Intelligent Classification and Prediction Method, System and Device for Oncology Care

Abstract

This application is applicable to the field of tumor care technology, and particularly relates to an intelligent classification prediction method, system and device for tumor care. The method includes: obtaining the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics of each target patient to comprehensively integrate the multi-dimensional medical information of the patients; constructing a scientifically quantified multi-dimensional prognosis stratification index set based on the correlation between molecular markers and clinical characteristics to form an accurate stratification evaluation framework; inputting the medical data into a pre-trained multi-layer neural network model, and generating a high-resolution tumor characterization map through cross-modal feature fusion technology; based on the dynamic matching of the map and the stratification index set, combining with a random forest model to calculate the stratification contribution degree parameter and the prognosis risk score, completing the accurate quantitative evaluation of the patient's condition, providing a highly reliable quantitative basis for clinical decision-making, effectively reducing the risk of over-treatment and improving the effect of patient prognosis management.

Description

Technical Field

[0001] This application belongs to the technical field of tumor care, and particularly relates to an intelligent classification prediction method, system and device for tumor care. Background Art

[0002] Intelligent classification prediction of tumor care is a key technology for clinical medical decision-making support, aiming to recommend precise diagnosis and treatment plans by integrating multi-dimensional medical data of patients. By analyzing static indicators such as tumor stage, histological type, and gene mutation status, it provides a reference basis for the selection of treatment plans and is widely used in scenarios such as personalized medication guidance, prognosis assessment, and efficacy monitoring.

[0003] Traditional tumor classification prediction methods have significant limitations, including insufficient data integration capabilities, limited model processing capabilities, low fitness of the classification system, and lack of effective classification prediction means, resulting in low clinical decision-making efficiency, improper allocation of medical resources, and at the same time restricting the optimization space of personalized treatment plans and affecting the survival quality and prognosis improvement effect of patients. Summary of the Invention

[0004] The embodiments of this application provide an intelligent classification prediction method, system and device for tumor care, which can solve the problems of low clinical decision-making efficiency, improper allocation of medical resources, and affecting the survival quality and prognosis improvement effect of patients due to the lack of effective classification prediction means in the process of tumor classification prediction.

[0005] In the first aspect, the embodiments of this application provide an intelligent classification prediction method for tumor care, including:

[0006] Obtain the medical data of each target patient; wherein, the medical data includes the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics;

[0007] Construct a multi-dimensional prognosis stratification index set based on the detection values of the tumor-related molecular markers and the clinical diagnosis and treatment characteristics;

[0008] Input the medical data into a pre-trained multi-layer neural network model for feature fusion to obtain the tumor characterization map of each target patient; wherein, the tumor characterization map is a visual data model generated by the deep learning model through feature fusion of medical data;

[0009] Calculate the stratification contribution degree parameter and prognosis risk score of each target patient according to the tumor characterization map and the multi-dimensional prognosis stratification index set;

[0010] Generate the personalized care plan classification result of each target patient based on the stratification contribution degree parameter and the prognosis risk score of each target patient.

[0011] In the embodiments of the present application, the above technical solutions have at least the following technical effects:

[0012] The intelligent classification prediction method for tumor care provided by the embodiments of the present application obtains the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics of each target patient, comprehensively integrates the multi-dimensional medical information of the patients, and provides a data basis for subsequent stratified prediction. According to the correlation between molecular markers and clinical characteristics, a scientifically quantified multi-dimensional prognostic stratification index set is constructed to form an accurate stratification evaluation framework. The medical data is input into a pre-trained multi-layer neural network model, and a high-resolution tumor characterization map is generated through cross-modal feature fusion technology to achieve a deep analysis of the biological characteristics of tumors. Based on the dynamic matching of the map and the stratification index set, combined with the random forest model, the stratification contribution degree parameter and the prognostic risk score are calculated to complete the accurate quantitative evaluation of the patient's condition. Finally, according to the collaborative analysis of the dynamic contribution degree and the risk score, a personalized classification result of the care plan is generated, significantly improving the objectivity and prediction efficiency of tumor stratification, optimizing the allocation of medical resources, and at the same time enhancing the interpretability of the model through the dynamic weight mechanism, providing a highly reliable quantitative basis for clinical decision-making, effectively reducing the risk of over-treatment and improving the effect of patient prognosis management.

[0013] In a second aspect, the embodiments of the present application provide an intelligent classification prediction system for tumor care, including:

[0014] An acquisition unit for acquiring the medical data of each target patient; wherein, the medical data includes the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics;

[0015] A stratification unit for constructing a multi-dimensional prognostic stratification index set according to the detection values of the tumor-related molecular markers and the clinical diagnosis and treatment characteristics;

[0016] A fusion unit for inputting the medical data into a pre-trained multi-layer neural network model for feature fusion to obtain a tumor characterization map; wherein, the tumor characterization map is a visual data model generated by performing feature fusion on medical data through a deep learning model;

[0017] A scoring unit for calculating the stratification contribution degree parameter and the prognostic risk score of each target patient according to the tumor characterization map and the multi-dimensional prognostic stratification index set;

[0018] A prediction unit for generating a personalized classification result of the care plan for each target patient based on the stratification contribution degree parameter and the prognostic risk score of each target patient.

[0019] In a third aspect, an embodiment of the present application provides an intelligent classification prediction device for tumor care, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method described in any one of the above aspects is implemented.

[0020] In a fourth aspect, an embodiment of the present application provides a computer program product. When the computer program product runs on an intelligent classification prediction device for tumor care, the intelligent classification prediction device for tumor care is enabled to execute the method described in any one of the above aspects.

[0021] It can be understood that the beneficial effects of the above second to fourth aspects can be referred to the relevant descriptions in the above aspects and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a schematic flow chart of an intelligent classification prediction method for tumor care provided by an embodiment of the present application;

[0024] Figure 2 It is a schematic operation diagram of an intelligent classification prediction method for tumor care provided by an embodiment of the present application;

[0025] Figure 3 It is a schematic structural diagram of an intelligent classification prediction system for tumor care provided by an embodiment of the present application;

[0026] Figure 4 It is a schematic structural diagram of an intelligent classification prediction device for tumor care provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] In the following description, specific details such as specific system structures and technologies are proposed for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0028] It should be understood that, as used in the specification of this application and the appended claims, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or their combinations.

[0029] It should also be understood that the term "and / or" as used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0030] As used in the specification of this application and the appended claims, the term "if" may be construed, depending on the context, as "when", "once", "in response to determining", or "in response to detecting". Similarly, the phrases "if determined" or "if the described condition or event is detected" may be construed, depending on the context, to mean "once determined", "in response to determining", "once the described condition or event is detected", or "in response to detecting the described condition or event".

[0031] In addition, in the description of the specification of this application and the appended claims, the terms "first", "second", "third", etc. are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0032] Reference to "one embodiment" or "some embodiments" or the like described in the specification of this application means that a particular feature, structure, or characteristic described in connection with that embodiment is included in one or more embodiments of this application. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in another way. The terms "comprising", "including", "having", and their variants all mean "including but not limited to", unless otherwise specifically emphasized in another way.

