Thyroid cancer metastasis risk prediction model and construction method and application thereof

Abstract

The invention relates to the field of disease risk prediction models, and discloses a thyroid cancer metastasis risk prediction model and a construction method and application thereof. The construction method of the model comprises the following steps: S1, collecting and preprocessing a sample; s2, gene expression quantity detection; s3, data set division and standardization processing; s4, constructing a deep learning model; s5, performing model evaluation; and the biomarker group for detecting the gene expression quantity in the S2 is composed of RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7 and RBP4. The prediction model can be used as a clinical auxiliary diagnosis tool and is used for guiding operation range selection, postoperative follow-up visit strategy making and early intervention of high-risk patients, and the precise diagnosis and treatment level of thyroid cancer is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disease risk prediction models, and more particularly, to a thyroid cancer metastasis risk prediction model and a construction method and application thereof. BACKGROUND

[0002] Papillary thyroid carcinoma (PTC) is the most common malignant tumor of the thyroid gland, accounting for more than 80% of all thyroid cancers. Although the prognosis of most PTC patients is good, a portion of patients will develop lymph node metastasis (LNM), especially central lymph node metastasis (CLNM) and lateral cervical lymph node metastasis (LLNM), which significantly affect the surgical extent, postoperative adjuvant therapy strategy and long-term prognosis.

[0003] Currently, the commonly used metastasis risk assessment methods in clinical practice mainly include imaging examinations such as ultrasound, CT, MRI and postoperative pathological analysis. However, the imaging methods have limited sensitivity in detecting micro-metastasis or early metastasis, especially in the central lymph node assessment, and have a high rate of missed diagnosis. Postoperative pathology is the gold standard, but it is a retrospective diagnosis and cannot be used for preoperative decision-making. In recent years, although some studies have attempted to predict risk based on single gene mutations such as BRAF and TERT, the prediction efficiency is limited, and there is a lack of multi-gene combination and systematic prediction model.

[0004] Therefore, there is still a lack of a molecular tool that can accurately, non-invasively and efficiently predict the risk of lymph node metastasis in thyroid cancer patients before surgery in the prior art, and it is urgent to combine multi-gene expression profiles and artificial intelligence technology to construct a prediction system with high sensitivity and specificity to assist clinical individualized treatment decisions. SUMMARY

[0005] The purpose of the present application is to provide a thyroid cancer metastasis risk prediction model based on 8-gene expression profiles and deep learning algorithms to solve the technical problems of insufficient micro-metastasis detection ability of imaging examinations, limited prediction efficiency of single molecular markers and lack of preoperative metastasis prediction models with high accuracy in the prior art. The present application realizes the accurate stratification of metastasis risk in patients before surgery by constructing an AI prediction system combined with multiple models, and provides reliable decision support for clinical practice.

[0006] In order to achieve the above-mentioned application purposes, the present application adopts the following technical solutions: In a first aspect, the application provides a method for constructing a thyroid cancer metastasis risk prediction model, comprising the following steps: S1, sample collection and pretreatment; S2, gene expression detection; S3, data set division and standardization processing; S4, deep learning model construction; S5, model evaluation. The biomarker group for gene expression detection in S2 consists of RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7 and RBP4.

[0007] Further, S1 comprises the following steps: Collect postoperative tumor tissue samples of patients diagnosed as thyroid papillary carcinoma by pathology; According to the results of postoperative pathological examination, label the metastasis state of each sample, including whether lymph node metastasis occurs, whether central lymph node metastasis occurs, and whether lateral neck lymph node metastasis occurs; Extract total RNA from tumor tissue, detect RNA concentration and purity, and detect RNA integrity.

[0008] Further, S2 comprises the following steps: Detect the gene expression of the biomarker group, and select GAPDH gene as an internal reference gene for data standardization; qPCR experiment, 3 repeats for each gene of each sample, record the Ct value of each reaction well; Calculate the average value of the Ct value of 3 repeats of each gene, and calculate the relative expression of each target gene, and then perform subsequent standardization processing, and finally obtain the standardized expression data which can be used for model input.

[0009] Further, S3 comprises the following steps: Arrange the standardized expression data of 8 genes of the sample and its corresponding three metastasis labels into a data matrix; Randomly divide the entire data set into a training set and a test set; Z-score standardization is performed on the gene expression data.

