Pan-cancer classification screening method based on blood test markers and ai model

Abstract

The application belongs to the technical field of classification based on AI models, and discloses a pan-cancer classification screening method based on blood test markers and AI models, which comprises the following steps: first, acquiring blood test data of an object to be evaluated and a medical center identifier, and generating a standardized feature vector through preprocessing; obtaining a shared feature representation through a shared feature extraction layer, and then outputting a steady state category by a pathological steady state evaluation branch; combining the center identifier and the steady state category, and generating a calibrated feature vector through a center calibration branch; determining a gating mode according to the steady state category and generating an expert sub-model weight distribution by a joint gating network; and finally, weighting and converging expert sub-model prediction outputs to obtain a pan-cancer classification result. Through the collaborative design of multiple modules, the method effectively avoids cross-center differences and non-steady physiological interference, improves the accuracy, cross-center generalization ability and medical interpretability of pan-cancer classification, and is suitable for large-scale screening scenarios.

Description

Technical Field

[0001] This application relates to the field of classification technology based on AI models, and more specifically, to a pan-cancer classification and screening method based on blood test biomarkers and AI models. Background Technology

[0002] Blood routine and blood biochemistry tests, as basic and widely used diagnostic tools, offer advantages such as convenience, low cost, and non-invasiveness. Changes in their indicators are closely related to the occurrence and development of tumors. Therefore, how to construct a pan-cancer screening model based on routine blood test data that can avoid central variability and non-stationary interference and has strong generalization ability, in order to achieve accurate and universally accessible early screening, has become an urgent technical problem to be solved. Summary of the Invention

[0003] In response, this application provides a pan-cancer classification and screening method based on blood test biomarkers and AI models to at least partially solve the above-mentioned technical problems.

[0004] This application provides a pan-cancer classification and screening method based on blood test biomarkers and AI models, including the following method steps: Obtain the blood test data of the object to be evaluated and the center identifier of the medical center to which it belongs, and preprocess the blood test data to generate a standardized feature vector; The standardized feature vector is input into the shared feature extraction layer for processing to obtain the shared feature representation; The shared feature representation and a subset of key indicators selected from the standardized feature vector are input into the pathological homeostasis assessment branch for processing, and the homeostasis category of the object to be assessed is output, which includes homeostasis, metastable and non-homeostasis. The corresponding center embedding vector is obtained based on the center identifier, and combined with the steady-state category, the normalized feature vector is calibrated through the center calibration branch to generate the calibrated feature vector. The shared feature representation, the center embedding vector, the steady-state category, and the calibrated feature vector are input into a joint gating network for processing. The joint gating network determines the gating mode based on the steady-state category and generates a weight distribution corresponding to multiple expert sub-models based on the shared feature representation, the center embedding vector, and the calibrated feature vector. Based on the weight distribution, the predicted outputs of the multiple expert sub-models, obtained from the shared feature representation and the calibrated feature vector, are weighted and converged to obtain the pan-cancer classification result of the object to be evaluated.

[0005] In one possible embodiment, outputting the steady-state category of the object to be evaluated includes triggering a non-steady-state determination based on whether the standardized value of the key indicator in the blood test data exceeds a preset extreme threshold; and / or grouping the indicators in the blood test data, calculating the consistency score of the indicators within the group and the inconsistency score between the indicators of different groups, and determining the steady-state category based on whether the inconsistency score exceeds a preset inconsistency threshold.

[0006] In one possible embodiment, the central calibration branch calibrates the standardized feature vector, specifically including: determining a calibration intensity coefficient based on the steady-state category, wherein the calibration intensity coefficient corresponding to the steady-state category is higher than the calibration intensity coefficient corresponding to the steady-state category is metastable or non-steady-state; calculating the calibration offset of the index in the standardized feature vector based on the statistical profile corresponding to the central identifier; and calibrating the standardized feature vector according to the calibration intensity coefficient and the calibration offset to obtain the calibrated feature vector.

[0007] In one possible embodiment, the joint gating network determines the gating mode based on the steady-state category. Specifically, if the steady-state category is steady-state, the gating mode is determined to be a discriminative mode, and a first sparsity parameter is set; if the steady-state category is metastable, the gating mode is determined to be a transitional mode, and a second sparsity parameter is set, which is greater than the first sparsity parameter; if the steady-state category is non-steady-state, the gating mode is determined to be a conservative mode, and a third sparsity parameter is set, which is greater than the second sparsity parameter; wherein, the sparsity parameter is used to control the number of expert sub-models with non-zero weights in the weight distribution.

[0008] In one possible embodiment, when the gating mode is conservative, after the joint gating network generates a weight distribution corresponding to multiple expert sub-models, it further performs a smoothing process on the weight distribution to make the weight distribution more uniform.

[0009] In one possible embodiment, the plurality of expert sub-models include common mechanism experts and cancer type-specific experts; wherein the common mechanism experts are used to learn consistent tumor-related physiological patterns across medical centers, and the cancer type-specific experts are used to learn classification patterns for specific cancer types or specific organ systems.

[0010] In one possible embodiment, during the training of the expert sub-model, a cross-center consistency constraint is applied, which is: for sample pairs from different medical centers with similar basic characteristics, the intermediate representation distance of the common mechanism experts is constrained to be less than a preset consistency threshold.

[0011] In another aspect, this application also provides a pan-cancer classification and screening system based on blood test biomarkers and an AI model, comprising: The data acquisition and preprocessing module is used to acquire the blood test data of the object to be evaluated and the center identifier of the medical center to which it belongs, and to preprocess the blood test data to generate a standardized feature vector. The shared feature extraction module is used to input the standardized feature vector into the shared feature extraction layer for processing to obtain the shared feature representation; The pathological homeostasis assessment module is used to process the shared feature representation and a subset of key indicators selected from the standardized feature vector into the pathological homeostasis assessment branch, and output the homeostasis category of the object to be assessed, which includes homeostasis, metastable and non-homeostasis. The center calibration module is used to obtain the corresponding center embedding vector according to the center identifier, and combine it with the steady-state category to calibrate the normalized feature vector through the center calibration branch to generate a calibrated feature vector; The joint gating module is used to input the shared feature representation, the center embedding vector, the steady-state category, and the calibrated feature vector into the joint gating network for processing; the joint gating network determines the gating mode according to the steady-state category, and generates a weight distribution corresponding to multiple expert sub-models based on the shared feature representation, the center embedding vector, and the calibrated feature vector. The classification result generation module is used to perform weighted aggregation of the prediction outputs obtained by the multiple expert sub-models based on the shared feature representation and the calibrated feature vector according to the weight distribution, so as to obtain the pan-cancer classification result of the object to be evaluated.

[0012] This application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the pan-cancer classification screening method based on blood test biomarkers and AI models as described above.

