A method and system for constructing a cancer early screening model based on difference statistics

Abstract

The application discloses a kind of based on the construction method and system of cancer early screening model of difference statistics, belong to cancer model training technical field, specifically include: first, the circulating free DNA original feature of normal sample and cancer sample is extracted, and the normal sample in high cancer proportion cluster is filtered after clustering integration as target group;By comparing the tree model feature importance analysis of target group and the remaining normal group, cancer group and the remaining normal group, obtain first feature set and second feature set;Common feature set is acquired, the common feature set is eliminated from the second undetermined feature set, to obtain target feature set, based on the initial classification model of target feature set is constructed;Incremental training is carried out to each common feature in turn, to obtain corresponding incremental model and performance score, and select the incremental model of performance score as final model, the method is compared by clustering guided sample screening and differentiating feature analysis, effectively improve the distinguishing ability and generalization performance of cancer detection model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cancer model training, and in particular to a method and system for constructing a cancer early screening model based on difference statistics. BACKGROUND

[0002] In vivo, with the process of apoptosis, DNA fragments in cells are released into the plasma to become cell-free DNA (cfDNA). In the early stage of cancer, when the patient has not yet shown obvious clinical symptoms, the state of intracellular DNA has already changed. These DNA are released into the plasma, so that the plasma cfDNA contains information related to cancer. By extracting and processing this information, non-invasive diagnosis of cancer can be achieved, and early diagnosis and treatment of cancer can be realized.

[0003] The mainstream idea of current cancer cfDNA related research is to infer the occurrence of cancer by using mutations of single gene or a few genes related to cancer on cfDNA. In early cancer screening, the effective combination of cfDNA fragmentomic features (including fragment size, fragment end base motif, fragment breakpoint base motif and nucleosome footprint, etc.) is crucial to improve the accuracy of screening.

[0004] Since cancer data usually has high-dimensional characteristics, that is, the number of features is much larger than the number of samples, the complexity of the model will be significantly increased, too many features will increase the cost of calculation, leading to problems such as dimension disaster, model overfitting and affecting the generalization ability of the model. Feature selection is mainly to select the most informative variables from high-dimensional data and eliminate redundancy and reduce noise, so feature selection is very important for cancer data analysis. Effective feature selection method can not only improve the classification performance of the classifier and the prediction accuracy of the survival analysis model, but also find some biomarkers closely related to the disease, providing a more reliable basis for decision-making medical problems.

[0005] However, there is a lack of intuitive and effective means to evaluate high-dimensional feature combinations. In the prior art, traditional feature selection methods cannot intuitively show the influence of feature combinations on sample classification, and cannot quickly judge the effectiveness of feature combinations from the perspective of visualization. Although t-SNE algorithm can be used for high-dimensional data visualization, in the specific application of cfDNA fragmentomic feature combination evaluation, there is a lack of systematic process and optimization strategy for this field, which leads to the inability to accurately evaluate the actual value of feature combinations in cancer early screening. SUMMARY

[0006] The purpose of the present application is to provide a method and system for constructing a cancer early screening model based on difference statistics, which solves the following technical problems:

[0007] Traditional feature selection methods are difficult to intuitively show the influence of feature combination on sample classification, and cannot quickly judge the effectiveness of feature combination from the visual angle.

[0008] The object of the application can be achieved by the following technical solutions:

[0009] A method for constructing a cancer early screening model based on difference statistics, comprising the following steps:

[0010] S1, obtaining a normal sample group and a cancer sample group, wherein the normal sample group and the cancer sample group respectively contain a plurality of sample data, and for each sample data, its corresponding circulating free DNA original feature set is extracted;

[0011] S2, integrating the original feature sets of the normal sample group and the cancer sample group to construct a joint data set containing all sample features, performing clustering on the joint data set to obtain a plurality of category clusters, statistically analyzing the proportion of cancer samples in any category cluster, sorting all category clusters according to the proportion, and selecting the top N category clusters as target category clusters; N is a preset number threshold;

[0012] S3, removing the target sample group from the normal sample group to obtain a comparison sample group, inputting the target sample group and the comparison sample group into a preset tree model to obtain an important score of any original feature, and marking the original feature with an important score greater than or equal to a preset threshold as a pending feature to obtain a first pending feature set; inputting the cancer sample group and the comparison sample group into the preset tree model and repeating the above process to obtain a second pending feature set;

[0013] S4, obtaining a common feature set of the first pending feature set and the second pending feature set, removing the common feature set from the second pending feature set to obtain a target feature set, and constructing an initial classification model based on the target feature set;

[0014] S5, based on the initial classification model, performing incremental training on each common feature in the common feature set to obtain an incremental model and a performance index corresponding to each common feature, the performance index including accuracy, AUC and loss value, determining the performance score of all incremental models based on the optimal solution distance method, and selecting the corresponding incremental model with the maximum performance score as the final model.

