Intelligent management platform and method for the entire course of tumor disease

Abstract

The present invention discloses an intelligent management platform and method for the entire course of tumor disease, and specifically relates to the technical field of medical model management, which collects multimodal data of patients; uses natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources; uses a feature selection algorithm to screen key features related to tumor type and treatment plan; evaluates the degree of redundancy hidden danger of the screened key features by analyzing the mutual information between the key features and the target variables and the mutual information between the key features; evaluates the semantic consistency of multimodal data from different sources by comparing and analyzing the predicted labels and actual labels of the multimodal data; and judges whether it is possible to construct a personalized treatment effect prediction model by comprehensively analyzing the degree of redundancy hidden danger of the screened key features and the semantic consistency of multimodal data from different sources.

Description

Technical Field

[0001] The present invention relates to the field of medical model management technology, and more specifically, to a platform and method for intelligent management of the entire course of tumor disease. Background Art

[0002] In modern medical practice, the concept of personalized treatment has been increasingly valued, especially in the field of tumor treatment. Traditional tumor treatment methods often adopt a relatively standardized treatment plan, which does not fully consider the biological differences between individual patients, such as differences in gene expression, biomarkers, and individual responses to drugs, which often leads to uncertainty in treatment effects and fluctuations in efficacy. Therefore, there is currently a personalized treatment effect prediction model to achieve full-course management of tumors. The personalized treatment effect prediction model is a model specifically designed to predict each patient's response to a specific treatment method. This model uses the patient's personal medical history, genetic information, biomarkers, clinical parameters, and other relevant data that may be included (such as lifestyle, environmental factors, etc.), through complex data analysis and machine learning methods, to predict the patient's possible response to different treatment options. The purpose of this type of model is to optimize treatment decisions and achieve more efficient and accurate medical services.

[0003] However, before building a personalized treatment effect prediction model, if the accuracy and reliability of the personalized treatment effect prediction model are insufficient, it may cause the personalized treatment effect prediction model to learn and predict based on erroneous or inconsistent information, thereby affecting the personalized customization of the tumor treatment plan and the ultimate treatment effect.

[0004] In order to solve the above problems, an intelligent management solution for the entire course of tumor is now provided. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a platform and method for intelligent management of the entire course of tumor disease.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The intelligent management method for the whole course of tumor disease includes the following steps:

[0008] S1: Collect multimodal data of patients, including genomic information, imaging data, electronic medical records, and laboratory test results;

[0009] S2: Use natural language processing technology and standardized tools to semantically parse and unify the format of multimodal data from different sources;

[0010] S3: Use feature selection algorithms to screen key features associated with tumor type and treatment regimen;

[0011] S4: By analyzing the mutual information between key features and target variables and the mutual information between key features, the degree of redundancy risk of the selected key features is evaluated; by comparing and analyzing the predicted labels and actual labels of multimodal data, the semantic consistency of multimodal data from different sources is evaluated;

[0012] S5: Through comprehensive analysis of the potential redundancy of key features and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed.

[0013] In a preferred embodiment, the multimodal data of the patient is collected, specifically:

[0014] S101: Identify and define the types and sources of multimodal data that need to be collected;

[0015] S102: Develop standardized collection protocols to ensure data integrity and consistency;

[0016] S103: Acquire and store the patient's genomic data, collect the patient's medical images, extract the patient's historical medical history and clinical records, and collect and store laboratory test results including the patient's blood and tissue samples.

[0017] In a preferred embodiment, natural language processing technology and standardized tools are used to perform semantic parsing and unified format processing on multimodal data from different sources, specifically:

[0018] S201: Clean and perform preliminary word segmentation on text data such as electronic medical records;

[0019] S202: Use natural language processing technology to extract key medical terms and perform semantic annotation;

[0020] S203: Unify the formats of data from different sources and map the information in different data sources into a unified data model.

[0021] In a preferred embodiment, a feature selection algorithm is used to screen key features related to tumor type and treatment regimen, specifically:

[0022] S301: Preliminary selection of potential features based on domain knowledge and data relevance;

[0023] S302: Use statistical methods to evaluate the importance of features:

[0024] For categorical variables, analysis of variance was used to assess the ability of different features to explain differences in tumor type or treatment regimen;

[0025] For categorical data, the chi-square test was used to assess the independence between features and the target variable;

[0026] For binary classification problems, a logistic regression model is used to evaluate the regression coefficient and significance level of the features;

[0027] Dealing with multicollinearity issues and reducing unimportant features through regularization methods;

[0028] S303: Apply feature selection algorithm to screen key features:

[0029] Recursive feature elimination repeatedly builds models and evaluates the importance of features, gradually eliminating the least important features and finally retaining the most important feature set;

[0030] Apply the RFE algorithm to recursively analyze the initially selected features, repeatedly train the model, and gradually reduce the number of features until the preset number of features or the optimal state of model performance is reached;

[0031] Use the random forest algorithm to evaluate the importance of features and select the features that contribute most to the model's predictive ability based on the importance scores of the features in the decision tree splitting process;

[0032] By improving the importance scores of features in the tree model, the feature set can be further screened and optimized;

[0033] S304: Verify the effectiveness of the selected features through cross-validation and model performance evaluation;

[0034] S305: Adjust the feature set according to the verification results to optimize the final feature selection.

