Esophageal cancer prognosis risk analysis method and system based on machine learning and medium

Abstract

The invention discloses an esophageal cancer prognosis risk analysis method and system based on machine learning, and a medium, and relates to the technical field of artificial intelligence technology and bioinformatics, and the method comprises the steps: 1, collecting multi-modal data of a patient, and carrying out the standardization processing, so as to construct a stable feature set; 2, constructing a plurality of machine learning models, realizing high and low risk group prediction of each machine learning model based on stable feature set modeling, and determining the machine learning model and an optimal feature set according to the high and low risk group prediction; and step 3, calculating SHAP interaction values among the features in the optimal feature set, drawing an interaction value curve according to the SHAP interaction values to obtain TopA interaction feature pairs, taking the interaction feature pairs as newly constructed features to be included in the original feature set in the step 1, and repeating the feature screening process in the step 1 and the step 2 to obtain an optimal machine learning model. Predicting high and low risk groups of patients; according to the prognosis risk analysis method, the detection specificity of high and low risk groups of local advanced esophageal cancer patients is greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and bioinformatics, and in particular to a method, system, and medium for esophageal cancer prognostic risk analysis based on machine learning. Background Technology

[0002] Esophageal cancer is a common malignant tumor, and locally advanced patients face significant treatment challenges and poor prognosis. Accurate risk stratification and prognostic assessment are critical issues that urgently need to be addressed in clinical practice. Currently, methods such as TNM staging, clinicopathological features, and serum biomarkers are relied upon. Among these, TNM staging is the most widely used system for tumor staging internationally and is the standard method for clinical staging of malignant tumors. However, these methods have limitations: TNM staging fails to reflect tumor biological behavior and individual differences; clinical indicators are highly subjective; and the sensitivity and specificity of single biomarkers are insufficient.

[0003] While radiomics can extract quantitative features from CT scans, current research is largely limited to single-modal data, resulting in unsystematic feature selection and insufficient model stability. Clinical and transcriptomic data contain important information but lack in-depth mining of multimodal associations. Therefore, a method is needed to integrate multimodal data, optimize features and models, and reveal macro- and micro-level correlations to achieve accurate prediction and analysis. Summary of the Invention

[0004] Based on the technical problems existing in the background technology, this invention proposes a method, system and medium for esophageal cancer prognostic risk analysis based on machine learning, which significantly improves the detection specificity for high-risk and low-risk groups of patients with locally advanced esophageal cancer.

[0005] The machine learning-based prognostic risk analysis method for esophageal cancer proposed in this invention includes: Step 1: Collect multimodal patient data and standardize it. Extract features from the standardized data to obtain radiomics features. Determine the stable feature set after screening by the original feature set composed of radiomics features and clinical features through a feature screening process. The multimodal data includes CT images and clinical features. Step 2: Construct multiple machine learning models, and based on the selected stable feature set, model and implement high-risk and low-risk group predictions for each machine learning model. Select the machine learning model with the best number of features and the highest comprehensive evaluation score, and determine the optimal feature set. Step 3: Calculate the SHAP interaction values ​​between each feature in the optimal feature set, and construct a feature interaction strength matrix after averaging and absolute processing. Plot the interaction value curve to obtain the Top A interaction feature pairs. Perform hierarchical normalization on the two features in each interaction feature pair, multiply them, and then take the square root. This result is used as a newly constructed feature and incorporated into the original feature set in Step 1. Repeat the feature selection process in Step 1 and Step 2 to obtain the optimal machine learning model, which is used to predict the high-risk and low-risk groups of patients for esophageal cancer prognostic risk analysis, where A is an integer.

[0006] Furthermore, the acquisition and standardization of patient multimodal data includes, specifically, the standardization of the CT images as follows: Collect CT images of the patient before radiotherapy, in which the clinical target area has been delineated by the doctor; Spatial standardization: B-spline interpolation is used to resample CT images to a uniform pixel spacing; for clinical target area masks, nearest neighbor interpolation is used for synchronous resampling to ensure that the spatial dimensions of the mask and the CT image match. Gray-scale normalization: The gray-scale values ​​of the CT image are normalized and scaled to a set standard range through linear transformation; Gray-level discretization: Set Q gray-level intervals to discretize continuous gray-level values ​​into discrete intervals; Image transformation preprocessing: Based on the grayscale discretized CT image, multimodal image representations are generated using wavelet transform, local binary mode, square transform, exponential transform, gradient transform, square root transform, and logarithmic transform methods, which serve as the standardized data.

[0007] Furthermore, the step of extracting features from the standardized data to obtain radiomics features specifically involves: A multi-process parallel processing framework was adopted, and a feature extractor was built based on the pyradiomics library. All feature categories were enabled and features were extracted from the standardized data to obtain radiomics features.

