Method and device for predicting ptc lymph node metastasis based on DNA methylation

Abstract

The application discloses a method and device for predicting PTC lymph node metastasis based on DNA methylation, and the method comprises the following steps: setting three-stage IFF-RFE-AE feature dimension reduction technology and an auto-encoder classification model to predict lymph node metastasis of PTC patients; combining an integrated feature filtering (IFF) processing method of correlation coefficient, mutual information and chi-square test to effectively eliminate redundant methylation sites; further using a linear support vector machine recursive feature elimination (SVM-RFE) method to automatically identify an optimal feature subset related to lymph node metastasis, which is helpful to identify features with the most information quantity and correlation in a classification task, and thus improves the accuracy of predicting lymph node metastasis of PTC patients; finally, an auto-encoder (AE) classification model is developed to obtain compressed data representation, reduce feature dimension, and use data enhancement strategies such as three times interpolation and random disturbance in preprocessing to improve the robustness of the classification model.

Description

Technical Field

[0001] This invention relates to machine learning technology, specifically to the field of machine learning that supports neural network models, and specifically to a technique for predicting lymph node metastasis in papillary thyroid carcinoma (PTC) using an end-to-end neural network model. Background Technology

[0002] Papillary thyroid carcinoma (PTC) is the most common type of endocrine malignancy, accounting for approximately 85% of all thyroid cancers. Its 10-year survival rate is approximately 93%, and it is generally considered a relatively indolent tumor. Although most PTC cases exhibit favorable clinicopathological behavior and improved prognosis, some cases demonstrate aggressive behavior and poor prognosis. Some cancer cells metastasize to lymph nodes surrounding the thyroid gland, including central and lateral cervical lymph nodes. Lymph node metastasis is a key indicator of the prognosis, extent, and surgical approach of PTC, and is also an important risk factor for high recurrence rates and low survival. Therefore, accurate prediction of lymph node metastasis in PTC is of great significance.

[0003] DNA methylation, a form of chemical modification of DNA, can alter gene expression without changing the DNA sequence. DNA methylation near and far from regulatory elements (i.e., promoters and enhancers) inhibits transcription by affecting the binding of transcription factors and / or the recruitment of chromatin-modifying enzymes. Changes in methylation in gene regions are closely associated with tumor progression. Hypermethylation in promoter regions can lead to the inactivation of tumor suppressor genes, while hypomethylation in gene regions may promote tumorigenesis and activate genes associated with tumor invasion or metastasis by acting on classical oncogenes. Changes in DNA methylation can be observed in the early stages of tumorigenesis, making them potential biomarkers for cancer detection. DNA methylation has been reported as a fundamental biomarker in various cancers, including thyroid, lung, prostate, and colorectal cancers.

[0004] Invasive histology, extrathyroidal extension, and lymph node metastasis frequently occur in patients with low DNA methylation levels in pneumothyroid carcinoma (PTC). The application of DNA methylation analysis in PTC has attracted significant attention for differentiating between tumor pathological subtypes and benign / malignant tumors. Existing technologies have seen numerous studies applying DNA methylation to predict PTC: 1. Studies have demonstrated that different subtypes of thyroid cancer originating from the same cell line exhibit different DNA methylation profiles, and a prognostic classifier based on 21 DNA methylation probes has been developed to predict adverse outcomes in patients with well-differentiated thyroid cancer; 2. Thyroid tissue-specific characteristics are closely associated with active enhancers and cancer-related genes, leading to the development of accurate and cost-effective epigenetic diagnostic methods based on DNA methylation patterns in benign and malignant thyroid nodules; 3. Genome-wide assessment of tumor-specific differentially methylated CpG sites in thyroid tumors has revealed that low DNA methylation levels of three markers (cg10705422, cg17707274, cg26849382) are significantly associated with recurrence or persistent disease and distant metastasis in patients with differentiated thyroid cancer, etc.

[0005] In summary, although some studies have preliminarily explored the correlation between DNA methylation and PTC lymph node metastasis, an effective predictive tool for PTC lymph node metastasis based on DNA methylation sites has not yet been developed. Furthermore, no method for predicting PTC lymph node metastasis based on DNA methylation sites has been analyzed and validated in terms of accuracy. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, this invention proposes a method and apparatus for predicting lymph node metastasis in patients with primary cancer (PTC) based on DNA methylation. DNA methylation site data of PTC patients are collected from the Cancer Genome Atlas (TCGA) database. Based on these DNA methylation sites, a three-stage IFF-RFE-AE feature reduction technique and an autoencoder classification model are developed to predict lymph node metastasis in PTC patients. First, an integrated feature filtering (IFF) method combining correlation coefficient, mutual information, and chi-square test is used to effectively eliminate redundant methylation sites. Second, a linear support vector machine recursive feature elimination (SVM-RFE) method is used to automatically identify the optimal subset of features associated with lymph node metastasis. Finally, an autoencoder (AE) classification model is developed, which obtains compressed data representation, reduces feature dimensionality, and uses data augmentation strategies such as cubic interpolation and random perturbation to improve model robustness.

[0007] In a first aspect, the present invention proposes a method for predicting PTC lymph node metastasis based on DNA methylation, the method comprising the following steps:

[0008] S1, the DNA methylation site data of m samples and the corresponding tags of the samples are combined into grid data, and data augmentation is performed to obtain enhanced grid data;

[0009] S2. The enhanced grid data is subjected to integrated feature filtering (IFF) to obtain the training dataset;

[0010] S3. Process the training dataset using Linear Support Vector Machine Recursive Feature Elimination (SVM-RFE) to obtain the target dataset;

[0011] S4. Set up an autoencoder classification model (AE) and train the autoencoder classification model on the target dataset;

[0012] S5. Input the target DNA methylation site data into the autoencoder classification model, and output the classification results of the lymph node metastasis.

