] Therefore, the present invention provides an anti-tumor treatment adverse reaction risk warning method and system, which can solve the problems mentioned in the background art.

[0006] To solve the above technical problems, the present invention provides the following technical solution: An anti-tumor treatment adverse reaction risk warning method, comprising: obtaining physiological index data of a patient, constructing an immediate risk assessment model, and analyzing the temporal change characteristics of the physiological index data through the immediate risk assessment model to obtain a short-term risk prediction result;

[0007] Obtain the clinical characteristic data of the patient, construct a susceptibility assessment model, analyze the correlation characteristics of the clinical characteristic data through the susceptibility assessment model, and obtain a baseline risk assessment result;

[0008] Calculate a comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result. When the comprehensive risk score exceeds the warning threshold corresponding to the adverse reaction type, trigger a warning signal.

[0009] As a preferred scheme of the method for warning of adverse reactions in anti-tumor treatment according to the present invention, wherein: the physiological index data includes physical sign monitoring data, organ function indexes, immune function indexes, hematopoietic function indexes, and metabolic level indexes;

[0010] Obtain the physiological index data of the patient, construct an immediate risk assessment model, analyze the time-series change characteristics of the physiological index data through the immediate risk assessment model, and obtain a short-term risk prediction result, including the following steps:

[0011] Group the physiological index data according to the organ system classification of the WHO adverse reaction grading standard, and calculate the co-variation degree of each group of indexes; if the dispersion degree of a certain group of indexes exceeds the mean level of this group of indexes, and there are two or more indexes in this group of indexes whose change ranges exceed the standard deviation range of their baseline values, then determine this group of indexes as the key monitoring category, otherwise determine this group of indexes as the routine monitoring category;

[0012] Input the physiological index data into the immediate risk assessment model. If there is an index group determined as the key monitoring category, the immediate risk assessment model preferentially analyzes the time-series correlation characteristics of each index under the determined key monitoring category and outputs the short-term risk prediction result. If all the physiological index data belong to the routine monitoring category, analyze the combined characteristics of all category indexes and then output the short-term risk prediction result.

[0013] As a preferred scheme of the method for warning of adverse reactions in anti-tumor treatment according to the present invention, wherein: the calculation of the co-variation degree includes the following steps:

[0014] Obtain the values of each index within the same organ system in three consecutive sampling periods, and calculate the standard deviation of each index;

[0015] Calculate the mean value of the standard deviations of all indexes, and divide the standard deviation mean value by the overall mean value of the indexes within this organ system to obtain the co-variation degree;

[0016] The time-series correlation characteristics are obtained by calculating the cross-correlation coefficient between indexes, wherein the synchrony is determined based on the cross-correlation coefficient at zero time delay, and the lag is determined based on the time delay corresponding to the maximum cross-correlation coefficient;

[0017] The combined features are obtained based on hierarchical clustering analysis among different organ system indicators, and a feature combination relationship matrix is constructed by calculating the Euclidean distance between the indicators;

[0018] The real-time risk assessment model adopts a dual-branch structure, and the dual-branch structure includes a time series analysis branch and a feature fusion branch; among them, the time series analysis branch adopts a multi-layer LSTM network structure, and each layer processes change patterns of different time scales; the feature fusion branch adopts an attention mechanism to adjust the weights of different organ system indicators.

[0019] As a preferred scheme of the method for warning of adverse reactions in anti-tumor treatment according to the present invention, wherein: the clinical feature data includes laboratory test data, imaging examination data, gene detection data, medication regimen parameters, and treatment cycle indicators;

[0020] Obtain the clinical feature data of the patient, construct a susceptibility assessment model, and analyze the correlation features of the clinical feature data through the susceptibility assessment model to obtain a baseline risk assessment result, including the following steps:

[0021] Construct a feature association network based on the clinical feature data, and calculate the association strength between clinical features in the feature association network; wherein, the association strength is the conditional mutual information of the clinical feature set; if the conditional mutual information of the clinical feature set is greater than the marginal mutual information of its constituent features, and the mutual information entropy between the organ system corresponding to the clinical feature set and the physiological index data is greater than the average mutual information entropy, then the clinical feature set is marked as a core feature set, otherwise it is marked as an ordinary feature set;

[0022] Input the feature association network into the susceptibility assessment model. If there is a clinical feature set marked as a core feature set, the susceptibility assessment model calculates the conditional probability distribution of the core feature set according to the Bayesian network and then outputs the baseline risk assessment result. If all the clinical feature sets are ordinary feature sets, the baseline risk assessment result is calculated based on the feature propagation algorithm of the graph neural network.