[0008] Furthermore, the original feature set composed of radiomics features and clinical features is used to determine a stable feature set after screening through a feature screening process, wherein the feature screening process specifically includes: (a1) Remove low-variance features with variance less than the set variance; (a2) Perform an independent samples t-test on the characteristics of the high-risk group and the low-risk group, retain the characteristics with p values ​​less than the preset p threshold, and screen out the first candidate feature set with statistically significant differences between the groups. The high-risk group is patients with a survival period of less than 6 months, and the low-risk group is patients with a survival period of more than 36 months. The survival period is the difference between the time of death and the date of end of radiotherapy. (a3) Calculate the Spearman rank correlation coefficient among the first candidate feature sets, construct the correlation matrix, remove features whose absolute value of the correlation coefficient is greater than the preset correlation threshold, and obtain the second candidate feature set accordingly; (a4) Based on the principle of maximum relevance and minimum redundancy, the correlation between the second candidate feature set and the target variable and the redundancy between features are quantified by mutual information. Features with high correlation with the label and low redundancy between features are selected to obtain the third candidate feature set. (a5) The third candidate feature set is selected by L1 regularization and the regularization parameter is optimized by cross-validation to retain features with non-zero coefficients, thereby obtaining the fourth candidate feature. (a6) By resampling a set number of times, repeat (a1) to (a6) to calculate the frequency of each feature in the screening and obtain a stable feature set after screening.

[0009] Further, in step two, multiple machine learning models are constructed. Based on the selected stable feature set, each machine learning model is used to predict high-risk and low-risk groups. The machine learning model with the optimal number of features and the highest comprehensive evaluation score is selected, and the optimal feature set is determined. Specifically: The various machine learning models are: Random Forest, Extremely Random Tree, Gradient Boosting Tree, Gradient Boosting Decision Tree, Support Vector Machine, and Ensemble Model. Optimize the hyperparameters of the random forest using grid search; The importance score of each feature in the stable feature set is scored and sorted in descending order using a random forest with optimized parameters, thereby obtaining an ordered feature set; Starting with the first feature in the ordered feature set, traverse the subsequent features and construct subsets in turn, obtaining a number of subsets equal to the total number of features in the ordered feature set. Each subset includes a single feature or consecutive features. Each machine learning model finds the balance point between "feature quantity and model performance" by traversing the range of feature quantity in the corresponding subset, thereby obtaining a machine learning model that can stably predict high-risk and low-risk groups of patients and the optimal feature set.

[0010] Furthermore, it is also used for patient prognostic assessment, specifically: By combining the optimal feature set with transcriptome data, gene expression-radiomics feature pairs can be identified, which can then be used to analyze patient prognosis and find prognostic biomarkers.

[0011] Furthermore, the joint analysis of the optimal feature set and transcriptome data to mine gene expression-radiomics feature pairs specifically involves: Collect transcriptomic data from patient tumor tissues to obtain a pivot gene expression dataset; The optimal feature set is preprocessed to remove non-feature columns and retain patient identifiers and radiomics feature columns. Using patient identifiers as the association key, the hub gene expression dataset is merged with the preprocessed optimal feature set to form a joint dataset containing patient identifiers, gene expression data, and radiomics features; Gene expression and radiomics feature columns were extracted from the joint dataset and used as input variables for correlation analysis. A correlation calculation function was defined, and the function was calculated for each gene expression-radiomics feature pair combination. Gene expression-radiomics feature pairs that met the conditions were selected based on the correlation threshold and the statistical significance threshold.

[0012] Furthermore, a correlation calculation function is defined, and at least three correlation calculation functions are set in parallel to calculate each gene expression-radiomics feature pair combination. A computer system includes a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the above-described method.

[0013] A computer-readable storage medium storing a plurality of classification programs, the plurality of classification programs being invoked by a processor to execute the method described above.

[0014] The advantages of the machine learning-based esophageal cancer prognostic risk analysis method, system, and medium provided by this invention are as follows: based on multimodal and multidimensional modeling and analysis, it solves the problem of the one-sidedness of single-index analysis; through the comparison of multiple machine learning models, the detection specificity of high-risk and low-risk groups of patients with locally advanced esophageal cancer is greatly improved; and through the integration of multidimensional data analysis, it has better interpretability of the prognosis of patients with locally advanced esophageal cancer. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 A schematic diagram of the Top 15 features retained after Bootstrap resampling feature filtering; Figure 3 The PDP image with the maximum probability is obtained by multiplying the inverse variance of the gray-level co-occurrence matrix of the wavelet-LLL layer by the gray-level co-occurrence matrix of the wavelet-HHL layer. Figure 4 A schematic diagram illustrating the optimal number of features and the ROC of the optimal model; Figure 5 This is a topological map showing the correlation between gene expression and radiomics features. Detailed Implementation