[0033] Traditional tumor classification prediction methods have significant limitations, including insufficient data integration capabilities, limited model processing capabilities, low adaptability of the classification system, and lack of effective classification prediction means, resulting in low clinical decision-making efficiency, improper allocation of medical resources, and at the same time restricting the optimization space of personalized treatment plans and affecting the survival quality and prognosis improvement effect of patients.

[0034] To solve the above problems, an embodiment of the present application provides a tumor care intelligent classification and prediction method, system, and device. In this method, by obtaining the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics of the target patient, the multi-dimensional medical information of the patient is comprehensively integrated to provide a data basis for subsequent stratified prediction. According to the correlation between molecular markers and clinical features, a scientifically quantified multi-dimensional prognostic stratification index set is constructed to form an accurate stratification evaluation framework. The medical data is input into a pre-trained multi-layer neural network model, and a high-resolution tumor characterization map is generated through cross-modal feature fusion technology to achieve in-depth analysis of the biological characteristics of the tumor. Based on the dynamic matching of the map and the stratification index set, combined with the random forest model, the stratification contribution degree parameter and the prognostic risk score are calculated to complete the accurate quantitative evaluation of the patient's condition. Finally, according to the collaborative analysis of the dynamic contribution degree and the risk score, a personalized care plan classification result is generated, significantly improving the objectivity and prediction efficiency of tumor stratification, optimizing the allocation of medical resources, and at the same time enhancing the interpretability of the model through a dynamic weight mechanism, providing a highly reliable quantitative basis for clinical decision-making, effectively reducing the risk of over-treatment and improving the effect of patient prognosis management.

[0035] The tumor care intelligent classification and prediction method provided by the embodiment of the present application can be applied to a tumor care intelligent classification and prediction device. At this time, the tumor care intelligent classification and prediction device is the execution subject of the tumor care intelligent classification and prediction method provided by the embodiment of the present application. The embodiment of the present application does not impose any restrictions on the specific type of the tumor care intelligent classification and prediction device.

[0036] For example, the tumor care intelligent classification and prediction device can be a mobile phone, a tablet computer, a notebook computer, an ultra-mobile personal computer (UMPC), a netbook, a desktop computer, a smart large screen, a smart TV, a handheld device with wireless communication function, a computing device, or other processing devices connected to a wireless modem, a vehicle-mounted device, a vehicle networking terminal, a computer, a laptop computer, a communication device, a computing device, etc.

[0037] To better understand the tumor care intelligent classification and prediction method provided by the embodiment of the present application, the following provides an exemplary introduction to the specific implementation process of the tumor care intelligent classification and prediction method provided by the embodiment of the present application.

[0038] Figure 1 FIG. shows a schematic flowchart of the tumor care intelligent classification and prediction method provided by the embodiment of the present application. Figure 2 FIG. shows a running flowchart of the tumor care intelligent classification and prediction method provided by the embodiment of the present application. The tumor care intelligent classification and prediction method includes:

[0039] S100, obtain the medical data of each target patient; wherein, the medical data includes the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics.

[0040] It can be understood that the medical data includes the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics. The detection values of tumor-related molecular markers may include molecular biological indexes such as the concentration of circulating tumor DNA (ctDNA), the expression level of PD-L1, the tumor mutation burden (TMB), and the microsatellite instability (MSI) status. The detection values of tumor-related molecular markers are obtained through laboratory tests and reflect the molecular characteristics and biological behaviors of tumors. The clinical diagnosis and treatment characteristics may include structured clinical data such as TNM staging, ECOG performance status, treatment response rate, pathological grade, and complication occurrence. TNM staging is a standard for evaluating the size of tumors, lymph node metastasis, and distant metastasis, while ECOG performance status evaluates the physical function status of patients. The medical data can be automatically retrieved from the hospital laboratory information system (LIS) and the electronic medical record system (EMR) through the API interface. For unstructured text data, such as doctors' diagnosis and treatment records, natural language processing techniques (such as the BERT model) can be used for entity recognition and structured conversion to ensure the integrity and availability of the data.

[0041] S200, construct a multi-dimensional prognostic stratification index set based on the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics.

[0042] It can be understood that by integrating the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics, an index set that has a significant impact on the prognosis of patients can be formed, that is, construct a multi-dimensional prognostic stratification index set. Factor analysis can be performed on the detection values of tumor-related molecular markers to obtain the molecular feature sets of each target patient after dimensionality reduction. Factor analysis is a statistical method used to identify the underlying structure hidden in a large number of variables, thereby reducing the dimensionality of the data. Hierarchical clustering is performed on the clinical diagnosis and treatment characteristics to generate the clinical feature classification groups of each target patient. Hierarchical clustering is a hierarchical clustering method that classifies similar features into one category by calculating the similarity between features. Finally, the molecular feature sets and clinical feature classification groups of each target patient are obtained, and then strongly associated feature combinations are screened out to construct a hierarchical classification framework to form a multi-dimensional prognostic stratification index set.

[0043] In a possible implementation manner, S200, construct a multi-dimensional prognostic stratification index set based on the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics, including:

[0044] S210, perform factor analysis on the detection values of tumor-related molecular markers to obtain the molecular feature sets of each target patient after dimensionality reduction.

[0045] It can be understood that factor analysis methods can be used to reduce the dimensionality of tumor-related molecular marker detection data. Factor analysis is a statistical method that reduces the number of variables by identifying latent factors in the data, retaining the most important features for prognostic analysis. For example, ctDNA concentration and TMB may be combined into a latent "tumor burden factor" through factor analysis, thus simplifying the data complexity while retaining key information. The set of molecular features after dimensionality reduction provides a more efficient input for subsequent analysis.

[0046] Exemplarily, the factor analysis process can condense high-dimensional molecular data through statistical methods. First, calculate the covariance matrix between each indicator to identify latent common factors. For example, PD-L1 expression and TMB may share the latent factor of "immune activity". Use varimax rotation to optimize the factor loading matrix so that each principal component clearly represents a specific biological process. The set of molecular features generated after dimensionality reduction retains 85% of the information in the original data and eliminates the problem of multicollinearity. For example, 12 original indicators are condensed into 5 principal components, where the first principal component reflects immune checkpoint-related features, and the second principal component represents tumor proliferation activity, etc.

[0047] S220, perform hierarchical clustering on clinical diagnosis and treatment features to generate clinical feature classification groups for each target patient.

[0048] It can be understood that clinical diagnosis and treatment features can be divided into different classification groups through hierarchical clustering algorithms. Hierarchical clustering is a hierarchical clustering method that generates a dendrogram to represent the grouping results of features by gradually merging or splitting data points. For example, ECOG performance status and TNM stage may be classified into categories such as "early low-risk" and "late high-risk" through clustering. Hierarchical clustering can group complex clinical data according to feature similarity, facilitating subsequent analysis and decision-making.

[0049] Optionally, S220, perform hierarchical clustering on clinical diagnosis and treatment features to generate clinical feature classification groups for each target patient, including:

[0050] S221, perform standardization processing on clinical diagnosis and treatment features to obtain a standardized clinical feature matrix.