[0023] As a preferred scheme of the method for warning of adverse reactions in anti-tumor treatment according to the present invention, wherein: the feature association network is constructed through the following steps:

[0024] Group the clinical feature data according to the organ system classification of the WHO adverse reaction grading standard to obtain an initial set of clinical features;

[0025] Calculate the conditional mutual information and marginal mutual information within the initial set of clinical features to generate the nodes of the feature association network;

[0026] Determine the connection relationship based on the mutual information entropy between the nodes; wherein, if the clinical features corresponding to two nodes belong to the same organ system, the conditional mutual information between the nodes is used as the connection weight, and if the clinical features corresponding to two nodes belong to different organ systems, a connection is established when the mutual information entropy between the nodes exceeds the average mutual information entropy within their respective organ systems, and the excess value is used as the connection weight; the update of the feature association network is achieved by calculating the change rate of the connection weight and iteratively propagating between the nodes.

[0027] As a preferred solution of the anti-tumor treatment adverse reaction risk warning method of the present invention, wherein: calculate the comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result, including:

[0028] Establish a risk fusion matrix according to the organ system classification, and calculate the risk weights at different time scales based on the short-term risk prediction result and the baseline risk assessment result;

[0029] Apply the risk weights to the risk fusion matrix to calculate the system risk score; if the short-term risk prediction result and the baseline risk assessment result correspond to the same organ system, the diagonal term of the system risk score is calculated through the coupling coefficient, and if they correspond to different organ systems, the non-diagonal term of the system risk score is calculated through the risk transfer coefficient;

[0030] The system risk score is obtained after normalizing the risk fusion matrix by the eigenvalue decomposition method;

[0031] The comprehensive risk score is calculated according to the following steps:

[0032] Construct a risk feature vector, and organize the short-term risk prediction result and the baseline risk assessment result into a matrix form according to the organ system correspondence relationship;

[0033] Calculate the eigenvalues and eigenvectors of the risk feature vector, and select the eigenvector corresponding to the largest eigenvalue as the main risk direction;

[0034] Take the projection of the risk feature vector in the main risk direction as the benchmark value of the comprehensive risk score;

[0035] Correct the benchmark value of the comprehensive risk score based on the coupling coefficient and the risk transfer coefficient.

[0036] As a preferred solution of the anti-tumor treatment adverse reaction risk warning method of the present invention, wherein: the risk weight is adjusted by a time decay factor;

[0037] If there is a predicted value of the key monitoring category in the short-term risk prediction result, then use the reciprocal of the exponential function of the predicted value as the time decay factor to update the diagonal elements of the risk fusion matrix;

[0038] If there is an evaluation value of the core feature set in the baseline risk assessment result, then use the reciprocal of the exponential function of the evaluation value as the time decay factor to update the non-diagonal elements of the risk fusion matrix;

[0039] Among them, the calculation period of the time decay factor is inversely proportional to the occurrence period of the corresponding adverse reaction.

[0040] To further solve the above technical problems, the present invention provides the following technical solution: An anti-tumor treatment adverse reaction risk warning system, including: a collection module for obtaining the physiological index data and clinical feature data of a patient;

[0041] A processing module for constructing an immediate risk assessment model and a susceptibility assessment model, analyzing the time series change characteristics of the physiological index data through the immediate risk assessment model to obtain a short-term risk prediction result, and analyzing the correlation characteristics of the clinical feature data through the susceptibility assessment model to obtain a baseline risk assessment result;

[0042] A warning module for calculating a comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result, and triggering a warning signal when the comprehensive risk score exceeds the warning threshold of the corresponding adverse reaction type.

[0043] A computer device, including a memory and a processor, the memory stores a computer program, characterized in that when the processor executes the computer program, the steps of the anti-tumor treatment adverse reaction risk warning method described above are implemented.

[0044] A computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the anti-tumor treatment adverse reaction risk warning method described above are implemented.

[0045] Advantages of the present invention: First, by introducing the concept of co-variation degree and the dual-channel analysis mechanism, the present invention overcomes the limitation that traditional single-index monitoring methods are difficult to reflect the overall functional state of organ systems, realizes the accurate identification of the co-variation of multiple indicators, and improves the response speed to high-risk situations. Second, the present invention constructs a feature association network based on organ system classification, and through differential connection strategies and double-layer architecture design, realizes the intelligent identification of complex association patterns between clinical features, and solves the problem of insufficient feature interaction analysis in existing risk warning models. Finally, the present invention adopts a risk scoring method based on time decay, and by introducing a coupling coefficient and a risk transmission coefficient, establishes a quantitative evaluation mechanism for risk propagation between multiple organ systems, breaking through the technical bottleneck that traditional warning systems are difficult to predict chain reactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0047] Figure 1 It is a schematic diagram of the overall process of a method for warning against adverse reactions to anti-tumor treatment proposed by the present invention;

[0048] Figure 2 It is a diagram of a computer device in a method for warning against adverse reactions to anti-tumor treatment proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings of the specification. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0050] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0051] Example 1, referring to Figure 1 This is an embodiment of the present invention, which provides a method for warning against adverse reactions to anti-tumor treatment.

