A method for predicting the risk of lymph node metastasis of thyroid papillary carcinoma based on an integrated machine learning model

Abstract

The application discloses a method for predicting the lymph node metastasis risk of thyroid papillary carcinoma based on an integrated machine learning model. Specifically, the application provides a training method for an integrated risk prediction model for the lymph node metastasis of thyroid papillary carcinoma, which comprises the following steps: (s1) providing a data set comprising the expression data of marker genes of thyroid cancer patients with and without lymph node metastasis; (s2) training a basic model: training a basic prediction model based on the data set; (s3) training a fusion layer: training a stacked ensemble model and a bagging ensemble model based on the output results of the basic prediction model; (s4) training a decision layer: training a weighted fusion model based on the output results of the stacked ensemble model and the bagging ensemble model, thereby obtaining the integrated risk prediction model. The model can realize accurate prediction of the lymph node metastasis risk of thyroid papillary carcinoma.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and artificial intelligence, specifically, it relates to a method for predicting the risk of lymph node metastasis in papillary thyroid carcinoma based on an integrated machine learning model. Background Technology

[0002] In existing technologies, the assessment of lymph node metastasis in papillary thyroid carcinoma primarily relies on preoperative imaging examinations, intraoperative pathological evaluation, and postoperative pathological results. Imaging methods mainly include ultrasound, CT, or MRI, used to evaluate the morphological characteristics of cervical lymph nodes. Pathological methods typically confirm lymph node metastasis through intraoperative frozen sections or postoperative paraffin sections.

[0003] In recent years, with the development of high-throughput sequencing technology and bioinformatics, some studies have attempted to use gene expression information, combined with statistical analysis or a single machine learning model, to predict the risk of tumor metastasis. These methods are usually based on a limited number of molecular markers or a single modeling strategy to assist in the determination of whether a patient has lymph node metastasis.

[0004] However, while the aforementioned techniques can be used to some extent to assess lymph node metastasis in papillary thyroid carcinoma, they still have certain limitations. Imaging methods primarily rely on morphological features, making it difficult to reflect the molecular biological differences of tumors. Their ability to identify early or micrometastases is limited, and they are easily influenced by the operator's experience and subjective judgment. Pathological assessments typically rely on tissue samples obtained intraoperatively or postoperatively, making it difficult to provide sufficient risk assessment evidence before surgery to determine whether lymph node dissection should be performed. Existing prediction methods based on gene molecular markers often employ single models or simple statistical analyses, limiting their ability to model nonlinear relationships between features, resulting in insufficient stability of prediction results across different patient populations.

[0005] Therefore, current technology lacks a method that can stably and accurately predict the risk of lymph node metastasis in papillary thyroid carcinoma based on molecular biological characteristics and through multi-model synergistic analysis.

[0006] Therefore, it is necessary to develop a new method that can stably, rapidly, and accurately predict the risk of lymph node metastasis in papillary thyroid carcinoma based on molecular biological characteristics. Summary of the Invention

[0007] This invention provides a method for predicting the risk of lymph node metastasis in papillary thyroid carcinoma based on an integrated machine learning model.

[0008] The purpose of this invention is to provide a method for predicting the risk of lymph node metastasis in papillary thyroid carcinoma based on molecular biological characteristics and combined with machine learning. This method analyzes the patient's molecular expression information to achieve preoperative assessment of lymph node metastasis risk. By employing multi-model collaborative analysis and fusing the prediction results of different models, this method improves the stability and accuracy of metastasis risk prediction, thereby providing a reliable auxiliary decision-making basis for clinical judgment on whether to perform lymph node dissection surgery.

[0009] In a first aspect of the invention, a method for training an integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma is provided, comprising the steps of: (s1) Provide a dataset that includes expression data of marker genes in thyroid cancer patients with and without lymph node metastasis; (s2) Basic model training: The expression data containing the expressed gene and the result of whether lymph node metastasis has occurred are respectively input into multiple basic prediction models and trained to obtain multiple trained basic prediction models; wherein, the multiple basic prediction models include: gradient boosting model, deep learning model, ensemble tree model and distance class model; (s3) Fusion layer training: The prediction results output by the trained multiple basic prediction models for the dataset are used as inputs to train the stacked ensemble model and the bagging ensemble model, thereby obtaining the trained first ensemble model and the second ensemble model. (s4) Decision layer training: The first and second fusion features output by the trained first fusion model and the second fusion model for the dataset are used as training inputs to train the weighted fusion model to obtain the trained weighted fusion model; thereby obtaining the trained integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma.

