Breast cancer axillary lymph node state prediction model and construction method

Abstract

The invention relates to the technical field of medical systems, in particular to a breast cancer axillary lymph node state prediction model and a construction method. The breast cancer axillary lymph node state prediction model is formed by integrating various hyper-parameter optimization algorithms and dynamic combination optimization algorithms, and adaptive decision making is performed on the decision threshold of the breast cancer axillary lymph node state prediction model through the genetic algorithm and based on the maximized specific evaluation index, so that the breast cancer axillary lymph node state prediction model is obtained. The breast cancer axillary lymph node state prediction method comprises the steps that a breast cancer axillary lymph node state prediction model is established, repeated sampling and statistical analysis are conducted through a self-service method, the confidence interval of the breast cancer axillary lymph node state prediction model is calculated, and the statistical uncertainty range of model performance indexes in features is determined based on the calculated confidence interval. And a foundation is laid for clinical auxiliary decision making and prognosis improvement of breast cancer axillary lymph nodes.

Description

Technical Field

[0001] This invention relates to the field of medical system technology, specifically to a predictive model and construction method for the status of axillary lymph nodes in breast cancer. Background Technology

[0002] Breast cancer is one of the most common malignant tumors, seriously threatening women's physical and mental health. According to statistics from the American Cancer Database, the 5-year relative survival rate for breast cancer is 90.8%. Among them, the 5-year relative survival rate for breast cancer that occurs only in the local breast and has not metastasized is 99.3%, while the 5-year relative survival rate for breast cancer that has metastasized to regional lymph nodes drops to 86.3%.

[0003] Therefore, accurate assessment of the axillary lymph node status in breast cancer patients is crucial in clinical practice. There are various methods for assessing the axillary lymph node status in breast cancer. Among them, ultrasound examination is non-invasive and cost-effective, making it the preferred method for assessing axillary lymph node involvement in breast cancer patients. Morphological criteria for axillary lymph nodes are more important than size criteria and are more indicative of axillary lymph node metastasis. Literature reports that using ultrasound morphological features to assess axillary lymph node metastasis has high accuracy, with a sensitivity range of 26%-76% and a specificity range of 88%-98%. However, there are no guidelines, consensus statements, or imaging report scoring systems for ultrasound assessment of lymph nodes, necessitating further improvement in assessment capabilities to better guide clinical practice.

[0004] For breast cancer patients with negative axillary lymph nodes on clinical evaluation, sentinel lymph node biopsy (SLNB) is the preferred surgical procedure for axillary surgery; however, SLNB results are negative in over 70% of patients. Approximately half of SLNB-positive cases do not develop further metastatic lymph nodes after routine axillary lymph node dissection (ALND). Studies have shown that early-stage breast cancer patients with 1-2 positive sentinel lymph nodes can be conditionally exempted from ALND, with no significant difference in prognosis. Therefore, accurate prediction of preoperative axillary lymph node status helps avoid ALNB and may even replace SLNB in ​​the future, reducing the risk of postoperative complications such as upper limb lymphedema, numbness, pain, and functional impairment, thus improving patients' quality of life. For patients expected to have a high lymph node burden (≥3 axillary lymph node metastases), further intensive treatment can provide a survival benefit.

[0005] Axillary lymph node burden in breast cancer is correlated with tumor size, histological grade, and clinicopathological factors such as Ki-67 expression. Larger tumors, higher histological grades, and higher Ki-67 expression levels are associated with a greater risk of metastasis. Vascular invasion is also considered a significant factor in breast cancer lymph node metastasis, potentially increasing the risk by promoting angiogenesis and metastatic colonization. Furthermore, several genes, such as BRCA1, nm23, p53, and the bcl-2 gene family, are also closely related to breast cancer lymph node metastasis.

[0006] In recent years, preoperative serological markers have been widely studied due to their low cost and ease of access. Increasing evidence suggests that inflammation is closely related to the occurrence, development, and metastasis of malignant tumors. Systemic inflammation promotes tumor progression and invasion by upregulating cytokines, inhibiting apoptosis, promoting angiogenesis, and initiating DNA degradation. Indicators reflecting systemic inflammation include the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (MLR), systemic immune-inflammation index (SII), and pan-immune-inflammation value (PIV). These indicators have been shown to be associated with lymph node metastasis in solid tumors such as gastric cancer, lung cancer, and liver cancer, but studies on their correlation with axillary lymph node burden in breast cancer are limited.