[0013] In another aspect, this application provides a storage medium having stored computer program instructions thereon, which can be executed by a processor to implement the pan-cancer classification screening method based on blood test biomarkers and AI models as described above.

[0014] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the pan-cancer classification screening method based on blood test biomarkers and an AI model as described above.

[0015] This application effectively addresses the technical challenges of cross-center differences, interference from non-steady-state physiological states, and the amplification of misjudgments due to the coupling of these factors through a collaborative design of a pathological homeostasis assessment branch, a central calibration branch, and a joint gating network. The pathological homeostasis assessment branch enables pre-judgment of the sample's physiological state, providing a reliable basis for subsequent processing. The central calibration branch dynamically adjusts the calibration strategy based on the homeostasis category, achieving effective alignment of data from different centers. The joint gating network combines homeostasis category and center information for collaborative decision-making, optimizing the selection and weight allocation of expert sub-models. Through the collaborative work of each module and the iterative training process of the system, this method significantly improves the accuracy, cross-center generalization ability, and medical interpretability of pan-cancer classification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of a pan-cancer classification and screening method based on blood test biomarkers and AI models, provided as an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the overall model structure provided for an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of the central calibration branch processing flow provided in the embodiments of this application.

[0020] Figure 4 This is a schematic diagram of the joint gating network processing flow provided in the embodiments of this application.

[0021] Figure 5 This is a schematic diagram of the structure of a pan-cancer classification and screening system based on blood test biomarkers and an AI model, provided as an embodiment of this application.

[0022] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this application are information and data authorized by the user or fully authorized by all parties.

[0025] This method is applicable to scenarios such as physical examination screening, outpatient and emergency department initial screening, and inpatient admission assessment. It can be deployed and executed on servers, cloud computing nodes, or hospital data centers. Implementation typically relies on the hospital or physical examination center's laboratory information system to ensure the acquisition of blood test data such as complete blood count and blood biochemistry, along with the corresponding medical center identification. The system must also have the basic hardware and software environment for data storage, model operation, and iterative updates.

[0026] The following detailed description, in conjunction with specific embodiments, illustrates the implementation process of the pan-cancer classification and screening method based on blood test biomarkers and AI models described in this application. It should be noted that these embodiments are merely for explaining this application and are not intended to limit the scope of protection of this application. Conventional adjustments or substitutions made by those skilled in the art to each step without departing from the concept of this application should be included within the scope of protection of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a schematic diagram of a pan-cancer classification and screening method based on blood test biomarkers and AI models, including the following method steps: S1, Obtain the blood test data of the object to be evaluated and the center identifier of the medical center to which it belongs, and preprocess the blood test data to generate a standardized feature vector; S2, the standardized feature vector is input into the shared feature extraction layer for processing to obtain the shared feature representation; S3, the shared feature representation is input into the pathological homeostasis assessment branch for processing, and the homeostasis category of the object to be assessed is output. The homeostasis category includes homeostasis, metastable and non-homeostasis. S4. Obtain the corresponding center embedding vector according to the center identifier, and combine it with the steady-state category to calibrate the normalized feature vector through the center calibration branch to generate the calibrated feature vector. S5, the shared feature representation, the center embedding vector, the steady-state category, and the calibrated feature vector are input into the joint gating network for processing; the joint gating network determines the gating mode according to the steady-state category, and generates a weight distribution corresponding to multiple expert sub-models based on the shared feature representation, the center embedding vector, and the calibrated feature vector; S6. Based on the weight distribution, the predicted outputs of the multiple expert sub-models obtained from the shared feature representation and the calibrated feature vector are weighted and converged to obtain the pan-cancer classification result of the object to be evaluated.

[0028] In some embodiments, the overall model structure of this method is a dual-constraint collaborative hybrid expert architecture, which realizes pan-cancer classification through the collaborative work between modules. The core lies in reconstructing the logical order and constraint relationship of gating decisions to ensure that the model performs expert judgment only under medically reliable and statistically comparable conditions.

[0029] Please see Figure 2 , Figure 2 This is a schematic diagram of the overall model structure provided in the embodiments of this application. The model structure, from top to bottom, includes an input layer, a data preprocessing module, a shared feature extraction layer, a pathological homeostasis assessment branch, a central calibration branch, a joint gating network, an expert sub-model set, a result aggregation output layer, and a model iteration module. The connections and functions of each module work together to form a complete pan-cancer classification chain.

[0030] The input layer is responsible for receiving the blood test data, basic information, and center identifier of the medical center to be evaluated. The blood test data includes routine blood test indicators and blood biochemistry indicators. The basic information includes age, gender, etc. The center identifier is a unique code of the medical center used to distinguish different data sources.

[0031] The data preprocessing module performs missing value processing, standardization, and combined feature construction on the data received from the input layer to generate standardized feature vectors, while retaining the missing mask and center identification information, providing high-quality, structured input data for subsequent modules.

[0032] The shared feature extraction layer receives standardized feature vectors and a missing mask, and generates a low-dimensional shared feature representation through joint encoding. This shared feature representation can effectively capture common information in blood test data, while being sensitive to missing mechanisms, avoiding misinterpretation of missing patterns as disease signals, and providing a unified feature input for the pathological homeostasis assessment branch and the center calibration branch. This layer adopts a multilayer perceptron structure, containing multiple hidden layers, and uses ReLU as the activation function.

[0033] The pathological homeostasis assessment branch, located after the shared feature extraction layer and before the joint gating network, employs a hybrid architecture combining rules and models. Its inputs include shared feature representations and a subset of key indicators extracted from standardized feature vectors. This branch first performs extreme threshold triggering based on the standardized values ​​of the key indicator subset, directly outputting a non-stationary category for obviously abnormal samples. For other samples, it calculates the statistical features of multiple indicators and fuses them with the shared feature representation, inputting this into a lightweight neural network, such as a fully connected network, for refined evaluation. The final output is a homeostasis confidence score and a homeostasis category, including homeostasis, metastable, and non-stationary. This branch acts as a pre-gating constraint, its core function being to determine whether the current blood indicators are within the pathological homeostasis range suitable for tumor pattern learning and inference. The homeostasis category output by the pathological homeostasis assessment branch serves as the control signal input to the joint gating network and the central calibration branch for downstream modules. Specifically, this homeostasis category determines the calibration strength coefficient α of the central calibration branch, as well as the gating mode and sparsity parameters of the joint gating network, thereby realizing a dynamic inference strategy based on physiological states.

[0034] The central calibration branch also receives the shared feature representation output from the shared feature extraction layer. Employing a fully connected neural network structure, it combines the central identifier with the steady-state category output from the pathological steady-state assessment branch to dynamically calibrate the standardized feature vector, generating a calibrated feature vector and calibration strength coefficients. The calibration process adjusts the calibration strength according to the steady-state category, achieving coordinated modulation of strong steady-state calibration and weak non-steady-state calibration, thus avoiding misjudgments caused by the coupling of non-steady-state anomalies and central differences.