[0052] Figure 1The figure shows a schematic diagram of the overall process of an anti-tumor treatment adverse reaction risk warning method, including the following steps:

[0053] S1: Obtain the physiological index data of the patient, construct an immediate risk assessment model, analyze the temporal variation characteristics of the physiological index data through the immediate risk assessment model, and obtain the short-term risk prediction result.

[0054] Specifically, the physiological index data includes physical sign monitoring data, organ function indicators, immune function indicators, hematopoietic function indicators, and metabolic level indicators.

[0055] S1.1: Group the physiological index data according to the organ system classification of the WHO adverse reaction grading standard, and calculate the co-variation degree of each group of indicators. Among them, the co-variation degree characterizes the change consistency of each indicator within the same organ system; if the dispersion degree of a certain group of indicators exceeds the mean level of this group of indicators, and there are two or more indicators in this group whose change ranges exceed the standard deviation range of their baseline values, then this group of indicators is determined as the key monitoring category, otherwise this group of indicators is determined as the routine monitoring category.

[0056] S1.2: Input the physiological index data into the immediate risk assessment model. If there is an indicator group determined as the key monitoring category, the immediate risk assessment model preferentially analyzes the temporal correlation characteristics of each indicator under this category and outputs the short-term risk prediction result. If all the physiological index data belongs to the routine monitoring category, the short-term risk prediction result is output after analyzing the combined characteristics of all category indicators.

[0057] Specifically, the calculation of the co-variation degree includes the following steps: First, obtain the values of each indicator within the same organ system in three consecutive sampling periods, calculate the standard deviation of each indicator, then calculate the mean of the standard deviations of all indicators, and finally divide the mean of the standard deviations by the overall mean of the organ system indicators to obtain the co-variation degree.

[0058] The time series correlation characteristics are obtained by calculating the mutual correlation coefficients between indicators, where synchronization is determined based on the mutual correlation coefficient of zero time delay, and hysteresis is determined based on the time delay corresponding to the maximum mutual correlation coefficient. The calculation process of the mutual correlation coefficient involves two time series indicator data. Taking liver function monitoring as an example, when analyzing the time series correlation characteristics of alanine aminotransferase ALT and aspartate aminotransferase AST: the synchronization calculation process is to first obtain the sequence data of the two indicators at the same time point, standardize the data at each time point, and eliminate the dimension effect. Then calculate the mutual correlation coefficient of zero time delay (that is, without considering time offset), specifically, multiply and sum the values ​​of the two standardized sequences at the corresponding time points, and then divide them by the product of their respective standard deviations. The mutual correlation coefficient reflects the degree of consistency of the changes of two indicators at the same time point. For example, if the zero-delay mutual correlation coefficient of ALT and AST is close to 1, it indicates that the two indicators rise or fall synchronously, which often indicates the occurrence of acute liver injury. The hysteresis calculation process is to gradually move one of the indicator sequences on the time axis (it can be moved forward or backward), calculate the correlation coefficient at each offset position, and find the time offset corresponding to the maximum correlation coefficient. This offset reflects the hysteresis relationship between the two indicators. For example, if it is found that the change in bilirubin level reaches its peak 24 hours later than the transaminase level, this time delay feature can be used to predict the progression of liver damage. This has important guiding significance for the timely adjustment of clinical intervention measures.

[0059] The combined features are obtained based on the hierarchical cluster analysis between the indicators of different organ systems, and the feature combination relationship matrix is ​​constructed by calculating the Euclidean distance between the indicators. The above mechanism adopts differentiated analysis strategies for different risk levels, which ensures the accuracy of early warning while improving the response speed to high-risk situations.

[0060] Furthermore, the real-time risk assessment model is a risk prediction model based on deep learning, specifically designed for the characteristics of adverse reactions to anti-tumor treatment. The model adopts a dual-branch structure: a time series analysis branch and a feature fusion branch. The time series analysis branch uses a multi-layer LSTM network structure, with each layer processing change patterns at different time scales. Taking liver function indicators as an example, the first layer of LSTM captures acute changes at the hourly level (such as a rapid increase in transaminases), the second layer captures cumulative effects at the daily level (such as a progressive change in bilirubin), and the third layer focuses on long-term trends at the weekly level (such as a slow decline in albumin levels). This multi-scale time series analysis can simultaneously monitor acute adverse reactions and chronic cumulative damage. The feature fusion branch uses an attention mechanism to dynamically adjust the weights of indicators in different organ systems. For example, when abnormal cardiac function indicators are detected, the model will automatically increase the attention to renal function indicators because there is often a close relationship between cardiac and renal functions. This dynamic weight adjustment between features enables the model to capture the interactive effects between organ systems. The output layer of the model is designed with a hierarchical prediction mechanism: for key monitoring categories, the probability prediction values of each specific adverse reaction type are directly output; for routine monitoring categories, a comprehensive risk score is output. This hierarchical output mechanism not only ensures a timely response to high-risk situations but also maintains a continuous assessment of the overall risk.