[0007] Given the unique biological behaviors and clinical characteristics of different molecular subtypes of breast cancer, in-depth analysis of the relevant factors of axillary lymph node metastasis in each subtype is crucial. The molecular classification criteria for breast cancer adopted at the 2013 St. Gallen International Breast Cancer Conference are currently the most widely used. Based on the expression of estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor-2 (HER-2), and Ki-67, it classifies breast cancer into Luminal A, Luminal B (including HER-2 negative and HER-2 positive), triple-negative, and other special types of breast cancer. However, different institutions have not yet reached a consensus on the critical threshold for Ki-67. The latest NCCN guidelines classify breast cancer into four categories based solely on hormone receptor and HER-2 receptor status: HR positive / HER2 negative, HR positive HER-2 positive, HR negative HER-2 positive, and HR negative HER-2 negative.

[0008] Currently, it is known that various factors influencing axillary lymph node burden in breast cancer may interact rather than act independently. Constructing a predictive model that integrates multiple indicators can help accurately predict axillary lymph node status and guide clinical practice. Currently, the Memorial Sloan Kettering Cancer Center (MSKCC) model is the most widely used. When predicting sentinel lymph node metastasis, the MSKCC model incorporates nine indicators: age, multifocality of the primary tumor, tumor quadrant, tumor size, ER status, PR status, pathological type, vascular invasion, and histological grade, with an area under the curve (AUC) of 0.754. When predicting non-sentinel lymph node metastasis, the MSKCC model incorporates predictive factors including: sentinel lymph node detection method, multifocality of the primary tumor, tumor size, histological type and grade, number of positive sentinel lymph nodes, number of negative sentinel lymph nodes, vascular invasion, and ER status. The AUC values ​​for this model in the modeling group and validation group were 0.76 and 0.77, respectively. However, both of MSKCC's predictive models are based on case data from Western breast cancer populations, and the validation results in the Chinese population are less than satisfactory.

[0009] Currently, most predictive models for axillary lymph node status are based on logistic regression. With the development of artificial intelligence and interdisciplinary research in medicine and engineering, more machine learning models are also being used for predicting axillary lymph node status. Machine learning is an important branch of artificial intelligence, possessing stronger capabilities for processing large-scale and high-dimensional data compared to traditional statistics, as well as better predictive performance and flexibility. Madekivi et al. used XGBoost and logistic regression models to predict high lymph node burden in breast cancer. Their results showed that the AUC values ​​of the XGBoost model in the training and validation groups were 0.80 (95% CI 0.71–0.89) and 0.80 (95% CI 0.65–0.92), respectively, while the AUC values ​​of the logistic regression model were 0.85 (95% CI 0.77–0.93) and 0.75 (95% CI 0.58–0.89), respectively. Compared to the logistic regression model, the XGBoost model maintained its discriminative power better in the validation cohort. In most studies, machine learning-based predictive models outperform regression-based models, but some studies have reported the opposite results. We still need to investigate the performance of machine learning in predicting the status of axillary lymph nodes in breast cancer. Summary of the Invention

[0010] In view of the shortcomings of the prior art, the present invention aims to provide a predictive model and construction method for the axillary lymph node status of breast cancer, so as to accurately predict the axillary lymph node status of breast cancer.

[0011] To solve the above problems, the present invention adopts the following technical solution:

[0012] In a first aspect, the present invention provides a method for constructing a predictive model for the status of axillary lymph nodes in breast cancer, comprising:

[0013] By integrating multiple hyperparameter optimization algorithms and dynamically combining optimization algorithms, a predictive model for the status of axillary lymph nodes in breast cancer is formed.

[0014] Using a genetic algorithm, an adaptive decision is made on the decision threshold of the breast cancer axillary lymph node status prediction model based on maximizing a specific evaluation index.

[0015] By employing a bootstrap method for repeated sampling and statistical analysis, the confidence interval of the breast cancer axillary lymph node status prediction model was calculated. Based on the calculated confidence interval, the range of statistical uncertainty of the model performance indicators in the features was determined.