[0015] As a further scheme of the application: in S1, the original feature set includes fragment size distribution parameters, terminal base sequence base frequency features, breakpoint base sequence base frequency features and nucleosome footprint features.

[0016] As a further scheme of the present application: in S2, if the total amount of samples in any target cluster is less than the preset total amount of samples, the target cluster is removed.

[0017] As a further scheme of the present application: in S3, the specific acquisition process of the importance score is as follows:

[0018] The target sample group and the contrast sample group are respectively labeled as different group labels, and together with the original feature data, form a labeled data set; the labeled data set is input into a preset tree model, the impurity reduction amount caused by each original feature at the tree node splitting is calculated according to the splitting rule of the tree model, the impurity reduction amounts of the same original feature in all trees are accumulated and normalized, and are arranged in descending order according to the numerical value, and the importance score corresponding to each original feature is obtained.

[0019] As a further scheme of the present application: the specific construction process of the initial classification model is as follows:

[0020] The feature filtering operation is performed on the joint data set, only the original features in the target feature set are retained, all other original features are removed, the filtered joint data set is divided into a training set and a test set according to a preset proportion, and the preset general model is trained based on the training set to obtain an initial classification model.

[0021] As a further scheme of the present application: in S5, the performance score corresponding to the initial classification model is obtained, if the performance score of any incremental model is greater than the performance score corresponding to the initial classification model, the common features corresponding to the incremental model are retained, and if the performance score of any incremental model is less than or equal to the performance score corresponding to the initial classification model, the common features corresponding to the incremental model are removed.

[0022] As a further scheme of the present application: the retained common features are combined by using the exhaustive method to generate a plurality of test groups, the initial classification model is used to perform incremental training on each test group to obtain an incremental model and a performance score corresponding to each test group, and the incremental model corresponding to the maximum performance score is selected as the final model.

[0023] As a further scheme of the present application: if the performance scores of all incremental models are less than or equal to the performance score corresponding to the initial classification model, the initial classification model is directly used as the final model.

[0024] The present application also includes a construction system of a cancer early screening model based on difference statistics, which is used to implement the construction method of the cancer early screening model based on difference statistics.

[0025] a data acquisition module configured to acquire a normal sample group and a cancer sample group, wherein the normal sample group and the cancer sample group each comprise a plurality of sample data, and a corresponding set of original circulating free DNA features is extracted for each sample data;

[0026] a data analysis module configured to integrate the sets of original features of the normal sample group and the cancer sample group, construct a joint data set comprising all sample features, perform clustering on the joint data set to obtain a plurality of category clusters, sort all category clusters according to the proportions of cancer samples in any category cluster, select the first N category clusters as target category clusters, and extract a data set of all normal samples in any target category cluster and label the data set as a target sample group; N is a preset threshold value;

[0027] a data optimization module configured to remove the target sample group from the normal sample group to obtain a comparison sample group, input the target sample group and the comparison sample group into a preset tree model to obtain an importance score of any original feature, label an original feature with an importance score greater than or equal to a preset threshold as a pending feature to obtain a first set of pending features, and input the cancer sample group and the comparison sample group into the preset tree model to repeat the above process to obtain a second set of pending features;

[0028] a model construction module configured to obtain a common feature set of the first set of pending features and the second set of pending features, remove the common feature set from the second set of pending features to obtain a target feature set, and construct an initial classification model based on the target feature set;

[0029] a model optimization module configured to perform incremental training on each common feature in the common feature set based on the initial classification model to obtain an incremental model and a performance indicator corresponding to each common feature, wherein the performance indicator comprises accuracy, AUC and loss value, determine a performance score of all incremental models based on a superior-inferior solution distance method, and select a corresponding incremental model with the largest performance score as a final model.