[0013] Furthermore, specifically, the data augmentation includes cubic interpolation and / or random perturbation.

[0014] Specifically, the data augmentation process of the cubic interpolation includes the following steps:

[0015] 1) Determine the interpolation point;

[0016] 2) For each feature's interpolation location, perform the following cubic spline interpolation:

[0017] a. Sort the data for this feature to ensure that the x value increases;

[0018] b. Calculate the difference h and Δy for each pair of adjacent data points, where h = x(i+1) - x(i) and Δy = y(i+1) - y(i);

[0019] c. Calculate the first derivative (slope) using the difference Δy / h;

[0020] d. Calculate the second derivative value, i.e., the concavity / convexity of the curve;

[0021] e. Calculate the coefficients a, b, c, and d of the cubic spline interpolation;

[0022] 3) Using the calculated coefficients and x values, the corresponding y values ​​can be generated, forming the interpolation curve for each interpolation point;

[0023] 4) For each column of features at the interpolation location points, repeat the above interpolation steps to generate the interpolation curve for each column of feature location points;

[0024] 5) By applying the interpolation curve to each feature location point of each of the n features of each sample, new data points are generated at each feature location point, thereby expanding the dataset;

[0025] And / or, the data augmentation process for the random perturbation specifically includes the following steps:

[0026] 1) Determine the range of random disturbances, including the intensity and range of values ​​of the disturbances;

[0027] 2) For each feature value of the sample, generate a random number, ensuring that the random number is within the selected range;

[0028] 3) Add random noise: Add the random number generated in step 2) to the position corresponding to the feature value;

[0029] 4) For the feature values ​​of one or more samples in the grid data, repeat steps 2 and 3 to add random noise to generate new expanded samples.

[0030] Furthermore, the integrated feature filtering (IFF) includes one or more of correlation coefficient, mutual information, and chi-square test.

[0031] Specifically, the feature filtering based on the correlation coefficient includes: calculating the Pearson correlation coefficient between each feature of each sample in the enhanced grid data and the target category; and filtering features whose Pearson correlation coefficient exceeds the preset threshold according to a predefined threshold.

[0032] Specifically, the mutual information feature filtering includes: calculating the mutual information score between each feature and the target category in the enhanced grid data; and filtering features whose mutual information scores exceed the preset threshold according to a predefined threshold.

[0033] Specifically, the chi-square test for feature filtering includes: calculating the chi-square test value between each feature in the enhanced grid data and the target category; and filtering features whose chi-square test values ​​exceed a preset threshold based on a predefined threshold.

[0034] Furthermore, the linear support vector machine recursive feature elimination (SVM-RFE) process for the training dataset specifically includes:

[0035] 1) Initialize the linear support vector machine model and train the linear support vector machine model using the training dataset;

[0036] 2) Calculate the weight or coefficient of each feature in the training dataset using the trained linear support vector machine model;

[0037] 3) Sort the features according to the weight or coefficient of each feature in the training dataset;

[0038] 4) Remove the least important features and update the feature set of the samples in the training dataset;

[0039] 5) Retrain the linear support vector machine model using the updated feature set;

[0040] 6) Termination condition check: If a sufficient number of features have been removed to meet the preset final feature count, or if a certain stopping condition is met, the algorithm terminates. Otherwise, return to step 3, sort and remove features from the new feature set, and then retrain the model.

[0041] Furthermore, the autoencoder classification model includes an autoencoder and a multilayer perceptron (MLP); the network structure of the multilayer perceptron (MLP) includes an input layer, a hidden layer, and an output layer;

[0042] The training of the autoencoder classification model on the target dataset specifically includes:

[0043] Using the target dataset, the autoencoder is trained by minimizing the reconstruction error; the encoder part maps the data to the latent space.

[0044] Using the latent space of the encoder portion of the autoencoder as input, the MLP is trained by minimizing the classification loss to construct a binary classification model.

[0045] Secondly, this invention proposes a device for predicting PTC lymph node metastasis based on DNA methylation, comprising:

[0046] The preprocessing module is used to combine the DNA methylation site data of m samples and the corresponding tags of the samples into grid data, and perform data augmentation to obtain enhanced grid data.

[0047] The feature filtering module is used to perform integrated feature filtering (IFF) on the enhanced grid data to obtain the training dataset;

[0048] The feature elimination module is used to process the training dataset using linear support vector machine recursive feature elimination (SVM-RFE) to obtain the target dataset;

[0049] The classification model building module is used to set up the autoencoder classification model and train the autoencoder classification model on the target dataset.

[0050] The output module is used to input target DNA methylation site data into the autoencoder classification model and output the classification results of the lymph node metastasis.

[0051] Thirdly, the present invention proposes an electronic device characterized by comprising a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs comprising steps for performing the above-described method for predicting PTC lymph node metastasis based on DNA methylation.

[0052] Fourthly, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting PTC lymph node metastasis based on DNA methylation.