[0010] In another preferred embodiment, the marker gene is selected from the group consisting of RPS4Y1, PKHD1L1, CRABP1, MTT1G, TIMP1, SLC26A7, NUPR1, SERPINA1, BMP8A, SLC34A2, APOE, or combinations thereof.

[0011] In another preferred embodiment, the marker genes include RPS4Y1, PKHD1L1, and CRABP1.

[0012] In another preferred embodiment, the weighted fusion model is an automatically learned weighted ensemble model.

[0013] In another preferred embodiment, the weighted fusion model includes: a weighted average model, a regularized linear model, and a linear combination model with total weight constraints.

[0014] In another preferred embodiment, the expression data refers to the mRNA expression level of the marker gene.

[0015] In another preferred embodiment, the expression data refers to the ΔCt value of the marker gene relative to the internal reference gene (i.e., the ΔCt value after internal reference correction).

[0016] In another preferred embodiment, the ΔCt value is negatively correlated with the gene expression level.

[0017] In another preferred example, the loss function during model training is the binary log loss function.

[0018] In another preferred embodiment, the output of the integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma is a binary classification result indicating whether or not there is a risk of metastasis.

[0019] In a second aspect of the invention, a lymph node metastasis risk prediction system for papillary thyroid carcinoma is provided, comprising: (a) An input module configured to input expression data of marker genes for lymph node metastasis in papillary thyroid carcinoma in a sample to be tested; (b) An evaluation module, configured to receive the expression data, input the expression data into an integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma constructed using the method described in the first aspect of the present invention for evaluation, and obtain an evaluation result of the risk of lymph node metastasis in the subject. (c) Output module, which is configured to output the evaluation result.

[0020] In another preferred embodiment, the system further includes (d) a control module configured to control the operation of the modules.

[0021] In another preferred embodiment, the system further includes (a0) a detection module configured to detect expression data of marker genes.

[0022] In another preferred embodiment, the expressed data is represented by a ΔCt value.

[0023] In another preferred embodiment, the ΔCt value is negatively correlated with the gene expression level.

[0024] In another preferred embodiment, the marker gene is selected from the group consisting of: RPS4Y1, PKHD1L1, CRABP1, MTT1G, TIMP1, SLC26A7, NUPR1, SERPINA1, BMP8A, SLC34A2, and APOE. In another preferred embodiment, the marker genes include RPS4Y1, PKHD1L1, and CRABP1.

[0025] In another preferred embodiment, the evaluation result is a binary classification result of whether the subject has a risk of lymph node metastasis.

[0026] In another preferred embodiment, the subject of the test is a patient with papillary thyroid carcinoma.

[0027] In another preferred embodiment, the system can be used to assist in diagnosing whether a subject has undergone lymph node metastasis.

[0028] In a third aspect of the invention, an electronic device is provided, the electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method as described in the first aspect of the invention.

[0029] In a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in the first aspect of the invention.

[0030] In another preferred embodiment, the computer-readable storage medium may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may include, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0031] In a fifth aspect of the invention, a computer program product is provided, comprising computer-executable instructions or a computer program that, when executed by a processor, implements the method described in the first aspect of the invention.

[0032] In another preferred embodiment, the computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0033] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0034] Figure 1 A flowchart of the method of the present invention is shown.

[0035] Figure 2 The process of building the model of this invention is shown.

[0036] Figure 3 The ROC curve results of the model of this invention during training and testing are shown. Detailed Implementation

[0037] Through extensive and in-depth research, the inventors have developed, for the first time, a method for predicting the risk of lymph node metastasis in papillary thyroid carcinoma based on an ensemble machine learning model. This method is based on expression data of three genes (RPS4Y1, PKHD1L1, and CRABP1) and a fusion model comprising a base layer, a fusion layer, and a decision layer. The base layer includes gradient boosting, deep learning, ensemble tree, and distance-based models; the fusion layer includes stacked ensemble and bagging ensemble; and the decision layer includes automatically learned weighted ensemble. Experiments show that the ensemble model prediction method constructed in this invention has the best prediction accuracy. Based on this, the present invention was completed.