[0016] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0017] like Figures 1 to 5 As shown, the machine learning-based prognostic risk analysis method for esophageal cancer proposed in this invention includes: Step 1: Collect multimodal patient data and standardize it. Extract features from the standardized data to obtain radiomics features. Determine the stable feature set after screening by the original feature set composed of radiomics features and clinical features through a feature screening process. The multimodal data includes CT images and clinical features. Step 2: Construct multiple machine learning models, and based on the selected stable feature set, model and implement high-risk and low-risk group predictions for each machine learning model. Select the machine learning model with the best number of features and the highest comprehensive evaluation score, and determine the optimal feature set. Step 3: Calculate the SHAP interaction values ​​between each feature in the optimal feature set, and construct a feature interaction strength matrix after averaging and absolute processing. Plot the interaction value curve to obtain the Top A interaction feature pairs. Perform hierarchical normalization on the two features in each interaction feature pair, multiply them, and then take the square root. This result is used as a newly constructed feature and incorporated into the original feature set in Step 1. Repeat the feature selection process in Step 1 and Step 2 to obtain the optimal machine learning model, which is used to predict the high-risk and low-risk groups of patients for esophageal cancer prognostic risk analysis, where A is an integer.

[0018] This embodiment is based on multimodal and multidimensional modeling and analysis to solve the problem of the one-sidedness of single-index analysis. By comparing various machine learning models, the detection specificity of high-risk and low-risk groups of patients with locally advanced esophageal cancer is greatly improved. By integrating multidimensional data analysis, it also has better interpretability for the prognosis of patients with locally advanced esophageal cancer.

[0019] This embodiment can mainly realize the risk stratification and risk assessment of the prognosis of locally advanced esophageal cancer. Among them, (1) the machine learning model obtained in step three is used for risk stratification, (2) the optimal feature set and transcriptome data are analyzed together to mine gene expression-radiomics feature pairs, thereby mining the relationship between images and genes, mining the micro nature through macro phenomena, analyzing the patient's prognosis, finding prognosis-related biomarkers, and thus conducting risk assessment. The details are as follows.

[0020] I. Risk stratification of prognosis in locally advanced esophageal cancer; Phase 1: Constructing predictive models for high-risk and low-risk groups of locally advanced esophageal cancer patients based on multimodal data; Specifically, it includes: S1. Data Acquisition: Pre-radiotherapy localization CT scans were performed on esophageal cancer patients while they were breathing freely. The scans were conducted on a Brilliance CT BigBore system manufactured by Philips Medical Systems in Cleveland, USA, using CT simulation localization technology. Preferably, the scan parameters were set to 120 kV, an effective milliampere-second range of 300-350, a beam width of 16 × 1.5 mm, an image resolution of 512 × 512, a pitch range of 0.9-1.0, and a rotation time range of 0.7-0.8 seconds. Each slice was 2-5 mm thick, integrating enhanced chest CT and arterial CT. The CT images were then imported into a CT simulation workstation to precisely delineate the region of interest.

[0021] On localization CT scans, the three-dimensional volumes of interest for the GTV and CTV are manually drawn layer by layer using the Elekta Monaco planning system V3.8.0 by a physician with five years of experience in esophageal imaging (e.g., a senior radiologist). These images are then reviewed by a radiation oncologist specializing in upper gastrointestinal malignancies. Any disagreements are resolved through discussion.

[0022] GTV (Gross Tumor Volume) and CTV (Clinical Target Volume) are core terms used in radiotherapy to define the treatment area. GTV refers to the tumor volume that is clearly visible through imaging examinations (such as CT and MRI), including the primary lesion and metastatic lymph nodes. It is the "visible target" that is actually irradiated in radiotherapy and usually corresponds to the part of the tumor with the highest density. CTV extends a certain range from GTV to cover any possible subclinical lesions (such as microscopic infiltration and edema). CTV ensures that all potentially malignant tissues are treated, and its range is usually larger than GTV. The specific extension amount depends on the type and location of the tumor (e.g., CTV for lung cancer is often extended by 5-10 mm).

[0023] Clinical data collection relies on the hospital's clinical big data system to collect patient age, BMI, clinical stage, past medical history, blood routine test indicators, and tumor marker indicators; senior clinicians follow up on patient survival time, which is the difference between the time of death and the end date of radiotherapy; record patient survival time (month) and whether the patient has died (die: 1 indicates death, 0 indicates survival).