[0053] Based on the present invention, a method and apparatus for predicting lymph node metastasis in PTC based on DNA methylation are provided. The method for predicting lymph node metastasis in PTC based on DNA methylation of the present invention sets up a three-stage IFF-RFE-AE feature dimensionality reduction technology and an autoencoder classification model to predict lymph node metastasis in PTC patients. First, an integrated feature filtering (IFF) method combining correlation coefficient, mutual information, and chi-square test was used to effectively eliminate redundant methylation sites while retaining key feature information. Furthermore, a linear support vector machine recursive feature elimination (SVM-RFE) method was used to automatically identify the optimal feature subset related to lymph node metastasis, which helps to identify the most informative and relevant features in the classification task, thereby improving the accuracy of predicting lymph node metastasis in PTC patients. Simultaneously, an autoencoder (AE) classification model was developed, obtaining compressed data representation and reducing feature dimensionality. Data augmentation strategies such as cubic interpolation and random perturbation were used in preprocessing to improve model robustness. This model also demonstrated a significant reduction in the risk of missed diagnoses. Accurate prediction of lymph node metastasis will contribute to targeted tumor treatment and clinical management. These findings are of great significance for the clinical prediction and treatment of PTC. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a schematic diagram of the overall process of the method provided in the embodiments of the present invention. Detailed Implementation

[0056] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the various embodiments of the following methods, apparatus, and systems, and the features in those embodiments, can be combined with each other.

[0057] In view of the shortcomings of existing technologies, this invention proposes a method and apparatus for predicting lymph node metastasis in papillary thyroid carcinoma (PTC) based on DNA methylation. DNA methylation site data of PTC patients were collected from the Cancer Genome Atlas (TCGA) database. Based on these DNA methylation sites, a three-stage IFF-RFE-AE feature dimensionality reduction technique and an autoencoder classification model were developed to predict whether lymph node metastasis has occurred in PTC patients. Figure 1 The diagram shows the overall flowchart of the method proposed in this invention. First, an integrated feature filtering (IFF) method combining correlation coefficient, mutual information, and chi-square test is used to effectively eliminate redundant methylation sites. Second, a linear support vector machine recursive feature elimination (SVM-RFE) method is used to automatically identify the optimal feature subset related to lymph node metastasis. Finally, an autoencoder (AE) classification model is developed, and a classification network composed of an autoencoder and a multilayer perceptron is trained to obtain optimal feature representation and classification of lymph node metastasis. This achieves compressed data representation, reduces feature dimensionality, and improves model robustness by using data augmentation strategies such as cubic interpolation and random perturbation in preprocessing.

[0058] In a specific embodiment, this invention collected DNA methylation site data from patients with papillary thyroid carcinoma (PTC). The DNA methylation data were all sourced from the Cancer Genome Atlas (TCGA) database, a readily available database for basic research. The dataset for this invention was pre-configured, removing some non-standard samples and discarding data with missing sites to form the method dataset required by this invention. DNA methylation site data from 510 tumor samples from the TCGA database (THCA) were used. Eight cases with metastasis, five cases with a history of neoadjuvant therapy, and eight cases pathologically diagnosed as non-capsular sclerosing carcinoma, eosinophilic adenocarcinoma, or follicular carcinoma were excluded. The remaining cases were screened for lymph node metastasis based on their clinical TNM (tumor, lymph node, and metastasis classification index). Specifically, cases with an N and M index of 0 and a T index of 1 to 4 were classified as non-metastatic (N0), while cases with an N index of 1 and a T index of 1 to 4 were classified as metastatic (N+). After these screening criteria, DNA methylation site data from 447 PTC cases were included in the data of this invention.

[0059] Table 1 summarizes the clinical and histological information of PTC patients involved in this invention. This invention uses β values ​​as a quantitative indicator of the methylation level of each probe. β values ​​range from 0 (completely unmethylated) to 1 (completely methylated). For each sample, any methylated probes containing missing values ​​were discarded.

[0060] Table 1. Clinical and histological characteristics of PTC patients participating in the study.

[0061]

[0062] In the following embodiments of the present invention, the specific exemplary technical solutions are for illustrative purposes only, and the implementation of the technical solutions is not limited thereto.

[0063] First Embodiment

[0064] In one embodiment, the present invention proposes a method for predicting PTC lymph node metastasis based on DNA methylation, the method specifically including the following steps.

[0065] In step S100, the DNA methylation site data of m samples and the corresponding tags of the samples are combined into grid data, and data augmentation is performed to obtain enhanced grid data.

[0066] In one embodiment, for ease of understanding, the DNA methylation site data of the present invention is illustrated using a cell data structure, with each case sample as a unit. This structure contains m samples, each occupying one row in the cell data. Each sample has n features and one corresponding label. The label indicates whether the sample has undergone lymph node metastasis, for example, by marking 0 or 1 to represent no lymph node metastasis, lymph node metastasis, or vice versa. Alternatively, it can be any labeling method that distinguishes between the two, indicating whether metastasis has occurred. This application is not limited to this. Each DNA methylation site data point is a floating-point number, meaning that each sample's features can be floating-point numbers. Furthermore, besides cell data, the DNA methylation data combination structure can also utilize other feature combination methods.

[0067] Preferably, in one embodiment, the DNA methylation site data of the sample can be normalized data or raw DNA methylation site data; the data enhancement can include cubic interpolation and / or random perturbation; the data enhancement method can also be selected specifically according to different data combination types, and other methods or combinations conventional in the art can also be used. In this invention, for the original samples in the grid data, including samples with transfer and samples without transfer, corresponding sample amplification and enhancement processing is performed respectively. In order to facilitate the differentiation of the sample labels corresponding to the samples, this invention preferably performs sample amplification and enhancement separately (divided into two categories) for samples with transfer and samples without transfer.