[0038] Specifically, this invention is based on the division of training, validation, and test sets, and uses multi-center data (1889 cases) for validation, resulting in a model with an AUROC of up to 0.9761. Comparative experiments (ablation experiments) demonstrate that the ensemble model has significant advantages over the single model.

[0039] the term To facilitate a clearer understanding of this disclosure, certain terms are first defined. As used herein, unless otherwise expressly specified herein, each of the following terms shall have the meaning given below. Other definitions are set forth throughout the application.

[0040] The term “about” can refer to a value or composition within an acceptable range of error for a particular value or composition as determined by a person skilled in the art, depending in part on how the value or composition is measured or determined. For example, as used herein, the expression “about 100” includes all values ​​between 99 and 101.

[0041] As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed. In other words, the terms also include “consistently made of” or “composed of”.

[0042] As used herein, unless otherwise stated, any concentration range, percentage range, proportion range, or integer range shall be understood to include any integer value within the range and, where appropriate, its fractional value (e.g., one-tenth and one-hundredth of an integer).

[0043] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0044] As used in this article, the term "sample" or "sample" refers to material specifically associated with a subject from which specific information relating to the subject can be determined, calculated, or inferred. A sample may consist wholly or partially of biological material from the subject.

[0045] As used herein, the term "expression" includes the production of mRNA from a gene or a gene segment, and includes the production of proteins encoded by the RNA or gene segment, as well as the presence of detection substances associated with expression. For example, the binding of cDNA, binding ligands (such as antibodies) to genes or other oligonucleotides, proteins, or protein fragments, and the chromogenic portion of the binding ligand are all included within the scope of the term "expression." Therefore, an increase in the density of half-points on an immunoblot such as a Western blot also falls within the scope of the biologically molecular-based term "expression."

[0046] As used in this invention, "Ct value" or "Cycle Threshold" refers to the number of amplification cycles required for the fluorescence signal of the amplified product to reach a set fluorescence threshold in quantitative PCR. Simply put, the Ct value represents the number of cycles required to amplify a certain amount of product from the initial template. A higher initial template concentration results in a lower Ct value; a lower initial template concentration results in a higher Ct value.

[0047] As used in this invention, "delta Ct value" or "ΔCt value" refers to the value of the gene to be tested (R gene) compared with the internal reference, for example, by adding or subtracting the Ct value of GAPDH (the internal reference selected in this invention). The ΔCt value of the R gene is Ct(R gene) - Ct(GAPDH) gene.

[0048] RPS4Y1, PKHD1L1 and CRABP1 RPS4Y1 is a component of the ribosome and participates in protein synthesis.

[0049] The PKHD1L1 gene is involved in regulating cell growth and differentiation, which is crucial for tissue development and normal organ function.

[0050] CRABP1 is involved in the metabolism of retinoic acid, a molecule that regulates cell differentiation and growth. Dysregulation of the retinoic acid pathway affects tumor behavior in various cancers, including thyroid cancer. Some studies suggest that altered CRABP1 expression may be associated with the development of thyroid cancer, potentially influencing tumor growth and treatment response.

[0051] PKHD1L1, CRABP1, and RPS4Y1 are candidate genes identified in recent studies as being associated with lymph node metastasis in thyroid cancer (PTC). Among them, independent studies have shown that low expression of PKHD1L1 is significantly associated with PTC lymph node metastasis (Zheng, Chen, et al. "Original tumor suppressor gene polycystic kidney and hepaticdisease 1-like 1 is associated with thyroid cancer cell progression." Oncology Letters 18.3 (2019): 3227-3235.). CRABP1 is associated with invasion in thyroid cancer (Hsu, Yi‐Chiung, et al. "Downregulation of cellular retinoic acid binding protein 1 fosters epithelial–mesenchymal transition in thyroid cancer." Molecular Carcinogenesis 62.12 (2023): 1935-1946.). RPS4Y1 is primarily used as a newly discovered predictive biomarker, demonstrating strong discriminative ability for metastatic status in multicenter cohorts (Liu, Wei, et al. "Clinical performance of a machine learning-based model for detecting lymph node metastasis in papillary thyroid carcinoma: A multicenter study." International Journal of Surgery 111.6(2025): 4062-4067.). Gradient boosting model, deep learning, ensemble tree model, distance class model, stacked ensemble, bagging ensemble, automatically learned weighted ensemble Gradient boosting model: The gradient boosting model is an ensemble learning method based on iterative optimization, which improves overall prediction performance by progressively building multiple weak learners. Each training round aims to minimize the residual or loss function of the previous model. The weak learner described in this invention preferably employs a gradient boosting decision tree model (LightGBM) based on the histogram algorithm, with 128 leaf nodes, a minimum number of samples per leaf node of 3, and a feature sampling ratio of 0.9. Structurally, the model contains approximately 421 decision trees, each with 128 leaf nodes, and an average tree depth of approximately 27.86.