[0061] In S1.1, the present invention selects the WHO adverse reaction grading standard as the basis for index grouping. Different from traditional single-index monitoring, it reflects the in-depth understanding of clinical practice by the present invention. Through organ system classification, the correlation between different physiological indicators is established, making the monitoring more in line with the overall physiological characteristics of the human body. The introduced concept of co-variation degree, by calculating the ratio of the standard deviation to the mean, not only avoids the problem of direct comparison of indicators with different dimensions but also can reflect the overall change trend of indicators within the same organ system, which is of great significance for early identification of organ function disorders. Especially in anti-tumor treatment, drug adverse reactions often first manifest as co-variations of multiple indicators within a certain organ system rather than significant abnormalities of a single indicator.

[0062] The dual-channel analysis mechanism in S1.2 solves the contradiction between response speed and accuracy in traditional warning methods. For key monitoring categories, it preferentially analyzes their time series correlation features, calculates the synchrony and lag between indicators through cross-correlation coefficients, and can timely capture early signs of adverse reactions. This analysis method is particularly applicable to the field of anti-tumor treatment because some severe adverse reactions (such as myocardial damage, liver function damage, etc.) often manifest as a chain reaction of multiple related indicators. For routine monitoring categories, hierarchical clustering analysis is used to construct a feature combination relationship matrix, and the Euclidean distance is used to measure the correlation degree between indicators in different organ systems, which not only ensures the comprehensiveness of the warning but also avoids waste of computing resources.

[0063] From the perspective of solving technical problems, the present invention solves the problem that traditional single-index monitoring cannot reflect the overall functional state of organ systems through a calculation method of co-variation degree. The dual-channel analysis mechanism overcomes the defect of insufficient recognition of multi-index co-variation in existing early warning systems. Especially in clinical practice, this grouping monitoring method based on organ systems can effectively reduce false positive early warnings and improve the accuracy of early warnings. For example, in the monitoring of liver function damage caused by chemotherapy drugs, traditional methods often rely on the threshold judgment of a single index (such as transaminase level), while the present invention can detect potential liver damage risks earlier by analyzing the co-variation of liver function-related indicators.

[0064] S2: Obtain the clinical characteristic data of the patient, construct a susceptibility assessment model, analyze the correlation characteristics of the clinical characteristic data through the susceptibility assessment model, and obtain the baseline risk assessment result.

[0065] Specifically, the clinical characteristic data includes laboratory test data, imaging examination data, gene detection data, medication regimen parameters, and treatment cycle indicators.

[0066] S2.1: Construct a feature association network based on the clinical characteristic data, and calculate the association strength between clinical characteristics in the feature association network.

[0067] Among them, the association strength is the conditional mutual information of the clinical characteristic set; if the conditional mutual information of the clinical characteristic set is greater than the marginal mutual information of its constituent features, and the mutual information entropy between the organ system corresponding to the clinical characteristic set and the physiological index data is greater than the mean mutual information entropy, then the clinical characteristic set is marked as the core characteristic set, otherwise it is marked as the ordinary characteristic set.

[0068] S2.2: Input the feature association network into the susceptibility assessment model. If there is a clinical characteristic set marked as the core characteristic set, the susceptibility assessment model outputs the baseline risk assessment result after calculating the conditional probability distribution of the core characteristic set according to the Bayesian network. If all clinical characteristic sets are ordinary characteristic sets, the baseline risk assessment result is calculated based on the feature propagation algorithm of the graph neural network.

[0069] Specifically, the susceptibility assessment model is a dual-channel risk analysis model based on graph neural network and Bayesian inference, and is designed specifically for the individual characteristics of adverse reactions to anti-tumor treatment. The model adopts a two-layer architecture: a feature representation layer and a risk inference layer.

[0070] The feature representation layer consists of two parallel branches:

[0071] 1. Medical knowledge encoding branch: Map clinical feature data into a pre-defined medical knowledge graph. Each feature node contains three types of attribute information: basic feature value, temporal change trend, and association rule strength. The basic feature value reflects the current state of the indicator, the temporal change trend captures the dynamic features of the indicator, and the association rule strength represents the interaction relationship with other features.

