Construction method, application and system of lung adenocarcinoma tumor invasiveness prediction model based on multi-modal data, electronic device and storage medium

Abstract

This invention provides a method, application, system, electronic device, and storage medium for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data, belonging to the field of artificial intelligence-assisted diagnostic technology. The method involves collecting peripheral blood samples from lung adenocarcinoma patients to obtain clinical phenotypic data, serum biochemical data, driver gene mutation status data, and serum metabolomics data. After missing value imputation, categorical variable encoding, outlier handling, and standardization, the optimal regularization parameter is determined using the LASSO logistic regression algorithm through cross-validation. Simultaneously, automatic variable selection and model parameter estimation are achieved to construct a tumor invasiveness prediction model. This invention integrates multi-level biological information, and the selected predictive factors have clear biological interpretability. All input data are obtained through peripheral blood, eliminating the need for histopathological specimens or imaging examinations, achieving minimally invasive detection, and is suitable for clinical scenarios such as preoperative invasiveness assessment and dynamic monitoring.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted diagnostic technology, specifically to a method, application, system, electronic device, and storage medium for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data. Background Technology

[0002] Lung cancer is one of the leading causes of cancer-related deaths worldwide, with lung adenocarcinoma being the most common pathological subtype of non-small cell lung cancer. With the widespread use of low-dose spiral CT in early lung cancer screening, an increasing number of lung nodules, especially early-stage lung adenocarcinoma, are being detected clinically. However, the biological behavior of early-stage lung adenocarcinoma is highly heterogeneous: some present as indolent carcinoma in situ or microinvasive carcinoma, progressing slowly and requiring only regular follow-up; while others are highly aggressive, exhibiting pathological features such as lymph node metastasis, airway dissemination, vascular invasion, bronchial wall infiltration, or breaching the pleural elastic lamina, making them highly prone to recurrence and metastasis, necessitating timely radical surgery and postoperative adjuvant therapy. Therefore, accurately assessing the invasiveness of lung adenocarcinoma before treatment is of crucial clinical significance for developing individualized precision treatment plans and avoiding overtreatment or undertreatment.

[0003] Currently, the gold standard for clinically assessing the invasiveness of lung adenocarcinoma still relies on histopathological specimens obtained after surgical resection or biopsy. However, this method is significantly invasive, and the sampling process causes pain and risks to patients. It is difficult to achieve a comprehensive assessment for patients who cannot tolerate biopsy or have multiple pulmonary nodules. In addition, biopsy sampling itself has sampling errors and may not fully reflect the heterogeneity of the tumor. Furthermore, although traditional imaging features can provide some clues to invasiveness, their predictive efficacy is limited, and there is a high misdiagnosis rate for cases with atypical imaging manifestations, making it difficult to accurately distinguish between invasive and non-invasive lesions.

[0004] In recent years, researchers have attempted to construct predictive models from different perspectives to assist in the diagnosis and prognostic assessment of lung adenocarcinoma, but existing technologies still have significant shortcomings in addressing the specific clinical need of predicting tumor invasiveness.

[0005] Chinese patent application CN 118888010 A discloses a method for predicting the risk of lung adenocarcinoma. This method uses RNA sequencing data from public databases such as TCGA and employs LASSO regression to screen five risk genes from differentially expressed calcium-related genes, constructing a linear risk scoring model to predict the survival prognosis of lung adenocarcinoma patients. However, the input data for this method is essentially limited to a single dimension: gene mRNA expression levels. It does not involve metabolomics data or serum biochemical indicators, and it requires RNA sequencing data from tumor tissue, still relying on surgery or biopsy to obtain tissue samples, making it difficult to meet the clinical needs of minimally invasive testing. More importantly, this method predicts long-term survival probability, rather than the current invasive biological characteristics of the tumor, and therefore cannot directly serve the formulation of preoperative treatment plans.

[0006] Chinese patent application CN 121171603 A discloses a prognostic risk assessment method for EGFR wild-type lung adenocarcinoma based on multi-omics and machine learning. This method integrates data from four omics—mRNA, lncRNA, miRNA, and DNA methylation—for consensus clustering molecular typing, and uses multi-algorithm machine learning to screen core genes and construct a prognostic risk scoring model. Although this method nominally achieves "multi-omics" fusion, its various omics data are all at the nucleic acid level—transcriptomics or epigenetics—and the data source is tissue-level sequencing data from public databases, failing to extend to metabolomics or serum biochemistry. Furthermore, it is only applicable to the specific subgroup of EGFR wild-type lung adenocarcinoma, limiting its coverage. The predictive target of this method is also survival prognosis rather than tumor invasiveness.