[0030] Advantages of the present application:

[0031] 1) By clustering the joint data set of normal and cancer samples, selecting target category clusters according to the proportion of cancer samples, and extracting “abnormal” normal samples (i.e., normal samples similar to cancer samples) from the target category clusters, the first set of original features causing feature confusion can be located. This process can effectively identify the feature confusion point of potential false negative samples, avoid misjudgment of cancer caused by abnormal features of normal samples, improve the ability of the model to distinguish true and false positives from the sample selection stage, and solve the problem of insufficient specificity of early screening caused by feature confusion in traditional methods, laying a precise data foundation for subsequent feature elimination and model optimization.

[0032] 2) From the second original feature set screened based on clustering and tree model, the first original feature set with feature confusion is directly removed to remove redundant features that cause sample classification ambiguity. This feature removal strategy based on difference statistics can not only retain key features specific to cancer samples (such as fragment size distribution, terminal base motif, etc.), but also avoid the interference of feature confusion on model classification, so that the model construction focuses more on cancer-specific features, effectively reduces the noise in high-dimensional data, and improves the diagnostic performance of the feature set for cancer samples, providing purer feature input for early screening models.

[0033] 3) After removing the confused features to construct the initial model, the potential useful features hidden in the second original feature set by the confused features can be mined by performing incremental training on the common feature set and combining multiple indicators such as accuracy, AUC, and optimal solution distance method evaluation. This process dynamically evaluates the impact of feature addition on model performance, avoids directly removing potentially useful features, and improves the generalization ability of the model through step-by-step optimization of feature combination, solves the limitations of traditional feature screening "either-or", realizes fine mining of cancer early screening features, and finally forms an optimal feature combination model that balances specificity and sensitivity.

[0034] 4) In the incremental training and feature combination optimization process, the reserved features are combined by exhaustive method, and the nonlinear interaction between features (such as the synergistic effect of fragment size distribution and nucleosome footprint features) is naturally captured based on the splitting rule of tree model, avoiding the loss of interaction information caused by independent evaluation of feature importance in traditional feature screening, thereby effectively mining the combination effect of cancer-related features, such as certain base motif features only showing diagnostic value under certain fragment size conditions, so that the model can more comprehensively depict the complex relationship between cfDNA features and cancer status, further improving the accuracy and robustness of the early screening model. BRIEF DESCRIPTION OF DRAWINGS

[0035] The application will be further described below with reference to the drawings.

[0036] Figure 1 is a construction method flow diagram of a cancer early screening model based on difference statistics of the application. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0038] Please refer to Figure 1As shown, the application is a method for constructing a cancer early screening model based on difference statistics, comprising the following steps:

[0039] S1, obtaining a normal sample group and a cancer sample group, wherein the normal sample group and the cancer sample group respectively contain a plurality of sample data, and for each sample data, its corresponding circulating free DNA original feature set is extracted;

[0040] S2, integrating the original feature sets of the normal sample group and the cancer sample group to construct a joint data set containing all sample features, performing clustering on the joint data set to obtain a plurality of category clusters, statistically calculating the proportion of cancer samples in any category cluster, sorting all category clusters according to the proportion, selecting the first N category clusters as target category clusters, and extracting the data set of all normal samples from any target category cluster and labeling it as a target sample group; N is a preset quantity threshold;

[0041] S3, removing the target sample group from the normal sample group to obtain a comparison sample group, inputting the target sample group and the comparison sample group into a preset tree model to obtain an important score of any original feature, labeling the original feature with an important score greater than or equal to a preset threshold as a pending feature to obtain a first pending feature set; inputting the cancer sample group and the comparison sample group into the preset tree model and repeating the above process to obtain a second pending feature set;

[0042] S4, obtaining a common feature set of the first pending feature set and the second pending feature set, removing the common feature set from the second pending feature set to obtain a target feature set, and constructing an initial classification model based on the target feature set;

[0043] S5, based on the initial classification model, performing incremental training on each common feature in the common feature set to obtain an incremental model and a performance index corresponding to each common feature, wherein the performance index includes accuracy, AUC and loss value, determining the performance score of all incremental models based on the superior-inferior solution distance method, and selecting the corresponding incremental model with the maximum performance score as the final model.

[0044] 1) By clustering the joint data set of normal and cancer samples, selecting target category clusters according to the proportion of cancer samples, and extracting "abnormal" normal samples (i.e. normal samples similar to cancer samples) from the target category clusters, the first original feature set causing feature confusion can be located. This process can effectively identify the feature confusion point of potential false negative samples, avoid cancer misjudgment caused by abnormal normal sample features, improve the ability of the model to distinguish true and false positives from the sample screening stage, solve the problem of insufficient specificity of early screening caused by feature confusion in traditional methods, and lay a precise data foundation for subsequent feature elimination and model optimization.