[0010] Further, S4 comprises the following steps: Three independent binary classification models are constructed using Python language and deep learning framework, including total metastasis prediction model, central lymph node metastasis prediction model and lateral neck lymph node metastasis prediction model.

[0011] Further, S5 comprises the following steps: After the model training is completed, the performance of the model is evaluated on an independent test set, and the following indicators are calculated: sensitivity, specificity, accuracy and area under the receiver operating characteristic curve.

[0012] In a second aspect, the application provides a thyroid cancer metastasis risk prediction model obtained by the construction method in the first aspect.

[0013] In a third aspect, the application provides use of the thyroid cancer metastasis risk prediction model in the second aspect in preparation of a thyroid cancer metastasis risk prediction product.

[0014] In a fourth aspect, the application provides a thyroid cancer metastasis risk prediction system, comprising: a data acquisition module: collecting postoperative tumor tissue sample related information of a thyroid papillary carcinoma patient to be tested, including the source of the sample, pathological diagnosis, and simultaneously collecting gene expression data of biomarker group RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7 and RBP4 in the sample; a data preprocessing module: performing standardization processing on the collected gene expression data; a data processing module: preconfigured with a trained total metastasis prediction model, a central lymph node metastasis prediction model and a lateral neck lymph node metastasis prediction model, processing the standardized gene expression data to obtain a total probability value P_total of lymph node metastasis of the patient, a probability value P_central of central lymph node metastasis and a probability value P_lateral of lateral neck lymph node metastasis; a result output module: comparing the obtained P_total, P_central and P_lateral with a preconfigured risk threshold value, if P_total is greater than the risk threshold value, it is determined that the patient has a high risk of lymph node metastasis, otherwise it is a low risk, and the same applies to P_central and P_lateral.

[0015] In a fifth aspect, the application provides a storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the construction method of the thyroid cancer metastasis risk prediction model in the first aspect, or part or all steps of the thyroid cancer metastasis risk prediction system in the fourth aspect.

[0016] In summary, the application uses 8 gene expression data (RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7, RBP4, and the reference gene is GAPDH), combined with a deep learning algorithm, to successfully construct a metastasis risk prediction system, which has the following beneficial effects: The present application solves the technical problems of insufficient detection ability of micro-metastasis in imaging examination, limited prediction performance of single molecular marker, and lack of preoperative metastasis prediction model with high accuracy. The prediction model of the present application can be used as a clinical auxiliary diagnosis tool to guide the selection of surgical range, the development of postoperative follow-up strategy, and the early intervention of high-risk patients, thereby improving the precision diagnosis and treatment level of thyroid cancer. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 : AUC curve of the total metastasis prediction model; Figure 2 : Prediction result confusion matrix of the total metastasis prediction model; Figure 3 : AUC curve of the central lymph node metastasis model; Figure 4 : Prediction result confusion matrix of the central lymph node metastasis model; Figure 5 : AUC curve of the lateral neck lymph node metastasis model; Figure 6 : Prediction result confusion matrix of the lateral neck lymph node metastasis model. DETAILED DESCRIPTION

[0018] The technical solutions and effects of the present application will be further described in detail below in combination with the embodiments and drawings. It can be understood that the specific embodiments described herein are only used to explain the present application, and not to limit the present application.

[0019] EMBODIMENT This embodiment details how to construct a deep learning model for predicting the risk of thyroid cancer metastasis step by step from the original sample. The whole process mainly includes the following steps: Step 1: Sample collection and pretreatment 1. Sample source: Collect postoperative tumor tissue samples of patients diagnosed as papillary thyroid carcinoma (PTC) by pathology. In this embodiment, a total of 198 patient samples were included, and all patients signed the informed consent form, and the study was approved by the ethics committee.

[0020] 2. Clinical data labeling: According to the postoperative pathological examination results, label the metastasis state label for each sample, including: a) whether lymph node metastasis occurs (yes / no); b) whether central lymph node metastasis occurs (yes / no); c) whether lateral neck lymph node metastasis occurs (yes / no).

[0021] 3. RNA Extraction and Quality Control: Total RNA was extracted from tumor tissue using a commercially available RNA extraction kit. RNA concentration and purity (A260 / A280 ratio) were measured using Nanodrop, and RNA integrity (RIN value) was measured using an Agilent 2100 bioanalyzer or a similar method to ensure that the RNA quality met the requirements for qPCR experiments.