[0051] It is understandable that the standardization process aims to eliminate the interference of different clinical feature dimensions on subsequent analyses. For example, age (0 - 100), tumor size (0 - 30 cm), and ECOG score (0 - 5) need to be standardized by Z-score transformation into a distribution with a mean of 0 and a standard deviation of 1. The mean and standard deviation of each clinical diagnosis and treatment feature can be calculated, and the original values can be transformed according to the formula to convert clinical features of different types and dimensions into a structured data matrix with a unified standard distribution. For missing values, the multiple imputation by chained equations (MICE) method is used to predict the missing ECOG score based on other features (such as the correlation between age and tumor stage). Each patient in the finally generated standardized matrix is represented as a vector v = [Zage, ZECOG, ZTNM,...], providing comparable input for clustering.

[0052] Exemplarily, robust processing methods can be adopted for clinical feature standardization. For continuous variables that may contain outliers (such as white blood cell count), a scaling method based on the median and interquartile range is used to avoid the interference of extreme values. For example, for a patient with a white blood cell count of 28×10^9 / L (normal range 4 - 10), after being processed by RobustScaler, it will not deviate too much from other samples. For categorical variables (such as TNM stage), one-hot encoding is first performed to generate a sparse vector, and then it is projected into a continuous space through an embedding layer to make the stages have computable semantic distances (such as the vector distance between stage III and stage IV is less than that between stage I and stage IV).

[0053] S222, calculate the mutual information matrix between each clinical diagnosis and treatment feature to obtain the hierarchical clustering distance matrix between each clinical diagnosis and treatment feature.

[0054] It is understandable that mutual information is used to quantify the non-linear association between features. The mutual information matrix is a symmetric matrix based on the difference measure between features, which is used to guide the hierarchical clustering algorithm to construct a dendrogram. The amount of information sharing between each pair of clinical diagnosis and treatment features can be calculated by the k-nearest neighbor estimation method. For example, the mutual information value between the number of chemotherapy cycles and neutropenia is relatively high, indicating a strong correlation between the two. A dendrogram is constructed according to the mutual information distance through the hierarchical clustering algorithm: in the initial stage, each feature forms a separate class, and the classes with the closest distance are gradually merged to form a hierarchical structure. The optimal number of clusters can be automatically determined by dynamic threshold segmentation through the gradient change of the merging distance, avoiding the bias caused by manual setting. For example, when the merging distance suddenly increases by 30%, it is determined that the current is a reasonable classification level.

[0055] Exemplarily, for example, there may be a U-shaped relationship rather than a linear correlation between age and the incidence of chemotherapy side effects. For continuous features (such as age) and categorical features (such as cancer subtypes), the continuous variable can be discretized into 5 intervals through adaptive binning, and then the normalized mutual information (NMI) can be calculated based on the joint probability distribution. Suppose the NMI value between "TNM stage" and "lymph node metastasis" is 0.8, indicating a high correlation between the two, while the NMI between "gender" and "pathological grade" being 0.1 is regarded as a weak association. The mutual information matrix is converted into a difference matrix according to 1 - NMI (difference value = 1 - NMI), and then the Ward variance minimization algorithm is used to calculate the hierarchical clustering distance to form the branch structure of the dendrogram.

[0056] S223, Based on the hierarchical clustering distance matrix, perform dynamic threshold segmentation to obtain the initial clinical feature groups of each target patient.

[0057] It can be understood that dynamic threshold segmentation determines the optimal cutting point by analyzing the branch height of the hierarchical clustering dendrogram. For example, calculate the height difference between adjacent merging steps in the dendrogram (such as when the merging height changes from 1.2 to 2.5, the difference is 1.3), and select the position corresponding to the maximum difference as the cutting threshold. If the maximum difference occurs when merging into 3 groups, the dendrogram is cut into 3 initial clinical feature groups (such as "low-risk group", "medium-risk group", "high-risk group"). For special scenarios (such as multiple differences with similar heights), the density peak detection (DPC algorithm) can be used to select the threshold corresponding to the local maximum density to avoid subjectively setting the number of groups.

[0058] Exemplarily, dynamic threshold segmentation can adopt an adaptive mechanism by real-time monitoring the change rate of the inter-class distance during the hierarchical clustering process. When a significant jump is detected (such as the spacing increase exceeding 2 standard deviations of the historical average), trigger the segmentation operation to generate new clusters. For example, divide "liver function indicators" and "renal function indicators" into independent clusters, while group "thrombocytopenia" and "anemia" which belong to the same hematological toxicity into one group.

[0059] S224, Evaluate the clustering quality of each initial clinical feature group through the silhouette coefficient, and merge the initial clinical feature groups with an overlap degree higher than the preset threshold to generate an optimized clinical feature classification group.

[0060] It can be understood that the silhouette coefficient evaluation optimizes the clustering quality by quantifying the intra-class compactness and inter-class separation. Calculate the similarity (a value) of each initial clinical feature group with the features in the same cluster and the difference degree (b value) with the nearest feature in a different cluster. When (b - a) / max(a, b) is close to 1, it is judged as an ideal classification. Set the merging threshold as the average silhouette coefficient difference < 0.15. For example, when the silhouette coefficients of two groups are 0.68 and 0.63 respectively, perform the merging. This mechanism can automatically eliminate over-segmentation, such as merging features with similar semantics like "gastrointestinal reaction" and "loss of appetite".

[0061] S230. Calculate the mutual information between the molecular feature sets of each target patient and the clinical feature classification groups, and screen out strongly associated feature combinations; wherein, the strongly associated feature combinations are used to reflect multi-modal feature sets greater than a preset threshold.

[0062] It can be understood that mutual information (MI) is used to capture the non-linear association between molecules and clinical features. For example, to calculate the MI value between "tumor metabolic factor" and "postoperative recurrence", the continuous molecular factor can be binned (e.g., into 10 intervals) first, and its joint probability distribution with the binary recurrence event can be calculated, and then the MI value can be obtained through the discrete entropy formula. If the MI value of a certain combination exceeds the preset threshold (e.g., 0.3), it is considered to have strong relevance (e.g., the correlation between "TP53 mutation + ECOG ≥ 2 points" and prognosis is retained). At the same time, redundant combinations (such as the MI value of "EGFR mutation + BMI" being lower than the threshold) are excluded, and finally 10 - 15 key cross-modal features are screened out.

[0063] S240. Construct a hierarchical classification framework based on the strongly associated feature combinations to obtain a multi-dimensional prognostic stratification index set.

[0064] It can be understood that hierarchical classification divides the patient population recursively to obtain a multi-dimensional prognostic stratification index set. For example, in the first layer, it is divided into early stage (Ⅰ - Ⅱ) and late stage (Ⅲ - Ⅳ) according to TNM staging; in the second layer, the early stage group is further subdivided according to the PD-L1 expression level (≥50% vs. <50%), while the late stage group is subdivided according to the ctDNA concentration (≥5% vs. <5%). Each splitting node can select the optimal feature based on the principle of minimizing Gini impurity (such as the largest decrease in the Gini coefficient of TNM), and finally generate a classification framework similar to a tree structure, with each level corresponding to specific prognostic criteria (such as "late stage - high positive ctDNA - immunotherapy resistant group"), forming 4 - 6 terminal nodes for clinical decision-making.

[0065] S300. Input the medical data into a pre-trained multi-layer neural network model for feature fusion to obtain a tumor characterization map; wherein, the tumor characterization map is a visual data model generated by the deep learning model through feature fusion of the medical data.