[0068] In one embodiment, cubic interpolation is a method for estimating intermediate values ​​between known data points using a smooth curve. It calculates the interval between adjacent data points and determines the position and corresponding value of the intermediate point based on a polynomial function. Taking a sample set without transitions as an example, the data augmentation process for samples without transitions using cubic interpolation specifically includes the following steps:

[0069] 1) Determine the interpolation points. To perform interpolation, it's necessary to determine where to perform the interpolation. Typically, interpolation can be performed at the boundary points and midpoints of the original data for each sample; these selected target points become the control points of the interpolation curve. Each sample has n feature values, and one or more feature values ​​at specific locations for each sample need to be selected for interpolation.

[0070] 2) Perform cubic spline interpolation. For each feature, treat it as a one-dimensional interpolation problem. For each column of features at the interpolation point location, perform cubic spline interpolation according to the following steps:

[0071] a. Sort the data for this column of features to ensure that the x values ​​increase sequentially.

[0072] b. Calculate the difference h and Δy for each pair of adjacent data points, where h = x(i+1) - x(i) and Δy = y(i+1) - y(i).

[0073] c. Calculate the first derivative value (slope), usually using the difference Δy / h.

[0074] d. Calculate the second derivative value, i.e. the concavity / convexity of the curve.

[0075] e. Calculate the coefficients a, b, c, and d for cubic spline interpolation. These coefficients will be used to determine the interpolation curve between adjacent data points.

[0076] 3) Generate interpolated data. Using the calculated coefficients and x values, corresponding y values ​​can be generated, forming an interpolation curve for each interpolation point. Multiple interpolation points can be generated to augment the dataset. In this case, interpolation can be performed only on each column of data at some feature points, or interpolation curves can be generated for each column of data at all feature points.

[0077] 4) Apply to all features. For each column of features at the interpolation point location, repeat the above interpolation steps to generate an interpolation curve at each column of feature location points.

[0078] 5) Data Augmentation. The dataset is augmented by applying the corresponding interpolation curve to each feature location of the n features in each sample, generating new data points at each feature location. This can be achieved by inserting interpolation points between the original data points of multiple samples at each feature location. This augments the sample dataset, resulting in enhanced grid data.

[0079] Specifically, the three interpolations described above are applied to the sample dataset with and without transfer, respectively. For example, when the three interpolations are applied to the data samples without transfer, the label of each sample in the resulting expanded sample dataset corresponds to the label of the sample without transfer, which is the same as the label representation method in the previous steps. Similarly, the above method is applied to the sample set with transfer to perform interpolation, resulting in an expanded sample set with transfer.

[0080] The enhanced grid data obtained through the cubic interpolation method described above includes both the original sample dataset and the enhanced dataset, with all grid data being floating-point numbers. By introducing the interpolation points of the cubic interpolation, the sample dataset in the grid data becomes denser, capturing more subtle variations and thus increasing data diversity. This enhancement technique helps alleviate the limitations of a limited dataset, providing more variation patterns for the network model in subsequent steps.

[0081] In one embodiment, specifically, random perturbation is a method of introducing random variations or noise into data. It adds random perturbation values ​​to existing data points, thereby creating variations and introducing variability into the dataset. The data augmentation process for generating each augmented sample of data using random perturbation specifically includes the following steps:

[0082] 1) Determine the range of random disturbances. This range includes the intensity and value range of the disturbance, which can determine that the amplitude of the disturbance does not exceed a certain percentage or absolute value of the characteristic value.

[0083] 2) Generate a random number for each sample feature. For each feature data point or feature value, generate a random number, ensuring that the random number is within the selected range.

[0084] 3) Add random noise: Add the random number generated in step 2) to the position corresponding to the feature value. Specifically, random noise can be generated by adding or multiplying the random number with the original data.

[0085] 4) Repeated application: For all data points or feature values ​​of one or more samples in the grid data, repeat steps 2 and 3 to add random noise to generate augmented samples.

[0086] The enhanced grid data obtained through the aforementioned cubic interpolation method includes both the original sample dataset and the sample dataset composed of the enhanced and expanded data. All resulting grid data are floating-point numbers. By setting the magnitude and range of the perturbation, a balance is struck between introducing sufficient noise to enhance learning and avoiding excessive noise that leads to unrealistic or distorted samples. Random perturbation helps the model learn more robustly and adapt to different variations and noise levels in real-world scenarios.

[0087] Step S200: The enhanced grid data is subjected to integrated feature filtering (IFF) to obtain the training dataset.

[0088] In one embodiment, specifically, integrated feature filtering, also known as comprehensive feature filtering technology, may, exemplarily, include one or more of correlation coefficient, mutual information, and chi-square test; however, the specific implementation of the present invention is not limited thereto, and may also be one or more combinations of other data filtering technologies well known in the art.

[0089] In one embodiment, the correlation coefficient may optionally be used for feature filtering. The correlation coefficient measures the linear correlation between a feature and a categorical variable. If the correlation coefficient is close to 1 or -1, it indicates a strong linear correlation, suggesting that the feature may have high predictive power for the categorical variable and is considered an important feature; specifically, the categorical variable here refers to the label value corresponding to whether lymph node metastasis has occurred, i.e., 0 or 1. Conversely, if the correlation coefficient is close to 0, the feature may have weak predictive power and is considered an unimportant feature. The correlation coefficient is calculated using formula (1). During feature selection, features can be ranked according to the correlation coefficient, and features with higher correlation coefficients can be selected as important features.