[0007] Chinese patent application CN 115458173 A discloses a multi-omics diagnostic model for differentiating benign and malignant pulmonary nodules based on serum metabolic fingerprinting and its construction method. This method utilizes MALDI-MS technology to obtain serum metabolic fingerprint profiles, combines CEA protein content and CT imaging features, and constructs a diagnostic model through a multi-level stacked machine learning approach to differentiate between benign and malignant pulmonary nodules. While this method involves serum metabolic testing, it uses unidentified full-spectrum metabolic fingerprint features, lacking clear biological interpretability. Furthermore, this method addresses the qualitative diagnostic question of "whether a nodule is malignant," which is a different level of clinical decision-making than the quantitative grading question of "the degree of invasiveness after a diagnosis of lung adenocarcinoma." In addition, this method does not incorporate driver gene mutation status and multiple serum biochemical indicators, and relies on CT imaging as one of the core modalities.

[0008] In summary, existing technologies for preoperative prediction of lung adenocarcinoma tumor invasiveness still suffer from the following common shortcomings: First, there is a lack of specialized predictive tools for the invasiveness of diagnosed lung adenocarcinoma. Existing models mostly focus on survival prognosis prediction or benign / malignant differential diagnosis, failing to cover the critical clinical need of invasiveness assessment. Second, the data modalities are singular or have limited fusion levels, failing to achieve cross-level systemic fusion of clinical phenotypic information, driver gene mutation status, serum biochemical indicators, and high-throughput metabolomics data. Third, most methods rely on obtaining tumor tissue specimens or imaging examinations, and cannot achieve minimally invasive detection through peripheral blood, limiting their application in clinical scenarios such as inoperable cases, multiple nodules, or cases requiring dynamic monitoring. Therefore, there is an urgent need for a minimally invasive assessment method that can integrate multi-dimensional peripheral blood detection data to accurately predict the invasiveness of lung adenocarcinoma. Summary of the Invention

[0009] The purpose of this invention is to provide a method, application, system, electronic device, and storage medium for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data. By integrating four types of multimodal data from the peripheral blood of lung adenocarcinoma patients—clinical phenotype, serum biochemistry, driver gene mutation status, and metabolomics—the invention uses the LASSO logistic regression algorithm to automatically screen key variables and construct a prediction model, thereby achieving minimally invasive and precise preoperative assessment of lung adenocarcinoma tumor invasiveness.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data includes the following steps: S1: Data Acquisition: Collect peripheral blood samples from patients diagnosed with lung adenocarcinoma to obtain multimodal data, including clinical phenotype data, serum biochemical data, driver gene mutation status data, and serum metabolomics data; S2: Data preprocessing: The multimodal data is imputed for missing values, categorical variable encoding is performed, outlier handling is performed, and standardization is performed to obtain a standardized dataset; S3: Data partitioning: The standardized dataset is randomly divided into a training set and a test set according to a preset ratio; S4: Model building and variable selection: Using tumor invasiveness as the dependent variable, the LASSO logistic regression algorithm with L1 regularization penalty term is used to fit the training set. The optimal regularization parameter λ is determined through cross-validation. At the same time, automatic variable selection and model parameter estimation are realized to obtain a lung adenocarcinoma tumor invasiveness prediction model containing variables with non-zero coefficients. S5: Model Evaluation: Evaluate the performance of the prediction model using the test set; The tumor invasiveness is determined comprehensively based on the patient's pathological invasiveness characteristics, which include one or more of the following: lymph node metastasis, airway dissemination, vascular invasion, bronchial wall infiltration, and breach of the pleural elastic layer.

[0011] Furthermore: the clinical phenotypic data includes BMI and smoking history; the serum biochemical data includes at least one of apolipoprotein A1, apolipoprotein B, triglycerides, high-density lipoprotein, low-density lipoprotein, and total cholesterol; the driver gene mutation status data includes at least one of the mutation statuses of EGFR, ALK, ROS1, KRAS, BRAF, RET, and C-MET genes; and the serum metabolomics data are obtained by high-throughput mass spectrometry detection of peripheral blood.

[0012] Further: The serum metabolomics data include the detection values ​​of the following metabolites: 17-Hydroxypregnenolone sulfate, 1-Methylnicotinamide, 2-Ketobutyric acid, 3-Chlorotyrosine, 3-Hydroxybutyrylcarnitine, N6,N6,N6-Trimethyllysine, Oxindole, Pregnenolone sulfate, Propionylcarnitine, Trimethylamine N-oxide, Serotonin, Tryptophan, Xanthine, Decadienoylcarnitine (C10:2), Dihydrothymine, Docosahexaenoic acid (C22:6n-3), Tiglylcarnitine (C5:1), Lactic acid, and Hypoxanthine.

[0013] Further: The non-zero coefficient variables obtained by screening using the LASSO logistic regression algorithm in step S4 include: smoking history, EGFR gene mutation status, ALK gene mutation status, 17-Hydroxypregnenolone sulfate, Docosahexaenoic acid (C22:6n-3), Hypoxanthine, Lactic acid, N6,N6,N6-Trimethyllysine, Pregnenolone sulfate, and Tiglylcarnitine (C5:1).