[0035] In a preferred embodiment, by analyzing the mutual information between the key features and the target variable and the mutual information between the key features, the degree of redundancy risk of the selected key features is evaluated, specifically:

[0036] Use mutual information to quantify the amount of information shared between each feature and the target variable; calculate the feature and The mutual information between them is expressed as: ;

[0037] Calculate the feature redundancy hidden danger assessment index, the expression is: ;in, is the feature redundancy hazard assessment index, is the size of the feature set, It is a feature set Medium Features With the target variable The mutual information of It is a feature set Medium Features and The mutual information between and are the actual values ​​of the two features in the dataset, and Represents two different feature columns in the dataset, and Respectively represent characteristics and the target variable The specific value of Usually represents a feature set, represents the target variable, and Are used as indexes.

[0038] In a preferred embodiment, the semantic consistency of multimodal data from different sources is evaluated by comparing and analyzing the predicted labels and actual labels of the multimodal data, specifically:

[0039] Extract features from data of various modalities; including text data, image data, and medical image annotation;

[0040] Use the extracted features to build a CRF model; define the label sequence as the semantic label of the data point; define the feature sequence as the extracted feature vector; the basic formula for using the CRF model is: ;in, is the conditional probability expression in the conditional random field model, Usually used to represent an index into a state or time step, and is the weight parameter in the CRF model, and is the characteristic function, is the normalization factor; the semantic consistency index is calculated, and its expression is: ;in, is the semantic consistency index, is the total number of test samples, It is with The weight of the mode corresponding to the data point, is the indicator function, It is The predicted labels for data points; is the actual label.

[0041] In a preferred embodiment, by comprehensively analyzing the potential risk of redundancy in the key features of the screening and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed, specifically:

[0042] The feature redundancy hidden danger assessment index and the semantic consistency index are normalized, and after the normalized feature redundancy hidden danger assessment index and the semantic consistency index are respectively assigned preset proportional coefficients, the model establishment judgment coefficient is calculated to determine whether the personalized treatment effect prediction model can be constructed;

[0043] Set the model building judgment threshold and compare the model building judgment coefficient with the model building judgment threshold:

[0044] When the model establishment judgment coefficient is greater than the model establishment judgment threshold, it is determined that the personalized treatment effect prediction model can be constructed;

[0045] When the model establishment judgment coefficient is less than or equal to the model establishment judgment threshold, it is determined that the construction of the personalized treatment effect prediction model cannot be carried out.

[0046] On the other hand, the present invention provides a tumor full-course intelligent management platform, including a modality data collection module, a modality data processing module, a key feature screening module, a redundant hidden danger assessment module, a semantic consistency assessment module, and a model construction judgment module;

[0047] Modal data collection module: collects multimodal data of patients, including genomic information, imaging data, electronic medical records and laboratory test results;

[0048] Modal data processing module: uses natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources;

[0049] Key feature screening module: Use feature selection algorithm to screen key features related to tumor type and treatment regimen;

[0050] Redundancy risk assessment module: By analyzing the mutual information between key features and target variables and the mutual information between key features, the redundancy risk level of the selected key features is assessed;

[0051] Semantic consistency assessment module: evaluates the semantic consistency of multimodal data from different sources by comparing and analyzing the predicted labels of multimodal data with the actual labels;

[0052] Model construction judgment module: Through comprehensive analysis of the degree of redundancy risk of key screened features and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed.

[0053] The technical effects and advantages of the tumor full-course intelligent management platform and method of the present invention are as follows:

[0054] The intelligent management method for the whole course of tumors of the present invention significantly improves the construction and application efficiency of personalized treatment effect prediction models by integrating and analyzing multimodal data. First, by using natural language processing technology and standardized tools to perform semantic parsing and unified format processing on data from multiple sources such as genomic information, imaging data, electronic medical records and laboratory test results, the method ensures the efficiency and accuracy of data processing. Then, the feature selection algorithm is used to screen out the key features most relevant to the tumor type and treatment plan, thereby reducing the redundancy and noise in model training and improving the accuracy and reliability of prediction. In addition, by evaluating the redundancy of key features and the semantic consistency of multimodal data, the method can effectively identify and optimize the key information in the data set, further improving the explanatory power and practical application value of the model. Finally, by comprehensively analyzing the consistency of these features and data, the present invention can not only determine whether it is suitable to construct a personalized treatment effect prediction model, but also ensure the efficiency and accuracy of the constructed model in clinical applications, which greatly promotes the development of precision medicine and provides patients with a more scientific and reasonable treatment plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a schematic diagram of the intelligent management method for the entire course of tumor disease according to the present invention;

[0056] Figure 2 It is a schematic diagram of the structure of the intelligent management platform for the entire course of tumor disease of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Embodiment 1: Figure 1 The present invention provides a method for intelligent management of the entire course of tumor disease, which includes the following steps:

[0059] S1: Collect multimodal data of patients, including genomic information, imaging data, electronic medical records, and laboratory test results.