[0066] It can be understood that molecular and clinical diagnosis and treatment features can be integrated through a deep learning model (such as Graph Convolutional Network, GCN). The molecular feature data is input into a fully connected layer after normalization, and the text-based diagnosis and treatment features are encoded into semantic vectors through a Long Short-Term Memory network (LSTM). Multimodal interactions (such as "the association between tumor mutational burden and the necrotic area in CT images") are captured through a Cross-Attention mechanism, and then t-SNE is used to project the high-order features into a two-dimensional space to form a tumor representation map, which is a visual data model generated by feature fusion of medical data through a deep learning model. Adjacent points in the map represent similar biological behaviors (such as the tendency to metastasize), and high-risk clusters can be identified through density-based spatial clustering of applications with noise (DBSCAN).

[0067] In a possible implementation, in S300, medical data is input into a pre-trained multi-layer neural network model for feature fusion to obtain a tumor representation map, including:

[0068] S310, standardize the detection values of tumor-related molecular markers to obtain the standardized molecular marker data of each target patient.

[0069] It can be understood that the detection values of tumor-related molecular markers can be standardized to eliminate the dimensional differences in the data and ensure the accuracy and stability of subsequent feature fusion. For example, the ctDNA concentration and TMB are respectively standardized to the same numerical range.

[0070] Exemplarily, an index with a skewed distribution (such as PD-L1 expression level) can be transformed by log2(x + 1) to make it approximately normally distributed; binary encoding (MSS = 0 / MSI-H = 1) is used for categorical data (such as microsatellite status). The standardized data is stored in an HDF5 file, retaining the original data, processing parameters, and version information to ensure the traceability of the results.

[0071] S320, perform feature embedding and null value compensation on the clinical diagnosis and treatment features to obtain the structured diagnosis and treatment data of each target patient.

[0072] It can be understood that clinical diagnosis and treatment features can be transformed into low-dimensional vector representations through feature embedding, such as converting text data into numerical vectors using Word2Vec or Transformer models. At the same time, missing values are compensated, and null values are processed using mean filling, interpolation, or machine learning methods (such as KNN imputation) to ensure the integrity of the data.

[0073] Exemplarily, a 768-dimensional context vector can be generated for text descriptions (such as "partial remission") to capture semantic nuances (e.g., the vector distance between "PR" and "CR" is greater than that between "PR" and "SD"). Missing value imputation uses conditional generative adversarial networks: the generator receives known features (such as age, gender) to generate reasonable estimated values (such as ECOG score), and the discriminator evaluates the authenticity of the generated data through the Wasserstein distance. The Transformer encoder can adopt masked language modeling (MLM) + contrastive learning and be fine-tuned through hierarchical learning rates (top layer lr = 5e-5, bottom layer lr = 1e-6). The conditional generative adversarial network uses a 5-layer fully connected layer generator architecture (1024-512-256-128-64), and the discriminator is designed with spectral normalization + gradient penalty (λ = 10), and the training parameters are Adam (β1 = 0.5, β2 = 0.999), batch = 64.

[0074] S330, combine the standardized molecular marker data and structured diagnosis and treatment data of each target patient into a joint input tensor.

[0075] It can be understood that the joint input tensor can fuse multimodal data through channel concatenation. Assume that the standardized molecular marker data is a 100-dimensional vector (such as 20 gene expression values, 5 protein markers, and 3 factor analysis results), and the structured diagnosis and treatment data is a 50-dimensional vector (such as age, stage, treatment response flag), then it is concatenated in the channel dimension into a 150-dimensional 2D tensor (shape [number of patients, 150]). To preserve time series features (such as multiple ctDNA detections), it can be extended to a 3D tensor ([number of patients, time steps, number of features]) for processing by a recurrent neural network. When the time dimension is missing, it is default filled with a single time step input to ensure that the model is compatible with heterogeneous data.

[0076] S340, extract local features of the joint input tensor through the convolutional layer of a multi-layer neural network model, and perform weighted fusion on the local features of the joint input tensor through an attention mechanism to obtain a fused cross-modal fusion feature tensor.

[0077] It can be understood that the local features of the joint input tensor are extracted through the convolutional layer of the multi-layer neural network model. A 1D convolutional kernel (width = 3, stride = 1) is used for the joint input tensor to extract local patterns (such as the gain effect of "EGFR mutation + concurrent radiotherapy" on prognosis) from adjacent molecular features (such as EGFR mutation status) and clinical features (such as concomitant treatment). The attention mechanism assigns higher weights to key features (such as driver gene mutations) by calculating the feature weight matrix. By generating query (Q), key (K), and value (V) matrices, and calculating the attention scores through Softmax, scaled dot-product attention is performed on the fused features. For the finally output cross-modal fused feature tensor, high-weight features are strengthened and noise features are suppressed.

[0078] Optionally, in step S340, the local features of the joint input tensor are weighted and fused through the attention mechanism to obtain the fused cross-modal fused feature tensor, including:

[0079] S341, generating a first feature projection based on the normalized molecular marker data and constructing a molecular marker query matrix.

[0080] It can be understood that the first feature projection is a feature transformation process based on the molecular marker data. The original data can be projected into a 32-dimensional semantic space through two fully connected layers. The first layer uses ReLU activation to extract non-linear features, and the second layer performs a linear transformation to generate the final query vector. The first fully connected layer maps the input normalized molecular marker data to a high-dimensional space through the ReLU activation function to capture non-linear features. The non-linear characteristic of ReLU enables the model to better fit complex biomarker data during the learning process. The second fully connected layer maps the features after ReLU activation to a 32-dimensional semantic space through a linear transformation to form the query vector. During the pre-training stage, a contrastive learning strategy can be adopted: the positive samples come from the temporal data augmentation of the same patient, and the negative samples are randomly mixed with the features of different patients, so that the query vectors of similar patients are clustered in the space.

[0081] S342, generating a second feature projection based on the structured diagnosis and treatment data and constructing a diagnosis and treatment key matrix and a diagnosis and treatment value matrix.

[0082] It can be understood that the second feature projection is a feature transformation process based on structured diagnosis and treatment data. Feature extraction of structured data can be performed by methods such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs). For example, a one-dimensional convolutional network is used to process time series data to extract key information during the diagnosis and treatment process. By encoding the extracted features (such as using a transformation layer), a "key matrix" is obtained, which plays a key role in the subsequent cross-modal attention mechanism and serves as the paired object for "query" data. The diagnosis and treatment value matrix contains the actual numerical values in the diagnosis and treatment data (such as treatment plans, treatment effect evaluations, etc.), which will be concatenated with the weighted features during fusion.

[0083] S343, According to the query matrix of the normalized molecular marker data and the diagnosis and treatment key matrix of the structured diagnosis and treatment data, calculate the cross-modal correlation weight matrix through dual-line attention.

[0084] It can be understood that the query matrix is derived from molecular marker data, and the key matrix comes from structured diagnosis and treatment data. The dual-line attention mechanism performs interactive calculations through the query matrix and the key matrix. Calculate the similarity between the query and the key, which can be calculated by the dot product score(q,k)=qTk, where q is the query vector and k is the key vector. Based on the calculated similarity score, the similarity is transformed into a probability distribution through the softmax function to obtain the attention weights. For each query vector, calculate its attention weights with all keys and perform weighted summation to generate weighted features. Finally, through the dual-line attention mechanism, a cross-modal correlation matrix is obtained. Each element represents the correlation between the query and the key, which helps to find the most relevant cross-modal features in subsequent steps.