[0014] Furthermore: the prediction function of the prediction model is:

[0015] Where p is the predicted probability that the tumor is invasive. The intercept is... to The regression coefficients are for smoking history, EGFR gene mutation status, ALK gene mutation status, 17-Hydroxypregnenolone sulfate, Docosahexaenoicacid (C22:6n-3), Hypoxanthine, Lactic acid, N6,N6,N6-Trimethyllysine, Pregnenolone sulfate, and Tiglylcarnitine (C5:1), respectively.

[0016] Further: In step S2: Missing values ​​for numerical variables are filled using the mean or median of the training set, while missing values ​​for categorical variables are filled using the mode. Smoking history, driver gene mutation status, and tumor invasiveness labels are encoded using one-hot encoding, where 0 indicates the absence of the feature and 1 indicates the presence of the feature. Outliers were identified and normalized using box plots and Z-score methods. For continuous variables, the Z-score standardization method is used. The standardization formula is:

[0017] in Let j be the original value of the j-th variable in the i-th sample. and denoted as the mean and standard deviation of the j-th variable in the training set, respectively.

[0018] Furthermore, the method for determining the optimal regularization parameter λ in step S4 includes: Perform multiple independent 10-fold cross-validation operations on the training set, recording the results that minimize the cross-validation error each time. value; Collected from multiple cross-validations The distribution of values ​​is analyzed, and the median is selected as the optimal regularization parameter λ for the final model.

[0019] Furthermore, step S4 also includes variable selection stability verification: the training set is resampled multiple times with replacement using Bootstrap, and the optimal regularization parameter λ is used to fit the LASSO logistic regression model each time. The frequency of each variable being selected is recorded, and variables with a selected frequency higher than a preset threshold are identified as stable predictors.

[0020] Further: Step S5 includes: Plot the receiver operating characteristic (ROC) curve and calculate the area under the curve (AUC). The optimal probability classification threshold is determined by maximizing the Youden exponent. Based on the optimal probability classification threshold, the predicted probability is converted into a binary classification result, a confusion matrix is ​​constructed, and the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value are calculated. The training set is resampled multiple times using Bootstrap. Each time, a model is built and predictions are made on the test set to obtain the median and 95% confidence interval of each performance metric.

[0021] Furthermore, step S5 is followed by a risk stratification step S6: based on the predicted probability output by the prediction model, the tumor invasiveness risk of lung adenocarcinoma patients is divided into three levels: low risk, medium risk, and high risk.

[0022] The present invention also provides the application of the prediction model constructed according to the above construction method in the preparation of a lung adenocarcinoma tumor invasiveness assessment system.

[0023] The present invention also provides a lung adenocarcinoma tumor invasiveness prediction system, comprising: The data acquisition module is used to collect multimodal data corresponding to peripheral blood samples from patients diagnosed with lung adenocarcinoma. The multimodal data includes clinical phenotype data, serum biochemical data, driver gene mutation status data, and serum metabolomics data. The data preprocessing module is used to perform missing value imputation, categorical variable encoding, outlier handling, and standardization on the multimodal data. The prediction module is used to input the preprocessed data into a pre-trained LASSO logistic regression prediction model and output the predicted probability of tumor invasiveness; wherein the prediction model is constructed using the above-described construction method. The result output module is used to output the prediction result of tumor invasiveness based on the predicted probability.

[0024] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described construction method.

[0025] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described construction method.

[0026] Compared with the prior art, the present invention has the following advantages: I. This invention systematically integrates clinical phenotypic data, serum biochemical data, driver gene mutation status data, and serum metabolomics data to provide a three-dimensional characterization of tumor invasiveness from three different biological levels: phenotype, gene, and metabolism. This overcomes the problems of insufficient information and limited predictive efficacy caused by relying on single-dimensional data in existing technologies. Compared to existing solutions based solely on gene expression levels or metabolic fingerprints, the multimodal fusion architecture of this invention can more comprehensively capture biological information related to tumor invasiveness, providing richer feature inputs for the predictive model.

[0027] II. This invention employs the LASSO logistic regression algorithm, which simultaneously achieves automatic variable selection and model parameter estimation when dealing with high-dimensional multimodal data. Through L1 regularization, the coefficients of unimportant features are compressed to zero, and 10 stable predictors are selected from dozens of candidate variables, including smoking history, EGFR and ALK gene mutation status, Hypoxanthine, 17-Hydroxypregnenolonesulfate, Lactic acid, and Tiglylcarnitine. The selected metabolites are associated with tumor biological pathways such as purine metabolism, steroid metabolism, glycolysis, and fatty acid oxidation, exhibiting clear biological interpretability and avoiding the credibility issues faced by "black box" models in clinical translation.