[0060] S2: Use natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources.

[0061] S3: Feature selection algorithms were used to screen key features associated with tumor type and treatment regimen.

[0062] S4: By analyzing the mutual information between key features and target variables and the mutual information between key features, the degree of redundancy risk of the selected key features can be evaluated; by comparing and analyzing the predicted labels and actual labels of multimodal data, the semantic consistency of multimodal data from different sources can be evaluated.

[0063] S5: Through comprehensive analysis of the potential redundancy of key features and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed.

[0064] Collect multimodal data from patients, specifically:

[0065] S101: Identify and define the types and sources of multimodal data that need to be collected:

[0066] Identification of data types includes:

[0067] Genomic information: including whole genome sequencing data (Whole Genome Sequencing, WGS), whole exome sequencing data (Whole Exome Sequencing, WES), gene expression data (RNA-seq), etc. These data can reveal the genetic background of the patient's tumor and potential targeted therapeutic targets.

[0068] Imaging data: including the patient's X-ray, magnetic resonance imaging (MRI), positron emission tomography (PET), computed tomography (CT) and other imaging data, which can provide the morphological characteristics of the tumor and its specific location and spread in the body.

[0069] Electronic medical record data: including the patient's hospitalization records, surgical records, pathology reports, drug prescriptions, treatment plans, and follow-up records, etc. These data can reflect the patient's entire course of medical management.

[0070] Laboratory test results: including blood biochemical indicators, tumor markers, histopathological examination results, immunohistochemical analysis, flow cytometry data, etc. These data can provide the molecular characteristics of the tumor and the physiological status of the patient.

[0071] The definition of data sources includes:

[0072] Hospitals and clinics: As the main data providers, hospitals and clinics will provide most of the medical imaging data, electronic medical record data, and laboratory test results.

[0073] Genetic testing companies: provide genomic information such as whole genome sequencing, exome sequencing, and gene expression data.

[0074] Patient self-management platform: Through the patient self-management platform, health data of patients in their daily lives (such as diet, exercise, and symptom records) can be collected. These data can supplement clinical data and provide information on patients' living habits and behavior patterns.

[0075] S102: Develop standardized collection protocols to ensure data integrity and consistency:

[0076] Genomic data collection standards: Based on internationally recognized sequencing standards, we will develop genomic data collection standards, including sequencing depth, coverage, accuracy and sensitivity requirements for variant detection. This will ensure that all genomic data have consistent quality and comparability across different sequencing platforms and laboratories.

[0077] Image data acquisition standards: Establish image data acquisition standards, including resolution, imaging time, and scanner parameters (such as voltage and current settings for CT scans) to ensure the uniformity of image data across different devices and institutions.

[0078] Electronic medical record data collection standards: Establish format standards for electronic medical record data, unify the structured way of medical record recording, and ensure that electronic medical record data from different sources can be interoperable and integrated.

[0079] Laboratory testing data collection standards: Develop standard operating procedures (SOPs) for laboratory testing, including the entire process of sample collection, processing, testing and recording to ensure the accuracy and repeatability of data.

[0080] Develop a data storage agreement to ensure that all data can be stored safely and effectively after collection. The storage agreement includes encrypted storage of data, access control, data backup strategy, and regulations on data retention and destruction.

[0081] S103: Acquire and store the patient's genomic data, collect the patient's medical images, extract the patient's historical medical records and clinical records, and collect and store laboratory test results including the patient's blood and tissue samples:

[0082] According to the genomic data type defined in S101, contact the genetic testing company or the molecular diagnostic laboratory of the hospital to arrange for the genetic sequencing of the patient's samples (such as blood, tumor tissue). After the sequencing is completed, the raw data and analysis results are stored in the designated data storage system according to the standards established in S102.

[0083] Collaborate with the radiology or medical imaging department responsible for the patient's imaging to collect all relevant medical imaging data of the patient. Ensure that the acquisition of imaging data complies with the acquisition standards in S102 and store the data in a picture archiving and communication system (PACS) or a dedicated medical imaging database.

[0084] Extract the patient's historical medical history and clinical records from the electronic medical record system, including hospitalization records, surgical records, pathology reports, etc. Use the format standards established by S102 to organize these data and store them in the electronic medical record database to ensure interoperability with other data types.

[0085] In the hospital or designated laboratory, collect laboratory test results of patients' blood and tissue samples, including but not limited to blood biochemical indicators, tumor markers, histopathological examination results, etc. Ensure that all test results meet the standards in S102 and store these data in a dedicated laboratory information management system (LIMS) or corresponding database.