[0090]

[0091] Where m represents the sample size, x i and y i Let represent the values ​​of the feature variable (i.e., feature value) and the category variable (transition label value) of the i-th sample, respectively. and These represent the mean of the feature variable and the mean of the categorical variable, respectively. The categorical variable consists of 0 and 1 values. The mean of the feature variable is the sum of the feature values ​​of the m samples divided by the total number of samples. The mean of the categorical variable is the number of samples with category 1 divided by the total number of samples.

[0092] Specifically, for a sample dataset of grid data, feature filtering of the feature correlation coefficients of the samples includes: calculating the Pearson correlation coefficient between each feature in the enhanced grid data and the target category; filtering features whose Pearson correlation coefficients exceed a preset threshold according to a predefined threshold; and forming a training dataset of samples from the filtered features.

[0093] In one embodiment, mutual information may be used for feature filtering. Mutual information is a measure of the correlation between two random variables. In feature selection, mutual information quantifies the degree of information sharing between each feature and the target category. A higher mutual information value indicates a stronger correlation between the feature and the target category, suggesting that the feature contributes more information to the prediction. During feature selection, features with high mutual information can be selected as important features. Mutual information is calculated using formula (2).

[0094]

[0095] Where X and Y are two random variables, P(x,y) is the probability that X=x and Y=y occur simultaneously, and P(x) and P(y) are the probabilities that X=x and Y=y occur respectively. Specifically, X is a set of features, Y is the target category, the marginal distribution probabilities p(x) and p(y) are calculated separately, and then the joint distribution probability p(x,y) is calculated to obtain I(X,Y). For example:

[0096] X=[0,1,2,3,1], Y=[0,1,0,1,0].

[0097] Marginal distribution probabilities: P(x=0)=1 / 5, p(x=1)=2 / 5, p(y=0)=3 / 5, p(y=1)=2 / 5.

[0098] Joint probability distribution: p(x=0,y=0)=1 / 5, p(x=1,y=0)=1 / 5.

[0099] Specifically, feature filtering using mutual information includes: calculating the mutual information score between each feature and the target category in the enhanced grid data; filtering features whose mutual information scores exceed a preset threshold according to a predefined threshold; and forming a training dataset from the filtered features.

[0100] In one embodiment, feature filtering may optionally be performed using the chi-square test. The chi-square test is a statistical method used to assess the correlation between two categorical variables. It calculates a chi-square statistic based on the difference between the observed and expected values ​​to determine whether the two variables are independent or correlated. By calculating the chi-square statistic and obtaining the corresponding p-value, the significance of the association between the feature and the target category can be assessed. A larger chi-square statistic and a smaller p-value indicate a higher correlation between the feature and the target category, suggesting that these features are important for predicting the target variable. The chi-square test is calculated using formula (3).

[0101]

[0102] Where O is the observed frequency, E is the expected frequency, and χ² is the value of the observed frequency. 2 It is the chi-square value. Specifically, for example, a contingency table is constructed, which includes the observation frequencies between each value x of feature X and each category of target variable Y (the sum of all observation frequencies is the total number of samples m), such as O 00 =10. Expected frequency represents the value of feature X under the independence assumption. i The category y of the target variable Y j The expected frequency between the values ​​is usually calculated based on the marginal sum of the contingency table, i.e., E = (row sum) * (column sum) / total number of samples, such as E 00 = (10+20)*(10+30) / 100.

[0103] The table below is an example contingency table, where X represents the feature value, Y represents the category label, and the total number of samples is 10+20+30+40=100.

[0104] Y＝0 Y＝1 X = x1 10 20 X = x2 30 40

[0105] In one embodiment, the chi-square test for feature filtering specifically includes: calculating the chi-square test value between each feature in the enhanced grid data and the target category; filtering features whose chi-square test values ​​exceed a preset threshold according to a predefined threshold; and forming a training dataset from the filtered features.

[0106] In one embodiment, the feature filtering technique may include two or three of the correlation coefficient, mutual information, and chi-square test; features in which both or all three of the correlation coefficient, mutual information score, or chi-square test value exceed the threshold are filtered according to a predefined threshold.

[0107] In gene expression profiling, the sheer number of genes often overwhelms meaningful information with irrelevant, redundant, or noisy data. Therefore, it is necessary to extract or select the most discriminative features from tens of thousands of genes to accurately classify lymph node metastasis and reduce computational costs in subsequent steps. Thus, feature filtering techniques such as correlation coefficients, mutual information, and chi-square tests are used to achieve accurate data simplification.

[0108] Step S300: Process the training dataset using Linear Support Vector Machine Recursive Feature Elimination (SVM-RFE) to obtain the target dataset.

[0109] In one embodiment, specifically, a linear support vector machine (SVM) algorithm and recursive feature elimination technique are used to iteratively eliminate feature data for each sample in the training dataset, and k-fold cross-validation is used to evaluate the importance of each sample's feature data. The process begins by building an initial SVM model using all features and training and predicting, then calculating an importance score for each feature based on the coefficients or weights assigned in the SVM model. Generally, the significance of a feature can be determined by the absolute value of its coefficients. In each iteration, the feature with the lowest score is removed from the feature set, and training and prediction are performed again using the remaining features. After each round of feature elimination, the performance of the feature subset is evaluated using an objective function (typically via k-fold cross-validation). This process continues until the desired number of features or the desired performance level is reached. By comparing the performance of different feature subsets, the best-performing feature subset is selected as the final result. This approach helps identify the most informative and relevant features in classification tasks.