[0085] S344, According to the modal correlation weight matrix, perform multi-scale attention operations in parallel through the multi-scale attention mechanism, and perform weighted concatenation of the cross-modal correlation weight matrix and the diagnosis and treatment value matrix to obtain a weighted concatenation result.

[0086] It can be understood that at different scales, data can be processed using convolutional kernels of different sizes or different context windows. For example, different time windows or spatial scales are used to capture features at different levels. The attention mechanism at each scale is calculated independently and outputs multiple weighted results. These results will be executed in parallel and then summarized. The results obtained from the multi-scale attention calculation can be concatenated with the diagnosis and treatment value matrix. After concatenation, a composite feature vector is formed, which contains weighted information from molecular markers, structured diagnosis and treatment data, and multi-scale attention. The form of weighted concatenation is to concatenate two matrices (the modal correlation weight matrix and the diagnosis and treatment value matrix) column-wise or row-wise and then merge them into a large feature tensor.

[0087] S345. Perform skip connection and batch normalization on the weighted splicing result after splicing to obtain the fused cross-modal fusion feature tensor.

[0088] It can be understood that skip connection directly transfers the information of the input layer to the output layer, skipping the intermediate layer, which helps to maintain the flow of information and accelerate the training process. Batch Normalization can be applied to the weighted result after splicing. The purpose is to standardize the data output by each layer so that its mean is 0 and variance is 1. Batch Normalization can make the training process more stable, accelerate convergence, and reduce the risk of overfitting.

[0089] S350. Project the fused cross-modal fusion feature into a two-dimensional space to obtain a tumor characterization map.

[0090] It can be understood that the fused cross-modal fusion feature can be projected into a two-dimensional space feature projection. The t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm is used to preserve the local structure of high-dimensional data. The specific parameters can be set as perplexity (perplexity = 30), learning rate (200), and number of iterations (1000). Input the cross-modal fusion tensor (such as 1024-dimensional features), and optimize the KL divergence between the target distribution and the low-dimensional embedding through gradient descent, so that adjacent points in the high-dimensional space (such as patients with high PD-L1 expression) are still closely distributed in the two-dimensional map. To enhance visualization, map the results of K-means clustering (k = 5) to different color clusters, and label the distribution area of patients with known good prognosis (such as the lower left quadrant is the "immunotherapy sensitive area") to assist doctors in intuitively evaluating tumor heterogeneity.

[0091] S400. Calculate the stratification contribution degree parameter and prognosis risk score of each target patient according to the tumor characterization map and the multi-dimensional prognosis stratification index set.

[0092] It can be understood that the stratification contribution degree parameter uses the SHAP value (Shapley Additive exPlanations) to quantify the impact of each stratification index on the individual patient. For example, for a certain patient, the SHAP value of TNM stage is 0.3, and that of PD-L1 is 0.5, indicating that the latter contributes more to its high-risk classification. The prognosis risk score is obtained by weighted summing of features through logistic regression (such as Score = 0.3×TNM_Ⅲ stage + 0.5×PD-L1_high expression + 0.2×ctDNA_positive), and is normalized to a threshold of 0-100 points (such as >70 points is high risk). At the same time, calibrate the score based on the topological features of the tumor characterization map (such as the median survival rate of the cluster where the patient is located) to make it consistent with the actual clinical outcome.

[0093] In a possible implementation, S400 calculates the stratification contribution degree parameter and the prognosis risk score of each target patient according to the tumor characterization atlas and the multi-dimensional prognosis stratification index set, including:

[0094] S410 extracts the topological connection features and probability distribution features from the tumor characterization atlas to form a composite feature vector.

[0095] It can be understood that the tumor characterization atlas can be transformed into a graph structure through topological feature extraction, with each pixel as a node, and edges are established between adjacent pixels and regions with similar intensities. The neighborhood information is aggregated through a three-layer graph convolutional network: the first layer captures local metabolic hotspots, the second layer identifies sub-region structures, and the third layer integrates global topological patterns. The mean of the output node embedding vectors is used as the final topological feature, reflecting the spatial heterogeneity of the tumor. Two types of features can be extracted from the tumor characterization atlas: one is the topological connection feature, including the degree centrality and betweenness centrality of the patient node in the atlas, capturing the non-Euclidean space association of neighborhood information through the graph convolutional network (GCN); the other is the probability distribution feature, using the kernel density estimation method (KDE) to quantify the cumulative probability density of patient features in the global space, where h is the window width coefficient, and K selects the Epanechnikov kernel function to reduce the edge smoothing error, and the two are fused to form a composite feature vector.

[0096] Exemplarily, to extract the spatial heterogeneity and local metabolic hotspots of the tumor, the tumor atlas is transformed into a graph structure. The pixel points of each image are regarded as the nodes of the graph, and edges are formed between adjacent pixels or regions with similar intensities. Then, a convolutional network (GCN) is used to process these graph data. The GCN gradually aggregates the neighborhood information of the nodes through multiple convolutional layers. Each time the node embedding (the output of each convolutional layer) is aggregated, a low-dimensional feature representation is obtained, which reflects the topological structure and spatial heterogeneity of the tumor atlas. By taking the mean of the embedding vectors of each node, the global topological feature of the entire tumor region can be obtained. The intensity distribution of the tumor in different regions can be calculated to obtain the intensity probability distribution feature of the tumor region, further enriching the feature vector.

[0097] S420 constructs a stratification weight allocation tensor according to the multi-dimensional prognosis stratification index set.

[0098] It can be understood that the hierarchical weight tensor constructs a three-dimensional dynamic structure: the first dimension corresponds to the molecular principal components (such as PC1-PC5), the second dimension is associated with the clinical feature clusters (such as the treatment response group), and the third dimension divides the time windows (0-3 months, 3-6 months, etc.). The initial weights are set according to the historical data distribution. For example, the weight of chemotherapy sensitivity is relatively high in the initial stage of treatment (0-3 months) and gradually decreases later. The weights are updated weekly through Kalman filtering to track the changes in treatment effects. The potential correlations between hierarchical metrics are modeled through a multi-task learning framework.

[0099] Exemplarily, define the weight tensor w∈ℝ^{K×D} (K is the number of hierarchical dimensions, D is the composite feature dimension), and each column corresponds to the dynamic weight channel of a certain hierarchical metric (such as T stage or gene mutation type). Its initialization uses the Xavier method to keep the variance consistent and is then optimized through multiple rounds of backpropagation: when updating the gradient, an L2 sparse constraint is added to the loss function of each metric to enable the learning of differential feature weights for each layer.

[0100] S430, input the composite feature vector into the pre-trained random forest model, and calculate the dynamic contribution degrees of each hierarchical metric based on the dynamic weight channels in the weight assignment tensor to obtain the hierarchical contribution degree parameters.