[0035] The joint gating network is the core collaborative module of the model, employing a multilayer perceptron structure. Its inputs include shared feature representations, center embedding vectors, steady-state categories, and calibrated feature vectors. The outputs correspond to the weight distribution, sparsity parameters, and gating pattern labels for multiple expert sub-models. This network follows a three-stage decision-making logic: steady-state constraint determination, center context selection, and expert consistency constraints. It achieves collaborative decision-making between steady-state perception and center adaptation, optimizing the selection and weight allocation of expert sub-models.

[0036] The expert sub-model ensemble comprises common mechanism experts and cancer-specific experts. Common mechanism experts learn consistent tumor-related physiological patterns across medical centers, while cancer-specific experts learn classification patterns for specific cancer types or organ systems. Each expert sub-model takes a shared feature representation and a calibrated feature vector as input and outputs a corresponding prediction. During training, it is subject to cross-center consistency and expert diversity constraints to ensure the model's cross-center generalization ability and the differentiated expression of the expert sub-models. Each expert can employ a multilayer perceptron architecture or a deep neural network architecture such as CNN, RNN, or Transformer.

[0037] The result aggregation output layer performs weighted aggregation of the predicted outputs of the expert sub-models based on the weight distribution generated by the joint gating network to obtain pan-cancer classification results. At the same time, it generates steady-state prompts and center adaptation prompts by combining steady-state categories and center calibration information, thereby improving the medical interpretability and engineering auditability of the model output.

[0038] The model iteration module is responsible for collecting data records generated during online inference, including input data, steady-state categories, weight distribution, pan-cancer classification results, and center identifiers. When the accumulated data records reach the preset conditions, the model is retrained, the parameters of each module are updated, and the model is continuously optimized and iterated.

[0039] In some embodiments, for step S1, data acquisition and preprocessing are key steps in generating standardized feature vectors, providing high-quality input data for subsequent model processing.

[0040] Specifically, the first step is to extract the blood test data of the subject to be evaluated and the center identifier of the affiliated medical center from the testing information system or data warehouse. The blood test data includes complete blood count (CBC) indicators and blood biochemistry indicators. CBC indicators include white blood cell count, red blood cell count, hemoglobin, platelet count, neutrophil percentage, and lymphocyte percentage. Blood biochemistry indicators include alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin, albumin, creatinine, urea, blood glucose, and lactate dehydrogenase (LDH). Specific indicators depend on the actual tests performed at the affiliated medical center. Basic information includes the subject's age and gender. The center identifier is a unique code for the medical center; if this identifier is missing, the subject is set to an unknown status.

[0041] The extracted blood test data is preprocessed, including missing value handling and standardization, and finally a standardized feature vector is generated.

[0042] Optionally, in handling missing values, a missing value mask is retained for each blood test indicator. The missing value mask is used to indicate whether the corresponding indicator is missing; for example, 0 indicates that the indicator is not missing, and 1 indicates that the indicator is missing. For missing indicator values, instead of using a single method of imputation with a uniform constant, the median of the indicator in the training set or the median of the medical center to which it belongs is filled in. This maintains the consistency of the input dimensions, and at the same time, the missing value mask enables the model to distinguish between the true normal indicator values ​​and the imputed values, avoiding the model misinterpreting the imputed values ​​as true physiological state signals.

[0043] Optionally, in terms of standardization, a preset standardization method is used to output standardized values ​​for each indicator. Standardization is performed using global parameters obtained statistically from the training sample set, such as the mean and standard deviation of the training set. The standardized parameters are obtained statistically from the training sample set during offline training and are then embedded into the model version to ensure the consistency and repeatability of the standardization process. The standardization formula can be expressed as:

[0044] in, For the first The standardized value of each indicator, For the first The original values ​​of each indicator The mean of the training set. denoted as the standard deviation of the training set.

[0045] Furthermore, to more comprehensively reflect the physiological state of the object being evaluated, combined features can be constructed. These combined features are formed based on derived ratios or combinations related to inflammation, immunity, and metabolic status, such as the ratio of neutrophils to lymphocytes, or the combination of albumin and inflammatory markers. The specific combination form can be determined according to the indicators in the training set. The constructed combined features are added as derived features to the standardized feature vector to form the final standardized feature vector. This standardized feature vector contains the original standardized indicators and derived features, and is associated with the corresponding missing mask and center identifier.

[0046] In some embodiments, for step S2, the generated standardized feature vector is input into the shared feature extraction layer for processing to obtain a shared feature representation.

[0047] The core function of the shared feature extraction layer is to jointly encode the numerical vectors in the standardized feature vectors with the missing mask to generate a low-dimensional shared feature representation. The algorithm logic of this layer must be sensitive to the missing mechanism; that is, missing patterns will affect the generation of the shared feature representation but will not be misinterpreted as disease signals. For example, the shared feature extraction layer can adopt a multilayer perceptron structure, with each layer containing a predetermined number of neurons. Nonlinear transformations are introduced through activation functions to enhance the model's ability to express features. During training, the shared feature extraction layer trains collaboratively with other modules of the model, adjusting network parameters through backpropagation to ensure that the generated shared feature representation can effectively capture key information in blood test data, providing a reliable feature foundation for subsequent pathological homeostasis assessment, center calibration, and expert sub-model inference.

[0048] Specifically, the input to the shared feature extraction layer is a numerical vector of standardized feature vectors. With missing mask First, the two are concatenated to obtain the input matrix. The input is then subjected to linear transformation and nonlinear activation via the first-layer perceptron. The linear transformation formula is as follows:

[0049] in, This is the weight matrix of the first-layer perceptron. This is the bias vector of the first-layer perceptron. This is the linear output of the first-layer perceptron. Non-linear activation uses the ReLU function, with the following formula:

[0050] in, This is the activation output of the first-layer perceptron.

[0051] Through subsequent iterative processing using a multilayer perceptron, the feature dimensions are gradually compressed, extracting more representative high-level features. The output of the final perceptron layer is the shared feature representation. Its linear transformation and activation process can be expressed as:

[0052]

[0053] in, The total number of layers in the perceptron. This is the weight matrix for the last layer of the perceptron. This is the bias vector for the last layer of the perceptron. This is the activation output of the penultimate perceptron layer. It is a Sigmoid activation function used to map shared feature representations to the [0,1] interval.

[0054] Finally, a fixed-dimensional shared feature representation is output, and this shared feature representation is recorded as an intermediate state for subsequent auditing and debugging.

[0055] In some embodiments, for step S3, the pathological homeostasis assessment branch solves the problem of model misjudgment caused by physiological non-homeostasis factors in the prior art. By pre-judging whether the physiological state of the object to be assessed is suitable for tumor pattern learning and reasoning, it provides a reliable state basis for subsequent model processing and significantly improves the accuracy of pan-cancer classification results.