[0110] In one embodiment, specifically, the step of SVM-RFE processing the training dataset to obtain the target dataset includes:

[0111] 1) Initialize the model and feature set. Initialize a linear support vector machine model and train it using all sample features from the training dataset. Alternatively, a feature set can be initialized for training, which assigns feature values ​​to each sample in the grid data; that is, this set contains all features.

[0112] 2) Feature weight evaluation: Calculate the weight or coefficient of each feature in the training dataset using the trained support vector machine model. These weights represent the importance of each feature to the support vector machine model.

[0113] 3) Feature ranking: Features are ranked according to their weights or coefficients. The ranked features are arranged from the most important to the least important.

[0114] 4) Feature Removal: Remove the least important features (features ranked last), or remove a preset number of features at the end of the sequence. Update the feature set, removing the removed features, so that each sample in each target dataset contains only the features remaining after removing the corresponding features.

[0115] 5) Model retraining: Retrain the linear support vector machine model using the updated feature set.

[0116] 6) Termination condition check: If a sufficient number of features have been removed to meet the preset final feature count, or if another stopping condition is met, the algorithm terminates. Otherwise, return to step 3, sort and remove features from the new feature set, and then retrain the model.

[0117] After the above steps, the target feature dataset obtained by linear support vector machine recursive feature elimination (SVM-RFE) includes the most important features of each sample, and the trained linear support vector machine model can also be better used to filter the target dataset.

[0118] Step S400: Set up an autoencoder classification model and train the autoencoder classification model on the target dataset.

[0119] In one embodiment, specifically, the autoencoder classification model includes an autoencoder and a multilayer perceptron (MLP). Specifically, an autoencoder is an artificial neural network used for semi-supervised and unsupervised learning tasks. Its main function is to learn a compact representation of input data (here, the input data is the target dataset) and reduce its dimensionality. By utilizing the concept of sparse coding, the autoencoder attempts to reconstruct the input data by recombining important features. In other words, the autoencoder learns an efficient representation that captures the essential features of the input. An autoencoder consists of two main parts: an encoder and a decoder. The encoder maps the high-dimensional input to a low-dimensional latent space, forcing the neural network to extract the most informative features. Conversely, the decoder reconstructs the latent variables back to the original input with the initial dimensions, aiming to minimize the reconstruction error and ensure that the output is highly similar to the input.

[0120] In one embodiment, specifically, setting the autoencoder classification model includes:

[0121] 1) Constructing the autoencoder: Initialize the autoencoder network structure, including the encoder and decoder; using the target dataset obtained above, train the autoencoder by minimizing the reconstruction error (typically using MSE loss). The encoder part maps the training data in the target dataset to the latent space. By iterating the autoencoder multiple times, the representation quality of the latent space can be gradually improved.

[0122] 2) Constructing the Multilayer Perceptron (MLP): Initialize the network structure of the MLP, including an input layer, hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the latent space. This setting is a conventional setting for neural network models, and any feasible implementation can be incorporated into the embodiments of this invention. Using the latent space obtained from the encoder part of an autoencoder as input, the MLP is trained by minimizing the classification loss to construct a binary classification model. The classification performance is gradually improved by iterating the MLP multiple times.

[0123] Specifically, in the autoencoder classification model, the fused feature set, as the input signal of the autoencoder, is converted into a latent signal through a combination of linear operations and nonlinear activation, and then converted back into an output signal. Furthermore, the latent signal is fed into the autoencoder classification network to predict whether lymph node metastasis has occurred. The network loss function is calculated using one of the formulas (4-6).

[0124]

[0125]

[0126] CE(y, y) true )=-(y true·log(y)+(1-y true )·log(1-y)) (6)

[0127] Where λ is the weighting factor for the two parts of the loss, n represents the dimension of the data, and x i and These are the input data and the i-th component of the reconstructed output, respectively, where y is the predicted label. true It's a real label.

[0128] In one embodiment, the above process is used to perform end-to-end training from input data (the target dataset obtained in step 300) to the classification result of the autoencoder latent space to the MLP; by minimizing the overall loss (the combination of the autoencoder's reconstruction loss and the MLP's classification loss), the autoencoder and the classification model are trained simultaneously, resulting in an autoencoder dimensionality reduction model and a binary classification MLP model obtained through end-to-end training.

[0129] Step S500: Input the target DNA methylation site data into the autoencoder classification model and output the classification result of the lymph node metastasis.

[0130] Specifically, the target DNA methylation site data can be either the PTC data being tested or the PTC methylation site data that needs to be predicted; the classification output can be the label value corresponding to the sample label, for example, an output of 0 or 1 can indicate no lymph node metastasis or lymph node metastasis, or vice versa, and it can be any output mode that distinguishes whether metastasis has occurred.

[0131] The present invention, using a DNA methylation-based method for predicting lymph node metastasis in patients with post-traumatic lung cancer (PTC), employs an IFF-RFE-AE feature reduction method and classification model that effectively utilizes patient DNA methylation data. Through stepwise feature screening, this method effectively removes redundant features while retaining key feature information, achieving high accuracy in predicting lymph node metastasis. The model also demonstrates a significant reduction in the risk of missed diagnoses. Accurate prediction of lymph node metastasis will contribute to targeted tumor treatment and clinical management, and these findings are of great importance to the clinical diagnosis and treatment of PTC.