[0101] It can be understood that the composite feature vector extracted from the tumor characterization atlas and the multi-dimensional prognostic stratification metric set (such as gene expression, clinical features, etc.) can be combined as the input of the model and input into the pre-trained random forest model. Random forest is an ensemble learning method that is trained through multiple decision trees, can effectively handle high-dimensional data and prevent overfitting. The random forest calculates the contribution degree of each feature according to the training data. The influence degree of each feature on the patient's prognostic risk score is measured by the feature gain value. According to the dynamic weights in the weight assignment tensor, calculate the contribution degrees of each hierarchical metric (such as treatment response, gene expression, etc.).

[0102] Exemplarily, based on the pre-trained random forest model, analyze the meaning of the composite features. The specific process includes: ① Input the pre-constructed training data set (including patient features and survival status labels), and define the cross-entropy loss function for initial training; ② Generate a decision tree ensemble through multiple rounds of iteration. During the construction of each tree, calculate the feature gain value and accumulate the gain value along the iteration rounds as the total gain of each metric; ③ Set a threshold according to the gain ranking (such as retaining the Top-15 metrics) to eliminate low-contribution features; ④ Normalize the filtered gain values (using the Softmax function) to output the hierarchical contribution degree parameters.

[0103] Optionally, in S430, input the composite feature vector into the pre-trained random forest model, and calculate the dynamic contribution degrees of each hierarchical metric based on the dynamic weight channels in the weight assignment tensor to obtain the hierarchical contribution degree parameters, including:

[0104] S431. Obtain a pre - constructed training dataset based on patient stratification, and define a classification error objective function based on the training dataset.

[0105] It can be understood that the training dataset is data constructed according to the stratification of patients (such as disease stage, treatment plan, etc.). The label of each patient can be its final prognosis result (such as survival period, recurrence rate, etc.). The objective function is usually a classification error (such as cross - entropy loss), which is used to measure the difference between the model prediction and the actual result. By optimizing the objective function, the model can learn the contribution degree of each stratification index.

[0106] S432. Generate a decision tree classifier through multiple rounds of iteration, and calculate the feature gain values of each classifier. In each round of iteration, update the contribution degree scores of each stratification index according to the feature gain values, and obtain the cumulative contribution degree scores of each stratification index.

[0107] It can be understood that the model can generate a decision tree classifier through multiple rounds of iteration, and calculate the gain values of each feature in each classifier. In each iteration, the decision tree updates the contribution degree scores of the stratification index according to the current feature gain value. Through multiple rounds of iteration, the contribution degree scores of the stratification index can be gradually optimized, and the cumulative contribution degree can be obtained to help understand the impact of each stratification index on the patient's prognosis.

[0108] S433. Sort the feature contribution degrees of the multi - dimensional prognosis stratification indexes based on the cumulative contribution degree scores, and perform feature screening on each stratification index in the multi - dimensional prognosis stratification index set by filtering out low - contribution degree indexes according to the preset pruning threshold based on the feature contribution degree sorting.

[0109] It can be understood that each stratification index can be sorted according to the cumulative contribution degree score. An index with a higher contribution degree score indicates a greater impact on the patient's prognosis. According to the preset pruning threshold (such as the contribution degree is lower than a certain percentile), filter out the indexes with a smaller contribution to the prognosis risk score, and retain the indexes with a greater impact on the prediction.

[0110] S434. Perform probability normalization processing on the cumulative contribution degree scores of the screened stratification indexes to generate stratification contribution degree parameters.

[0111] It can be understood that the contribution degree scores of the screened features can be probabilistically normalized so that the sum of the contribution degrees of all indexes is 1. This can make the contribution degrees of different indexes comparable and can be used for subsequent weighted calculations. Finally, a normalized stratification contribution degree parameter is obtained, which reflects the importance of each stratification index in predicting the patient's prognosis risk.

[0112] S440, weighted aggregation is performed on the composite feature vector and the hierarchical contribution degree parameter, and the prognosis risk score of each target patient is converted through a logical function.

[0113] It can be understood that the composite feature vector and the hierarchical contribution degree parameter can be weighted aggregated to form a new comprehensive feature vector. The weighted features are transformed through a logical function (such as logistic regression or Sigmoid function) to obtain the prognosis risk score of each patient. The prognosis risk score represents the prognosis risk level of the patient, usually ranging from 0 to 1, and the larger the value, the higher the risk.

[0114] Exemplarily, the composite feature vector and the contribution degree parameter are weighted by tensor product, and its calculation form is: , and then the linear combination value is mapped to the 0-100 score interval through a logical function. The final score is adjusted based on topological calibration. The cluster center to which the target patient belongs is retrieved from the atlas, and the median actual survival rate of the patients to which it belongs (such as the median survival time of a certain cluster is 18 months) is compared with the model prediction value. Through the weighted moving average method (α = 0.7 determined according to the learning of the validation set), the strict consistency between the risk score and the clinical observation outcome is achieved.

[0115] S500, based on the hierarchical contribution degree parameter and the prognosis risk score of each target patient, generate the classification result of the personalized care plan for each target patient.

[0116] It can be understood that the generation of the treatment plan can combine evidence-based medicine rules and resource constraint optimization. For example, for high-risk patients (prognosis risk > 70), the "immunotherapy + chemotherapy" combination plan recommended by the guidelines is preferentially matched, but it needs to be adjusted according to the hospital drug inventory (such as the availability of PD-1 inhibitors). The particle swarm optimization algorithm (PSO) takes maximizing the treatment effect (based on the survival gain model in historical data) and minimizing the cost (number of chemotherapy cycles, bed occupancy time) as the objective function, and iteratively searches for the optimal plan combination. The final output is a structured treatment plan label (such as "high risk - priority Keytruda + platinum - need to monitor liver toxicity") and resource scheduling suggestions (such as "need to reserve an ICU bed for 3 days"), and a visualization interface is used to assist doctors in making decisions.

[0117] In a possible implementation manner, S500, based on the hierarchical contribution degree parameter and the prognosis risk score of each target patient, generate the classification result of the personalized care plan for each target patient, including:

[0118] S510, based on the prognosis risk score of each target patient, obtain the tumor risk level of each target patient.

[0119] It can be understood that patients can be classified into different risk levels based on the prognostic risk score of each patient. Generally, these levels can be divided into low risk, medium risk and high risk. For example, if the prognostic risk score is greater than 70, the patient is classified as a high-risk group. The division of risk levels is the basis for the formulation of treatment plans and provides guidance for subsequent treatment decisions. Patients in different risk levels require different intensities and types of therapeutic interventions.

[0120] S520, matching a preset treatment pathway template according to the tumor risk level of each target patient to obtain a basic treatment decision for each target patient; wherein the basic treatment decision includes multiple intervention measures.

[0121] It can be understood that based on the patient's tumor risk level, the patient's basic treatment decision is generated by matching the preset treatment pathway template. For example, for high-risk patients, the treatment pathway template may recommend a combination of immunotherapy and chemotherapy; for low-risk patients, a single targeted therapy or monitoring program may be recommended. These treatment pathway templates take into account evidence-based medical guidelines and are personalized based on the patient's specific circumstances (such as age, comorbidities, etc.). Each treatment pathway contains multiple interventions, such as drug therapy, surgery, radiotherapy, nursing intervention, etc., to form a preliminary treatment plan.