[0024] Based on the patients' survival time after radiotherapy, patients with a survival time of less than 6 months were included in the high-risk group (label 1), and patients with a survival time of more than 36 months were included in the low-risk group (label 0). The standardized processing of CT images strictly follows the International Initiative for Standardization in Radiomics (IBSI) guidelines, specifically including: Spatial standardization: B-spline interpolation is used to resample CT images to a uniform pixel spacing (e.g., 1mm×1mm×1mm) to ensure spatial scale consistency of images from different patients; for clinical target volume (CTV) masks, nearest neighbor interpolation is used for synchronous resampling to ensure strict spatial dimension matching between the mask and CT images; Gray-scale normalization: The gray-scale values ​​of CT images are normalized by scaling them to a set standard range (e.g., 0-100) through linear transformation, eliminating gray-scale shifts caused by different scanning devices and parameters, and enhancing feature comparability; Gray-level discretization: Set Q (e.g., 25) gray-level intervals to discretize continuous gray-level values ​​into discrete intervals, reduce image noise interference, and stabilize texture feature calculation. Q is an integer, and its specific value is determined as needed. Image transformation preprocessing: Based on the grayscale discretized CT images, eight transformation methods are used to generate multimodal image representations, including wavelet transform, local binary mode (LBP2D / LBP3D), square transform, exponential transform, gradient transform, square root transform, and logarithm transform, to maximize the mining of tumor image information, which is then used as the standardized data. S2. The original feature set composed of radiomics features and clinical features is used to determine the stable feature set after screening through a feature screening process: A multi-process parallel processing framework is adopted, and a feature extractor is built based on the pyradiomics library, enabling all feature categories (including first-order statistical features, shape features, gray-level co-occurrence matrix, gray-level running length matrix, etc.). pyradiomics is an open-source Python library specifically designed for extracting high-throughput quantitative radiomics features from medical images (such as MRI, CT, PET, etc.).

[0025] The feature selection process is a progressive selection process of "significance test → redundancy filtering → correlation optimization → sparse compression → stability verification", with specific steps including (a1) to (a6): (a1) The first step is to remove low-variance features with variance less than the set variance (e.g., 1e-6) (to avoid interference from features with no discriminative power) and retain the initial feature set with practical information value. (a2) The second step is to perform an independent samples t-test (Welch test, without assuming equal variance) on the features of the high-risk group and the low-risk group (label 1 / 0), retain the features with p values ​​less than the preset p threshold (e.g., threshold 0.05), screen out candidate features with statistically significant differences between groups, and exclude noisy features with no significant differences between groups, as the first candidate feature set; (a3) The third step is to calculate the Spearman rank correlation coefficient between the first candidate feature sets and construct the correlation matrix; remove features whose absolute correlation coefficient is greater than the preset correlation threshold (the threshold is 0.9), eliminate highly collinear redundant features, and retain the subset with high independence between features to obtain the second candidate feature set; (a4) The fourth step is based on the principle of maximum relevance and minimum redundancy (mRMR). Mutual information (MIC) is used to quantify the correlation between the second candidate feature set and the target variable and the redundancy between features. Features with high correlation with the label and low redundancy between features are selected (Formula 1) to obtain the third candidate feature set. The number of features retained is 1 / 10 of the sample size (maximum 19), balancing the amount of feature information and computational efficiency. Formula 1: Where S represents the feature subset in the second candidate set, These represent the average correlation (maximization term) between the feature subset S and the target variable, and the average redundancy (minimization term) among features within the feature subset S, respectively. A greedy strategy is used to obtain the third candidate feature set during the calculation.

[0026] (a5) The fifth step uses L1 regularization (Lasso) for further screening (Formula 2). The regularization parameter (alpha) is optimized through 10-fold cross-validation to retain features with non-zero coefficients. By utilizing the sparsity of L1 penalty, the feature space is compressed to the maximum extent while ensuring predictive ability, and finally a highly recognizable core feature set is obtained. Formula 2: ; in, The first term is the regularization parameter, and the second term is the L1 penalty term. This represents the coefficient vector of the Lasso model (sparse regression model). For the number of samples (patients), Index of the sample (patient) (from 1- ), For the first The label value corresponding to each sample (patient). For the intercept term of the Lasso model, Index of features The number of features in the third candidate feature set. For the first The coefficient vector corresponding to each feature.

[0027] This embodiment is based on the principle that "L1 penalty will cause some coefficients to be affected". The mathematical property of "shrinking to 0" (sparse solution characteristics in convex optimization) preserves the non-zero coefficient feature.

[0028] (a6) By resampling a set number of times, repeat (a1) to (a6) to calculate the frequency of each feature in the screening and obtain a stable feature set after screening.

[0029] As one implementation method, the screening process (a1) to (a6) is repeated 500 times using Bootstrap resampling (sampling with replacement) to calculate the frequency of each feature in the screening process, such as... Figure 2 As shown, this ensures the stability and robustness of the screening results.