[0016] As one possible implementation method, the integrated optimization algorithm with multiple hyperparameters and the dynamic combination optimization algorithm include:

[0017] Establish an optimization algorithm library, and determine the applicable conditions for each algorithm based on the algorithms in the optimization algorithm library;

[0018] Based on the applicable conditions of each algorithm, feature extraction is performed on the hyperparameter optimization problem to generate a problem feature vector;

[0019] A decision model is used to calculate the weights of each algorithm based on the problem's feature vector;

[0020] A collaborative optimization strategy is adopted to construct a dynamic optimization algorithm combination. Based on each algorithm and its weight in the dynamic optimization algorithm combination, the optimal hyperparameter combination of each algorithm is summarized, and the hyperparameter with the best performance is selected as the hyperparameter of the dynamic optimization algorithm combination.

[0021] As one possible implementation, the adaptive decision-making based on maximizing a specific evaluation index using a genetic algorithm to optimize the decision threshold includes:

[0022] S21. Within the search range of the decision threshold, multiple initial thresholds are randomly generated to form an initial population;

[0023] S22. For each initial threshold in the initial population, calculate the fitness of each initial threshold according to the selected specific evaluation index;

[0024] S23. Based on fitness, select the initial threshold corresponding to a fitness value greater than the preset value as the parent generation;

[0025] S24. Perform single-point crossover, double-point crossover, or uniform crossover operations on the selected parent generation to generate offspring.

[0026] S25. Perform mutation operations on the generated offspring, adopt an elite strategy, retain the multiple offspring with the highest fitness after the mutation operation, and add the offspring to the population to form a new generation of population.

[0027] S26. If the maximum number of iterations or the fitness value converges, the optimization is complete. Then, select the individual with the highest fitness from the new generation of population as the optimal decision threshold. If the maximum number of iterations or the fitness value does not converge, repeat S22-S25.

[0028] As one possible implementation, the method of performing repeated sampling and statistical analysis using a bootstrap method to calculate the confidence interval of the breast cancer axillary lymph node status prediction model, and determining the statistical uncertainty range of the model performance indicators in the features based on the calculated confidence interval, includes:

[0029] Based on features as the dataset, sample data is extracted from the dataset using a bootstrap random sampling method with replacement to generate multiple bootstrap sample sets.

[0030] Based on each self-service sample set, the target performance evaluation index of the breast cancer axillary lymph node status prediction model is calculated, forming an index distribution based on the self-service sample set;

[0031] Arrange the indicators of the indicator distribution in ascending order, and determine the lower and upper limits of the confidence interval according to the required confidence level;

[0032] The calculated confidence intervals are applied to model performance evaluation, providing the statistical uncertainty range of model performance indicators and assisting in model optimization and selection.

[0033] As one possible implementation, the present invention also includes standardizing the medical diagnostic dataset, which includes negative and positive medical diagnostic data;

[0034] Construct a multilayer perceptron generator, with a 100-dimensional random noise vector as input and a synthetic sample with the same dimensions as the real sample as output.

[0035] Construct a multilayer perceptron discriminator, which takes a real sample or a synthetic sample as input and outputs a probability value representing the likelihood that a sample is a real sample;

[0036] The generator and discriminator are optimized separately using the Adam optimizer, and the cross-entropy loss function is used for training. During the training process, the parameters of the generator and discriminator are updated alternately to train the GAN model.

[0037] Based on the feedback from the training of the GAN model, if the recognition rate of the GAN model on the anomaly class does not meet the standard, more high-quality synthetic real samples will be generated.

[0038] The generated synthetic samples are scored by a discriminator, and samples with scores higher than 0.9 are selected.

[0039] As one possible implementation, the present invention also includes multiple imputation processing for missing values ​​in the medical diagnostic dataset;

[0040] Features are generated using polynomial features, logarithmic transformations, power transformations, and combined features;

[0041] Calculate the correlation coefficient between each generated feature and whether or not the disease is present, and select features with a correlation greater than 0.3;

[0042] Principal component analysis was performed on the selected features, and the top 10 features were retained.

[0043] Secondly, the present invention provides a method for constructing a breast cancer axillary lymph node status prediction model, which constructs the breast cancer axillary lymph node status prediction model.

[0044] Thirdly, the present invention provides a breast cancer axillary lymph node status prediction system, including a data acquisition module and a prediction module;

[0045] The data acquisition module is used to collect the medical diagnostic data of the test subject.

[0046] The prediction module is used to input the medical diagnostic data into the breast cancer axillary lymph node status prediction model to predict and evaluate the breast cancer axillary lymph node status of the test subject.