[0022] Step 2: Gene Expression Level Detection 1. Gene Selection: The expression levels of the following eight target genes were detected: RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7, and RBP4. Simultaneously, the GAPDH gene was selected as an internal reference gene for data normalization.

[0023] 2. qPCR experiment: a) Reverse transcription: Reverse transcription of high-quality RNA into cDNA.

[0024] b) qPCR amplification: Amplification was performed using the TaqMan probe method on a real-time quantitative PCR instrument. Three technical replicates were set up for each gene in each sample.

[0025] c) Data acquisition: Record the Ct value for each reaction well.

[0026] Table A8 gene expression primer and probe sequences

[0027] 3. Expression Level Calculation: First, the average Ct value of the three technical replicates for each gene is calculated. Then, the relative expression level of each target gene is calculated using the 2^(-ΔΔCt) method. That is, the Ct value of the internal reference gene GAPDH is subtracted from the Ct value of the target gene to obtain ΔCt. Subsequent standardization processing is then performed to obtain standardized expression level data that can be used as model input.

[0028] Step 3: Dataset Partitioning and Preprocessing 1. Data processing: The standardized expression levels of 8 genes from 198 samples and their corresponding three transfer tags were processed into a data matrix (rows of samples, columns of gene expression values).

[0029] 2. Data Splitting: The entire dataset is randomly divided into a training set and a test set, typically in a ratio of 7:3 or 8:2. In this embodiment, triple-fold cross-validation is used for model training and parameter tuning on the training set, and the final model performance is reported on a separate test set.

[0030] 3. Data standardization: Perform Z-score standardization on gene expression data (features), which is to subtract the mean of each gene expression level from its mean in the training set and then divide by its standard deviation, so that each feature follows a normal distribution with a mean of 0 and a standard deviation of 1.

[0031] Step 4: Deep Learning Model Construction and Training Build three independent binary classification models using the Python language and deep learning frameworks such as TensorFlow / Keras or PyTorch.

[0032] Model 1: Total Transfer Prediction Model I network structure: Input layer (8 nodes, corresponding to 8 genes) → Single hidden layer (20 neurons) → Output layer (1 neuron, sigmoid activation function).

[0033] II. Hidden layer activation function: Rectifier (ReLU) combined with Dropout regularization, with the Dropout rate set to 0.4 to prevent overfitting.

[0034] III. Output and Loss Function: The output layer uses the Sigmoid function to map the output value to the [0,1] interval, representing the probability of a transition. The loss function is binary crossentropy.

[0035] IV Optimizer: Use an adaptive learning rate optimizer (such as Adam or Nadam).

[0036] V Training Strategy: Enable class balancing (e.g., class_weight='balanced') to address potential imbalances in the number of transition and non-transition samples in the data. Train the model on the training set and monitor performance on the validation set for early stopping.

[0037] Model 2: Central Lymph Node Metastasis Prediction Model I network structure: Input layer (8 nodes) → 3 hidden layers (100 neurons per layer) → Output layer (1 neuron, Sigmoid).

[0038] The remaining settings (activation function, Dropout, loss function, optimizer, etc.) are similar to those of Model 1. The specific hyperparameters can be fine-tuned based on the performance on the validation set.

[0039] Model 3: Lateral Cervical Lymph Node Metastasis Prediction Model I network structure: Input layer (8 nodes) → 2 hidden layers (100 neurons per layer) → Output layer (1 neuron, Sigmoid).

[0040] II. The remaining settings are similar to those of Models 1 and 2.

[0041] Step 5: Model Evaluation After the model is trained, its performance is evaluated on an independent test set. The following metrics are calculated: sensitivity (recall), specificity, accuracy, and area under the receiver operating characteristic curve (AUROC).

[0042] This invention significantly improves the accuracy of predicting metastasis risk in patients with papillary thyroid carcinoma using deep learning methods, and is particularly suitable for preoperative stratified diagnosis. The method was validated in 50 trials based on qPCR data from 198 thyroid cancer patients. These results (experimental data are shown in Tables A, B, and C) demonstrate the model's strong robustness and support the solidity and reliability of its conclusions.

[0043] Table B: Validation Results of Overall Transfer Prediction

[0044] Overall metastasis prediction model: Sensitivity 0.795, Specificity 0.909, AUROC 0.904, Accuracy 0.833. The AUC curve and confusion matrix of the prediction results are shown below. Figure 1 and Figure 2 As shown.