[0030] S3. This modeling process employs a multi-model comparison and system optimization strategy, combined with data preprocessing, feature importance ranking, and cross-validation, to achieve accurate prediction of high-risk and low-risk groups in patients with locally advanced esophageal cancer. The specific steps are as follows (b1) to (b6): (b1) Load radiomics features and clinical features, filter the preset core feature subset (including radiomics features such as wavelet-LLL layer_grayscale co-occurrence matrix_inverse variance (wavelet_LLL_glcm_InverseVariance), and clinical features such as lymphocyte percentage (Lymphocyte Percentage)), remove non-feature columns such as patient ID and invalid data rows to ensure the validity of the input data; (b2) Feature standardization: The standardization tool StandardScaler is used to scale the features to a distribution with a mean of 0 and a standard deviation of 1 to eliminate the influence of dimensional differences on the model; (b3) Class imbalance handling: The minority class samples are synthesized by the SMOTE algorithm to balance the sample ratio of the high-risk group and the low-risk group (label 0 / 1) and avoid the model from being biased towards the majority class. SMOTE (Synthetic Minority Over-sampling Technique) is an oversampling algorithm used to solve the problem of data class imbalance. It balances the dataset by synthesizing minority class samples, thereby improving the classification model's ability to identify the minority class. (b4) Dataset partitioning: The resampled data is split into a training set (for model training) and a test set (for preliminary performance evaluation) in an 8:2 ratio to ensure the randomness and stability of the data partitioning (the random seed is fixed at 42). (b5) Construct 6 machine learning models; Six classifiers were constructed, including Random Forest, Extremely Random Tree, Gradient Boosting Tree (XGBoost), Gradient Boosting Decision Tree (LightGBM), Support Vector Machine (SVM), and Ensemble Model (VotingClassifier, which uses soft voting to fuse the probability prediction results of Random Forest and Extremely Random Tree). The hyperparameters of the random forest are optimized using grid search (5-fold cross-validation) (the total number of decision trees n_estimators is set to 50 / 100 / 200, and the maximum depth of the decision trees max_depth is set to None / 10 / 20).

[0031] (b6) Calculate feature importance based on the random forest method (quantify contribution through tree node splitting); The importance score of each feature in the stable feature set is scored and sorted in descending order using a random forest with optimized parameters, thereby obtaining an ordered feature set.

[0032] (b7) Starting with the first feature in the ordered feature set, traverse the subsequent features and construct subsets in turn to obtain a number of subsets equal to the total number of features in the ordered feature set. Each subset includes a single feature or consecutive features. It should be noted that, for example, if there are 5 features in the ordered feature set, then the first feature is used as the starting point to construct a subset, and a total of 5 subsets are constructed, specifically: {1}, {1,2}, {1,2,3}, {1,2,3,4}, {1,2,3,4,5}.

[0033] (b8) Each machine learning model finds the balance point of "feature quantity - model performance" by traversing the range of feature quantity of the corresponding subset, thereby obtaining a machine learning model that can stably predict the high-risk group and low-risk group of patients and the optimal feature set.

[0034] It should be noted that, within the range of the number of features traversed in the subset, the random forest specifically refers to the parameter-optimized random forest.

[0035] Example 1: Evaluation of 6 machine learning models in the first phase; The model is evaluated using a multi-dimensional metric, including AUC (area under the ROC curve, reflecting discrimination ability), precision, recall, and F1 score (all using macro averages to adapt to imbalanced data). The best model is selected by the comprehensive score (mean) of the four metrics, thus finding the optimal machine learning model.

[0036] Load a dataset containing radiomics features, survival time (month), and outcome events (die, 1 indicates an event occurred, 0 indicates censoring); Kaplan-Meier survival analysis and log-rank test were used to quantify the association between radiomics features and patient survival prognosis. Kaplan-Meier is a nonparametric statistical method widely used in survival analysis, mainly for estimating survival functions and comparing survival rates between different groups. The quantiles of the eigenvalues ​​(25%, 50%, 75%) are used to set the grouping thresholds, dividing patients into "high value group" (eigenvalue > threshold) and "low value group" (eigenvalue ≤ threshold), thus achieving binary grouping of the samples; When determining the threshold, it is important to ensure that the sample ratio of the "high value group" and the "low value group" is close to 1:1. The study found that all 12 features in the optimal feature set were significantly correlated with patient survival, with p-values ​​all <0.05, further demonstrating the robustness and stability of the feature set.

[0037] The comparison revealed that the AUC predicted by image fusion clinical data analysis for high-risk and low-risk patients was about 7% higher than that of the single image model and about 13% higher than that of the single clinical model.

[0038] The second stage: Based on the optimal feature set identified in the first stage, analyze the "AND" relationships between the features in the optimal feature set, specifically including: Based on the optimal machine learning model obtained after the first stage of training, the SHAP interaction values ​​(shap_interaction_values) between image features in the optimal feature set are calculated using shap.TreeExplainer (shap.TreeExplainer is a tool in the SHAP library specifically for interpretability analysis of tree models), as shown in Formula 3. Formula 3: ; in, This represents all image features in the optimal feature set. for Not included Feature subset, This is the optimal machine learning model. Indicates that only a subset is retained. The output of the features, These are feature indexes, for Features in a feature subset for The total number of features in the feature subset. , , These represent the added features. Add features and simultaneously add features and The predicted value at that time Features in the sample With features The interaction value.