[0045] 2) From the second original feature set screened based on clustering and tree model, the first original feature set with feature confusion is directly removed to remove redundant features that cause sample classification ambiguity. This feature removal strategy based on difference statistics can not only retain key features specific to cancer samples (such as fragment size distribution, terminal base motif, etc.), but also avoid the interference of feature confusion on model classification, so that the model construction focuses more on cancer-specific features, effectively reduces the noise influence in high-dimensional data, and improves the diagnostic efficiency of the feature set for cancer samples, providing purer feature input for early screening model.

[0046] 3) After removing the confused features to construct the initial model, the potential useful features hidden in the second original feature set by the confused features are mined by performing incremental training on the common feature set, combining multiple indicators such as accuracy, AUC, and optimal solution distance method evaluation. This process dynamically evaluates the impact of feature addition on model performance, which can avoid directly removing potentially useful features, and can improve the generalization ability of the model by gradually optimizing the feature combination, solve the limitations of traditional feature selection "either-or", realize the fine mining of cancer early screening features, and finally form an optimal feature combination model that takes into account specificity and sensitivity.

[0047] 4) In the incremental training and feature combination optimization process, the reserved features are combined by exhaustive method, and the nonlinear interaction between features (such as the synergistic effect of fragment size distribution and nucleosome footprint features) is naturally captured based on the splitting rule of tree model, avoiding the loss of interaction information caused by independent evaluation of feature importance in traditional feature selection, thereby effectively mining the combination effect of cancer-related features, for example, certain base motif features only show diagnostic value under certain fragment size conditions, so that the model can more comprehensively describe the complex relationship between cfDNA features and cancer status, further improving the accuracy and robustness of the early screening model.

[0048] In a preferred embodiment of the present application, the S1, the original feature set includes fragment size distribution parameters, terminal base sequence frequency features, breakpoint base sequence frequency features, and nucleosome footprint features.

[0049] In another preferred embodiment of the present application, the S2 further includes if the total amount of samples in any target cluster is less than the pre-set total amount of samples, the target category cluster is removed.

[0050] After clustering is performed on the joint dataset, a plurality of category clusters are obtained, and for each category cluster, the total number of samples contained therein is counted. A threshold of the total number of samples is preset, and the threshold is set to ensure that there are sufficient sample quantities in the category cluster to reflect the feature distribution rule. For example, according to a common statistical sampling principle, if the sample quantity is too small, the feature distribution may have contingency and cannot represent the overall situation. When the total number of samples in a target category cluster is less than the preset threshold, it indicates that the sample quantity in the cluster is insufficient to reliably reflect the feature mode of the category cluster, and therefore the cluster is removed.

[0051] The retained target category cluster is ensured to have sufficient sample quantity, so that the target sample group extracted from the target category cluster is more representative. Because only the sample quantity is sufficient, it can be ensured that the feature distribution of the category cluster is stable and not formed accidentally. This helps the final purpose of the scheme in avoiding the deviation of the extracted target sample group due to the insufficient sample quantity of the category cluster, and then affecting the accuracy of the subsequent feature importance analysis, and ensuring that the cancer early screening model constructed based on the samples and features can more reliably identify cancer samples and normal samples, and improve the stability and accuracy of the model.

[0052] In another preferred embodiment of the present application, in S3, the specific acquisition process of the importance score is:

[0053] The target sample group and the contrast sample group are respectively labeled as different group labels, and together with the original feature data, a labeled dataset is formed; the labeled dataset is input into a preset tree model, the impurity reduction amount caused by each original feature at the time of tree node splitting is calculated according to the splitting rule of the tree model, the impurity reduction amounts of the same original feature in all trees are accumulated and normalized, and are arranged in descending order according to the numerical value, and the importance score corresponding to each original feature is obtained.