[0086] Use natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources, specifically:

[0087] S201: Clean and perform preliminary word segmentation on text data such as electronic medical records:

[0088] First, the text data extracted from the electronic medical record system is cleaned. The cleaning step includes removing irrelevant characters (such as special symbols, formatting tags), invalid information (such as headers and footers, additional advertisements), and repeated information to ensure the purity of the text data.

[0089] Automatically correct spelling errors in text. Through dictionary matching and context analysis, it identifies and corrects common spelling errors and input errors to ensure the accuracy of medical terminology and key data.

[0090] For Chinese text, use word segmentation algorithms (such as word segmentation methods based on hidden Markov models) to segment continuous text streams into independent word units. For English text, use spaces and punctuation marks for word segmentation. The text after word segmentation is easier to perform subsequent semantic analysis.

[0091] During the word segmentation process, special attention is paid to identifying medical terms and proper nouns, such as "malignant tumor", "gene mutation", "targeted therapy", etc., and ensuring that these terms are not mis-segmented or omitted.

[0092] S202: Use natural language processing technology to extract key medical terms and perform semantic annotation:

[0093] The high-frequency medical terms in the electronic medical record text are identified through word frequency statistics. Combined with medical dictionaries (such as ICD-10, MeSH, etc.), the medical meaning and importance of these terms are further confirmed.

[0094] Use named entity recognition technology to automatically extract key medical entities such as disease names, drug names, symptoms, treatment plans, etc. from electronic medical record texts. Improve the accuracy and coverage of entity recognition through pre-trained medical corpus models (such as BioBERT).

[0095] Based on medical ontologies (such as UMLS and SNOMED CT), a semantic labeling system is established to map the extracted medical terms and entities to corresponding standardized concepts and categories. For example, "malignant tumor" is labeled as "disease" and "targeted drug" is labeled as "treatment plan".

[0096] Perform contextual analysis on the annotated text to understand the semantic relationships between terms, such as causal relationships ("gene mutations lead to malignant tumors"), time sequence ("after surgery, the patient received radiotherapy"), etc. Through semantic network construction, complex medical knowledge graphs can be identified.

[0097] S203: Unify the formats of data from different sources and map the information in different data sources into a unified data model:

[0098] Convert unstructured data in electronic medical record text (such as doctors' free descriptions) into structured data. Use template tools and rule engines to map text information to predefined structured fields (such as diagnosis results, treatment plans, drug dosages, etc.).

[0099] Use format conversion tools (such as HL7 and FHIR standards) to convert data from different systems to ensure data format consistency. For example, convert the electronic medical record data format of hospital A to a standard format consistent with that of hospital B.

[0100] According to the predefined medical data model (such as the FHIR resource model), data from different sources are mapped into a unified data structure. For example, genomic data is mapped to the "gene mutation" field, and the descriptive text of imaging data is mapped to the "imaging diagnosis result" field.

[0101] After completing data mapping, perform data integration and verification to ensure that all data can be seamlessly integrated into a unified data model. Use verification tools to check the consistency and completeness of the data to ensure that the final data integration results meet expectations.

[0102] A feature selection algorithm was used to screen for key features associated with tumor type and treatment regimen, specifically:

[0103] S301: Preliminary selection of potential features based on domain knowledge and data relevance:

[0104] Based on existing clinical studies and medical literature, potential features associated with specific tumor types and treatment options are identified. For example, for breast cancer, potential features that may be selected include HER2 gene status, hormone receptor status (ER / PR), tumor grade, etc.

[0105] Refer to current clinical diagnosis and treatment guidelines (such as NCCN guidelines) and standards to select key variables related to treatment decisions. These variables may include the patient's age, gender, pathological stage, molecular marker status, etc.

[0106] Perform correlation analysis on the collected multimodal data to calculate the correlation coefficient between each variable and the target variable (such as tumor type or treatment effect). You can use the Pearson correlation coefficient or the Spearman correlation coefficient, and choose the appropriate method based on the distribution characteristics of the data.

[0107] Based on the results of correlation analysis, the features significantly correlated with tumor type and treatment regimen were preliminarily screened out. The features with low correlation or no correlation were excluded to reduce the data dimension and improve the calculation efficiency.

[0108] S302: Use statistical methods to evaluate the importance of features:

[0109] For categorical variables, analysis of variance was used to assess the ability of different features to explain differences in tumor type or treatment regimen. The F value of the feature was calculated, and the higher the F value, the more important the feature.

[0110] For categorical data, the chi-square test is used to assess the independence between features and the target variable. A high chi-square value indicates that the feature has a strong explanatory power for the classification result.

[0111] For binary classification problems, a logistic regression model is used to evaluate the regression coefficient and significance level of the feature. The lower the significance level (P value), the more important the feature is to the classification result.

[0112] Apply ridge regression or LASSO regression to deal with multicollinearity issues and reduce unimportant features through regularization methods.