[0039] Average and absolute value of SHAP interaction. (Formula 5), ​​and remove diagonal elements (only retain the interaction effects between features off-diagonal, and exclude the main effects of a single feature), and construct the feature interaction intensity matrix; Formula 4: ; in, This refers to the sample size (i.e., the number of patients). For the first Features in a sample (patient) and The interaction value.

[0040] Create an interaction heatmap: Use color gradients to display the interaction intensity of all feature pairs, intuitively presenting the global interaction pattern; Plot the interaction value curve: Mark significant inflection points and define the threshold boundary of significant interactions. That is, the interaction threshold is derived from the interaction value curve. The selected threshold is the value when the curve has the largest rate of change - the inflection point of the curve. That is, the threshold boundary is determined to be 0.0068. Retain the top A interaction feature pairs, for example, obtain the top 5 interaction feature pairs, where A is an integer.

[0041] The five interaction feature pairs include: Wavelet-LLL layer_grayscale dependency matrix_dependence variance (wavelet-LLL_gldm_DependenceVariance) × Lymphocyte Percentage: capturing the synergistic effect of "tumor macrostructural heterogeneity" and "body's immune status"; Lymphocyte Percentage × Wavelet-LLL Layer Gray-Level Dependence Matrix Large Dependence High Gray-Level Emphasis: Focusing on the combined effects of "immune status" and "range of highly active tumor regions"; Wavelet-LLL layer_gray-level dependency matrix_large-dependence-high-gray-level-emphasis (wavelet-LLL_gldm_LargeDependenceHighGrayLevelEmphasis) × Wavelet-HHL layer_gray-level co-occurrence matrix_maximum-probability (wavelet-HHL_glcm_MaximumProbability): captures the synergistic effect of tumor "macrostructural activity" and "local typical patterns", reflecting the "global-local" structural consistency of tumor; Exponential_glrlm_RunEntropy × Wavelet-HHL Layer_Gray-Level Co-occurrence Matrix_MaximumProbability: Focuses on the combined effect of tumor "local texture disorder" and "typical structural pattern", reflecting the relationship between the "degree of disorder" and "dominant pattern" of tumor microstructure; Wavelet-LLL layer_gray-level co-occurrence matrix_inversevariance (wavelet_LLL_glcm_InverseVariance) × Wavelet-HHL layer_gray-level co-occurrence matrix_maximumprobability (wavelet_HHL_glcm_MaximumProbability): Reflects the synergistic effect of tumor "macro-heterogeneity" and "local canonical pattern".

[0042] Generate dependency plots: For the top 5 interactive feature pairs, show the trend of the SHAP value of one feature changing with the value of another feature, visualizing the influence of the combined effect of the two features on the prediction results. A total of 5 dependency plots are obtained, each corresponding to one of the top 5 interactive feature pairs. The purpose of plotting is to visualize the interaction relationships between the interactive features. One dependency plot is shown below. Figure 3 As shown: Wavelet-LLL layer_gray-level co-occurrence matrix_inverse variance × Wavelet-HHL layer_gray-level co-occurrence matrix_maximum probability PDP image; The five selected interactive feature pairs are multiplied after hierarchical normalization, and then the square root is taken as the newly constructed feature and incorporated into the original feature set of the first stage.

[0043] The same feature selection and modeling process was followed in the first stage; The wavelet-LLL layer_gray-level co-occurrence matrix_inverse variance × wavelet-HHL layer_gray-level co-occurrence matrix_the most probable interaction feature pair is the only one retained in the Bootstrap resampling (sampling with replacement) feature selection and ranks second in importance in the model (calculated by the random forest algorithm).

[0044] Wavelet-LLL layer_gray-level co-occurrence matrix_inverse variance characterizes "uniformity of global contour"; Wavelet-HHL layer - gray-level co-occurrence matrix - maximum probability characterization of "dominant mode of local details"; When these two features interact, the model can simultaneously focus on the "regularity of the global form" and the "typical patterns of local details," capturing collaborative information that a single feature cannot cover.

[0045] The results show that adding interactive features improved model performance by approximately 2% in AUC, and approximately 3% in Precision and Recall. Figure 4 As shown, this further improves the interpretability of predictions for high-risk and low-risk patient groups.

[0046] II. Risk assessment of prognosis in locally advanced esophageal cancer; The third stage involves jointly analyzing the optimal feature set with transcriptomic data to uncover gene expression-radiomics feature pairs, thereby analyzing patient prognosis and identifying prognostic biomarkers. The specific process includes: Collect transcriptome data from patients' tumor tissues; The Top20 hub genes were obtained using four methods: betweenness centrality (BC), closeness centrality (CC), degree centrality (DC), and maximal clique centrality (MCC).