[0054] First, different group labels are assigned to the target sample group and the contrast sample group, such as the target sample group is labeled as "group A" and the contrast sample group is labeled as "group B", and these labels are used to distinguish the attributes of the two types of samples. Then, the group labels are combined with the original feature data (such as fragment size distribution parameters, terminal base sequence base frequency features, etc.) corresponding to each sample to form a labeled data set, so that the tree model can know which group the features of each sample belong to. Next, the labeled data set is input into the preset tree model, and the tree model will split nodes according to the ability of the features to classify samples during construction, and each split hopes to reduce the impurity of the samples in the child nodes after the split. For each original feature, when it is used for tree node splitting, the impurity reduction amount caused by this split is calculated, such as a certain feature splits the node, and the classes of the samples in the child nodes are more concentrated, and the impurity is reduced by a certain degree. Since the tree model can be composed of multiple trees, the impurity reduction amounts generated by the same original feature in each split in all trees need to be accumulated and then normalized, and finally the importance scores are sorted from large to small to obtain the importance scores of each original feature.

[0055] Through the splitting mechanism of the tree model, the importance of each original feature in distinguishing the target sample group and the contrast sample group is objectively quantified. The advantage is that the characteristics of the tree model that are naturally suitable for processing high-dimensional features and can automatically evaluate the importance of the features are used to avoid subjective judgment of the value of the features, making the feature screening process more scientific. For the ultimate purpose of the scheme, after accurately obtaining the feature importance scores, the features that are more critical in distinguishing the "abnormal" group (target sample group) from ordinary normal samples (contrast sample group) can be screened out. These features may reflect the part of the normal samples that is confused with the characteristics of the cancer samples, laying a foundation for subsequent removal of confused features and retention of features that can truly distinguish cancer from normal, so that the constructed early screening model of cancer can more accurately identify sample categories, reduce misjudgments caused by feature confusion, and improve the accuracy and reliability of early screening.

[0056] In another preferred embodiment of the present application, the specific construction process of the initial classification model is as follows:

[0057] Perform a feature filtering operation on the joint data set, only retain the original features in the target feature set, remove all other original features, divide the filtered joint data set into a training set and a test set according to a preset proportion, train the preset general model based on the training set, and obtain an initial classification model.

[0058] When performing feature filtering operations on the joint dataset, the original features in the target feature set are first identified, for example, assuming that the target feature set includes fragment size distribution parameters, specific terminal base sequence base frequency features, etc. Then, for each sample in the joint dataset, only the features in the target feature set are retained, and all other original features not in the target feature set are removed. For example, a sample originally has 100 original features, of which 20 belong to the target feature set, only the 20 features are retained, and the remaining 80 features are removed. Next, the filtered joint dataset is divided into a training set and a test set according to a predetermined ratio (such as the commonly used 7:3 or 8:2). When dividing, random sampling is usually used to ensure that the sample distribution of the training set and the test set is similar to that of the original dataset. Then, the predetermined general model (such as a logistic regression model, a support vector machine model, etc.) is trained based on the training set. During the training process, the model learns the mapping relationship between the target features and the sample categories (normal or cancer), and by continuously adjusting the model parameters, the model can better fit the data on the training set, and finally an initial classification model is obtained. For example, use 70% of the filtered data to train a logistic regression model to enable the model to learn to judge the sample category based on the retained features, and obtain a preliminary classification model.

[0059] By reducing the interference of redundant features through feature filtering, the model focuses on learning the more critical features in the target feature set for distinguishing cancer and normal samples, thereby improving the training efficiency and generalization ability of the model. Dividing the training set and the test set is to evaluate the performance of the model after the model is trained using the test set that did not participate in the training, to ensure that the model has truly learned the feature rules of the samples rather than simply memorizing the training data. Training the general model based on the training set to obtain the initial classification model provides a basis for subsequent incremental training and model optimization. This helps the final goal of the scheme in that the model calculation cost is reduced after reducing the feature dimension, avoiding the curse of dimensionality and overfitting problems caused by high-dimensional features, and by focusing on key features, the initial classification model can more accurately perform preliminary classification on samples, and the evaluation of the test set can ensure the applicability of the model on new data, laying a solid foundation for further optimizing the model performance and building a more accurate cancer early screening model.

[0060] In another preferred embodiment of the present application, in S5, the performance score corresponding to the initial classification model is also obtained, and if there is any incremental model whose performance score is greater than the performance score corresponding to the initial classification model, the common features corresponding to the incremental model are retained; if there is any incremental model whose performance score is less than or equal to the performance score corresponding to the initial classification model, the common features corresponding to the incremental model are excluded.