[0047] The beneficial effects of this invention are as follows: By integrating multiple hyperparameter optimization algorithms and dynamically combining optimization algorithms, this invention forms a predictive model for the status of axillary lymph nodes in breast cancer, laying the foundation for clinical auxiliary decision-making and prognostic improvement of axillary lymph nodes in breast cancer. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 The top 10 features retained after principal component analysis;

[0050] Figure 2 The graph shows the F1 score and ROC and AUC metrics. Detailed Implementation

[0051] The present invention will be further described in detail below with reference to specific embodiments.

[0052] It should be noted that these embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Simple improvements to the method under the premise of the present invention are all within the scope of protection claimed by the present invention.

[0053] The number of axillary lymph node metastases ≥3 was defined as the high lymph node burden group, and the number of axillary lymph node metastases <3 was defined as the low lymph node burden group. Factors related to high lymph node burden in breast cancer were analyzed. The relevant factors are shown in Table 1, and these factors were used as the basis for collecting patient medical diagnostic data.

[0054] Table 1. Value Assignment Table for Relevant Factors

[0055] variable definition Side Left=1, Right=0 age ≤40 years old = 1, >60 years old = 0 BMI ＞24=1，24=0 gender Male=1, Female=0 Marital status Unmarried = 1, Married = 0 Menstrual status Premenopausal = 1, Menopausal = 0 Family history Present = 1, Absent = 0 Pathological types Non-special type breast cancer = 1, Special type breast cancer = 0 Tumor size >2cm=1, ≤2cm=0 Vascular invasion Yes = 1, No = 0 ER Positive = 1, Negative = 0 PR Positive = 1, Negative = 0 HER-2 Positive = 1, Negative = 0 KI-67 ≥14%=1，＜14%=0 Molecular typing HR positive HER2 negative = 0 HR positive HER2 positive = 1 HR negative HER2 positive = 2 HR negative HER2 negative = 3 Histological grading G3=1, G1 / G2=0 NLR Numeric variables PLR Numeric variables MLR Numeric variables SII Numeric variables PIV Numeric variables Hb hemoglobin Numeric variables WBC (White Blood Cell) Numeric variables LYM lymphocytes Numeric variables EOS eosinophils Numeric variables HCT (Hematocrit) Numeric variables Mean Corpuscular Hemoglobin (MCH) Numeric variables RDW_CV (Red Cell Distribution Width Coefficient of Variation) Numeric variables Mean Platelet Volume (MPV) Numeric variables RBC (Red Blood Cell) Numeric variables NEUT neutrophils Numeric variables MONO monocytes Numeric variables BASO basophils Numeric variables Mean Corpuscular Volume (MCV) Numeric variables Mean Corpuscular Hemoglobin Concentration (MCHC) Numeric variables PLT (platelet) Numeric variables Alanine aminotransferase (ALT) Numeric variables AST aspartate aminotransferase Numeric variables Total Protein (TP) Numeric variables ALB albumin Numeric variables Total Bilirubin (TBIL) Numeric variables DBIL (Direct Bilirubin) Numeric variables Total Bile Acids (TBA) Numeric variables ALP (Alkaline Phosphatase) Numeric variables GGTγ-Glutamyl Transferase Numeric variables GLU (glucose) Numeric variables BUN (Blood Urea Nitrogen) Numeric variables Creatinine (Cr) Numeric variables Total cholesterol (TC) Numeric variables TG triglycerides Numeric variables LDH lactate dehydrogenase Numeric variables HDL (High-density lipoprotein) Numeric variables LDL (low-density lipoprotein) Numeric variables

[0056] Using medical diagnostic data from 6162 breast cancer patients as a sample, patients with bilateral breast cancer, those with distant metastases, those with breast carcinoma in situ, those who underwent neoadjuvant chemotherapy before surgery, those with serious underlying diseases, and those with incomplete data for important cases were excluded according to the exclusion criteria. Finally, 4530 eligible cases were selected. Among them, based on postoperative pathology reports, 3280 cases (72.41%) were in the low lymph node burden group and 1250 cases (27.59%) were in the high lymph node burden group. The clinicopathological factors associated with high lymph node burden were investigated, and a predictive model was constructed.

[0057] A method for constructing a predictive model for the status of axillary lymph nodes in breast cancer includes:

[0058] (1) Integrate multiple hyperparameter optimization algorithms and dynamically combine optimization algorithms to form a breast cancer axillary lymph node status prediction model.