[0052] Deep learning: Deep learning is a data modeling method based on a multi-layer neural network structure, used to automatically extract multi-level feature representations from data. This method trains network parameters using the backpropagation algorithm to approximate complex nonlinear mapping relationships. The network structure described in this invention employs a multi-layer fully connected neural network, including four hidden layers with a hidden layer size of 2^13. It uses the ELU activation function and sets a Dropout probability of approximately 0.119 to reduce the risk of overfitting.

[0053] Ensemble Tree Model: An ensemble tree model is a type of learning model built from multiple decision trees to improve the accuracy and stability of predictions. This model effectively models complex data distributions by aggregating the outputs of multiple decision trees. The ensemble method described in this invention employs a random forest approach, containing 300 decision trees. Node partitioning is performed using the Gini index, with a maximum of 15,000 leaf nodes. Feature selection is based on square root sampling, and a bootstrap mechanism is used. Structurally, the average number of nodes per tree is approximately 588.81, the average number of leaf nodes is approximately 294.90, and the average tree depth is approximately 21.34.

[0054] Distance-based models: Distance-based models are learning methods that predict based on the distance or similarity between samples. This model determines the category or numerical output of a sample by calculating the distance relationship between the sample to be predicted and the training samples. This invention uses the k-nearest neighbor method, where the number of neighbors is set to 5, a uniform weighting strategy is employed, and the distance metric used is Euclidean distance.

[0055] Stacked ensemble: Stacked ensemble is a multi-level model fusion method that trains a meta-model by using the outputs of multiple base models as input. This method can combine the advantages of different models, thereby improving overall prediction performance. The meta-model described in this invention is a linear model combined with regularization constraints.

[0056] Bagging ensemble: Bagging ensemble is an ensemble method that constructs multiple sub-models by repeatedly sampling training data randomly. Each sub-model is trained independently on different subsets of samples, and their outputs are merged through averaging or voting. This method can effectively reduce model variance and improve generalization ability.

[0057] Automatically Learned Weighted Ensemble: Automatically learned weighted ensemble is a method that fuses prediction results by learning the weights of each base model. This method automatically determines the optimal weight allocation for each model based on validation data or an optimization objective function. The weights described in this invention satisfy non-negativity constraints and are normalized to improve fusion stability.

[0058] Specific construction process of the model of this invention The detailed construction process of the integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma provided by this invention is detailed in [link to relevant documentation]. Figure 2 The basic model layer of the ensemble model of this invention includes: a gradient boosting model (preferably LightGBM), a deep learning model (multi-layer fully connected neural network), an ensemble tree model (random forest), and a distance-based model (KNN); the fusion layer includes: stacking ensemble and bagging ensemble; and the decision layer includes: an automatically learned weighted ensemble model.

[0059] (1) The training process of the base layer model includes the following steps: The model input features are denoted as follows: The input features for each sample are: in, These represent the expression values ​​(preferably delta Ct values) of genes RPS4Y1, PKHD1L1, and CRABP1, respectively.

[0060] No. The labels for each sample are: In this context, 1 indicates the presence of lymph node metastasis, and 0 indicates no metastasis has occurred.

[0061] For the The basic model has the following prediction function: The output predicted probability is: The basic model set is as follows: Where K=4.

[0062] 1) Data methods involved in gradient boosting models In the basic model of this invention, the gradient boosting model adopts the LightGBM framework, and its prediction function can be expressed as: in, For the first A decision tree weak learner, For learning rate, This indicates that the input after feature sampling is... The feature subset of the tree, with a feature sampling ratio of 0.9.