[0028] Third, all input data in this invention can be obtained by collecting a small amount of peripheral blood, without relying on histopathological specimens obtained through surgical resection or puncture biopsy, nor on imaging examinations such as CT scans, thus achieving a truly minimally invasive detection pathway. This feature makes this invention particularly suitable for patients who cannot tolerate invasive procedures, patients with multiple pulmonary nodules requiring separate evaluation, and patients who need dynamic monitoring during treatment, significantly expanding the clinical applicability of tumor invasiveness assessment. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the method for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data according to the present invention. Figure 2 This is a structural block diagram of the lung adenocarcinoma tumor invasiveness prediction system of the present invention; Figure 3 This is a path diagram showing the variation of the LASSO regression coefficient with the regularization parameter λ in an embodiment of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] This invention provides a method for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data. The core idea is to collect peripheral blood samples from patients diagnosed with lung adenocarcinoma to obtain multimodal data covering four dimensions: clinical phenotype, serum biochemistry, driver gene mutation status, and serum metabolomics. After preprocessing and standardization, the LASSO logistic regression algorithm is used to simultaneously complete variable selection and model construction, ultimately obtaining a model capable of predicting lung adenocarcinoma tumor invasiveness. All input data required by this method can be obtained through a small amount of peripheral blood, making it a minimally invasive detection approach that does not rely on surgical or biopsy-obtained histopathological specimens.

[0032] like Figure 1 As shown, the steps of the method of the present invention will be described in detail below.

[0033] S1: Data Acquisition: Peripheral blood samples were collected from patients diagnosed with lung adenocarcinoma to obtain multimodal data, including clinical phenotypic data, serum biochemical data, driver gene mutation status data, and serum metabolomics data. Peripheral blood samples were collected from patients clinically and pathologically diagnosed with lung adenocarcinoma to obtain multimodal clinical data related to these patients. The multimodal data includes four categories: The first category is clinical phenotype data, including basic information such as the patient's BMI and smoking history; the second category is serum biochemical data, including lipid metabolism-related indicators such as apolipoprotein A1, apolipoprotein B, triglycerides, high-density lipoprotein, low-density lipoprotein, and total cholesterol; the third category is driver gene mutation status data, including the mutation status of common driver genes in lung adenocarcinoma such as EGFR, ALK, ROS1, KRAS, BRAF, RET, and C-MET, recorded in a binary classification format, i.e., with or without mutation; the fourth category is serum metabolomics data, obtained through high-throughput mass spectrometry detection of peripheral blood, specifically including the detection values ​​of the following 18 metabolites: 17-Hydroxypregnenolonesulfate, 1-Methylnicotinamide, 2-Ketobutyric acid... The metabolites listed include 3-Chlorotyrosine, 3-Hydroxybutyrylcarnitine, N6,N6,N6-Trimethyllysine, Oxindole, Pregnenolonesulfate, Propionylcarnitine, Trimethylamine N-oxide, Serotonin, Tryptophan, Xanthine, Decadienoylcarnitine (C10:2), Dihydrothymine, Docosahexaenoic acid (C22:6n-3), Tiglylcarnitine (C5:1), Lactic acid, and Hypoxanthine. These metabolites cover multiple metabolic pathways closely related to tumor biology, including purine metabolism, steroid metabolism, tryptophan metabolism, fatty acid oxidation, and glycolysis, providing rich information for assessing tumor invasiveness at the metabolic level.

[0034] Simultaneously, the tumor invasiveness of each patient needs to be labeled as a dependent variable for model training. Tumor invasiveness is determined comprehensively based on the patient's pathological invasiveness characteristics, including lymph node metastasis, airway dissemination, vascular invasion, bronchial wall infiltration, and breach of the pleural elastic layer. When a patient exhibits one or more of these pathological invasiveness characteristics, they are marked as invasive and coded as 1; when a patient does not exhibit any of these pathological invasiveness characteristics, they are marked as non-invasive and coded as 0.

[0035] S2: Data Preprocessing: The multimodal data undergoes missing value imputation, categorical variable encoding, outlier handling, and standardization to obtain a standardized dataset. Systematic preprocessing is required for the collected multimodal data to ensure data quality and consistency with model input. Specifically, for missing value imputation, numerical variables such as BMI, serum biochemical indicators, and metabolite detection values ​​are imputed using the mean or median of the corresponding variables in the training set; categorical variables such as smoking history and gene mutation status are imputed using the mode. It should be noted that the statistics for missing value imputation are calculated based on the training set to avoid information leakage from the test set. For categorical variable encoding, smoking history, mutation status of each driver gene, and tumor invasiveness labels are encoded using one-hot encoding, where 0 represents no smoking history, no gene mutation, or no invasiveness, and 1 represents a smoking history, gene mutation, or invasiveness, thus ensuring the consistency and comparability of model input. For outlier handling, a combination of box plot and Z-score methods was used to identify extreme values. The box plot method used a threshold of 1.5 times the interquartile range, while the Z-score method used an absolute value greater than 3 as the criterion. Identified extreme values ​​were then normalized according to their specific circumstances. For data standardization, continuous variables such as BMI, serum biochemical indicators, and metabolite test values ​​were standardized using the Z-score method. The standardization formula is as follows:

[0036] in For the first i The first sample j The original values ​​of each variable, and The training set j The mean and standard deviation of each variable. For variables with a standard deviation of zero, set... To maintain numerical stability. The purpose of standardization is to eliminate dimensional differences between different variables, make the variables comparable, and improve the convergence speed and interpretability of the coefficients in the LASSO model.