[0061] After obtaining the initial classification model, its performance score needs to be calculated first. Specifically, the initial model is evaluated using the previously divided test set by calculating performance indicators such as accuracy, AUC value, and loss value on the test set, and then these indicators are converted into a performance score using the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS). For example, the initial model has an accuracy of 80%, an AUC of 0.85, and a loss value of 0.3 on the test set. After TOPSIS processing, a comprehensive score is obtained. Next, each feature in the common feature set is incrementally trained, i.e., each time a common feature is added to the initial model, an incremental model is retrained, and then the performance indicators of each incremental model are evaluated using the test set and converted into corresponding performance scores. For example, there are features A and B in the common feature set. First, feature A is added to train incremental model 1, and after evaluation, a score is obtained. Then feature B is added to train incremental model 2, and after evaluation, another score is obtained. Then compare the performance score of each incremental model with the performance score of the initial model. If the score of a certain incremental model is higher than that of the initial model, it means that adding this common feature improves the performance of the model, so the feature is retained. If the score of a certain incremental model is less than or equal to that of the initial model, it means that the performance of the model does not improve or even decreases after adding this feature, so the feature is removed. For example, the score of incremental model 1 is higher than that of the initial model, so feature A is retained. The score of incremental model 2 is equal to that of the initial model, so feature B is removed.

[0062] By comparing the performance of the incremental model with the initial model, the actual value of each feature in the common feature set to the model is determined, and the features that can truly improve the performance of the model are selected. The advantage is to avoid blindly retaining all common features, and to determine the retention of features through quantitative performance evaluation, to ensure that the finally retained features are all positively affecting the model, and to avoid introducing redundant or ineffective features. This helps the final purpose of the scheme, as the features selected in this way can optimize the feature combination of the model, enabling the model to more accurately identify cancer samples in subsequent training and application, improving the classification performance and generalization ability of the model, and thus building a more accurate and reliable cancer early screening model, improving the accuracy and effectiveness of cancer early screening, and providing stronger support for early diagnosis and treatment of cancer.

[0063] In another preferred embodiment of the present application, the retained common features are combined using the exhaustive method to generate several experimental groups, and incremental training is performed on each experimental group based on the initial classification model to obtain an incremental model and a performance score corresponding to each experimental group. The corresponding incremental model with the largest performance score is selected as the final model.

[0064] When the common features are preliminarily screened and some features are reserved by comparing the performance scores of the incremental model and the initial model, the reserved common features are combined by using the exhaustive method. The principle of the exhaustive method is to arrange and combine all the reserved features in full, to generate all possible feature combination conditions. For example, if three features A, B and C are reserved, the exhaustive method will generate {A}, {B}, {C}, {A, B}, {A, C}, {B, C}, {A, B, C} these different test groups. After generating several test groups, the incremental training is performed on each test group based on the initial classification model, that is, the features in each test group are added to the initial model each time, the model is retrained using the training set, and the incremental model corresponding to each test group is obtained. Then, the test set is used to evaluate each incremental model, and the performance indicators such as accuracy, AUC value and loss value are calculated, and then the indicators are comprehensively converted into the performance score of each incremental model by using the distance method of superior and inferior solutions. Finally, the performance scores of all incremental models are compared, and the incremental model with the highest performance score is selected as the final model. For example, if the incremental model corresponding to the test group {A, B} has the highest performance score, the model is determined as the final model for early screening of cancer.

[0065] By using the exhaustive method to comprehensively explore all possible combinations of the reserved features, the feature combination mode with the most significant performance improvement of the model is found. Because the improvement of a single feature to the model may be limited, and there may be interactions between different features, the combination of different features may produce a synergistic effect, thereby further improving the classification performance of the model. The advantage is that the exhaustive method can ensure that no effective feature combination is missed, and by systematically evaluating each combination, the optimal feature combination scheme is determined from a quantitative point of view, avoiding the limitations of selecting feature combinations by experience or subjective judgment. This helps the final purpose in that by finding the optimal feature combination, the final model can more fully utilize the feature information in the cfDNA, capture more complex feature correlations between cancer samples and normal samples, thereby improving the classification accuracy and generalization ability of the model, so that the constructed early screening model of cancer can more accurately identify cancer samples in actual application, reduce the missed diagnosis and misdiagnosis, and provide more reliable technical support for early screening of cancer.

[0066] In another preferred embodiment of the present application, if the performance scores of all the incremental models are less than or equal to the performance score corresponding to the initial classification model, the initial classification model is directly used as the final model.