[0132] In one embodiment, tumor samples obtained from the TCGA dataset are divided into training and test sets in an 8:2 ratio. A 5-fold cross-validation strategy is employed during the training of the prediction model. In this process, tumor samples with lymph node metastasis are considered positive cases, while those without lymph node metastasis are considered negative cases. The proposed solution is based on an autoencoder architecture, employing a 4-layer fully connected network to construct the encoder and decoder. Subsequently, a two-layer perceptron is constructed based on the autoencoder's latent space to implement the classification model. During the 5-fold cross-validation training process, cubic interpolation and random perturbation techniques are used to generate similar but slightly different training samples, thereby expanding the training data and reducing the model's dependence on specific attributes, ultimately improving its generalization performance.

[0133] This invention also implements several optimization measures to further improve model performance. First, a Batch Normalization layer is added after each fully connected layer to accelerate model convergence and improve stability during training. Furthermore, the nonlinearity of the model is increased by using the Tanh function, which helps capture more complex patterns in the data. To mitigate overfitting, this invention employs dropout regularization and weight decay techniques. Dropout randomly deactivates a subset of neurons during training, preventing them from becoming overly dependent on specific features and improving generalization ability. Weight decay introduces a penalty term to suppress large weight values, resulting in a simpler and more generalizable model. To optimize the learning process, this invention applies a cosine annealing strategy to dynamically adjust the learning rate during training. This technique helps prevent the model from getting stuck in local optima by periodically reducing the learning rate and exploring different regions of the loss landscape. To determine the optimal parameters for the ensemble feature filtering method and the classification model, this invention also performs a grid search, combining three parameters: the number of features retained (the number of features retained by each filtering method in the ensemble feature filtering), the learning rate, and the latent space dimension. For each parameter combination, this invention conducts two experiments and takes the average. The optimal parameter values ​​are 300, 0.001, and 10.

[0134] In this invention, 5-fold cross-validation and independent testing were used to evaluate the performance of our model. The area under the receiver operating characteristic (AUC) curve was used as an indicator to evaluate the model's discriminative ability. Simultaneously, this invention calculated accuracy, precision, recall, and F1 score to evaluate the predictive performance for lymph node metastasis. These indicators can be calculated using formulas (7-10), where TN, TP, FN, and FP represent true negative, true positive, false negative, and false positive, respectively.

[0135]

[0136]

[0137]

[0138]

[0139] Example 2

[0140] The present invention also provides another embodiment, proposing a device for predicting PTC lymph node metastasis based on DNA methylation, comprising:

[0141] The preprocessing module is used to combine the DNA methylation site data of m samples and the corresponding tags of the samples into grid data, and perform data augmentation to obtain enhanced grid data.

[0142] The feature filtering module is used to perform integrated feature filtering (IFF) on the enhanced grid data to obtain the training dataset;

[0143] The feature elimination module is used to process the training dataset using linear support vector machine recursive feature elimination (SVM-RFE) to obtain the target dataset;

[0144] The classification model building module is used to set up the autoencoder classification model and train the autoencoder classification model on the target dataset.

[0145] The output module is used to input target DNA methylation site data into the autoencoder classification model and output the classification results.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the steps, processing methods, devices, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0147] Example 3

[0148] The present invention also provides another embodiment, which proposes an electronic device including: a processor 1 and a memory 2.

[0149] The memory 2 is used to store computer programs.

[0150] The memory 2 includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.

[0151] The processor 1 is connected to the memory 2 and is used to execute the computer program stored in the memory 2 so that the device for predicting PTC lymph node metastasis based on DNA methylation performs the above-described method for predicting PTC lymph node metastasis based on DNA methylation.

[0152] Preferably, the processor 1 may be a central processing unit (CPU); or it may be an application-specific integrated circuit (ASIC).

[0153] Example 4

[0154] The present invention also provides another embodiment, namely, a computer-readable storage medium storing a computer program that can be executed by at least one processor to cause the at least one processor to perform the steps of the method for predicting PTC lymph node metastasis based on DNA methylation as described above.

[0155] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0156] Those skilled in the art will recognize that the modules, units, and / or method steps of the various embodiments described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0157] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another device or system, or some features may be ignored or not executed.

[0158] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0159] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting PTC lymph node metastasis based on DNA methylation, characterized in that, The method includes the following steps: S1. The DNA methylation site data of m samples and the corresponding tags of the samples are combined into grid data, and data augmentation is performed to obtain enhanced grid data; wherein, each sample occupies one row of data in the grid data, each sample has n features, and each sample corresponds to 1 tag. S2. The enhanced grid data is subjected to integrated feature filtering (IFF) to obtain a training dataset; the integrated feature filtering (IFF) includes one or more of correlation coefficient, mutual information, and chi-square test; S3. Process the training dataset using Linear Support Vector Machine Recursive Feature Elimination (SVM-RFE) to obtain the target dataset, specifically including: 1) Initialize the linear support vector machine model and train the linear support vector machine model using the training dataset; 2) Calculate the weight or coefficient of each feature in the training dataset using the trained linear support vector machine model; 3) Sort the features according to the weight or coefficient of each feature in the training dataset; 4) Remove the least important features and update the feature set of the samples in the training dataset; 5) Retrain the linear support vector machine model using the updated feature set; 6) Termination condition check: If a sufficient number of features have been removed to meet the preset final number of features, the algorithm terminates; otherwise, return to step 3, sort and remove features from the new feature set, and then retrain the model. S4. Set up an autoencoder classification model (AE) and train the autoencoder classification model on the target dataset; the autoencoder classification model includes an autoencoder and a multilayer perceptron (MLP). S5. Input the target DNA methylation site data into the autoencoder classification model and output the classification result of the lymph node metastasis; wherein, the target DNA methylation site data is used to obtain the latent spatial signal through the autoencoder, and the latent spatial signal is input into the multilayer perceptron (MLP) to obtain a prediction of whether the PTC lymph node metastasis has occurred.