[0037] S3: Data Partitioning: The standardized dataset is randomly divided into training and testing sets according to a preset ratio. In this embodiment, a 7:3 partitioning ratio is used, meaning 70% of the data is used for model training and 30% for independent testing. During partitioning, it is necessary to ensure that the distribution of the dependent variable, i.e., tumor invasiveness, is basically consistent between the training and testing sets to avoid model bias caused by uneven data partitioning.

[0038] S4: Model Construction and Variable Selection: Using tumor invasiveness as the dependent variable, the LASSO logistic regression algorithm with an L1 regularization penalty term is used to fit the training set. The optimal regularization parameter λ is determined through cross-validation, simultaneously achieving automatic variable selection and model parameter estimation, resulting in a lung adenocarcinoma tumor invasiveness prediction model containing variables with non-zero coefficients. This invention uses the LASSO logistic regression algorithm for model construction, which can simultaneously achieve automatic variable selection and model parameter estimation in high-dimensional data environments. LASSO is an abbreviation for Least Absolute Shrinkage and Selection Operator. Its core idea is to add an L1 regularization penalty term to the loss function of ordinary logistic regression. By applying L1 norm constraints to the regression coefficients, the coefficients of unimportant features are compressed to exactly zero, thereby achieving automatic feature selection and obtaining a sparse model. For binary classification problems, LASSO logistic regression estimates the coefficients by minimizing the penalized negative log-likelihood function. Its objective function is:

[0039] in For the first i The values ​​of the dependent variable for each sample. For the first i The vector of independent variables for each sample. For the intercept term, Let N be the regression coefficient vector, and N be the number of training samples. λ For regularization parameters, for The L1 norm. Regularization parameter. λ Controlling the severity of punishment, λ The larger the coefficient, the more coefficients are compressed to zero, and the sparser the model becomes; λ The smaller the value, the more variables are retained, and the more complex the model becomes. Therefore, λ The choice of [aspect name] is crucial to model performance.

[0040] To determine the optimal regularization parameter λ This invention employs a robust selection method combining 10-fold cross-validation and the Bootstrap approach. The specific steps are as follows: Perform 50 independent 10-fold cross-validations on the training set. Each time, the training set is randomly divided into 10 parts, and 9 parts are used alternately as the training subset and 1 part as the validation subset. Calculate the average cross-validation error for different λ values ​​and record the result that minimizes the cross-validation error. λ Value, denoted as After 50 independent 10-fold cross-validation trials, 50 results were collected. Calculate the distribution characteristics of the values, and select these 50. The median of the values ​​is used as the optimal regularization parameter λ for the final model. For example... Figure 3 As shown, Figure 3 The graph shows the path of the LASSO regression coefficients as a function of the regularization parameter λ. The red dashed line marks the final selected value of λ, at which point the model achieves the best balance between sparsity and predictive power. The median was chosen instead of the mean because the median is more robust to extreme values ​​and can reduce the interference of individual outlier cross-validation results on parameter selection, thereby improving the robustness of parameter selection.

[0041] After determining the optimal λ, a LASSO logistic regression model is fitted on the complete training set using this parameter to obtain the regression coefficients for each variable. Variables with zero coefficients are automatically excluded, while variables with non-zero coefficients are the predictors selected by the model, forming the final lung adenocarcinoma tumor invasiveness prediction model.

[0042] To further evaluate the stability and reliability of the selected variables, this invention also performs variable selection stability verification. The training set is subjected to 200 bootstrap resampling cycles with replacement. Each time, an equal number of samples are drawn from the training set with replacement. Then, the optimal λ value determined above is used to fit a LASSO logistic regression model, and the selected variables (those with non-zero coefficients) are recorded each time. The frequency of each variable being selected in the 200 bootstrap cycles is counted, and variables with a selection frequency higher than 0.5 are considered stable predictors. This method can quantify the uncertainty of variable selection, identify metabolites and clinical indicators that have a stable contribution to invasiveness prediction, and enhance the reliability of the model.