[0056] This branch employs a hybrid decision architecture, located after the shared feature extraction layer and before the joint gating network, with the shared feature representation as its input. and optional subset vectors of key indicators extracted from the standardized feature vectors. The output is the steady-state confidence score of the object to be evaluated. With steady-state category The steady-state categories include steady state, metastable state, and non-steady state. The steady-state confidence score ranges from [0,1] and is determined by the steady-state confidence score through a threshold.

[0057] Among them, steady state can be used to indicate that the blood test indicators of the subject under evaluation are in a stable physiological range, without acute interfering factors such as acute infection, surgery, or severe inflammation; metastable state can be used to indicate that the blood test indicators have not shown extreme abnormalities, but there is some incoordination in the changes of indicators of different physiological functions; there may be mild non-tumor interfering factors such as mild inflammation, fatigue, and dietary fluctuations, and the physiological state is between stable and abnormal, requiring the interpretation of classification results in conjunction with the clinical context; nonstable state can be used to indicate that the blood test indicators are affected by acute interfering factors, showing significant abnormalities or systematic fluctuations. The changes in indicators are mostly non-tumor-related acute disturbances such as acute infection, postoperative recovery period, and severe organ damage, which cannot reflect the true tumor-related physiological pattern and are prone to false positives in classification.

[0058] The specific hybrid judgment process includes two sequentially executed stages: the first stage is a rule-based extreme threshold-triggered judgment, which first establishes a global extreme threshold for each key indicator. The global extreme threshold can be based on the mean of the training set. ± Double standard deviation Confirmed, among which This is a preset coefficient. The formula for calculating the global extreme threshold can be expressed as:

[0059] Then it is determined whether the sample contains more than the preset number. The standardized value of the index satisfies ,in For the first The standardized values ​​of each indicator. If there are more than a preset number of such indicators, the steady-state category of the sample is directly set to non-steady-state, and a very low steady-state confidence score is assigned to it, such as... Because multiple indicators simultaneously exceeding statistical extreme thresholds usually indicate the presence of an acute event, the changes in blood indicators at this time are mostly non-tumor-related systemic disturbances and are not suitable for tumor classification. If no more than a preset number of such indicators are found, the corresponding sample proceeds to the second stage.

[0060] The second stage involves refined evaluation based on feature fusion and neural networks. First, the indicators in the blood test data are grouped, for example, into inflammation and immune categories. Metabolism of liver and kidneys Injury and stress The groups were divided based on the physiological functions and pathological mechanisms reflected by the indicators. Consistency scores within each group were then calculated. Inconsistency score with inter-group indicators .

[0061] The calculation logic for the intragroup consistency score is as follows: for each group... The indicators are used to calculate the direction of change of the standardized value of each indicator. If all indicators within a group show an increasing or decreasing trend, the intra-group consistency score is high. The specific calculation formula is as follows:

[0062] in, Group The number of indicators in For the first The sign function of the standardized value of an indicator. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0063] The calculation logic for inter-group incoordination scores is as follows: for different groups and The degree of conflict in the direction of change of standardized values ​​between two groups of indicators is calculated. For example, when inflammatory and immune indicators are elevated while metabolic liver and kidney indicators are decreased, the inter-group inconsistency score is higher. The specific calculation formula is as follows:

[0064] in, Group The number of indicators in Group The number of indicators in the data.

[0065] The shared feature representation H is concatenated with all calculated statistical features, namely all intra-group consistency scores and inter-group inconsistency scores, to form a fused feature vector. This fused vector is then input into a lightweight neural network, such as a fully connected network.

[0066] The neural network outputs a steady-state confidence score. Subsequently, based on the threshold optimized on the validation set... and ( Mapping scores to steady-state categories: like If it is determined to be in a steady state, If it is determined to be metastable, This is considered non-steady-state. Regarding the steady-state score threshold setting, during the offline training phase, the output of the pathological steady-state assessment branch is fitted using labeled steady-state and non-steady-state samples. These labeled samples are formed through initial rule screening and manual review. Two thresholds are selected. and , such that when the steady-state confidence score When, it is determined to be in a steady state; when When, it is determined to be a metastable state; when When this occurs, it is determined to be an unsteady state. Threshold and The settings need to be optimized on the validation set to ensure that the non-steady-state misclassification rate is lower than the preset upper limit, while taking into account the steady-state coverage, so that most samples in the true steady state can be accurately identified.

[0067] The steady-state category output by the pathological steady-state assessment branch serves as the control signal input to the joint gating network and central calibration branch of downstream modules, rather than a common feature. Its core function is not to directly provide a disease conclusion, but rather to determine whether the current blood indicators are within a pathological steady-state range suitable for tumor pattern learning and inference, providing crucial information for subsequent central calibration and expert sub-model selection. For example, when a sample is determined to be non-steady-state, the intensity of subsequent central calibration is automatically weakened, and the joint gating network enters a conservative mode to avoid misclassifying non-tumor-related acute abnormalities as tumor features. When a sample is determined to be steady-state, central calibration is fully implemented, and the joint gating network enters a discriminative mode to achieve more accurate pan-cancer classification.

[0068] In some embodiments, for step S4, the center calibration branch addresses the problem of decreased model generalization performance caused by differences across medical centers. By dynamically adjusting the calibration strategy in conjunction with the steady-state category, it achieves effective alignment of data from different centers while avoiding the misjudgment amplification effect caused by the coupling of non-steady-state samples and center differences. This branch's processing requires combining center identifiers and steady-state categories to calibrate the standardized feature vectors, generating calibrated feature vectors.

[0069] Please see Figure 3 , Figure 3 This is a schematic diagram of the center calibration branch processing flow provided in an embodiment of this application. In S301, the corresponding center embedding vector is obtained according to the center identifier. Statistical profile of the center The center embedding vector is a low-dimensional vector representation of the center, learned through offline training, and can effectively characterize the testing characteristics of different medical centers; the center statistical profile includes the central median of each indicator for that center. Statistical information such as the center quantile range and missing rate vector are used to reflect the distribution characteristics of the test data for the center. If the center identifier is unknown, the closest center cluster is estimated through center identification logic. Specifically, the distance between the standardized feature vector of the sample and the statistical profile of each center is calculated, for example, the weighted distance based on the differences in the distribution of key indicators. The center cluster with the smallest distance is selected as the surrogate center, and the center embedding vector and the center statistical profile of the surrogate center are output, and the center inference label is recorded.

[0070] The formula for calculating the center distance is:

[0071] in, The distance between the sample and a certain center. The number of key indicators, For the first The weight of each key indicator, For the first Standardized values ​​of key indicators For the center's first The median of each of the key indicators. As learnable parameters, they are optimized together with other parameters during model training. The training objective is to minimize the distance between known center samples and their true centers, while maximizing the distance to other centers, thereby automatically learning the optimal weight allocation.