[0059] Integrating multiple hyperparameter optimization algorithms, the dynamically combined optimization algorithms include:

[0060] 1) Establish an optimization algorithm library and determine the applicable conditions for each algorithm based on the algorithms in the optimization algorithm library.

[0061] The optimization algorithm library includes Optuna, Particle Swarm Optimization (PSO), and Bayesian Optimization (BO).

[0062] 2) Based on the applicable conditions of each algorithm, feature extraction is performed on the hyperparameter optimization problem to generate a problem feature vector.

[0063] Before performing hyperparameter optimization, feature analysis is performed on the specific optimization problem to extract key features such as the search space dimension, the complexity of the objective function, the existence of multiple local optima, and computational resource constraints.

[0064] 3) Use a decision model to calculate the weights of each algorithm based on the problem feature vector.

[0065] Based on the problem feature vector F, decision model rules are used to evaluate the applicability and expected performance of each optimization algorithm on the current problem.

[0066] Each optimization algorithm is assigned a weight, representing its priority or importance in the current problem. The sum of the weights is 1.

[0067] 4) Adopt a collaborative optimization strategy to construct a dynamic optimization algorithm combination. Based on each algorithm and its weight in the dynamic optimization algorithm combination, summarize the optimal hyperparameter combination of each algorithm and select the hyperparameter with the best performance as the hyperparameter of the dynamic optimization algorithm combination.

[0068] Weight allocation formula: Based on the problem feature vector F and the algorithm applicability evaluation, calculate the weights w of m optimization algorithms. i:

[0069] ;

[0070] in The weight adjustment parameters for algorithm i control its responsiveness to feature vectors.

[0071] Based on the assigned weights, a combined optimization strategy is determined. For example, a weighted rotation strategy, a parallel operation strategy, or a collaborative optimization strategy can be adopted.

[0072] Weighted rotation strategy: Different optimization algorithms are called sequentially according to the weight allocation ratio for iterative optimization.

[0073] Parallel execution strategy: Run multiple optimization algorithms simultaneously, each independently searching for the optimal solution, and finally sum up their respective optimal results.

[0074] Collaborative optimization strategy: Different optimization algorithms share information or parameters during the search process. For example, PSO references a Bayesian optimization probability model during particle position update.

[0075] Optimize process control:

[0076] Set the total number of iterations for optimization or the convergence condition for the optimization objective.

[0077] In each iteration, the weight allocation is adjusted based on the feedback from the optimization algorithm to dynamically adapt to the optimization progress.

[0078] The fitness function is used to evaluate the performance of hyperparameter combinations and determine the search direction of the optimization algorithm.

[0079] Fitness function: Defines the fitness of a combination of hyperparameters, representing the model's performance metrics (such as F1 score, ROC, AUC, etc.) under that combination.

[0080] ;

[0081] θ represents a combination of hyperparameters.

[0082] Optimization Algorithm Combination Formula: In a collaborative optimization strategy, when updating the hyperparameter combination, the search directions of different algorithms are considered:

[0083] ;

[0084] in, , where represents the weight of each optimization algorithm in the combination.

[0085] Determining the optimal hyperparameters:

[0086] After the optimization process is completed, the optimal hyperparameter combinations of each optimization algorithm are summarized, and the hyperparameter with the best performance is selected as the final result.

[0087] The final optimal hyperparameters can be determined by weighted averaging, voting mechanisms, or other integration methods.

[0088] This embodiment provides the specific implementation steps of the above process:

[0089] Initialization and algorithm library setup:

[0090] Establish an optimization algorithm library that includes Optuna, Particle Swarm Optimization (PSO), and Bayesian Optimization (BO).

[0091] Define the applicable conditions for each algorithm: Optuna: suitable for medium- to high-dimensional continuous search spaces, with an early stopping mechanism. PSO: suitable for high-dimensional non-convex search spaces, with good parallelism. Bayesian optimization: suitable for low- to medium-dimensional continuous search spaces, especially suitable for situations with limited computational resources.

[0092] Problem Feature Analysis:

[0093] For specific hyperparameter optimization problems, feature extraction is performed to generate feature vector F:

[0094]

[0095] Optimization algorithm selection and weight allocation:

[0096] Using a pre-trained decision model, the weights of each optimization algorithm are calculated based on the feature vector F.