[0043] S5: Model Evaluation. A comprehensive performance evaluation of the constructed predictive model is conducted using an independent test set. The evaluation includes the following aspects: plotting the Receiver Operating Characteristic (ROC) curve and calculating the Area Under the Curve (AUC). An AUC value closer to 1 indicates a stronger discriminative ability. The optimal probability classification threshold is determined using the Youden exponent maximization principle. The Youden exponent is defined as sensitivity plus specificity minus 1; the probability value maximizing the Youden exponent is selected as the optimal classification threshold. Based on this optimal threshold, the predicted probabilities output by the model are converted into binary classification results, a confusion matrix is ​​constructed, and classification performance indicators such as accuracy, sensitivity, specificity, positive predictive value, and negative predictive value are calculated. Furthermore, to objectively evaluate the robustness of the model performance, the training set is resampled 500 times using Bootstrap. Each time, a LASSO logistic regression model is constructed and predictions are performed on the test set. Performance indicators are calculated, and finally, the median and 95% confidence interval of each indicator are extracted as robust estimates of the model performance.

[0044] S6: The lung adenocarcinoma tumor invasiveness prediction model is further stratified by risk. Based on the predicted probabilities output by the model, the tumor invasiveness risk of lung adenocarcinoma patients is divided into three levels: low risk, intermediate risk, and high risk, for clinical application. Risk level classification can be based on the quantile method of the predicted probability distribution in the training set, or on a high / low threshold method derived from the Youden threshold. For example, a predicted probability below 0.3 is considered low risk, above 0.7 is high risk, and between the two is intermediate risk. The distribution of actual invasive positive and negative patients in each risk level is statistically analyzed, the positive rate of each level is calculated, and a risk stratification table is formed, providing an intuitive reference for clinicians' diagnostic and treatment decisions.

[0045] The effects of the method of the present invention will be explained below with reference to specific implementation examples.

[0046] In this embodiment, a cohort of patients clinically and pathologically diagnosed with lung adenocarcinoma was collected. Peripheral blood samples were collected from each patient, and complete multimodal data, including BMI, smoking history, apolipoprotein A1, apolipoprotein B, triglycerides, high-density lipoprotein, low-density lipoprotein, total cholesterol, EGFR gene detection, ALK gene detection, ROS1 gene detection, KRAS gene detection, BRAF gene detection, RET gene detection, C-MET gene detection, and 18 serum metabolite values, were obtained according to the above-mentioned method. Simultaneously, the invasive characteristics of each patient were recorded based on the pathological results. The invasiveness of the tumor was determined by comprehensively considering pathological manifestations such as lymph node metastasis, airway dissemination, vascular invasion, bronchial wall infiltration, and breach of the pleural elastic layer.

[0047] After preprocessing and standardizing the data as described above, the training and test sets were randomly divided in a 7:3 ratio. The LASSO logistic regression algorithm was used to build the model, and the optimal regularization parameter λ was determined through 50 independent 10-fold cross-validations. The final model was then fitted on the complete training set. After LASSO variable selection, the model retained 10 non-zero coefficient variables, and the variables and their corresponding regression coefficients are shown in Table 1.

[0048] Table 1. Non-zero coefficients and regression coefficients of the LASSO logistic regression model.

[0049] The final prediction function of the model is expressed as:

[0050] in p To predict the probability of a tumor being invasive, each β The values ​​are the regression coefficients corresponding to each variable in Table 1.

[0051] The screening results show that the selected variables cover three dimensions: clinical phenotype, driver gene mutation status, and serum metabolites, validating the necessity of multimodal data fusion. Among them, smoking history had the largest positive coefficient, indicating that smoking history is an important risk factor for tumor invasiveness. Hypoxanthine and Lactic acid also had large positive coefficients; the former is associated with abnormal purine metabolism, while the latter reflects the Warburg effect of tumor cells, i.e., the metabolic characteristics of aerobic glycolysis, both closely related to the malignant biological behavior of tumors. 17-Hydroxypregnenolone sulfate showed a negative coefficient, suggesting that elevated levels of this steroid metabolite may be associated with lower invasiveness. Tiglylcarnitine (C5:1) had a positive coefficient associated with the fatty acid oxidation metabolic pathway. These results demonstrate that the combination of variables selected for the model has clear biological interpretability.

[0052] The stability of variable selection was verified by 200 Bootstrap resampling operations. The selection frequency of each selected variable was higher than 0.5, and the results are shown in Table 2.

[0053] Table 2. Bootstrap selection frequency for each selected variable

[0054] As shown in Table 2, the selection frequency of all selected variables exceeded the stability threshold of 0.5. Among them, the selection frequencies of Hypoxanthine, smoking history, and 17-Hydroxypregnenolone sulfate all exceeded 0.9, demonstrating extremely high selection stability.

[0055] The model was independently validated on a test set containing 46 samples, including 23 invasive positives and 23 negatives. The confusion matrix of the model is shown in Table 3.

[0056] Table 3. Confusion matrix of the prediction model on the test set

[0057] Note: TN indicates a true negative, FN indicates a false negative, FP indicates a false positive, and TP indicates a true positive.