[0072] In S302, the calibration intensity coefficient is determined according to the steady-state category. The calibration intensity coefficients are set according to the principle of strong calibration in steady state and weak calibration in unsteady state, where the calibration intensity coefficient corresponding to the steady state category is... Higher than the metastable state or the corresponding non-steady state ,satisfy The calibration intensity parameters need to be optimized on the offline validation set. The optimal parameter combination that minimizes false positives and achieves best cross-center consistency is searched on both steady-state and non-steady-state subsets and then fixed into the model configuration for direct use during online inference. The principle behind this design is that changes in blood indicators in steady-state samples are more likely to reflect the true physiological and pathological state; sufficient calibration allows data from different centers to be mapped to a unified medical semantic space, improving the model's cross-center generalization ability. Conversely, changes in blood indicators in non-steady-state samples are often acute perturbations; weakening the calibration intensity can prevent such acute abnormalities from being calibrated as tumor signals, thus preventing the coupling of acute abnormalities with strong center calibration to form a false positive amplifier.

[0073] In S303, based on the statistical profile corresponding to the center identifier, the calibration offset of each indicator in the standardized feature vector is calculated. The calibration offset can be calculated, for example, as the difference between the central median and the global median, i.e.:

[0074] in, For the first The global median of each indicator. By calibrating the offset, the differences in indicator distribution between different centers can be effectively quantified.

[0075] In S304, the standardized eigenvector is calibrated based on the calibration intensity coefficient and calibration offset to obtain the calibrated eigenvector. The calculation formula for the calibration process is as follows:

[0076] in, For the first The calibrated values ​​of each indicator. For the first The standardized values ​​of each indicator are used. For missing indicators, the missing marker remains unchanged, and the numerical bits are still filled with padding values, but in subsequent gating processing, a missing mask will be used for targeted processing.

[0077] In some embodiments, for step S5, the joint gating network reconstructs the decision-making logic of the traditional gating network. By combining the steady-state category and the center information, it achieves collaborative decision-making, which solves the problems of insufficient generalization performance and misjudgment caused by expert judgment on all samples in the prior art. It significantly improves the robustness and classification accuracy of the model in complex scenarios.

[0078] The network's input includes shared feature representations. , center embedding vector Steady-state categories and the calibrated feature vector The output is the weight distribution corresponding to multiple expert sub-models. The number of experts allowed to be activated, i.e., the sparsity parameter. And gating mode markers, which include discrimination mode, transition mode and conservative mode.

[0079] Please see Figure 4 , Figure 4 This is a schematic diagram of the joint gating network processing flow provided in an embodiment of this application. The specific processing follows a three-stage decision-making logic, namely, steady-state constraint determination, central context selection, and expert consensus constraint.

[0080] In S401, steady-state constraint determination is performed. The joint gating network determines the gating mode based on the input steady-state category and sets the corresponding sparsity parameter. If the steady-state category is steady state, the gating mode is determined to be the discrimination mode, and the first sparsity parameter is set. If the steady-state category is metastable, then the gating mode is determined to be a transitional mode, and a second sparsity parameter is set. ,and If the steady-state category is unsteady, then the gating mode is determined to be conservative, and a third sparsity parameter is set. ,and The sparsity parameter controls the number of expert sub-models with non-zero weights in the weight distribution. For example, the first sparsity parameter can be configured to 2 or 3, emphasizing sparse selection and interpretability; the second sparsity parameter can be configured to 3 or 4 to avoid overfitting by a single expert; the third sparsity parameter can be further increased or forced smoothing to make the weight distribution more uniform and limit any single expert from dominating the output. The sparsity threshold needs to be optimized in offline validation, constrained by the upper limit of the false positive rate of the non-steady-state subset and determined by the sensitivity of the steady-state subset, to ensure that the gating mode and the sparsity parameter settings can adapt to different steady-state states.

[0081] In S402, center context selection is performed. Based on the gating mode and steady-state category, a suitable expert portfolio strategy is selected using the center embedding vector. In the discriminative mode (corresponding to a steady-state category), the center embedding vector is fully utilized to select expert portfolios that have historically performed stably for that center. This is because, after sufficient calibration of test data from different centers in the steady-state state, center-specific expert portfolios can better capture the tumor classification characteristics of that center. In the conservative mode (corresponding to a non-steady-state category), the influence of center differences is mitigated, and experts with common mechanisms and high cross-center consistency are prioritized. This is because the index changes in non-steady-state samples are affected by both center differences and acute perturbations, and experts with common mechanisms can better avoid such interference and reduce misjudgments. In the transitional mode (corresponding to a metastable steady-state category), an expert selection strategy between the discriminative and conservative modes is adopted to balance center specificity and cross-center consistency.

[0082] In S403, expert consistency constraints are applied. Constraints are imposed on the output or representation of expert sub-models based on the steady-state category. For steady-state samples, the expert representations of samples with similar physiological states from different centers are forced to remain consistent. Specifically, in the same training batch, sample pairs from different centers but with similar ranges of key demographics and baseline indicators are selected, and their representation distance among experts on common mechanisms is constrained to be less than a preset consistency threshold. The formula for calculating the representation distance between sample pairs is:

[0083] in, and Samples With sample In the Intermediate representation of a common mechanism expert The value is L2 norm. This constraint enhances the model's cross-center generalization ability and suppresses the learning of center-specific irrelevant features by expert sub-models. For non-stationary samples, it allows for disagreements between expert sub-models, but restricts the output of any single expert sub-model from dominating, thus avoiding bias in the overall classification results due to misjudgment by a single expert.

[0084] In S404, the weight distribution is generated. Regarding weight distribution generation, the joint gating network first outputs the original weight vector through a multilayer perceptron. The input to the multilayer perceptron is a concatenated matrix of shared feature representation, center embedding vector, steady-state class encoding, and calibrated feature vectors. ,in This is the one-hot encoding vector for the steady-state category. The calculation process for the original weight vector is as follows:

[0085]

[0086] in, Here is the weight matrix of the gated network perceptron. For bias vectors, This is the Softmax activation function, used to map the original output to a probability distribution.

[0087] Then based on the sparsity parameter Before keeping The expert submodel with the largest weight is selected, and the weights of the remaining expert submodels are reset to zero and renormalized to obtain a sparse weight distribution. The normalization formula is:

[0088] in, For the first Ranking of experts by their original weights. This is an indicator function.

[0089] When the gating mode is conservative, the sparse weight distribution also needs to be smoothed. The formula for smoothing is:

[0090] in, For the smoothed first The weight of each expert, The smoothing coefficient can be selected from a preset value or derived from the steady-state confidence score. Mapped to obtain, This represents the total number of expert sub-models. Smoothing processes can make the weight distribution of non-steady-state samples more uniform, further reducing the risk of misjudgment caused by a single expert sub-model dominating the output.