[0067] After the performance scores of the incremental models corresponding to each feature in the common feature set are obtained, the performance scores of all the incremental models are compared with the performance score of the initial classification model one by one. If the performance scores of all the incremental models do not exceed the performance score of the initial classification model, that is, the score of each incremental model is less than or equal to the score of the initial model, it indicates that none of the currently screened common features or the addition of any single feature can improve the performance of the model, and may even cause the performance to remain unchanged or decrease. At this time, based on the principle of "not doing invalid optimization", the initial classification model is directly determined as the final model.

[0068] Under the premise of ensuring that the model performance does not decrease, meaningless feature addition and model adjustment are avoided, and redundant features are prevented from being introduced or the model from being over-fitted due to blind optimization. The advantage is that when the incremental training cannot improve the performance of the model, unnecessary optimization steps are stopped in time, the stability and reliability of the initial model are preserved, resource waste is avoided, and the final model used for early screening of cancer can at least maintain the initial classification ability, so that the model will not reduce the distinguishing effect of cancer samples and normal samples due to the forced addition of invalid features, thereby ensuring that the early screening model maintains the required accuracy and stability in actual application, and providing reliable technical support for early screening of cancer.

[0069] The application also includes a system for constructing a cancer early screening model based on difference statistics, which is used to implement the method for constructing a cancer early screening model based on difference statistics.

[0070] The data acquisition module is configured to obtain a normal sample group and a cancer sample group, wherein the normal sample group and the cancer sample group each include a plurality of sample data, and for each sample data, a corresponding circulating cell-free DNA original feature set is extracted.

[0071] The data analysis module is configured to integrate the original feature sets of the normal sample group and the cancer sample group, construct a joint data set containing all sample features, perform clustering on the joint data set, obtain a plurality of category clusters, statistically analyze the proportion of cancer samples in any category cluster, sort all category clusters according to the proportion, select the first N category clusters as target category clusters, and extract a data set of all normal samples from any target category cluster and label the data set as a target sample group. N is a preset quantity threshold.

[0072] The data optimization module is configured to remove the target sample group from the normal sample group to obtain a comparison sample group, input the target sample group and the comparison sample group into a preset tree model to obtain an importance score of any original feature, label an original feature with an importance score greater than or equal to a preset threshold as a pending feature to obtain a first pending feature set, and input the cancer sample group and the comparison sample group into the preset tree model to repeat the above process to obtain a second pending feature set.

[0073] a model construction module configured to obtain a common feature set of the first set of undetermined features and the second set of undetermined features, remove the common feature set from the second set of undetermined features to obtain a target feature set, and construct an initial classification model based on the target feature set;

[0074] a model optimization module configured to perform incremental training on each common feature in the common feature set based on the initial classification model to obtain an incremental model corresponding to each common feature and a performance indicator of each common feature, wherein the performance indicator comprises an accuracy, an AUC, and a loss value, determine a performance score of all the incremental models based on a distance between superior and inferior solutions, and select the incremental model corresponding to the maximum performance score as a final model.

[0075] The above describes one embodiment of the present application in detail, but the content is only a preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made within the scope of the present application should still belong to the patent coverage of the present application.

Claims

1. A method for constructing a cancer early screening model based on differential statistics, characterized in that, Includes the following steps: S1, Obtain normal sample group and cancer sample group, wherein the normal sample group and cancer sample group each contain multiple sample data, and extract the corresponding original feature set of circulating cell-free DNA for each sample data; S2, integrate the original feature sets of the normal sample group and the cancer sample group to construct a joint dataset containing the features of all samples, perform clustering on the joint dataset to obtain several category clusters; count the proportion of cancer samples in any category cluster, sort all category clusters according to the proportion, select the top N category clusters as target category clusters, extract the dataset of all normal samples from any target category cluster and label it as the target sample group; N is a preset quantity threshold; S3, remove the target sample group from the normal sample group to obtain the comparison sample group, input the target sample group and the comparison sample group into the preset tree model to obtain the importance score of any original feature, and label the original features with an importance score greater than or equal to a preset threshold as undetermined features to obtain the first undetermined feature set; Input the cancer sample group and the comparison sample group into the preset tree model, repeat the above process, and obtain the second undetermined feature set; S4, obtain the common feature set of the first undetermined feature set and the second undetermined feature set, remove the common feature set from the second undetermined feature set to obtain the target feature set, and construct an initial classification model based on the target feature set; S5. Based on the initial classification model, perform incremental training on each common feature in the common feature set to obtain the incremental model and performance index corresponding to each common feature. The performance index includes accuracy, AUC and loss value. Determine the performance score of all incremental models based on the best-in-best solution distance method, and select the incremental model with the largest performance score as the final model.