[0113] S303: Apply feature selection algorithm to screen key features:

[0114] Recursive feature elimination (RFE) repeatedly builds models and evaluates the importance of features, gradually eliminating the least important features and finally retaining the most important feature set.

[0115] The RFE algorithm is applied to recursively analyze the initially selected features, repeatedly train the model, and gradually reduce the number of features until the preset number of features or the optimal state of model performance is reached.

[0116] The importance of features is evaluated using the random forest algorithm, and the features that contribute most to the model's predictive ability are selected based on their importance scores (such as the Gini coefficient or information gain) during the decision tree splitting process.

[0117] Apply gradient boosting tree algorithms such as XGBoost or LightGBM to further filter and optimize the feature set by improving the importance scores of features in the tree model.

[0118] S304: Verify the effectiveness of selected features through cross-validation and model performance evaluation:

[0119] Divide the dataset into K subsets, use K-1 subsets for training in turn, and use the remaining subset for validation. Calculate the model performance indicators (such as accuracy, AUC value) for each validation to evaluate the stability and generalization ability of the selected features.

[0120] When the amount of data is small, the leave-one-out cross-validation (LOO-CV) is used to validate each sample individually to ensure the robustness of feature selection.

[0121] The model performance is evaluated using metrics such as accuracy, recall, precision, F1 value, AUC, etc. to determine whether the model based on the selected features performs well in the validation set.

[0122] The model built based on the selected features is compared with the models built with other feature sets (such as the initial feature set or a random feature set) to verify whether the effect of feature selection is significantly better than other methods.

[0123] S305: Adjust the feature set according to the verification results and optimize the final feature selection:

[0124] If you find during the cross-validation process that some features do not contribute much to the model performance or are highly correlated with other features, consider removing them from the feature set to reduce the model complexity.

[0125] Based on the validation results and the error analysis of the model, consider introducing other new features that may affect tumor types or treatment options to further optimize the feature set.

[0126] After multiple rounds of feature adjustment and model validation, a set of key feature sets was finally determined that can significantly improve the model's prediction accuracy, robustness, and generalization ability.

[0127] The final selected feature set is locked as the standard feature set for subsequent model training, validation, and clinical applications.

[0128] By analyzing the mutual information between key features and target variables and the mutual information between key features, the degree of redundancy risk of the selected key features is evaluated, specifically:

[0129] In the whole course of tumor management, the target variable can be a classification label such as tumor type, treatment response, etc. The feature set may include various biomarkers and clinical indicators extracted from multimodal data sources such as gene expression, clinical parameters, imaging data, laboratory test results, etc.

[0130] Target variable: A label of tumor type or treatment effect, usually a categorical variable such as {response, non-response}.

[0131] Feature set: gene variation frequency, protein expression level, MRI image features, CT value, etc.

[0132] Data preprocessing includes normalization, missing value filling, and outlier processing to ensure that the data quality is suitable for further analysis.

[0133] Mutual Information (MI) is used to quantify the amount of information shared between each feature and the target variable. Mutual Information is a nonlinear correlation measure that can capture any relationship (linear or nonlinear) between variables.

[0134] The mutual information formula is: ;in, It is a feature and the target variable The mutual information between is the joint probability density function (representing and probability of simultaneous occurrence), They are and The marginal probability density function of .

[0135] By calculating the mutual information between all features and the target variable, we can evaluate the contribution of each feature to the prediction of the target variable.

[0136] Mutual information between features is used to evaluate the redundancy between features. If two features share a lot of information, one feature may be sufficient to represent the other. and The mutual information between them is expressed as: .

[0137] Through this step, redundant features in the dataset can be identified and reduced, which is crucial to avoid model overfitting and improve model efficiency.

[0138] The mRMR algorithm aims to select a feature set whose features are highly correlated with the target variable and whose redundancy among features is minimal.

[0139] Calculate the feature redundancy hidden danger assessment index, the expression is: ;in, is the size of the feature set, It is a feature set Medium Features With the target variable The mutual information of It is a feature set Medium Features and The mutual information between them.

[0140] It is the feature redundancy hidden danger assessment index (the score calculated by the maximum correlation minimum redundancy algorithm), which is used to evaluate the quality of the selected feature set. The higher the feature redundancy hidden danger assessment index, the higher the correlation between the selected features and the target variable, and the lower the redundancy between the features. It is used to balance the correlation and redundancy of features to help select the optimal feature set for model training. The higher the feature redundancy hidden danger assessment index, the lower the risk of redundancy in the selected key features; the lower the feature redundancy hidden danger assessment index, the higher the risk of redundancy in the selected key features.

[0141] in: and are the actual values ​​of the two features in the dataset. In the formula, these values ​​represent the performance or value of a specific feature in a specific sample. For example, in a study on tumor data, It could be the gene expression level of a patient's tumor, It could be the expression level of another gene in the same patient.