[0091] In some embodiments, for step S6, the expert sub-model set and result aggregation process is the core execution link for realizing pan-cancer classification. By reasonably dividing expert roles and weighted aggregation strategies, the advantages of each expert sub-model are fully utilized to ensure the accuracy and reliability of pan-cancer classification results.

[0092] Multiple expert sub-models include common mechanism experts and cancer-specific experts. Common mechanism experts are used to learn consistent tumor-related physiological patterns across medical centers, such as common physiological mechanisms related to tumor development and progression, such as inflammation, immunity, and metabolism. These expert sub-models have strong cross-center robustness. Cancer-specific experts are used to learn classification patterns for specific cancer types or specific organ systems, such as the specific classification characteristics of common cancers like lung cancer, gastric cancer, and colorectal cancer, or the tumor classification patterns of organ systems such as the respiratory and digestive systems. These expert sub-models have strong class discrimination.

[0093] The input to each expert sub-model is a shared feature representation. With the calibrated feature vector Some expert sub-models can select a subset of input features based on their role. For example, a common mechanism expert can focus on a subset of features related to inflammation, immunity, and metabolism, while a cancer-specific expert can focus on a subset of features related to the corresponding cancer type. The structure of expert sub-models can employ a multilayer perceptron architecture or a deep neural network structure such as CNN, RNN, or Transformer. Taking the common mechanism expert as an example, its inference process is as follows:

[0094]

[0095]

[0096]

[0097] in, For the first The input feature vector of each expert sub-model and The first The expert sub-model Layer weight matrix and bias vector and The first The linear output and activation output of the layer, The number of layers in the expert sub-model. For the first The predicted output of each expert sub-model.

[0098] During training, the expert sub-models are trained collaboratively with the shared feature extraction layer, the pathological homeostasis assessment branch, the center calibration branch, and the joint gating network, while simultaneously applying cross-center consistency constraints and expert diversity constraints. The cross-center consistency constraint, as previously described, is used to enhance the model's cross-center generalization ability; the expert diversity constraint is used to limit the correlation between the outputs or representations of different expert sub-models to below a preset upper limit. The formula for calculating expert relevance is:

[0099] in, The Pearson correlation coefficient is used as a constraint to prevent expert sub-models from collapsing into the same function, ensuring that each expert sub-model can capture tumor classification features from different perspectives.

[0100] Each expert sub-model processes the input features and outputs the corresponding prediction. The prediction output can be a multi-class probability vector or a risk score vector. The multi-class probability vector is used to represent the probability that the object to be evaluated belongs to each type of cancer or health status, while the risk score vector is used to represent the pan-cancer risk level of the object to be evaluated.

[0101] Based on the weight distribution generated by the joint gating network The predicted outputs of multiple expert sub-models are weighted and aggregated to obtain the pan-cancer classification results of the object to be evaluated. The formula for weighted aggregation is:

[0102] in, For the first The weights of each expert sub-model For the first The predicted output of each expert sub-model. This represents the total number of expert sub-models. Weighted aggregation can integrate the prediction results of each expert sub-model, fully utilizing their respective strengths to improve the accuracy and reliability of pan-cancer classification results.

[0103] In one embodiment, the model also includes explanatory information output. This explanatory information output enhances the model's medical interpretability and engineering auditability, enabling physicians or relevant personnel to better understand and trust the pan-cancer classification results, thus providing a more comprehensive reference for clinical decision-making.

[0104] This step generates steady-state indications associated with the pan-cancer classification results based on the steady-state category. If the steady-state category is non-steady-state, the output result is marked as affected by non-steady-state factors, and the interpretation level of the risk result is downgraded to "requiring re-examination and follow-up reference," indicating to the physician that the classification result may be affected by non-tumor factors such as acute infection or surgery, and re-examination should be performed after the physiological state of the person being evaluated has returned to steady-state. If the steady-state category is metastable, the output suggests interpreting the result in conjunction with the clinical status, indicating that the physician needs to comprehensively interpret the classification result in conjunction with the specific clinical symptoms, medical history, and other information of the person being evaluated. If the steady-state category is steady-state, the output is a normal interpretation level, indicating that the classification result has high reliability.

[0105] In one embodiment, a model training process is also included. The model training process includes offline training and continuous iterative training. The core objective is to optimize the parameters of each module through multi-center sample data, so that the model has accurate pan-cancer classification ability, cross-center generalization ability, and resistance to non-steady-state interference.

[0106] The offline training phase first constructs a multi-center training sample set. This set comprises blood routine and blood biochemistry test data and corresponding clinical diagnostic information from multiple medical centers, covering healthy individuals undergoing checkups and patients diagnosed with various cancers, ensuring the diversity and representativeness of the sample set. Simultaneously, center codes and center statistical profiles are defined. Center codes are obtained through embedding learning, while center statistical profiles are generated based on the sample data from each center, including information such as median, quantile range, and missing rate.

[0107] The parameters of each module are then trained sequentially. The shared feature extraction layer adopts a fully connected neural network structure, taking the preprocessed standardized feature vector and the missing mask as input, and optimizing the weights and bias parameters through backpropagation, so that the generated shared feature representation can effectively capture common information across samples, while being sensitive to the missing mechanism.

[0108] The training of the pathological homeostasis assessment branch is based on labeled homeostasis and non-homeostasis samples. The branch network parameters are optimized so that the output homeostasis confidence score and homeostasis category can accurately reflect the physiological state of the sample. During the training process, extreme threshold triggering conditions and multi-category index consistency inconsistency discrimination logic are combined to ensure the reliability and interpretability of homeostasis judgment.

[0109] The training of the central calibration branch learns the central embedding vector and calibration parameters through multi-center sample data, so that the calibrated data can form a unified medical semantic space across different centers. At the same time, the calibration intensity is adjusted according to the different steady-state categories, and the calibration intensity coefficient is optimized during the training process to ensure sufficient calibration of steady-state samples and weak calibration effect of non-steady-state samples.

[0110] The training of the joint gating network combines steady-state categories, center embedding vectors, and calibrated feature vectors to optimize the gating network parameters, enabling the weight distribution and gating mode to adapt to different steady-state states and center scenarios. A three-stage decision logic constraint is introduced during the training process to ensure the rationality and synergy of gating decisions.

[0111] The training of the expert sub-model set is optimized separately for common mechanism experts and cancer type-specific experts. Common mechanism experts take the cross-center consistent tumor physiological pattern as the learning target, while cancer type-specific experts take the classification characteristics of specific cancer types as the learning target. Cross-center consistency constraints and expert diversity constraints are applied during the training process to avoid overfitting or collapse of the expert sub-models.