2. The method for constructing a cancer early screening model based on differential statistics according to claim 1, characterized in that, In S1, the original feature set includes fragment size distribution parameters, terminal base sequence fundamental frequency features, breakpoint base sequence fundamental frequency features, and nucleosome footprint features.

3. The method for constructing a cancer early screening model based on differential statistics according to claim 1, characterized in that, S2 further includes removing a target category cluster if the total number of samples in any target cluster is less than the preset total number of samples.

4. The method for constructing a cancer early screening model based on differential statistics according to claim 1, characterized in that, In S3, the specific process for obtaining important scores is as follows: The target sample group and the comparison sample group are labeled with different group labels, and together with the original feature data, they form a labeled dataset. The labeled dataset is input into a preset tree model. According to the splitting rules of the tree model, the amount of impurity reduction caused by the splitting of each original feature at the tree node is calculated. The amount of impurity reduction of the same original feature in all trees is accumulated and normalized. The data are then sorted in descending order according to the numerical value to obtain the importance score corresponding to each original feature.

5. The method for constructing a cancer early screening model based on differential statistics according to claim 1, characterized in that, The specific construction process of the initial classification model is as follows: A feature filtering operation is performed on the joint dataset, retaining only the original features in the target feature set and removing all other original features. The filtered joint dataset is then divided into a training set and a test set according to a preset ratio. A preset general model is trained based on the training set to obtain an initial classification model.

6. The method for constructing a cancer early screening model based on differential statistics according to claim 1, characterized in that, S5 further includes obtaining the performance score corresponding to the initial classification model. If the performance score of any incremental model is greater than the performance score corresponding to the initial classification model, the common feature corresponding to the incremental model is retained. If the performance score of any incremental model is less than or equal to the performance score of the initial classification model, then the common features corresponding to that incremental model are removed.

7. The method for constructing a cancer early screening model based on differential statistics according to claim 6, characterized in that, The common features retained are combined using an exhaustive method to generate several experimental groups. Incremental training is performed on each experimental group based on the initial classification model to obtain the incremental model and performance score for each experimental group. The incremental model with the highest performance score is selected as the final model.

8. The method for constructing a cancer early screening model based on differential statistics according to claim 6, characterized in that, If the performance scores of all incremental models are less than or equal to the performance score of the initial classification model, then the initial classification model is directly used as the final model.

9. A system for constructing a cancer early screening model based on differential statistics, used to implement the method for constructing a cancer early screening model based on differential statistics as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire normal sample groups and cancer sample groups, wherein each normal sample group and cancer sample group contains multiple sample data, and the corresponding original feature set of circulating cell-free DNA is extracted for each sample data. The data analysis module is used to integrate the original feature sets of the normal sample group and the cancer sample group to construct a joint dataset containing the features of all samples. Clustering is performed on the joint dataset to obtain several class clusters. The proportion of cancer samples in any class cluster is counted, and all class clusters are sorted according to the proportion. The top N class clusters are selected as target class clusters. The dataset of all normal samples is extracted from any target class cluster and labeled as the target sample group. N is a preset quantity threshold. The data optimization module is used to remove the target sample group from the normal sample group to obtain the comparison sample group, input the target sample group and the comparison sample group into the preset tree model to obtain the importance score of any original feature, and label the original features with an importance score greater than or equal to a preset threshold as undetermined features to obtain the first undetermined feature set. Input the cancer sample group and the comparison sample group into the preset tree model, repeat the above process, and obtain the second undetermined feature set; The model building module is used to obtain the common feature set of the first undetermined feature set and the second undetermined feature set, remove the common feature set from the second undetermined feature set to obtain the target feature set, and build an initial classification model based on the target feature set; The model optimization module is used to perform incremental training on each common feature in the common feature set based on the initial classification model, to obtain the incremental model and performance index corresponding to each common feature. The performance index includes accuracy, AUC and loss value. The performance score of all incremental models is determined based on the best-in-best solution distance method, and the incremental model with the largest performance score is selected as the final model.

Images