[0097] For example, assuming m = 3, the calculation yields:

[0098]

[0099] Dynamic algorithm combinatorial optimization:

[0100] A collaborative optimization strategy is adopted to update the hyperparameter combination according to the weight allocation:

[0101]

[0102] In each iteration, the optimization algorithms are updated according to their respective search mechanisms and share some information to guide the overall search direction.

[0103] Optimize process control:

[0104] Set the maximum number of iterations to 100, or terminate the optimization when the fitness change is less than 0.001.

[0105] In each iteration, the weights of each optimization algorithm are dynamically adjusted based on the current fitness evaluation results to adapt to the optimization progress.

[0106] Determining the optimal hyperparameters:

[0107] After optimization, the optimal hyperparameter combinations of Optuna, PSO, and BO are summarized, and the one with the highest fitness is selected as the final optimal hyperparameter.

[0108] (2) Using a genetic algorithm, based on maximizing a specific evaluation index, adaptive decision-making is performed on the decision threshold of the optimization algorithm / breast cancer axillary lymph node status prediction model; specifically including:

[0109] 1) Within the search range of the decision threshold, multiple initial thresholds are randomly generated to form an initial population;

[0110] 2) For each initial threshold in the initial population, calculate the fitness of each initial threshold according to the selected specific evaluation index;

[0111] 3) Based on fitness, select the initial threshold corresponding to a fitness value greater than the preset value as the parent generation;

[0112] 4) Perform single-point crossover, double-point crossover, or uniform crossover operations on the selected parent generation to generate offspring;

[0113] 5) Perform mutation operations on the generated offspring, adopt an elitist strategy, retain the multiple offspring with the highest fitness after mutation operations, and add the offspring to the population to form a new generation of population;

[0114] 6) If the maximum number of iterations or the fitness value converges, the optimization is complete, and the individual with the highest fitness is selected from the new generation of the population as the optimal decision threshold; if the maximum number of iterations or the fitness value does not converge, repeat steps 2)-5).

[0115] This embodiment provides the specific implementation steps of the above process:

[0116] Initialize the population: Randomly generate 20 threshold individuals within the range of 0.0 to 1.0 to form the initial population.

[0117] Fitness evaluation: For each threshold individual, the threshold is used to perform classification prediction on the validation set, and the F1 score is calculated as the fitness value.

[0118] Selection operation: A roulette wheel selection method is used to select 10 parent individuals based on their fitness values. The roulette wheel selection method involves converting the fitness values ​​of individuals into selection probabilities to achieve performance-based random selection. The steps include calculating fitness, determining probabilities, constructing the roulette wheel, and randomly selecting individuals. For example:

[0119] The population has 5 individuals with fitness values ​​(based on F1 scores) of 10, 20, 30, 25, and 15; the total fitness F = 10 + 20 + 30 + 25 + 15 = 100; the selection probabilities of each individual are 0.1, 0.2, 0.3, 0.25, and 0.15, respectively. A random number r = 0.45 is generated. r = 0.45 falls between 0.3 and 0.55 (corresponding to individual 3), therefore individual 3 is selected as the parent. This process is repeated until enough parents are selected for crossover and mutation.

[0120] Crossover operation: Perform a single-point crossover on the selected parent individuals to generate 10 child individuals.

[0121] Mutation operation: Gaussian mutation is performed on 5% of the offspring individuals, with a perturbation threshold.

[0122] Population renewal: Adding offspring individuals to the population to form a new generation. For example, using an elitist strategy, retaining the two individuals with the highest fitness from the previous generation.

[0123] Iterative optimization: Repeat steps 2 to 6 until the maximum number of iterations (e.g., 100 times) is reached or the fitness value converges (e.g., the fitness change is less than 0.001 for 10 consecutive iterations).

[0124] Determine the optimal threshold: After optimization, select the individual with the highest fitness value as the optimal decision threshold.

[0125] (3) The confidence interval of the breast cancer axillary lymph node status prediction model is calculated by repeated sampling and statistical analysis using the bootstrap method, and the statistical uncertainty range of the model performance index in the features is determined based on the calculated confidence interval.

[0126] 1) Based on the features as the dataset, sample data is extracted from the dataset using a self-service random sampling method with replacement to generate multiple self-service sample sets.

[0127] 2) Based on each self-help sample set, calculate the target performance evaluation index of the breast cancer axillary lymph node status prediction model to form an index distribution based on the self-help sample set.