[0072] 2. Feature association analysis branch: Analyze the interaction patterns between features based on the feature association network. For the core feature set, use the Bayesian network to calculate the conditional probability distribution and establish the causal relationship chain between features; for the general feature set, extract the combined representation of features through the message passing mechanism of the graph neural network.

[0073] The risk inference layer fuses the outputs of the two branches to achieve accurate assessment of the risk level:

[0074] 1. For the core feature set, calculate the posterior probabilities of different adverse reactions through Bayesian inference and verify and interpret them based on the medical knowledge graph.

[0075] 2. For the general feature set, adopt the attention mechanism of the graph neural network to dynamically adjust the weights according to the importance of feature nodes in the knowledge graph.

[0076] Specifically, the feature association network is constructed through the following steps: Group the clinical feature data according to the organ system classification of the WHO adverse reaction grading standard to obtain the initial set of clinical features, calculate the conditional mutual information and marginal mutual information within the initial set of clinical features to generate the nodes of the feature association network; Determine the connection relationship based on the mutual information entropy between nodes. Among them, if the clinical features corresponding to two nodes belong to the same organ system, use the conditional mutual information between the nodes as the connection weight. If the clinical features corresponding to two nodes belong to different organ systems, establish a connection when the mutual information entropy between the nodes exceeds the average mutual information entropy within their respective organ systems, and use the exceeded value as the connection weight; The update of the feature association network is achieved by calculating the change rate of the connection weight and iteratively propagating between nodes.

[0077] The conditional mutual information is obtained by calculating the probability dependence relationship between the clinical feature set and the adverse reaction label, and the marginal mutual information is the probability dependence relationship between a single feature in the clinical feature set and the adverse reaction label; the conditional probability distribution in the Bayesian network is obtained by the maximum likelihood estimation method, and the feature propagation algorithm of the graph neural network uses the message passing mechanism to update the node features of the feature association network; the mutual information entropy is calculated based on the co-variation degree between the clinical feature set and the physiological index data. This analysis method based on information theory and probabilistic graphical models constructs a quantitative relationship between clinical feature data and physiological index data. For example, when evaluating the myelosuppression risk of chemotherapy drugs, the mutual information entropy between gene detection data and hematopoietic function indicators reflects the susceptibility of patients to myelosuppression; when predicting the cardiotoxicity of targeted therapy, the topological relationship between functional indicators and imaging features is captured through the feature propagation algorithm, and combined with the co-variation degree of cardiac function indicators, more accurate risk assessment is provided.

[0078] The present invention solves the technical problem of insufficient analysis of clinical feature relevance in the prior art risk warning model. The prior art usually uses simple feature combinations or linear models for risk assessment, and it is difficult to capture complex feature interaction patterns. While the present invention enables the model to adaptively establish the association strength between features by constructing a feature association network and designing a differentiated connection strategy based on organ system classification. For example, when evaluating chemotherapy-related cardiotoxicity, traditional methods often consider electrocardiogram changes or myocardial enzyme spectrum changes separately, while ignoring the interaction effects between these indicators and drug dosage, gene polymorphisms. The feature association network of the present invention establishes a complete set of feature association evaluation systems by calculating the conditional mutual information of features within the same organ system and the mutual information entropy of cross-system features. This mechanism shows unique advantages in practical applications: when certain feature combinations are detected to have strong predictive ability, the model will automatically increase the weights of relevant features, thereby improving the accuracy of early warning. Especially when predicting chronic cumulative toxicity, the present invention can capture the potential laws of the evolution of clinical features over time through the feature association network, which is difficult to achieve simply relying on feature combinations or statistical analysis. In addition, the differentiated analysis strategy of the present invention (using a Bayesian network for the core feature set and a graph neural network for the ordinary feature set) achieves a good balance between computational efficiency and early warning timeliness, which is of great significance in clinical practice that requires rapid response.

[0079] The present invention realizes the intelligent recognition of complex association patterns between clinical features through the differentiated connection strategy of the feature association network and the hierarchical analysis mechanism based on organ systems, not only improving the accuracy of risk assessment, but also providing an interpretable basis for risk assessment for doctors. This method overcomes the limitations of fragmented feature analysis and insufficient relevance analysis in the prior art and plays an important role in the individualized risk warning of anti-tumor treatment.

[0080] S3: Calculate a comprehensive risk score based on the short-term risk prediction result and the baseline risk assessment result. When the comprehensive risk score exceeds the warning threshold corresponding to the adverse reaction type, trigger a warning signal.

[0081] S3.1: Establish a risk fusion matrix according to the organ system classification, and calculate the risk weights at different time scales based on the short-term risk prediction result and the baseline risk assessment result.