[0058] The performance indicators calculated based on the confusion matrix are shown in Table 4.

[0059] Table 4 Summary of performance metrics of the prediction model on the test set

[0060] Table 4 shows that the overall prediction accuracy of the model is 82.61%, and the lower limit of the 95% confidence interval is higher than 0.5, indicating that the model has significant predictive ability. The no-information rate is 0.5, and the p-value for comparing the model accuracy with the no-information rate is 0.0002678, indicating that the model accuracy is significantly better than random guessing. The Kappa coefficient is 0.6522, indicating a strong consistency between the model prediction and the true class. The McNemar test p-value is 0.0772, suggesting that the model's ability to identify the two classes of samples is basically symmetrical, and there is no statistically significant difference.

[0061] The sensitivity was 69.57%, indicating that the model can identify most patients with invasive pulmonary nodules. The specificity was as high as 95.65%, indicating that the model performed excellently in identifying negative samples, effectively avoiding unnecessary invasive examinations and overtreatment for patients with non-invasive nodules. The positive predictive value was 94.12%, meaning that when the model predicts a positive result, the probability that the patient actually has invasive nodules is very high, helping clinicians make reliable intervention decisions based on positive results. The negative predictive value was 75.86%, suggesting that negative results still need to be combined with other clinical information for comprehensive judgment.

[0062] The above performance metrics were robustly verified through 500 Bootstrap resampling tests. The median of each metric was in good agreement with the results of a single test, and the 95% confidence interval was narrow, indicating that the model has good generalization ability and stability.

[0063] In summary, the predictive model constructed in this embodiment demonstrates significant advantages in terms of high specificity and high positive predictive value, effectively reducing the misclassification rate of non-invasive lung adenocarcinoma nodules and providing a reliable auxiliary tool for preoperative clinical assessment of tumor invasiveness. Although there is still room for improvement in sensitivity, the overall performance of the model has met the preliminary requirements for clinical auxiliary diagnosis. Further optimization of the model performance can be achieved by expanding the sample size, introducing more feature variables, or employing ensemble learning methods.

[0064] like Figure 2As shown, this invention also provides a lung adenocarcinoma tumor invasiveness prediction system, which includes a data acquisition module, a data preprocessing module, a prediction module, and a result output module. The data acquisition module collects multimodal data from peripheral blood samples of patients diagnosed with lung adenocarcinoma, including clinical phenotype data, serum biochemical data, driver gene mutation status data, and serum metabolomics data. The data preprocessing module performs missing value imputation, categorical variable encoding, outlier handling, and standardization on the collected multimodal data to ensure the data meets the format and quality requirements of the model input. The prediction module inputs the preprocessed data into a pre-trained LASSO logistic regression prediction model, which is trained using the above-described construction method, and outputs the predicted probability of tumor invasiveness. The result output module outputs the predicted tumor invasiveness result based on the predicted probability. This result can be a binary classification result based on the optimal classification threshold, or a risk stratification result mapping the predicted probability to three levels: low risk, medium risk, and high risk.

[0065] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described construction method. The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described construction method. The computer-readable storage medium includes, but is not limited to, read-only memory, random access memory, flash memory, magnetic disk, or optical disk.

[0066] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a lung adenocarcinoma tumor invasiveness prediction model based on multimodal data, characterized in that, Includes the following steps: S1: Data Acquisition: Collect peripheral blood samples from patients diagnosed with lung adenocarcinoma to obtain multimodal data, including clinical phenotype data, serum biochemical data, driver gene mutation status data, and serum metabolomics data; S2: Data preprocessing: The multimodal data is imputed for missing values, categorical variable encoding is performed, outlier handling is performed, and standardization is performed to obtain a standardized dataset; S3: Data partitioning: The standardized dataset is randomly divided into a training set and a test set according to a preset ratio; S4: Model building and variable selection: Using tumor invasiveness as the dependent variable, the LASSO logistic regression algorithm with L1 regularization penalty term is used to fit the training set. The optimal regularization parameter λ is determined through cross-validation. At the same time, automatic variable selection and model parameter estimation are realized to obtain a lung adenocarcinoma tumor invasiveness prediction model containing variables with non-zero coefficients. S5: Model Evaluation: Evaluate the performance of the prediction model using the test set; The tumor invasiveness is determined comprehensively based on the patient's pathological invasiveness characteristics, which include one or more of the following: lymph node metastasis, airway dissemination, vascular invasion, bronchial wall infiltration, and breach of the pleural elastic layer.

2. The construction method according to claim 1, characterized in that, The clinical phenotypic data include BMI and smoking history; the serum biochemical data include at least one of apolipoprotein A1, apolipoprotein B, triglycerides, high-density lipoprotein, low-density lipoprotein, and total cholesterol; the driver gene mutation status data include at least one of the mutation statuses of EGFR, ALK, ROS1, KRAS, BRAF, RET, and C-MET genes; and the serum metabolomics data are obtained by high-throughput mass spectrometry detection of peripheral blood.