[0047] Load the hub gene expression dataset (gene_df) and the optimal feature set respectively.

[0048] The optimal feature set is preprocessed to remove non-feature columns (such as the label column) and retain the patient identifier (patient_id) and radiomics feature columns.

[0049] Using the patient identifier (patient_id) as the association key, the gene expression dataset is merged with the selected optimal feature set to form a joint dataset (merged_df) containing patient identifiers, gene expression data, and radiomics features, ensuring data alignment at the sample level.

[0050] Gene expression columns (excluding patient_id) and radiomics feature columns (excluding patient_id) were extracted from the joint dataset and used as input variables for correlation analysis.

[0051] Define a correlation calculation function to calculate the correlation for each gene expression-radiomics feature pair combination: Pearson correlation coefficient and corresponding p-value (measures the degree of linear correlation). Spearman correlation coefficient and corresponding p-value (measures the degree of monotonic correlation, applicable to nonlinear relationships); Kendall correlation coefficient and corresponding p-value (rank-based correlation metric, more robust to outliers).

[0052] A correlation threshold (absolute value > 0.7) and a statistical significance threshold (p value < 0.05) were set to screen gene expression-radiomics feature pairs that met the criteria.

[0053] Create a correlation heatmap of gene expression-radiomics feature pairs to visualize the relationships, such as... Figure 5 As shown; Identify multiple pairs of gene expression-radiomics feature pairs, among which: Original shape major axis length (ITGB2) is positively correlated with genes and images; Wavelet-HHL layer_gray-level co-occurrence matrix_maximum probability (IL1B---wavelet-HHL_glcm_MaximumProbability), genes are negatively correlated with images; Wavelet-HHH layer_gray-level co-occurrence matrix_second-order moment of inertia (IL1B---wavelet-HHH_glcm_Imc2), genes are negatively correlated with images; three gene-image feature pairs satisfy three correlation calculation methods and can be used as consensus gene expression-image feature pairs.

[0054] Plotting a topological map of the correlation between gene expression and radiomics features, such as... Figure 5 ; Based on Kaplan-Meier survival analysis and log-rank test analysis of the relationship between gene expression-radiomics features and patient prognosis, the following findings were found: A low ITGB2 level may indicate a better prognosis; Low IL1B expression is associated with better prognosis (which is consistent with clinical logic).

[0055] The results of the above three stages fully demonstrate that radiomics combined with multi-indicator clinical data can accurately predict the high and low risk of locally advanced esophageal cancer patients, with better results than single imaging or single clinical data. In-depth research into the "AND" relationship between imaging and clinical features reveals that interactive feature pairs can explain the high and low risk stratification and prognosis of patients from a higher dimension. Joint analysis of radiomics and transcriptomics data can uncover the subtle relationship between imaging and genes, thereby connecting macroscopic phenomena with microscopic essences and achieving multi-dimensional evaluation of patient prognosis. Therefore, this invention's method for predicting and analyzing the high and low risk groups of locally advanced esophageal cancer based on multimodal data is progressively more detailed, moving from macroscopic to microscopic levels, solving the current limitations of single-modality analysis, difficulties in connecting multimodal data, and poor interpretability.

[0056] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based method for prognostic risk analysis of esophageal cancer, characterized in that, include: Step 1: Collect multimodal patient data and standardize it. Extract features from the standardized data to obtain radiomics features. Determine the stable feature set after screening by the original feature set composed of radiomics features and clinical features through a feature screening process. The multimodal data includes CT images and clinical features. Step 2: Construct multiple machine learning models. Based on the stable feature set selected in Step 1, model and implement high-risk and low-risk group predictions for each machine learning model. Select the machine learning model with the best number of features and the highest comprehensive evaluation score, and determine the optimal feature set. Step 3: Calculate the SHAP interaction values ​​between each feature in the optimal feature set, and construct a feature interaction strength matrix after averaging and absolute processing. Plot the interaction value curve to obtain the Top A interaction feature pairs. Perform hierarchical normalization on the two features in each interaction feature pair, multiply them, and then take the square root. This result is used as a newly constructed feature and incorporated into the original feature set in Step 1. Repeat the feature selection process in Step 1 and Step 2 to obtain the optimal machine learning model, which is used to predict the high-risk and low-risk groups of patients for esophageal cancer prognostic risk analysis, where A is an integer.