[0082] The risk weights are adjusted by a time decay factor; if there is a predicted value of a key monitoring category in the short-term risk prediction result, the reciprocal of the exponential function of the predicted value is used as the time decay factor to update the diagonal elements of the risk fusion matrix. If there is an evaluation value of the core feature set in the baseline risk assessment result, the reciprocal of the exponential function of the evaluation value is used as the time decay factor to update the non-diagonal elements of the risk fusion matrix. Among them, the calculation period of the time decay factor is inversely proportional to the occurrence period of the corresponding adverse reaction.

[0083] S3.2: Apply the risk weights to the risk fusion matrix to calculate the system risk score; if the short-term risk prediction result and the baseline risk assessment result correspond to the same organ system, the diagonal term of the system risk score is calculated through a coupling coefficient. If they correspond to different organ systems, the non-diagonal term of the system risk score is calculated through a risk transfer coefficient.

[0084] Among them, the system risk score is obtained by normalizing the risk fusion matrix through the eigenvalue decomposition method.

[0085] Specifically, the comprehensive risk score is calculated according to the following steps:

[0086] 1. Construct a risk feature vector, and organize the short-term risk prediction result and the baseline risk assessment result into a matrix form according to the organ system correspondence.

[0087] 2. Calculate the eigenvalues and eigenvectors of the risk feature vector, and select the eigenvector corresponding to the largest eigenvalue as the main risk direction.

[0088] 3. Take the projection of the risk feature vector in the main risk direction as the reference value of the comprehensive risk score.

[0089] 4. Correct the reference value of the comprehensive risk score based on the coupling coefficient and the risk transfer coefficient.

[0090] The coupling coefficient reflects the degree of consistency between short-term risk and baseline risk within the same organ system, and is obtained by calculating the mutual information entropy of the risk prediction value and the evaluation value; the risk transmission coefficient characterizes the risk transmission characteristics between different organ systems, and is obtained by calculating the conditional probability after constructing a Bayesian network based on the temporal association pattern of organ function damage in historical case data. For example, when monitoring chemotherapy-related cardio-renal toxicity, if abnormal cardiac function indicators are detected, the model will predict the probability of renal function damage through the risk transmission coefficient, so as to achieve joint early warning of multiple organ systems. This risk scoring method based on feature decomposition not only ensures the interpretability of the scoring results, but also realizes the accurate quantification of the risks of complex systems.

[0091] Further, S3.3: Establish an early warning threshold adaptive system based on historical clinical data, including the following steps:

[0092] 1. Construct an early warning threshold matrix according to the organ system classification of the WHO adverse reaction grading standard. For the key monitoring categories, determine the initial threshold based on the diagonal elements of the risk fusion matrix and the time decay factor; for the routine monitoring categories, determine the initial threshold based on the ROC curve analysis.

[0093] 2. For each type of adverse reaction, calculate the benchmark parameter for threshold adjustment by combining the historical occurrence frequency, the weighted score of the severity, and the coupling coefficient of the corresponding organ system; for the adverse reaction types with a core feature set, further correct the threshold through the risk transmission coefficient.

[0094] 3. Periodically update the threshold matrix based on the newly added clinical data. The update rule is as follows: if the false negative rate of a certain type of adverse reaction in the most recent monitoring period exceeds the set value, reduce the early warning threshold of this type of adverse reaction through the time decay factor; if the false positive rate of a certain type of adverse reaction in the most recent monitoring period exceeds the set value, correspondingly increase the early warning threshold of this type of adverse reaction. The threshold update period is inversely proportional to the occurrence period of the corresponding adverse reaction.

[0095] S3.4: Design a hierarchical early warning signal triggering mechanism, specifically including:

[0096] 1. When the amplitude of the comprehensive risk score exceeding the early warning threshold of the corresponding adverse reaction type is different, trigger early warning signals of different levels:

[0097] For the key monitoring categories, when the diagonal term of the system risk score exceeds the early warning threshold but does not reach 1.2 times, trigger a first-level early warning;

[0098] When it exceeds 1.2 times but does not reach 1.5 times of the early warning threshold, or when the non-diagonal term of the risk score of the organ system with a risk transmission relationship exceeds the early warning threshold, trigger a second-level early warning;

[0099] When it exceeds 1.5 times the warning threshold, a third-level warning is triggered;

[0100] 2. The information content included in the warning signals of different levels increases sequentially:

[0101] The first-level warning includes risk scores, abnormal index analysis, and co-variation degree data;

[0102] The second-level warning adds the analysis results of the core feature set, risk transmission prediction, and preliminary intervention suggestions;

[0103] The third-level warning further supplements the analysis results of the feature association network, detailed medical intervention plans, and multi-department collaborative disposal suggestions.

[0104] 3. The warning signals are sent in a multi-modal manner: system pop-up prompts, push notifications to the mobile terminals of medical staff, audible and visual alerts on the diagnosis and treatment workstations, etc., to ensure that medical staff can receive warning information in a timely manner.