[0063] Each decision tree satisfies the following constraints: in, Indicates the first The set of leaf nodes of a tree Leaf nodes The number of samples in the sample, i.e., the minimum number of leaf node samples, is 3.

[0064] After training, the average tree depth of each decision tree can be expressed as: in, Indicates the first The trees are deep.

[0065] 2) Mathematical methods involved in deep learning models In the basic model of this invention, the deep learning model employs a multi-layer fully connected neural network structure. Let the input features be three-dimensional vectors. The neural network includes four hidden layers, and its forward propagation process can be represented as follows: in, , Indicates the layer number, Indicates the first The output vector of each hidden layer Indicates the first The weight matrix of the layer, Indicates the first The bias vector of the layer.

[0066] The output layer is: The activation function of the output layer is shown, preferably the Sigmoid function.

[0067] In another preferred embodiment, the number of neurons in each hidden layer is 213, that is: The output layer weight matrix is: The activation function is an exponential linear unit (ELU), and its expression is as follows: During training, the Dropout mechanism is introduced to reduce overfitting, which can be expressed as: in, Indicates the first The output vector of the layer after Dropout processing. For a random variable that follows a Bernoulli distribution, the Dropout probability is 0.119. This represents the Hadamard product of corresponding elements of a vector.

[0068] The model uses the Adam optimization algorithm for parameter updates with a learning rate of 0.01.

[0069] 3) Mathematical methods involved in the ensemble tree model In the basic model of this invention, the ensemble tree model adopts the random forest method, and its prediction function can be expressed as: in, Indicates the first There are 300 decision trees, where B=300 represents the number of decision trees.

[0070] During training, each decision tree is trained based on a training subset generated by bootstrapping, which can be represented as: Represents the original training dataset. Indicates that through self-sampling from The first generated in A training subset, where Bootstrap represents random sampling with replacement.

[0071] In another preferred example, the number of leaf nodes in each decision tree satisfies: After training, each decision tree satisfies the following statistical characteristics: in, Indicates the average tree depth. This represents the average number of leaf nodes.

[0072] In classification tasks, the output probability of a random forest can be expressed as: 4) Mathematical methods involved in distance-based models In the basic model of this invention, the distance-based model employs the K-nearest neighbor method. Let the input sample be a three-dimensional vector. This corresponds to the expression characteristics of three marker genes.

[0073] The distance between samples is measured using Euclidean distance, which is expressed as follows: In another preferred embodiment, the closest distance is selected. Each neighbor constitutes a neighborhood set: Where K=5.

[0074] In classification tasks, when a uniform weighting strategy is used for prediction, the output probability can be expressed as: in, This represents the category label of the neighboring samples.

[0075] In another preferred embodiment, a weighted form can be used: In this embodiment This is called uniform weighting.

[0076] (2) Mathematical methods involved in the fusion layer In a preferred embodiment of the present invention, the fusion layer includes two mechanisms: stacking and bagging.

[0077] For the For each sample, the feature vector is first constructed from the prediction results output by the base model layer: in, K represents the number of basic models.

[0078] In stack integration In this process, a meta-learning model is used to learn the features, and its output can be expressed as: In another preferred embodiment, the metamodel is a linear model with regularization constraints, expressed as: in, For learnable parameters, This is the Sigmoid function.

[0079] In another preferred embodiment, to avoid information leakage, the Obtained through cross-validation, i.e.: Meanwhile, in bagging integration In this process, multiple sub-models are trained by performing multiple bootstrapping operations on the training data: And by averaging its output, we get: Ultimately, the fusion layer output consists of both stacked integration and bagged integration: (3) Mathematical methods involved in decision-making In a preferred embodiment of the present invention, the decision layer employs a weighted fusion model to make a comprehensive decision based on the output of the fusion layer. For the first... For each sample, the output of the fusion layer is: in, Indicates stacked integrated output. This indicates the integrated output of bagging.

[0080] The output of the decision-making level can be expressed as: In another preferred embodiment, the output of the decision layer can be expressed as: in, These are the weight parameters to be learned.

[0081] In another preferred embodiment, the weights satisfy the following constraints: The weight parameters are learned by minimizing the following objective function: In another preferred example, a regularization term is introduced to improve the model's generalization ability: In classification tasks, the final prediction result is determined by the following rules: Among them, threshold The preferred value is 0.3108.