[0128] 3) Arrange the indicators in the indicator distribution in ascending order, and determine the lower and upper limits of the confidence interval according to the required confidence level.

[0129] 4) The calculated confidence intervals are applied to model performance evaluation, providing the statistical uncertainty range of model performance metrics to assist in model optimization and selection. Model performance metrics include accuracy, precision, recall, F1 score, AUC, etc.

[0130] This embodiment provides the specific implementation steps of the above process:

[0131] Sample data preparation: Suppose there is a test set containing 100 samples to evaluate the F1 score of the prediction model.

[0132] The bootstrap sampling process involves 1000 samplings with replacement. Each time, 100 samples are randomly selected from 100 samples (each sample can be understood as the combined indicator data of a patient) to generate 1000 bootstrap sample sets. For each bootstrap sample set, the F1 score is calculated, resulting in the distribution of the 1000 F1 scores.

[0133] Confidence interval estimation: Sort the 1000 F1 scores in ascending order. Calculate the F1 scores of the 25th (2.5%) and 975th (97.5%) scores, and use them as the lower and upper limits of the 95% confidence interval, respectively.

[0134] Application of confidence intervals: For example, the calculation results show that the 95% confidence interval is [0.80, 0.90], which means that at a 95% confidence level, the model's F1 score is between 0.80 and 0.90.

[0135] The present invention also includes:

[0136] To address class imbalance, a generative adversarial network (GAN)-based intelligent class imbalance handling method is introduced. This method leverages GAN models to generate high-quality minority class samples, achieving data augmentation and thereby improving the predictive model's recognition ability and overall performance on imbalanced datasets. For example:

[0137] Dataset Analysis and Preparation:

[0138] A medical diagnostic dataset contains two classes of samples: positive and negative. There are more than 4,530 negative samples and 1,250 positive samples.

[0139] Standardize the dataset to ensure that the feature values ​​are within the same scale range.

[0140] Generative Adversarial Network Construction:

[0141] Generator design: Construct a multilayer perceptron (MLP) generator with a 100-dimensional random noise vector as input and a synthetic sample with the same dimensions as the real sample as output.

[0142] Discriminator design: Construct a multilayer perceptron discriminator, take a real sample or a generated sample as input, and output a probability value representing the likelihood that the sample is a real sample.

[0143] GAN Training and Optimization:

[0144] The generator and discriminator were optimized separately using the Adam optimizer with a learning rate of 0.0002.

[0145] The cross-entropy loss function is used for training, and the parameters of the generator and discriminator are updated alternately during the training process.

[0146] Train for 10,000 iterations and periodically evaluate the quality of the generated samples to ensure that the discriminator and generator are in balance.

[0147] Intelligent mechanisms for handling class imbalance:

[0148] Dynamic generation strategy: Based on the feedback from model training, the quantity and quality of generated samples are dynamically adjusted. Through cross-validation and multi-index evaluation, the effectiveness of data augmentation methods in improving model performance is verified.

[0149] Quality assessment and screening: The generated synthetic samples are scored by a discriminator, and samples with scores higher than 0.9 are selected to ensure the authenticity and diversity of the generated samples.

[0150] Model training and evaluation:

[0151] Train a random forest classifier using the augmented balanced dataset.

[0152] The F1 score, ROC, and AUC of the model were calculated using 5-fold cross-validation and compared with a model without data augmentation.

[0153] The present invention also includes:

[0154] By employing custom feature engineering methods and specific data transformation and combination strategies, diverse and high-quality features are automatically generated, thereby improving the performance and generalization ability of the prediction model. For example:

[0155] Data preprocessing:

[0156] The medical diagnostic dataset was used, which contains multiple continuous and discrete features.

[0157] Standardize all continuous features to ensure that their mean is 0 and their standard deviation is 1.

[0158] Missing values ​​are imputed multiple times to ensure the integrity of the dataset.

[0159] Feature generation: Features are generated using various methods such as polynomial features, logarithmic transformation, power transformation, and combined features.

[0160] Feature selection:

[0161] Calculate the correlation coefficient between each new feature and the target variable "whether or not the disease is present", and select features with a correlation greater than 0.3.