[0112] The model's total loss function includes cross-entropy loss, cross-center consistency constraint loss, and expert diversity constraint loss, and the formula is:

[0113] in, Cross-entropy loss is used to optimize classification accuracy. This is a cross-center consistency constraint loss, used to enhance the model's cross-center generalization ability; The expert diversity constraint loss is used to ensure the differentiation of expert sub-models; and The weighting coefficients for constraining the loss are determined through optimization on the validation set.

[0114] The training uses a stochastic gradient descent optimizer, and the learning rate is dynamically adjusted using a cosine annealing strategy. During training, the model performance is evaluated periodically on the validation set. When the performance metrics stabilize or reach the preset target, training is stopped, the model parameters are determined, and the model version is released after the threshold and stratification strategies are set.

[0115] The continuous iterative training phase is based on data records collected during online inference. When the accumulated number of samples reaches a preset threshold or the model performance index drops to a preset lower limit, model retraining is triggered. During retraining, newly collected sample data is added to the training set, and the parameters of each module are updated using incremental training to maintain the model's adaptability and performance stability. At the same time, version management ensures the model's traceability and auditability.

[0116] The model iteration process enables the model to continuously adapt to new medical environments and data distributions, constantly improving the model's generalization ability and classification performance, and ensuring the model's reliability and effectiveness in long-term use.

[0117] During online inference, data records are collected, including key information such as input data, steady-state categories, weight distribution, pan-cancer classification results, and explanatory information. These data records are stored in a designated database for subsequent model iterations.

[0118] When the accumulated data records reach a preset condition, model retraining is triggered. The preset condition, for example, is that the accumulated number of data records reaches a preset threshold. Or the model deployment time reaches the preset cycle. Or, the model's performance metrics during online inference, such as accuracy, recall, and AUC, are lower than a preset threshold. .

[0119] During model retraining, the accumulated data records are used to update the training sample set, and the parameters of the shared feature extraction layer, pathological homeostasis assessment branch, center calibration branch, joint gating network, and expert sub-model are retrained. The training process uses a cross-entropy loss function combined with cross-center consistency constraint loss and expert diversity constraint loss, and the total loss function is consistent with that of the offline training stage.

[0120] Training employs a stochastic gradient descent optimizer, with the learning rate dynamically adjusted using cosine annealing. Valid parameter information from previous training is retained during training, and incremental training is used to update model parameters. After training, the new model is validated offline on the test set to assess its performance metrics. If the performance metrics meet preset requirements, a new model version is released, replacing the original version, thus achieving continuous iterative optimization. If the performance metrics do not meet preset requirements, the reasons are analyzed, the training strategy is adjusted, and training is repeated.

[0121] This method effectively addresses the technical challenges of cross-center differences, interference from non-steady-state physiological states, and the amplification of misjudgments due to the coupling of these factors through the collaborative design of a pathological homeostasis assessment branch, a central calibration branch, and a joint gating network. The pathological homeostasis assessment branch enables preliminary discrimination of the physiological state of samples, providing a reliable basis for subsequent processing. The central calibration branch dynamically adjusts the calibration strategy based on the homeostasis category, achieving effective alignment of data from different centers. The joint gating network combines homeostasis category and center information for collaborative decision-making, optimizing the selection and weight allocation of expert sub-models. Through the collaborative work of each module and the iterative training process of the system, this method significantly improves the accuracy, cross-center generalization ability, and medical interpretability of pan-cancer classification.

[0122] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0123] Please see Figure 5 , Figure 5 A schematic diagram of a pan-cancer classification and screening system based on blood test biomarkers and an AI model, provided as an embodiment of this application. As shown in the figure, the system 500 includes: The data acquisition and preprocessing module 501 is used to acquire the blood test data of the object to be evaluated and the center identifier of the medical center to which it belongs, and to preprocess the blood test data to generate a standardized feature vector. The shared feature extraction module 502 is used to input the standardized feature vector into the shared feature extraction layer for processing to obtain the shared feature representation; The pathological homeostasis assessment module 503 is used to input the shared feature representation into the pathological homeostasis assessment branch for processing and output the homeostasis category of the object to be assessed, wherein the homeostasis category includes homeostasis, metastable and non-homeostasis. The center calibration module 504 is used to obtain the corresponding center embedding vector according to the center identifier, and combine it with the steady state category to calibrate the standardized feature vector through the center calibration branch to generate a calibrated feature vector; The joint gating module 505 is used to input the shared feature representation, the center embedding vector, the steady-state category, and the calibrated feature vector into the joint gating network for processing; the joint gating network determines the gating mode according to the steady-state category, and generates a weight distribution corresponding to multiple expert sub-models based on the shared feature representation, the center embedding vector, and the calibrated feature vector. The classification result generation module 506 is used to perform weighted aggregation of the prediction outputs obtained by the multiple expert sub-models based on the shared feature representation and the calibrated feature vector according to the weight distribution, so as to obtain the pan-cancer classification result of the object to be evaluated.

[0124] Those skilled in the art will clearly understand that the technical solutions of the embodiments of this application can be implemented by means of software and / or hardware. In this specification, "unit" and "module" refer to software and / or hardware that can independently complete or cooperate with other components to complete a specific function, wherein the hardware may be, for example, a field-programmable gate array (FPGA), an integrated circuit (IC), etc.

[0125] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0126] Please see Figure 6 It shows a schematic diagram of the structure of an electronic device according to an embodiment of this application, which can be used to implement... Figure 1 The method in the illustrated embodiment. (As shown) Figure 6 As shown, the electronic device may include: The system includes at least one processor 601, at least one network interface 604, a user interface 603, a memory 605, and at least one communication bus 602. The communication bus 602 is used to enable connection and communication between the components. The user interface 603 may include buttons, and optionally include a standard wired or wireless interface. The network interface 604 may include, but is not limited to, a Bluetooth module, an NFC module, a Wi-Fi module, etc.

[0127] The processor 601 may include one or more processing cores and connect to various parts within the electronic device 600 via various interfaces and lines. It implements the various functions and data processing of the device 600 by running or executing instructions, programs, code sets, or instruction sets stored in the memory 605, and by accessing data in the memory 605. Optionally, the processor 601 may be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor 601 may also integrate one or more combinations of CPU, GPU, and modem.

[0128] Memory 605 may include random access memory (RAM) or read-only memory (ROM). Optionally, memory 605 includes a non-transitory computer-readable medium for storing instructions, programs, code, code sets, or instruction sets. Memory 605 may be divided into a program storage area and a data storage area, wherein the program storage area can be used to store instructions for implementing an operating system and instructions for implementing the foregoing method embodiments; the data storage area can be used to store data related to the relevant method embodiments. Memory 605 may also be at least one storage device located remotely from processor 601. Figure 6 As shown, the memory 605, which serves as a computer storage medium, may contain an operating system, a network communication module, a user interface module, and program instructions.