[0082] During the training process of this invention, the input data for the model consists of the delta Ct values ​​of three genes, and the loss function is binary log loss. The batch size is set to 256, the number of training epochs to 100, and an early stopping strategy is employed (with a patience value of 20 epochs).

[0083] The main advantages of this invention include: (a) By fusing the prediction results of multiple models, the advantages of different models can be combined to effectively improve the overall prediction accuracy and robustness. (b) Employing ensemble learning and weighted fusion strategies reduces the risk of overfitting by a single model, making the model's performance on unknown data more stable. (c) It can flexibly connect to different types of base models and optimize them through automatic learning of weights, making it suitable for various data scenarios and task requirements. The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions, such as those described in Sambrook et al., Molecular Cloning: A Laboratory Manual (New York: Cold Spring Harbor Laboratory Press, 1989), or as recommended by the manufacturer. Unless otherwise stated, percentages and parts are weight percentages and parts by weight.

[0084] Example 1 like Figure 1 As shown, the present invention provides a method for predicting the risk of lymph node metastasis in papillary thyroid carcinoma based on a three-layer ensemble machine learning model, the specific implementation process of which includes the following steps.

[0085] First, obtain the molecular biological characteristic data of the patient to be analyzed. The molecular biological characteristic data is the mRNA expression data of three genes (RPS4Y1, PKHD1L1 and CRABP1) associated with the metastasis of papillary thyroid carcinoma, and use the mRNA expression data as the input layer data of the model.

[0086] Second, the input mRNA expression data is simultaneously fed into multiple basic prediction models in the basic model layer (these basic prediction models are trained with training set data and can be directly used for prediction in clinical practice) for parallel analysis. The basic model layer includes different types of machine learning models used to learn gene expression characteristics and determine risks from different modeling perspectives.

[0087] The aforementioned basic prediction models include: gradient boosting model, deep learning model, ensemble tree model, and distance-based model.

[0088] Third, the prediction results output by the base model layer are fused to form a fusion layer. This fusion layer employs stacked ensemble and bagging ensemble (the parameters of the fusion layer are pre-trained using the training set data; when used, predictions can be made directly) to hierarchically integrate the prediction results of multiple base models. Stacked ensemble is used to learn the combination relationships between the outputs of different models, while bagging ensemble is used to reduce the volatility of model prediction results, thereby improving the stability and consistency of the overall prediction results.

[0089] Fourth, the output of the fusion layer is input into the decision layer, where the decision layer automatically learns and weights the fusion results to generate the final risk assessment result. By assigning adaptive weights to different fusion results, the model can automatically optimize the contribution of each prediction result to the final decision based on the training data.

[0090] Fifth, based on the output of the decision-making level, a prediction of whether the patient has developed lymph node metastasis of papillary thyroid carcinoma can be obtained, providing an auxiliary basis for clinical judgment on whether to perform lymph node dissection surgery.

[0091] Model pre-training, validation, and testing results The training input for the model consisted of ΔCt values ​​obtained from qPCR detection of three genes. For the neural network model, the input features were normalized. During training, the Adam optimizer was used for the neural network model with an initial learning rate of 0.01, while histogram-based training was employed for the gradient boosting and ensemble tree models. Binary log loss was used as the loss function. Training hyperparameters were set to a batch size of 256, 100 epochs, and an early stopping mechanism with a patience value of 20 epochs.

[0092] Post-trained model performance: accuracy 0.9092, sensitivity 0.8921, specificity 0.9262, AUROC 0.9613 on the training set.

[0093] The datasets used for training, validating, and testing the model of this invention are shown in Table 1. Each dataset includes both metastatic and non-metastatic patients. The total number of patients with papillary thyroid carcinoma is 1889.

[0094] This invention employs an ensemble learning and weighted fusion strategy to reduce the risk of overfitting by a single model, making the model's performance more stable on unknown data. This performance is demonstrated on multi-center validation and test sets.

[0095] Table 1 The final result data is as follows: Figure 3 And as shown in Table 2 below: Table 2 Example 2: Comparative or Ablation Experiment In this embodiment, the inventors verified the overall performance data of the final model when the base prediction model layer has only a single small model, two small models, or three small models, and compared the performance with that of a model containing all four small models (i.e., gradient boosting model, deep learning model, ensemble tree model, and distance class model). The datasets used are test set 1 and test set 2 in Table 1.