[0142] and Represents two different feature columns in the dataset. These symbols usually represent entire variables in mathematical descriptions, rather than individual instances of variables. They are used to describe how the algorithm treats these two features across all observations in the entire dataset.

[0143] and Respectively represent characteristics and the target variable The specific value of .

[0144] Usually refers to a feature set, which can be a single feature or a collection of multiple features. In the context of mRMR, represents all the features considered.

[0145] The target variable, such as tumor type or treatment response, is the main variable that you are trying to predict or classify, and all features are selected to better predict or understand the .

[0146] Often used as an index to represent a specific feature or data point.

[0147] is an index used to indicate a specific feature or data point. In a formula or algorithm description, Used to iterate over the elements in a feature or dataset.

[0148] By comparing and analyzing the predicted labels of multimodal data with the actual labels, the semantic consistency of multimodal data from different sources is evaluated, specifically:

[0149] Extract features from data of various modalities that can represent the core information and uniqueness of each data.

[0150] include:

[0151] Text data (T_Data): includes patients’ medical records, diagnostic reports, etc. Feature extraction includes extracting keywords, phrases, and entities using natural language processing techniques.

[0152] Image data (I_Data): involves features extracted from medical images (such as MRI, CT scans), such as texture, contours, key points, etc.

[0153] Medical image annotation (A_Data): This includes professional medical annotations of images, such as marking of tumor location and size.

[0154] For text data, use the TF-IDF method or the BERT model to extract text features.

[0155] For image data, apply pre-trained convolutional neural networks (CNNs) such as VGG or ResNet to extract deep features.

[0156] For medical image annotation, structured annotation data is extracted based on image processing technology.

[0157] The extracted features are used to build a CRF model that is able to identify and exploit dependencies between different data sources to predict the semantic label of each data.

[0158] Define the tag sequence ( ) are semantic labels for data points, such as tumor type labels or treatment responses.

[0159] Define the feature sequence ( ) is the feature vector extracted above.

[0160] The basic formula for using the CRF model is: ;in, It is the conditional probability expression in the conditional random field model.

[0161] Often used to represent an index into a state or time step, where and Often used to indicate a specific position or state in a sequence, while Might be used to represent another dimension, such as another state or feature.

[0162] and It is the weight parameter in the conditional random field (CRF) model, which is used to weight the feature functions in the model, thereby affecting the calculation of the final state sequence probability. Each weight is associated with a specific feature function, adjusting the importance of the feature in the model decision.

[0163] and are feature functions, which are the core components of the CRF model. These functions are used to calculate scores associated with a specific state transition or the relationship between a state and an observation. may be associated with a single state or observation, while Can involve state pairs or more complex dependency structures.

[0164] is a normalization factor, sometimes called a partition function, which ensures that the sum of the probabilities of all possible state sequences given a given observation sequence is 1. This factor is calculated by taking a weighted sum of the probabilities of all possible label sequences and is a key component in probabilistic models to ensure that probabilities are correctly defined.

[0165] Among them, CRF is a conditional random field.

[0166] The model parameters are learned using a training dataset with known labels, and the weights are adjusted using gradient descent or other appropriate optimization algorithm to maximize the log-likelihood function.

[0167] On the test set, the consistency between the predicted label and the true label of each data point is used to calculate the semantic consistency index, which is the ratio of all correctly predicted labels to the total number of data points multiplied by the consistency weighting factor of different modal data.

[0168] The expression of semantic consistency index is: ;in, is the semantic consistency index, is the total number of test samples, It is with The weights of the modes corresponding to the data points are predetermined based on the reliability and importance of the modes; is an indicator function, if the predicted label is the same as the actual label, the result is 1, otherwise it is 0; It is The predicted labels for data points; is the actual label.

[0169] The larger the semantic consistency index, the higher the semantic consistency of multimodal data from different sources, which means that various data sources (such as text, images, and medical image annotations) are more consistent and coordinated in representing and conveying relevant medical information, thereby providing more reliable and accurate support in the intelligent management of the entire course of tumor disease.

[0170] Through comprehensive analysis of the degree of redundancy in the key features of screening and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed. Specifically:

[0171] The feature redundancy hidden danger assessment index and the semantic consistency index are normalized, and the normalized feature redundancy hidden danger assessment index and the semantic consistency index are respectively assigned preset proportional coefficients, and then the model establishment judgment coefficient is calculated. The model establishment judgment coefficient is used to determine whether a personalized treatment effect prediction model can be constructed.

[0172] The expression of the model establishment judgment coefficient is: ;in, Establish the judgment coefficient for the model, are the preset proportional coefficients of the feature redundancy risk assessment index and the semantic consistency index, respectively, and Both are greater than 0.

[0173] The larger the model establishment judgment coefficient is, the more likely it is that a personalized treatment effect prediction model can be constructed; the smaller the model establishment judgment coefficient is, the less likely it is that a personalized treatment effect prediction model can be constructed.