[0122] S530, rank the sensitivity of the intervention measures in the basic treatment decision according to the stratified contribution parameters of each target patient, balance the medical resource constraints through the particle swarm optimization algorithm, and generate the classification results of the personalized treatment plan for each target patient.

[0123] It can be understood that each intervention in the basic treatment decision can be ranked by sensitivity based on the stratified contribution parameters. The purpose of sensitivity ranking is to evaluate the contribution of each intervention to the treatment effect. For example, for high-risk patients, immunotherapy may be more critical than chemotherapy, so immunotherapy interventions will be ranked first. The particle swarm optimization algorithm (PSO) can be used for further optimization to generate personalized treatment plan classification results, which significantly improves the objectivity and prediction efficiency of tumor stratification, optimizes the allocation of medical resources, and enhances the interpretability of the model through a dynamic weight mechanism, providing a highly reliable quantitative basis for clinical decision-making, effectively reducing the risk of overtreatment and improving patient prognosis management.

[0124] Exemplarily, PSO is an optimization algorithm that simulates the movement of particle swarms in nature and aims to find the optimal solution. In the generation of treatment plans, the goal of particle swarm optimization is to balance the treatment effect and medical resource constraints. Specifically, the algorithm is optimized under the following objectives: According to the survival gain model in historical data, select treatment combinations to maximize the survival period and quality of life of patients. According to the resource limitations of the hospital (such as drug inventory, beds, medical equipment, etc.), the algorithm needs to control the treatment cost and resource consumption when optimizing the treatment plan. For example, if the drug inventory is limited, the treatment plan will be adjusted to select available drugs or treatment plans. Through iterative optimization, the particle swarm algorithm can ultimately generate personalized treatment plans and resource scheduling suggestions. For example, for high-risk patients, the possible output treatment plan could be: "High-risk - Prioritize Keytruda + Platinum - Need to monitor liver toxicity", and the recommended resource scheduling could be: "Need to reserve an ICU bed for 3 days".

[0125] Corresponding to the tumor care intelligent classification and prediction method in the above embodiments, the embodiments of the present application also provide a tumor care intelligent classification and prediction system, and each unit of this system can implement each step of the tumor care intelligent classification and prediction method. Figure 3 The structural block diagram of the tumor care intelligent classification and prediction system provided by the embodiments of the present application is shown. For the sake of convenience of description, only the parts related to the embodiments of the present application are shown.

[0126] Referring to Figure 3 , this tumor care intelligent classification and prediction system includes:

[0127] An acquisition unit, configured to acquire the medical data of each target patient; wherein, the medical data includes the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics;

[0128] A stratification unit, configured to construct a multi-dimensional prognostic stratification index set based on the detection values of the tumor-related molecular markers and the clinical diagnosis and treatment characteristics;

[0129] A fusion unit, configured to input the medical data into a pre-trained multi-layer neural network model for feature fusion to obtain a tumor characterization map; wherein, the tumor characterization map is a visual data model generated by performing feature fusion on medical data through a deep learning model;

[0130] A scoring unit, configured to calculate the stratification contribution degree parameter and prognostic risk score of each target patient according to the tumor characterization map and the multi-dimensional prognostic stratification index set;

[0131] A prediction unit, based on the stratification contribution degree parameter and the prognostic risk score of each target patient, generates a personalized care plan classification result for each target patient.

[0132] It should be noted that for the information interaction, execution process, etc. between the above-mentioned systems / units, since they are based on the same concept as the method embodiments of this application, their specific functions and the technical effects brought about can be specifically referred to in the method embodiment section, and will not be elaborated here.

[0133] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above-mentioned division of each functional unit and module is used as an example. In actual applications, the above-mentioned functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the system is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into a processing unit, or each unit module exists physically alone, or two or more unit modules are integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated here.

[0134] The embodiment of this application also provides a tumor care intelligent classification prediction device. Figure 4 It is a schematic structural diagram of a tumor care intelligent classification prediction device provided by an embodiment of this application. As Figure 4 shown, the tumor care intelligent classification prediction device 6 of this embodiment includes: at least one processor 60 ( Figure 4 only one is shown in the figure), at least one memory 61 ( Figure 4 only one is shown in the figure), and a computer program 62 stored in the at least one memory 61 and executable on the at least one processor 60. When the processor 60 executes the computer program 62, the tumor care intelligent classification prediction device 6 implements the steps in any of the above-mentioned tumor care intelligent classification prediction method embodiments, or enables the tumor care intelligent classification prediction device 6 to implement the functions of each unit in the above-mentioned system embodiments.

[0135] Exemplarily, the computer program 62 can be divided into one or more units. The one or more units are stored in the memory 61 and executed by the processor 60 to complete this application. The one or more units can be a series of computer program instruction segments capable of completing specific functions, and these instruction segments are used to describe the execution process of the computer program 62 in the tumor care intelligent classification prediction device 6.

[0136] The intelligent classification and prediction device 6 for tumor care can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server, etc. The intelligent classification and prediction device for tumor care may include, but is not limited to, a processor 60 and a memory 61. Those skilled in the art can understand that Figure 4 This is only an example of the intelligent classification and prediction device 6 for tumor care, and does not constitute a limitation on the intelligent classification and prediction device 6 for tumor care. It may include more or fewer components than those shown in the figure, or combine some components, or different components. For example, it may also include input / output devices, network access devices, buses, etc.

[0137] The processor 60 may be a central processing unit (CPU), and the processor 60 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0138] In some embodiments, the memory 61 may be an internal storage unit of the intelligent classification and prediction device 6 for tumor care, such as the hard disk or memory of the intelligent classification and prediction device 6 for tumor care. In some other embodiments, the memory 61 may also be an external storage device of the intelligent classification and prediction device 6 for tumor care, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the intelligent classification and prediction device 6 for tumor care. Further, the memory 61 may also include both the internal storage unit and the external storage device of the intelligent classification and prediction device 6 for tumor care. The memory 61 is used to store an operating system, application programs, a boot loader, data, and other programs, such as the program code of the computer program, etc. The memory 61 may also be used to temporarily store data that has been output or will be output.

[0139] The embodiment of the present application also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0140] An embodiment of the present application provides a computer program product. When the computer program product runs on a tumor care intelligent classification prediction device, the tumor care intelligent classification prediction device implements the steps in any of the above method embodiments.

[0141] If the integrated unit 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 this understanding, to implement all or part of the processes in the above method embodiments of the present application, a computer program can be used to instruct the relevant hardware to complete. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps in the above method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the tumor care intelligent classification prediction device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc, etc. In some jurisdictions, according to legislation and patent practice, the computer-readable medium cannot be an electrical carrier signal and a telecommunication signal.

[0142] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0143] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0144] In the embodiments provided in the present application, it should be understood that the disclosed intelligent classification prediction system for tumor care / intelligent classification prediction device and method for tumor care can be implemented in other ways. For example, the embodiments of the intelligent classification prediction system for tumor care / intelligent classification prediction device described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces, and the indirect coupling or communication connection of devices or units can be in electrical, mechanical or other forms.