[0096] The specific test results are shown in Table 3.

[0097] Table 3 The results above show that when the base layer uses all four small models selected in this invention, the final model has the best performance.

[0098] discuss By employing multiple basic models to analyze gene mRNA expression data in parallel, this invention can extract molecular features related to lymph node metastasis in papillary thyroid carcinoma from different modeling perspectives. Through stacking and bagging ensembles to fuse the prediction results of multiple models, the prediction bias caused by a single model is effectively reduced, improving the stability and consistency of the prediction results. An automatic learning weighted ensemble method is used to optimize the fusion results, making the final metastasis risk assessment more accurate and reliable. Experimental results show that the method of this invention has high accuracy, sensitivity, and specificity in predicting lymph node metastasis in papillary thyroid carcinoma, and can provide effective auxiliary decision support for clinical decision-making regarding whether to perform lymph node dissection surgery before surgery, demonstrating good clinical application value.

[0099] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A training method for an integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma, characterized in that, Including the following steps: (s1) Provide a dataset comprising expression data of marker genes in thyroid cancer patients with and without lymph node metastasis; wherein the marker genes include: RPS4Y1, PKHD1L1 and CRABP1; (s2) Basic model training: The expression data of the marker gene and the result of whether lymph node metastasis has occurred are respectively input into multiple basic prediction models and trained to obtain multiple trained basic prediction models; wherein, the multiple basic prediction models include: gradient boosting model, deep learning model, ensemble tree model and distance class model; (s3) Fusion layer training: The prediction results output by the trained multiple basic prediction models for the dataset are used as inputs to train the stacked ensemble model and the bagging ensemble model, thereby obtaining the trained first ensemble model and the second ensemble model. (s4) Decision layer training: The first fusion feature and the second fusion feature output by the trained first fusion model and the second fusion model for the dataset are used as training inputs to train the weighted fusion model to obtain the trained weighted fusion model; thereby obtaining the trained integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma. During the training of the model, the loss function is the binary log loss function; In step (s2), the expression data of the marker gene is represented by the following formula: In the formula, X i For the first i Expression data of each sample These represent the expression values ​​of the marker genes RPS4Y1, PKHD1L1, and CRABP1, respectively, where each expression value is a ΔCt value relative to the internal reference gene; and the... i The label corresponding to each sample In this context, 1 indicates the presence of lymph node metastasis, and 0 indicates no metastasis. For the The basic model has the following prediction function: The output predicted probability is: ; In step (s3), the prediction results output from the base model layer are used to construct the feature vector: K represents the number of basic models, and K=4; In stack integration In this process, a meta-learning model is used to learn the feature vector, and its output can be expressed as: The meta-learning model is a linear model with regularization constraints, expressed as follows: ;in, For learnable parameters, For the Sigmoid function; The weighted fusion model is an automatically learned weighted ensemble model, which is a method to achieve the fusion of prediction results by learning the weights of each base model. This method automatically determines the optimal weight allocation of each model based on validation data or an optimization objective function.

2. The method as described in claim 1, characterized in that, The output of the integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma is a binary classification result indicating whether the subject has a risk of metastasis.

3. The method as described in claim 1, characterized in that, During model training, the batch size was set to 256, the number of training epochs was set to 100, and an early stopping strategy was adopted with a patience value of 20 epochs.

4. A system for predicting the risk of lymph node metastasis in papillary thyroid carcinoma, characterized in that, include: (a) An input module configured to input expression data of marker genes for lymph node metastasis in papillary thyroid carcinoma in a sample to be tested; (b) An evaluation module, configured to receive the expression data, input the expression data into an integrated model for predicting the risk of lymph node metastasis in papillary thyroid carcinoma constructed using the method of claim 1 for evaluation, and obtain an evaluation result of the risk of lymph node metastasis in the subject. (c) Output module, which is configured to output the evaluation result.

5. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in claim 1.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in claim 1.

7. A computer program product comprising computer-executable instructions or a computer program, characterized in that, The computer-executable instructions or computer program, when executed by a processor, implement the method of claim 1.

Images