2. The analytical method according to claim 1, characterized in that, The process of acquiring and standardizing patient multimodal data includes, specifically, the standardization of the CT images: Collect CT images of the patient before radiotherapy, in which the clinical target area has been delineated by the doctor; Spatial standardization: B-spline interpolation is used to resample CT images to a uniform pixel spacing; for clinical target area masks, nearest neighbor interpolation is used for synchronous resampling to ensure that the spatial dimensions of the mask and CT images match. Gray-scale normalization: The gray-scale values ​​of the CT image are normalized and scaled to a set standard range through linear transformation; Gray-level discretization: Set Q gray-level intervals to discretize continuous gray-level values ​​into discrete intervals; Image transformation preprocessing: Based on the grayscale discretized CT image, multimodal image representations are generated using wavelet transform, local binary mode, square transform, exponential transform, gradient transform, square root transform, and logarithmic transform methods, which serve as the standardized data.

3. The analytical method according to claim 1, characterized in that, The process of extracting features from the standardized data to obtain radiomics features specifically involves: A multi-process parallel processing framework was adopted, and a feature extractor was built based on the pyradiomics library. All feature categories were enabled and features were extracted from the standardized data to obtain radiomics features.

4. The analytical method according to claim 1, characterized in that, The original feature set, composed of radiomics features and clinical features, is used to determine a stable feature set after screening through a feature screening process. Specifically, the feature screening process is as follows: (a1) Remove low-variance features with variance less than the set variance; (a2) Perform an independent samples t-test on the characteristics of the high-risk group and the low-risk group, retain the characteristics with p values ​​less than the preset p threshold, and screen out the first candidate feature set with statistically significant differences between the groups. The high-risk group is patients with a survival period of less than 6 months, and the low-risk group is patients with a survival period of more than 36 months. The survival period is the difference between the time of death and the date of end of radiotherapy. (a3) Calculate the Spearman rank correlation coefficient among the first candidate feature sets, construct the correlation matrix, remove features whose absolute value of the correlation coefficient is greater than the preset correlation threshold, and obtain the second candidate feature set accordingly; (a4) Based on the principle of maximum relevance and minimum redundancy, the correlation between the second candidate feature set and the target variable and the redundancy between features are quantified by mutual information. Features with high correlation with the label and low redundancy between features are selected to obtain the third candidate feature set. (a5) The third candidate feature set is selected by L1 regularization and the regularization parameter is optimized by cross-validation to retain features with non-zero coefficients, thereby obtaining the fourth candidate feature. (a6) By resampling a set number of times, repeat (a1) to (a6) to calculate the frequency of each feature in the screening and obtain a stable feature set after screening.

5. The analytical method according to claim 1, characterized in that, Step 2: Construct multiple machine learning models. Based on the selected stable feature set, model each machine learning model to predict high-risk and low-risk groups. Select the machine learning model with the optimal number of features and the highest comprehensive evaluation score, and determine the optimal feature set. Specifically: The various machine learning models are: Random Forest, Extremely Random Tree, Gradient Boosting Tree, Gradient Boosting Decision Tree, Support Vector Machine, and Ensemble Model. Optimize the hyperparameters of the random forest using grid search; The importance score of each feature in the stable feature set is scored and sorted in descending order using a random forest with optimized parameters, thereby obtaining an ordered feature set; Starting with the first feature in the ordered feature set, traverse the subsequent features and construct subsets in turn, obtaining a number of subsets equal to the total number of features in the ordered feature set. Each subset includes a single feature or consecutive features. Each machine learning model finds the balance point between "feature quantity and model performance" by traversing the range of feature quantity in the corresponding subset, thereby obtaining a machine learning model that can stably predict high-risk and low-risk groups of patients and the optimal feature set.

6. The analytical method according to claim 1, characterized in that, It is also used for patient prognostic assessment, specifically: By combining the optimal feature set with transcriptome data, gene expression-radiomics feature pairs can be identified, which can then be used to analyze patient prognosis and find prognostic biomarkers.

7. The analytical method according to claim 6, characterized in that, The method of jointly analyzing the optimal feature set with transcriptome data to uncover gene expression-radiomics feature pairs specifically involves: Collect transcriptomic data from patient tumor tissues to obtain a pivot gene expression dataset; The optimal feature set is preprocessed to remove non-feature columns and retain patient identifiers and radiomics feature columns. Using patient identifiers as the association key, the hub gene expression dataset is merged with the preprocessed optimal feature set to form a joint dataset containing patient identifiers, gene expression data, and radiomics features; Gene expression and radiomics feature columns were extracted from the joint dataset and used as input variables for correlation analysis. A correlation calculation function was defined, and the function was calculated for each gene expression-radiomics feature pair combination. Gene expression-radiomics feature pairs that met the conditions were selected based on the correlation threshold and the statistical significance threshold.

8. The analytical method according to claim 6, characterized in that, Define correlation calculation functions and use at least three correlation calculation functions set in parallel to calculate the correlation coefficient of each gene expression-radiomics feature pair combination.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of classification programs, which are used by a processor to execute the method as described in any one of claims 1-8.

Images