[0145] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0146] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An intelligent classification and prediction method for tumor care, characterized in that, include: Obtaining medical data of each target patient; wherein the medical data includes tumor-related molecular marker detection values ​​and clinical diagnosis and treatment characteristics; Constructing a multi-dimensional prognostic stratification indicator set based on the tumor-related molecular marker detection values ​​and the clinical diagnosis and treatment characteristics; The medical data is input into a pre-trained multi-layer neural network model for feature fusion to obtain a tumor characterization map, including: standardizing the detection values ​​of the tumor-related molecular markers to obtain the normalized molecular marker data of each target patient; embedding and compensating the clinical diagnosis and treatment features to obtain the structured diagnosis and treatment data of each target patient; combining the normalized molecular marker data of each target patient with the structured diagnosis and treatment data into a joint input tensor; extracting the local features of the joint input tensor through the convolution layer of the multi-layer neural network model, and weightedly fusing the local features of the joint input tensor through the attention mechanism to obtain a fused cross-modal fusion feature tensor; projecting the fused cross-modal fusion features into a two-dimensional space to obtain a tumor characterization map; wherein the tumor characterization map is a visual data model generated by performing feature fusion on medical data through a deep learning model; Calculating the stratification contribution parameter and the prognostic risk score of each of the target patients according to the tumor characterization map and the multi-dimensional prognostic stratification indicator set; Generating a personalized treatment plan classification result for each target patient based on the stratified contribution parameter and the prognostic risk score of each target patient; Wherein, the multi-dimensional prognostic stratification indicator set is constructed based on the tumor-related molecular marker detection values ​​and the clinical diagnosis and treatment characteristics, including: Performing factor analysis on the tumor-related molecular marker detection values ​​to obtain a set of molecular features of each target patient after dimensionality reduction; Performing spectral clustering on the clinical diagnosis and treatment characteristics to generate clinical characteristic classification groups for each of the target patients; Perform mutual information calculation on the molecular feature set and the clinical feature classification group of each target patient to screen out a strongly correlated feature combination; wherein the strongly correlated feature combination is used to reflect a multimodal feature set greater than a preset threshold; A hierarchical classification framework is constructed based on the combination of the strongly correlated features to obtain a multi-dimensional prognostic stratification indicator set; Calculating the stratification contribution parameter and the prognostic risk score of each target patient according to the tumor characterization map and the multi-dimensional prognostic stratification indicator set includes: Extracting topological connection features and probability distribution features from the tumor representation map to form a composite feature vector; Constructing a stratified weight allocation tensor according to the multi-dimensional prognostic stratification indicator set; Inputting the composite feature vector into a pre-trained random forest model, calculating the dynamic contribution of each stratification index based on the dynamic weight channel in the weight allocation tensor, and obtaining a stratification contribution parameter; The composite feature vector and the stratification contribution parameter are weighted and aggregated, and converted into a prognostic risk score for each target patient through a logistic function.

2. The intelligent classification and prediction method for tumor care as described in claim 1, wherein, Performing pedigree clustering on the clinical diagnosis and treatment features to generate clinical feature classification groups for each of the target patients, including: Performing standardization processing on the clinical diagnosis and treatment features to obtain a standardized clinical feature matrix; Calculating the mutual information matrix between each of the clinical diagnosis and treatment features to obtain a hierarchical clustering distance matrix between each of the clinical diagnosis and treatment features; Performing dynamic threshold segmentation based on the hierarchical clustering distance matrix to obtain initial clinical feature groups for each of the target patients; Evaluating the clustering quality of each of the initial clinical feature groups through the silhouette coefficient, and merging the initial clinical feature groups with an overlap degree higher than a preset threshold to generate an optimized clinical feature classification group.

3. The intelligent classification and prediction method for tumor care as described in claim 1, wherein Performing weighted fusion on the local features of the joint input tensor through an attention mechanism to obtain a fused cross-modal fusion feature tensor, including: Generating a first feature projection based on the normalized molecular marker data and constructing a molecular marker query matrix; Generating a second feature projection based on the structured diagnosis and treatment data and constructing a diagnosis and treatment key matrix and a diagnosis and treatment value matrix; Calculating a cross-modal association weight matrix through bi-linear attention according to the query matrix of the normalized molecular marker data and the diagnosis and treatment key matrix of the structured diagnosis and treatment data; Performing multi-level attention operations in parallel through a multi-scale attention mechanism according to the modal association weight matrix, and performing weighted splicing on the cross-modal association weight matrix and the diagnosis and treatment value matrix to obtain a weighted splicing result; Performing skip connection and batch normalization on the spliced weighted splicing result to obtain a fused cross-modal fusion feature tensor.

4. The intelligent classification prediction method for tumor care as described in claim 1, wherein Inputting the composite feature vector into a pre-trained random forest model, and calculating the dynamic contribution degree of each hierarchical index based on the dynamic weight channels in the weight assignment tensor to obtain a hierarchical contribution degree parameter, including: Obtaining a pre-constructed training data set based on patient stratification, and defining a classification error objective function based on the training data set; Generating a decision tree classifier through multiple rounds of iteration, and calculating the feature gain value of each classifier. In each round of iteration, updating the contribution degree score of each hierarchical index according to the feature gain value to obtain the cumulative contribution degree score of each hierarchical index; Performing feature contribution degree ranking on the multi-dimensional prognosis stratification indexes based on the cumulative contribution degree score, and filtering out low contribution degree indexes through a preset pruning threshold according to the feature contribution degree ranking to perform feature screening on each hierarchical index in the multi-dimensional prognosis stratification index set; Performing probability normalization processing on the cumulative contribution degree scores of the screened hierarchical indexes to generate a hierarchical contribution degree parameter.

5. The intelligent classification prediction method for tumor care as claimed in claim 1, wherein Generating a personalized medical care plan classification result for each of the target patients based on the hierarchical contribution degree parameter and the prognosis risk score of each of the target patients, including: Obtaining the tumor risk level of each of the target patients based on the prognosis risk score of each of the target patients; Matching a preset treatment path template according to the tumor risk level of each of the target patients to obtain a basic treatment decision for each of the target patients; wherein, the basic treatment decision includes multiple intervention measures. Rank the interventions in the basic treatment decision according to the stratification contribution parameters of each target patient, and balance the medical resource constraints through the particle swarm optimization algorithm to generate the classification results of the personalized care plans for each target patient.

6. An intelligent classification and prediction system for tumor care, characterized in that, For implementing the method according to any one of claims 1 to 5, the intelligent classification prediction system for tumor care includes: An acquisition unit, configured to acquire the medical data of each target patient; wherein, the medical data includes the detection values of tumor-related molecular markers and clinical diagnosis and treatment characteristics; A stratification unit, configured to construct a multi-dimensional prognostic stratification index set according to the detection values of the tumor-related molecular markers and the clinical diagnosis and treatment characteristics; A fusion unit, configured to input the medical data into a pre-trained multi-layer neural network model for feature fusion to obtain a tumor characterization map; wherein, the tumor characterization map is a visual data model generated by performing feature fusion on the medical data through a deep learning model; A scoring unit, configured to calculate the stratification contribution parameter and the prognostic risk score of each target patient according to the tumor characterization map and the multi-dimensional prognostic stratification index set; A prediction unit, configured to generate the classification results of the personalized care plans for each target patient based on the stratification contribution parameter and the prognostic risk score of each target patient.

7. An intelligent classification and prediction device for tumor care, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

Images