[0105] In summary, the present invention has the following advantages: First, by introducing the concept of co-variation degree and the dual-channel analysis mechanism, the present invention overcomes the limitation that traditional single-index monitoring methods are difficult to reflect the overall functional state of organ systems, realizes the accurate identification of the co-variation of multiple indicators, and improves the response speed to high-risk situations. Second, the present invention constructs a feature association network based on organ system classification, and through differential connection strategies and double-layer architecture design, realizes the intelligent identification of complex association patterns between clinical features, and solves the problem of insufficient feature interaction analysis in existing risk warning models. Finally, the present invention adopts a risk scoring method based on time decay, and by introducing a coupling coefficient and a risk transmission coefficient, establishes a quantitative evaluation mechanism for risk propagation between multiple organ systems, breaking through the technical bottleneck that traditional warning systems are difficult to predict chain reactions.

[0106] Embodiment 2 is an embodiment of the present invention, which provides an anti-tumor treatment adverse reaction risk warning system, including:

[0107] An acquisition module for obtaining the physiological index data and clinical feature data of the patient;

[0108] A processing module for constructing an immediate risk assessment model and a susceptibility assessment model, analyzing the time-series change characteristics of the physiological index data through the immediate risk assessment model to obtain short-term risk prediction results, and analyzing the correlation characteristics of the clinical feature data through the susceptibility assessment model to obtain baseline risk assessment results;

[0109] A warning module for calculating a comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result, and triggering a warning signal when the comprehensive risk score exceeds the warning threshold corresponding to the adverse reaction type.

[0110] Embodiment 3, refer toFigure 2 This is an embodiment of the present invention. What is different from the previous embodiment is that when the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., all kinds of media that can store program codes.

[0111] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0112] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), optical fiber devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as necessary, and then storing it in a computer memory.

[0113] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0114] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for warning of the risk of adverse reactions to anti-tumor treatment, characterized in that, Including: Obtain the physiological index data of the patient, construct an immediate risk assessment model, analyze the temporal variation characteristics of the physiological index data through the immediate risk assessment model, and obtain a short-term risk prediction result; Obtain the clinical characteristic data of the patient, construct a susceptibility assessment model, analyze the correlation characteristics of the clinical characteristic data through the susceptibility assessment model, and obtain a baseline risk assessment result; Calculate a comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result. When the comprehensive risk score exceeds the warning threshold corresponding to the adverse reaction type, trigger a warning signal; The physiological index data includes physical sign monitoring data, organ function indicators, immune function indicators, hematopoietic function indicators, and metabolic level indicators; Obtain the physiological index data of the patient, construct an immediate risk assessment model, analyze the temporal variation characteristics of the physiological index data through the immediate risk assessment model, and obtain a short-term risk prediction result, including the following steps: Group the physiological index data according to the organ system classification of the WHO adverse reaction grading standard, and calculate the co-variation degree of each group of indicators; if the dispersion degree of a certain group of indicators exceeds the mean level of this group of indicators, and there are two or more indicators in this group of indicators whose change ranges exceed the standard deviation range of their baseline values, then determine this group of indicators as the key monitoring category, otherwise determine this group of indicators as the regular monitoring category; Input the physiological index data into the immediate risk assessment model. If there is an indicator group determined to be the key monitoring category, the immediate risk assessment model preferentially analyzes the temporal correlation characteristics of each indicator under the key monitoring category and outputs the short-term risk prediction result. If all the physiological index data belong to the regular monitoring category, analyze the combined characteristics of all category indicators and then output the short-term risk prediction result; The clinical characteristic data includes laboratory test data, imaging examination data, gene detection data, medication regimen parameters, and treatment cycle indicators; Obtain the clinical characteristic data of the patient, construct a susceptibility assessment model, analyze the correlation characteristics of the clinical characteristic data through the susceptibility assessment model, and obtain a baseline risk assessment result, including the following steps: Construct a feature association network based on the clinical characteristic data, and calculate the association strength between clinical characteristics in the feature association network; where the association strength is the conditional mutual information of the clinical characteristic set; if the conditional mutual information of the clinical characteristic set is greater than the marginal mutual information of its constituent features, and the mutual information entropy between the organ system corresponding to the clinical characteristic set and the physiological index data is greater than the mean mutual information entropy, then mark the clinical characteristic set as the core feature set, otherwise mark it as the ordinary feature set; Input the feature association network into the susceptibility assessment model. If there is a clinical characteristic set marked as the core feature set, the susceptibility assessment model calculates the conditional probability distribution of the core feature set according to the Bayesian network and then outputs the baseline risk assessment result. If all the clinical characteristic sets are ordinary feature sets, calculate the baseline risk assessment result based on the feature propagation algorithm of the graph neural network.