[0129] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by processor 601, it performs the functions defined in the methods of this application.

[0130] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application.

[0131] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A pan-cancer classification and screening method based on blood test biomarkers and AI models, characterized in that, include: Obtain the blood test data of the subject to be evaluated and the center identifier of the medical center to which it belongs, and preprocess the blood test data to generate a standardized feature vector; The standardized feature vector is input into the shared feature extraction layer for processing to obtain the shared feature representation; The shared feature representation and a subset of key indicators selected from the standardized feature vector are input into the pathological homeostasis assessment branch for processing, and the homeostasis category of the object to be assessed is output. The homeostasis category includes homeostasis, metastable state, and non-homeostasis. The output of the homeostasis category of the object to be assessed includes: a first-stage determination: triggering a non-homeostasis determination based on whether the standardized value of the key indicators in the blood test data exceeds a preset extreme threshold; a second-stage determination: for samples that are not determined to be non-homeostasis in the first stage, the indicators in the blood test data are grouped, and the consistency score of the indicators within the group and the inconsistency score between the indicators of different groups are calculated. Based on the inconsistency score and the shared feature representation, the category is determined to be homeostasis or metastable state. The corresponding center embedding vector is obtained based on the center identifier, and combined with the steady-state category, the standardized feature vector is calibrated through the center calibration branch to generate a calibrated feature vector. Specifically, this includes: determining a calibration intensity coefficient based on the steady-state category, wherein the calibration intensity coefficient corresponding to the steady-state category is higher than the calibration intensity coefficient corresponding to the steady-state category is metastable or non-steady-state; calculating the calibration offset of the index in the standardized feature vector based on the statistical profile corresponding to the center identifier; and calibrating the standardized feature vector according to the calibration intensity coefficient and the calibration offset to obtain the calibrated feature vector. The shared feature representation, the center embedding vector, the steady-state category, and the calibrated feature vector are input into a joint gating network for processing. The joint gating network determines the gating mode based on the steady-state category and generates a weight distribution corresponding to multiple expert sub-models based on the shared feature representation, the center embedding vector, and the calibrated feature vector. Specifically, if the steady-state category is steady-state, the gating mode is determined to be a discriminative mode, and a first sparsity parameter is set. If the steady-state category is metastable, the gating mode is determined to be a transitional mode, and a second sparsity parameter is set, which is greater than the first sparsity parameter. If the steady-state category is non-steady-state, the gating mode is determined to be a conservative mode, and a third sparsity parameter is set, which is greater than the second sparsity parameter. The sparsity parameter controls the number of expert sub-models with non-zero weights in the weight distribution. Based on the weight distribution, the predicted outputs of the multiple expert sub-models, obtained from the shared feature representation and the calibrated feature vector, are weighted and converged to obtain the pan-cancer classification result of the object to be evaluated.

2. The pan-cancer classification and screening method based on blood test biomarkers and AI models according to claim 1, characterized in that, When the gating mode is conservative, after the joint gating network generates the weight distribution corresponding to multiple expert sub-models, it also performs a smoothing process on the weight distribution to make the weight distribution more uniform.

3. The pan-cancer classification and screening method based on blood test biomarkers and AI models according to claim 1, characterized in that, The multiple expert sub-models include common mechanism experts and cancer type-specific experts; wherein, the common mechanism experts are used to learn consistent tumor-related physiological patterns across medical centers, and the cancer type-specific experts are used to learn classification patterns for specific cancer types or specific organ systems.

4. The pan-cancer classification and screening method based on blood test biomarkers and AI models according to claim 3, characterized in that, During the training of the expert sub-model, a cross-center consistency constraint is applied. The cross-center consistency constraint is as follows: for sample pairs from different medical centers with similar basic characteristics, the intermediate representation distance of the sample pairs in the common mechanism experts is constrained to be less than a preset consistency threshold.

5. A pan-cancer classification and screening system based on blood test biomarkers and AI models, characterized in that, include: The data acquisition and preprocessing module is used to acquire the blood test data of the object to be evaluated and the center identifier of the medical center to which it belongs, and to preprocess the blood test data to generate a standardized feature vector. The shared feature extraction module is used to input the standardized feature vector into the shared feature extraction layer for processing to obtain the shared feature representation; The pathological homeostasis assessment module is used to process the shared feature representation and a subset of key indicators selected from the standardized feature vector into the pathological homeostasis assessment branch, and output the homeostasis category of the object to be assessed. The homeostasis category includes homeostasis, metastable state, and non-homeostasis. The output of the homeostasis category of the object to be assessed includes: a first-stage determination: triggering a non-homeostasis determination based on whether the standardized value of the key indicators in the blood test data exceeds a preset extreme threshold; a second-stage determination: for samples that are not determined to be non-homeostasis in the first stage, the indicators in the blood test data are grouped, and the consistency score of the indicators within the group and the inconsistency score between the indicators of different groups are calculated. Based on the inconsistency score and the shared feature representation, the category is determined to be homeostasis or metastable state. The center calibration module is used to obtain the corresponding center embedding vector according to the center identifier, and, in conjunction with the steady-state category, calibrate the standardized feature vector through the center calibration branch to generate a calibrated feature vector. Specifically, it includes: determining a calibration intensity coefficient according to the steady-state category, wherein the calibration intensity coefficient corresponding to the steady-state category is higher than the calibration intensity coefficient corresponding to the steady-state category is metastable or non-steady-state; calculating the calibration offset of the index in the standardized feature vector based on the statistical profile corresponding to the center identifier; and calibrating the standardized feature vector according to the calibration intensity coefficient and the calibration offset to obtain the calibrated feature vector. A joint gating module is used to input the shared feature representation, the center embedding vector, the steady-state category, and the calibrated feature vector into a joint gating network for processing. The joint gating network determines the gating mode based on the steady-state category and generates a weight distribution corresponding to multiple expert sub-models based on the shared feature representation, the center embedding vector, and the calibrated feature vector. Specifically, if the steady-state category is steady-state, the gating mode is determined to be a discriminative mode, and a first sparsity parameter is set; if the steady-state category is metastable, the gating mode is determined to be a transitional mode, and a second sparsity parameter is set, which is greater than the first sparsity parameter; if the steady-state category is non-steady-state, the gating mode is determined to be a conservative mode, and a third sparsity parameter is set, which is greater than the second sparsity parameter. The sparsity parameter controls the number of expert sub-models with non-zero weights in the weight distribution. The classification result generation module is used to perform weighted aggregation of the prediction outputs obtained by the multiple expert sub-models based on the shared feature representation and the calibrated feature vector according to the weight distribution, so as to obtain the pan-cancer classification result of the object to be evaluated.

6. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 4.

7. A storage medium storing computer program instructions thereon, characterized in that, The computer program instructions can be executed by a processor to implement the method as described in any one of claims 1 to 4.

Images