3. The construction method according to claim 2, characterized in that, The serum metabolomics data included the detection values ​​of the following metabolites: 17-Hydroxypregnenolone sulfate, 1-Methylnicotinamide, 2-Ketobutyric acid, 3-Chlorotyrosine, 3-Hydroxybutyrylcarnitine, N6,N6,N6-Trimethyllysine, Oxindole, Pregnenolone sulfate, Propionylcarnitine, Trimethylamine N-oxide, Serotonin, Tryptophan, Xanthine, Decadienoylcarnitine (C10:2), Dihydrothymine, Docosahexaenoic acid (C22:6n-3), Tiglylcarnitine (C5:1), Lactic acid, and Hypoxanthine.

4. The construction method according to claim 1, characterized in that, The non-zero coefficient variables obtained by screening using the LASSO logistic regression algorithm in step S4 include: smoking history, EGFR gene mutation status, ALK gene mutation status, 17-Hydroxypregnenolone sulfate, Docosahexaenoic acid (C22:6n-3), Hypoxanthine, Lactic acid, N6,N6,N6-Trimethyllysine, Pregnenolone sulfate, and Tiglylcarnitine (C5:1).

5. The construction method according to claim 4, characterized in that, The prediction function of the prediction model is: ； Where p is the predicted probability that the tumor is invasive. The intercept is... to The regression coefficients are for smoking history, EGFR gene mutation status, ALK gene mutation status, 17-Hydroxypregnenolone sulfate, Docosahexaenoic acid (C22:6n-3), Hypoxanthine, Lactic acid, N6,N6,N6-Trimethyllysine, Pregnenolonesulfate, and Tiglylcarnitine (C5:1), respectively.

6. The construction method according to claim 1, characterized in that, In step S2: Missing values ​​for numerical variables are filled using the mean or median of the training set, while missing values ​​for categorical variables are filled using the mode. Smoking history, driver gene mutation status, and tumor invasiveness labels are encoded using one-hot encoding, where 0 indicates the absence of the feature and 1 indicates the presence of the feature. Outliers were identified and normalized using box plots and Z-score methods. For continuous variables, the Z-score standardization method is used. The standardization formula is: ； in For the first i The first sample j The original values ​​of each variable, and The training set j The mean and standard deviation of each variable.

7. The construction method according to claim 1, characterized in that, In step S4, the optimal regularization parameter is determined. λ The methods include: Perform multiple independent 10-fold cross-validation operations on the training set, recording the results that minimize the cross-validation error each time. value; Collected from multiple cross-validations The distribution of values ​​is analyzed, and the median is selected as the optimal regularization parameter for the final model. λ .

8. The construction method according to claim 7, characterized in that, The S4 step also includes variable selection stability verification: the training set is resampled multiple times with replacement using Bootstrap, and the optimal regularization parameter λ is used to fit the LASSO logistic regression model each time. The frequency of each variable being selected is recorded, and variables with a selected frequency higher than a preset threshold are identified as stable predictors.

9. The construction method according to claim 1, characterized in that, Step S5 includes: Plot the receiver operating characteristic (ROC) curve and calculate the area under the curve (AUC). The optimal probability classification threshold is determined by maximizing the Youden exponent. Based on the optimal probability classification threshold, the predicted probability is converted into a binary classification result, a confusion matrix is ​​constructed, and the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value are calculated. The training set is resampled multiple times using Bootstrap. Each time, a model is built and predictions are made on the test set to obtain the median and 95% confidence interval of each performance metric.

10. The construction method according to claim 1, characterized in that, The step S5 is followed by a risk stratification step S6: based on the predicted probability output by the prediction model, the tumor invasiveness risk of lung adenocarcinoma patients is divided into three levels: low risk, medium risk, and high risk.

11. The application of the prediction model constructed by the construction method according to any one of claims 1 to 10 in the preparation of a lung adenocarcinoma tumor invasiveness assessment system.

12. A lung adenocarcinoma tumor invasiveness prediction system, characterized in that, include: The data acquisition module is used to collect multimodal data corresponding to peripheral blood samples from patients diagnosed with lung adenocarcinoma. The multimodal data includes clinical phenotype data, serum biochemical data, driver gene mutation status data, and serum metabolomics data. The data preprocessing module is used to perform missing value imputation, categorical variable encoding, outlier handling, and standardization on the multimodal data. The prediction module is used to input preprocessed data into a pre-trained LASSO logistic regression prediction model and output the predicted probability of tumor invasiveness; wherein the prediction model is constructed using the construction method of any one of claims 1 to 10. The result output module is used to output the prediction result of tumor invasiveness based on the predicted probability.

13. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the construction method according to any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the construction method according to any one of claims 1 to 10.

Images