2. The anti-tumor treatment adverse reaction risk warning method according to claim 1, wherein: The calculation of the co-variation degree includes the following steps: Obtain the values of each index within the same organ system in three consecutive sampling periods, and calculate the standard deviation of each index; Calculate the mean of the standard deviations of all indices, and divide the mean of the standard deviations by the overall mean of the indices within the organ system to obtain the co-variation degree; The time-series correlation features are obtained by calculating the cross-correlation coefficients between indices, where the synchrony is determined based on the cross-correlation coefficient at zero time delay, and the lag is determined based on the time delay corresponding to the maximum cross-correlation coefficient; The combined features are obtained based on hierarchical clustering analysis between indices of different organ systems, and a feature combination relationship matrix is constructed by calculating the Euclidean distance between indices; The immediate risk assessment model adopts a dual-branch structure, and the dual-branch structure includes a time-series analysis branch and a feature fusion branch; among them, the time-series analysis branch adopts a multi-layer LSTM network structure, and each layer processes the change patterns at different time scales; the feature fusion branch adopts an attention mechanism to adjust the weights of indices of different organ systems.

3. The method for warning of the risk of adverse reactions in anti-tumor treatment according to claim 2, characterized in that: The feature association network is constructed through the following steps: Group the clinical feature data according to the organ system classification of the WHO adverse reaction grading standard to obtain an initial set of clinical features; Calculate the conditional mutual information and marginal mutual information within the initial set of clinical features to generate the nodes of the feature association network; Determine the connection relationship based on the mutual information entropy between the nodes; among them, if the clinical features corresponding to two nodes belong to the same organ system, the conditional mutual information between the nodes is used as the connection weight, and if the clinical features corresponding to two nodes belong to different organ systems, a connection is established when the mutual information entropy between the nodes exceeds the average mutual information entropy within their respective organ systems, and the excess value is used as the connection weight; the update of the feature association network is achieved by calculating the change rate of the connection weight and iteratively propagating between the nodes.

4. The anti-tumor treatment adverse reaction risk warning method according to claim 3, wherein: Calculate the comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result, including: Establish a risk fusion matrix according to the organ system classification, and calculate the risk weights at different time scales based on the short-term risk prediction result and the baseline risk assessment result; Apply the risk weights to the risk fusion matrix to calculate the system risk score; if the short-term risk prediction result and the baseline risk assessment result correspond to the same organ system, the diagonal term of the system risk score is calculated through a coupling coefficient, and if they correspond to different organ systems, the off-diagonal term of the system risk score is calculated through a risk transfer coefficient; The system risk score is obtained by normalizing the risk fusion matrix through eigenvalue decomposition; The comprehensive risk score is calculated according to the following steps: Construct a risk feature vector, and organize the short-term risk prediction result and the baseline risk assessment result into a matrix form according to the organ system correspondence; Calculate the eigenvalues and eigenvectors of the risk feature vector, and select the eigenvector corresponding to the largest eigenvalue as the main risk direction; Take the projection of the risk feature vector in the main risk direction as the benchmark value of the comprehensive risk score; Modify the benchmark value of the comprehensive risk score based on the coupling coefficient and the risk transfer coefficient.

5. The anti-tumor treatment adverse reaction risk warning method according to claim 4, wherein: The risk weight is adjusted by a time decay factor; If there is a predicted value of a key monitoring category in the short-term risk prediction result, use the reciprocal of the exponential function of the predicted value as the time decay factor to update the diagonal elements of the risk fusion matrix; If there is an evaluation value of the core feature set in the baseline risk assessment result, use the reciprocal of the exponential function of the evaluation value as the time decay factor to update the non-diagonal elements of the risk fusion matrix; Wherein, the calculation period of the time decay factor is inversely proportional to the occurrence period of the corresponding adverse reaction.

6. An anti-tumor treatment adverse reaction risk warning system, based on the anti-tumor treatment adverse reaction risk warning method according to any one of claims 1 to 5, characterized in that: Including, A collection module for obtaining the physiological index data and clinical feature data of the patient; A processing module for constructing an immediate risk assessment model and a susceptibility assessment model, analyzing the time series change characteristics of the physiological index data through the immediate risk assessment model to obtain a short-term risk prediction result, and analyzing the correlation characteristics of the clinical feature data through the susceptibility assessment model to obtain a baseline risk assessment result; An early warning module for calculating a comprehensive risk score according to the short-term risk prediction result and the baseline risk assessment result, and triggering an early warning signal when the comprehensive risk score exceeds the early warning threshold of the corresponding adverse reaction type.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the anti-tumor treatment adverse reaction risk early warning method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the anti-tumor treatment adverse reaction risk early warning method according to any one of claims 1 to 5.

Images