[0174] Set the model building judgment threshold and compare the model building judgment coefficient with the model building judgment threshold:

[0175] When the model establishment judgment coefficient is greater than the model establishment judgment threshold, it is determined that a personalized treatment effect prediction model can be constructed.

[0176] When the model establishment judgment coefficient is less than or equal to the model establishment judgment threshold, it is determined that the construction of the personalized treatment effect prediction model cannot be carried out.

[0177] When the model building judgment coefficient is higher than the model building judgment threshold, it means that the quality of the selected features and data sets is sufficient to support effective model building, which is usually associated with higher prediction performance and model reliability. Conversely, a model building judgment coefficient below the model building judgment threshold indicates that the existing features and data quality may not be sufficient to support a stable and accurate prediction model, so it is recommended not to build a model to avoid potential high risks and low efficiency. This threshold-based decision-making mechanism makes the model development process more objective, clear and easy to manage.

[0178] Among them, the setting of the preset proportionality coefficient is based on the assessment of the importance of the feature redundancy hidden danger assessment index and the semantic consistency index. These coefficients are usually set by data scientists or domain experts based on previous experience, performance feedback of the model, and the characteristics of a specific data set. Generally speaking, if the impact of feature redundancy on model performance is considered to be decisive, then the proportionality coefficient of the feature redundancy hidden danger assessment index will be set higher; conversely, if the semantic consistency of the data has a greater impact on the accuracy of the final prediction result, then the proportionality coefficient of the semantic consistency index will be higher. In actual operation, these coefficients may be adjusted according to specific application scenarios, such as optimizing these coefficients through methods such as cross-validation to ensure that the final model establishment judgment coefficient can effectively reflect the actual situation of the model's predictive ability.

[0179] Setting the judgment threshold for model building is a key step, which defines the criteria for whether a personalized treatment effect prediction model can be built. This threshold is usually set based on historical data and pre-conducted model tests. By analyzing the performance of previous models at different thresholds, a threshold is selected that maximizes the model effect (such as accuracy, recall, and other relevant indicators). In addition, the setting of the threshold must also consider the risk tolerance and expected goals in practical applications. For example, in high-risk medical applications, it may be tempting to set a higher threshold to ensure that the model is only applied when it is sufficiently reliable. The final determination of the threshold usually involves statistical analysis and expert review to ensure its scientificity and practicality.

[0180] Example 2: The difference between Example 2 of the present invention and Example 1 is that this example introduces an intelligent management platform for the entire course of tumor disease.

[0181] Figure 2 A structural schematic diagram of the intelligent management platform for the entire course of tumor disease of the present invention is given, which includes a modal data collection module, a modal data processing module, a key feature screening module, a redundant hidden danger assessment module, a semantic consistency assessment module and a model construction judgment module.

[0182] Modal data collection module: collects patients' multimodal data, including genomic information, imaging data, electronic medical records and laboratory test results.

[0183] Modal data processing module: uses natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources.

[0184] Key feature screening module: Use feature selection algorithms to screen key features related to tumor type and treatment regimen.

[0185] Redundancy risk assessment module: By analyzing the mutual information between key features and target variables and the mutual information between key features, the degree of redundancy risk of the selected key features is assessed.

[0186] Semantic consistency evaluation module: It evaluates the semantic consistency of multimodal data from different sources by comparing the predicted labels of multimodal data with the actual labels.

[0187] Model construction judgment module: Through comprehensive analysis of the degree of redundancy risk of key screened features and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed.

[0188] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0189] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0190] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0191] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0192] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0193] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0194] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0195] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0196] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0197] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent management method for the entire course of tumor disease, characterized in that: The steps include: S1: Collect multimodal data of patients, including genomic information, imaging data, electronic medical records, and laboratory test results; S2: Use natural language processing technology and standardized tools to semantically parse and unify the format of multimodal data from different sources; S3: Use feature selection algorithms to screen key features associated with tumor type and treatment regimen; S4: By analyzing the mutual information between key features and target variables and the mutual information between key features, the degree of redundancy risk of the selected key features is evaluated; by comparing and analyzing the predicted labels and actual labels of multimodal data, the semantic consistency of multimodal data from different sources is evaluated; S5: Through comprehensive analysis of the degree of redundancy in the key features of screening and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed; A feature selection algorithm was used to screen for key features associated with tumor type and treatment regimen, specifically: S301: Preliminary selection of potential features based on domain knowledge and data relevance; S302: Use statistical methods to evaluate the importance of features: For categorical variables, analysis of variance was used to assess the ability of different features to explain differences in tumor type or treatment regimen; For categorical data, the chi-square test was used to assess the independence between features and the target variable; For binary classification problems, a logistic regression model is used to evaluate the regression coefficient and significance level of the features; Dealing with multicollinearity issues and reducing unimportant features through regularization methods; S303: Apply feature selection algorithm to screen key features: Recursive feature elimination repeatedly builds models and evaluates the importance of features, gradually eliminating the least important features and finally retaining the most important feature set; Apply the RFE algorithm to recursively analyze the initially selected features, repeatedly train the model, and gradually reduce the number of features until the preset number of features or the optimal state of model performance is reached; Use the random forest algorithm to evaluate the importance of features and select the features that contribute most to the model's predictive ability based on the importance scores of the features in the decision tree splitting process; By improving the importance scores of features in the tree model, the feature set can be further screened and optimized; S304: Verify the effectiveness of the selected features through cross-validation and model performance evaluation; S305: Adjust the feature set according to the verification results to optimize the final feature selection.