[0045] Table C: Validation Results of Central Lymph Node Metastasis Prediction

[0046] Central lymph node metastasis prediction model: sensitivity 0.728, specificity 0.944, AUROC 0.858, accuracy 0.758. The AUC curve and confusion matrix of the prediction results are shown below. Figure 3 and Figure 4 As shown.

[0047] Table D: Validation Results of Cervical Lymph Node Metastasis Prediction

[0048] Lateral cervical lymph node metastasis prediction model: sensitivity 0.971, specificity 0.439, AUROC 0.797, accuracy 0.576. The AUC curve and confusion matrix of the prediction results are shown below. Figure 5 and Figure 6 As shown.

[0049] The experimental results above demonstrate that this technology significantly improves specificity while retaining high sensitivity, showing promising clinical application prospects and research value. It can be used as an auxiliary diagnostic tool for preoperative risk assessment, demonstrating significant social and economic benefits.

[0050] This specific embodiment is merely an explanation of this application and is not intended to limit it. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they fall within the scope of the claims of this application.

Claims

1. A method for constructing a thyroid cancer metastasis risk prediction model, characterized in that, Includes the following steps: S1, Sample collection and preprocessing; S2, Gene expression level detection; S3, Dataset partitioning and standardization; S4, Construction of deep learning model; S5, Model evaluation; The biomarker group for gene expression detection in S2 consists of RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7 and RBP4.

2. The construction method according to claim 1, characterized in that, S1 includes the following steps: Postoperative tumor tissue samples were collected from patients pathologically diagnosed with papillary thyroid carcinoma. Based on the postoperative pathological examination results, each sample was labeled with a metastasis status tag, including whether lymph node metastasis occurred, whether central lymph node metastasis occurred, and whether lateral cervical lymph node metastasis occurred. Total RNA was extracted from tumor tissue, and its concentration, purity, and integrity were detected.

3. The construction method according to claim 2, characterized in that, S2 includes the following steps: The gene expression levels of the biomarker group were detected, and the GAPDH gene was selected as an internal reference gene for data standardization. In the qPCR experiment, three replicates were set up for each gene in each sample, and the Ct value of each reaction well was recorded. The average Ct value of three replicates for each gene was calculated, and the relative expression level of each target gene was calculated. Then, the normalization process was performed to obtain the normalized expression level data that can be used as input to the model.

4. The construction method according to claim 3, characterized in that, S3 includes the following steps: The normalized expression levels of the eight genes in the sample and their corresponding three transfer tags were organized into a data matrix; The entire dataset is randomly divided into a training set and a test set; Z-score normalization was performed on gene expression data.

5. The construction method according to claim 4, characterized in that, S4 includes the following steps: Three independent binary classification models were built using Python and a deep learning framework, including a total metastasis prediction model, a central lymph node metastasis prediction model, and a lateral cervical lymph node metastasis prediction model.

6. The construction method according to claim 5, characterized in that, S5 includes the following steps: After the model is trained, its performance is evaluated on an independent test set, and the following metrics are calculated: sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve.

7. A thyroid cancer metastasis risk prediction model obtained by the construction method according to any one of claims 1-6.

8. The application of the thyroid cancer metastasis risk prediction model according to claim 7 in the preparation of thyroid cancer metastasis risk prediction products.

9. A thyroid cancer metastasis risk prediction system, characterized in that, include: Data acquisition module: Collects relevant information on postoperative tumor tissue samples from patients with papillary thyroid carcinoma, including the source of the samples and pathological diagnosis. At the same time, it collects gene expression data of biomarkers RPS4Y1, PKHD1L1, CRABP1, KRT18P8, AGPAT4, CPQ, SLC26A7 and RBP4 in the samples. Data preprocessing module: Standardizes the collected gene expression data; Data processing module: It has a pre-trained total metastasis prediction model, central lymph node metastasis prediction model and lateral cervical lymph node metastasis prediction model. It processes the standardized gene expression data to obtain the overall probability value P_total, the probability value P_central and the probability value P_lateral of lymph node metastasis in the patient. Results output module: The obtained P_total, P_central, and P_lateral are compared with the preset risk threshold. If P_total > risk threshold, the patient is determined to have a high risk of lymph node metastasis, otherwise it is considered low risk. The same applies to P_central and P_lateral.

10. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for constructing a thyroid cancer metastasis risk prediction model according to any one of claims 1 to 6, or some or all of the steps in the thyroid cancer metastasis risk prediction system according to claim 9.

Images