2. The method for predicting PTC lymph node metastasis based on DNA methylation according to claim 1, characterized in that, The data augmentation includes cubic interpolation and / or random perturbation.

3. The method for predicting PTC lymph node metastasis based on DNA methylation according to claim 2, characterized in that, The data augmentation process for the cubic interpolation specifically includes the following steps: 1) Determine the interpolation point; 2) For each feature's interpolation location, perform the following cubic spline interpolation: a. Sort the data by column characteristics to ensure that the x values ​​increase in ascending order; b. Calculate the difference h and Δy for each pair of adjacent data points, where h = x(i+1) - x(i) and Δy = y(i+1) - y(i); c. Calculate the first derivative (slope) using the difference Δy / h; d. Calculate the second derivative value, i.e., the concavity / convexity of the curve; e. Calculate the coefficients a, b, c, and d of the cubic spline interpolation; 3) Using the calculated coefficients and x values, the corresponding y values ​​can be generated, forming the interpolation curve for each interpolation point; 4) For each column of features at the interpolation location points, repeat the above interpolation steps to generate the interpolation curve for each column of feature location points; 5) By applying the interpolation curve to each feature location point of each of the n features of each sample, new data points are generated at each feature location point, thereby expanding the dataset; And / or, the data augmentation process for the random perturbation specifically includes the following steps: 1) Determine the range of the random disturbance, which includes the intensity and range of values ​​of the disturbance; 2) For each feature value of the sample, generate a random number, ensuring that the random number is within the selected range; 3) Add random noise: Add the random number generated in step 2) to the position corresponding to the feature value; 4) For the feature values ​​of one or more samples in the grid data, repeat steps 2 and 3 to add random noise to generate new expanded samples.

4. The method for predicting PTC lymph node metastasis based on DNA methylation according to claim 1, characterized in that, The feature filtering based on the correlation coefficient includes: calculating the Pearson correlation coefficient between each feature of each sample in the enhanced grid data and the target category; filtering features whose Pearson correlation coefficient exceeds a preset threshold based on a predefined threshold; and / or The mutual information feature filtering includes: calculating the mutual information score between each feature and the target category in the enhanced grid data; filtering features whose mutual information scores exceed a predefined threshold; and / or The chi-square test for feature filtering includes: calculating the chi-square test value between each feature and the target category in the enhanced grid data; and filtering features whose chi-square test values ​​exceed a preset threshold based on a predefined threshold.

5. The method for predicting PTC lymph node metastasis based on DNA methylation according to claim 1, characterized in that, The network structure of the multilayer perceptron (MLP) includes an input layer, a hidden layer, and an output layer. The training of the autoencoder classification model on the target dataset specifically includes: Using the target dataset, the autoencoder is trained by minimizing the reconstruction error; the encoder part maps the data to the latent space. Using the latent space obtained from the encoder part of the autoencoder as input, the MLP is trained by minimizing the classification loss to construct a binary classification model.

6. A device for predicting PTC lymph node metastasis based on DNA methylation, comprising: The preprocessing module is used to combine the DNA methylation site data of m samples and the corresponding tags of the samples into grid data, and perform data augmentation to obtain enhanced grid data; wherein, each sample occupies one row of data in the grid data, each sample has n features, and each sample corresponds to 1 tag. The feature filtering module is used to perform integrated feature filtering (IFF) on the enhanced grid data to obtain the training dataset. The feature elimination module is used to process the training dataset using linear support vector machine recursive feature elimination (SVM-RFE) to obtain the target dataset, specifically including: 1) Initialize the linear support vector machine model and train the linear support vector machine model using the training dataset; 2) Calculate the weight or coefficient of each feature in the training dataset using the trained linear support vector machine model; 3) Sort the features according to the weight or coefficient of each feature in the training dataset; 4) Remove the least important features and update the feature set of the samples in the training dataset; 5) Retrain the linear support vector machine model using the updated feature set; 6) Termination condition check: If a sufficient number of features have been removed to meet the preset final number of features, the algorithm terminates; otherwise, return to step 3, sort and remove features from the new feature set, and then retrain the model. The classification model building module is used to set up an autoencoder classification model and train the autoencoder classification model on the target dataset; the autoencoder classification model includes an autoencoder and a multilayer perceptron (MLP). The output module is used to input target DNA methylation site data into the autoencoder classification model and output the classification result of the lymph node metastasis; wherein, the target DNA methylation site data is used to obtain the latent spatial signal through the autoencoder, and the latent spatial signal is input into the multilayer perceptron (MLP) to obtain a prediction of whether the PTC lymph node metastasis has occurred.

7. An electronic device, characterized in that, It includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, wherein the one or more programs include steps for performing the method for predicting PTC lymph node metastasis based on DNA methylation as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for predicting PTC lymph node metastasis based on DNA methylation as described in any one of claims 1-5.

Images