2. The method for intelligent management of the entire course of tumor according to claim 1, characterized in that: Collect multimodal data from patients, specifically: S101: Identify and define the types and sources of multimodal data that need to be collected; S102: Develop standardized collection protocols to ensure data integrity and consistency; S103: Acquire and store the patient's genomic data, collect the patient's medical images, extract the patient's historical medical history and clinical records, and collect and store laboratory test results including the patient's blood and tissue samples.

3. The method for intelligent management of the entire course of tumor according to claim 2, characterized in that: Use natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources, specifically: S201: Cleaning and preliminary word segmentation of text data; S202: Use natural language processing technology to extract key medical terms and perform semantic annotation; S203: Unify the formats of data from different sources and map the information in different data sources into a unified data model.

4. The method for intelligent management of the entire course of tumor according to claim 3, characterized in that: By analyzing the mutual information between key features and target variables and the mutual information between key features, the degree of redundancy risk of the selected key features is evaluated, specifically: Use mutual information to quantify the amount of information shared between each feature and the target variable; Calculate features and The mutual information between them is expressed as: ； Calculate the feature redundancy hidden danger assessment index, the expression is: ； in, is the feature redundancy hazard assessment index, is the size of the feature set, It is a feature set Medium Features With the target variable The mutual information of It is a feature set Medium Features and The mutual information between and are the actual values ​​of the two features in the dataset, and Represents two different feature columns in the dataset, and Respectively represent characteristics and the target variable The specific value of represents the feature set, represents the target variable, and Are used as indexes.

5. The method for intelligent management of the entire course of tumor according to claim 4, characterized in that: By comparing and analyzing the predicted labels of multimodal data with the actual labels, the semantic consistency of multimodal data from different sources is evaluated, specifically: Extract features from data of various modalities; Including text data, image data and medical image annotation; Use the extracted features to build a CRF model; define the label sequence as the semantic label of the data point; define the feature sequence as the extracted feature vector; The basic formula for using the CRF model is: ； in, is the conditional probability expression in the conditional random field model, Used to represent an index into a state or time step, and is the weight parameter in the CRF model, and is the characteristic function, is the normalization factor; Calculate the semantic consistency index, the expression is: ； in, is the semantic consistency index, is the total number of test samples, It is with The weight of the mode corresponding to the data point, is the indicator function, It is The predicted labels for data points; is the actual label.

6. The method for intelligent management of the entire course of tumor according to claim 5, characterized in that: Through comprehensive analysis of the degree of redundancy in the key features of screening and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed. Specifically: The feature redundancy hidden danger assessment index and the semantic consistency index are normalized, and after the normalized feature redundancy hidden danger assessment index and the semantic consistency index are respectively assigned preset proportional coefficients, the model establishment judgment coefficient is calculated to determine whether the personalized treatment effect prediction model can be constructed; Set the model building judgment threshold and compare the model building judgment coefficient with the model building judgment threshold: When the model establishment judgment coefficient is greater than the model establishment judgment threshold, it is determined that the personalized treatment effect prediction model can be constructed; When the model establishment judgment coefficient is less than or equal to the model establishment judgment threshold, it is determined that the construction of the personalized treatment effect prediction model cannot be carried out.

7. A tumor full-course intelligent management platform, used to implement the tumor full-course intelligent management method according to any one of claims 1 to 6, characterized in that: It includes modal data collection module, modal data processing module, key feature screening module, redundant hidden danger assessment module, semantic consistency assessment module and model building judgment module; Modal data collection module: collects multimodal data of patients, including genomic information, imaging data, electronic medical records and laboratory test results; Modal data processing module: uses natural language processing technology and standardized tools to perform semantic analysis and unified format processing on multimodal data from different sources; Key feature screening module: Use feature selection algorithm to screen key features related to tumor type and treatment regimen; Redundancy risk assessment module: By analyzing the mutual information between key features and target variables and the mutual information between key features, the redundant risk level of the selected key features is assessed; Semantic consistency assessment module: evaluates the semantic consistency of multimodal data from different sources by comparing and analyzing the predicted labels of multimodal data with the actual labels; Model construction judgment module: Through comprehensive analysis of the degree of redundancy risk of key screened features and the semantic consistency of multimodal data from different sources, it is determined whether a personalized treatment effect prediction model can be constructed.

Images