[0162] Principal component analysis (PCA) was performed on the selected features, retaining the top 10 principal components to reduce feature dimensionality, such as... Figure 1 As shown in the figure. ULTRASOUND is ultrasound examination, HER-2: human epidermal growth factor receptor 2, ER is estrogen receptor, SIZE is tumor size, MOLECULAR TYPE is molecular subtyping, MENSTRUATION is menstrual status, PATHOLOGICAL TYPE is pathological type, KI-67 is anti-proliferating cell nuclear antigen KI67, PR is progesterone receptor, HISTORY is family history, SIDE is laminar aberration, LDL is low-density lipoprotein, MCH is erythrocyte hemoglobin content, and NLR is the neutrophil-to-lymphocyte ratio.

[0163] Feature evaluation and optimization:

[0164] A random forest classifier was trained using 10-fold cross-validation to evaluate the impact of different feature combinations on model performance.

[0165] Based on the F1 score and ROC and AUC metrics, see Figure 2 We adjusted the feature generation and selection strategies to further optimize the feature engineering process.

[0166] The present invention also includes:

[0167] By incorporating incremental learning methods, the predictive model can efficiently update and optimize its parameters when new data arrives, thereby improving its adaptability, computational efficiency, and long-term performance.

[0168] Initialize the model:

[0169] The hyperparameters of the breast cancer axillary lymph node status prediction model were obtained using a trained dynamic combinatorial optimization algorithm.

[0170] Real-time data reception and processing:

[0171] The system continuously receives new patient data samples, one batch at a time. New samples are standardized to ensure that feature values ​​are within the same scale range.

[0172] Applications of incremental learning algorithms:

[0173] The model parameters are updated using an online gradient descent algorithm. A knowledge distillation mechanism is introduced, randomly selecting 200 previously saved historical samples and using them together with the current batch of samples for model updates.

[0174] Model evaluation and adjustment:

[0175] After each batch of data is received and processed, the F1 score, ROC, and AUC of the model are evaluated using an independent validation set.

[0176] If the F1 score does not improve significantly over 10 consecutive batches (change less than 0.01), reduce the learning rate. :

[0177] .

[0178] Model saving and deployment:

[0179] After processing every 100 batches of data, save the current model parameters. Deploy the updated model for real-time classification and prediction to ensure that the model always reflects the latest data trends.

[0180] A breast cancer axillary lymph node status prediction system includes a data acquisition module and a prediction module.

[0181] The data acquisition module is used to collect medical diagnostic data from the subjects being tested.

[0182] The prediction module is used to input medical diagnostic data into the breast cancer axillary lymph node status prediction model to predict and assess the breast cancer axillary lymph node status of the test subject.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for constructing a predictive model for the status of axillary lymph nodes in breast cancer, characterized in that, include: By integrating multiple hyperparameter optimization algorithms and dynamically combining optimization algorithms, a predictive model for the status of axillary lymph nodes in breast cancer is formed. Using a genetic algorithm, an adaptive decision is made on the decision threshold of the breast cancer axillary lymph node status prediction model based on maximizing a specific evaluation index. By employing a bootstrap method for repeated sampling and statistical analysis, the confidence interval of the breast cancer axillary lymph node status prediction model was calculated. Based on the calculated confidence interval, the range of statistical uncertainty of the model performance indicators in the features was determined.

2. The method for constructing a breast cancer axillary lymph node status prediction model according to claim 1, characterized in that, The integrated optimization algorithm for multiple hyperparameters and the dynamically combined optimization algorithm include: Establish an optimization algorithm library, and determine the applicable conditions for each algorithm based on the algorithms in the optimization algorithm library; Based on the applicable conditions of each algorithm, feature extraction is performed on the hyperparameter optimization problem to generate a problem feature vector; A decision model is used to calculate the weights of each algorithm based on the problem's feature vector; A collaborative optimization strategy is adopted to construct a dynamic optimization algorithm combination. Based on each algorithm and its weight in the dynamic optimization algorithm combination, the optimal hyperparameter combination of each algorithm is summarized, and the hyperparameter with the best performance is selected as the hyperparameter of the dynamic optimization algorithm combination.

3. The method for constructing a breast cancer axillary lymph node status prediction model according to claim 2, characterized in that, The adaptive decision-making process using a genetic algorithm, based on maximizing a specific evaluation metric, to optimize the decision threshold of the algorithm includes: S21. Within the search range of the decision threshold, multiple initial thresholds are randomly generated to form an initial population; S22. For each initial threshold in the initial population, calculate the fitness of each initial threshold according to the selected specific evaluation index; S23. Based on fitness, select the initial threshold corresponding to a fitness value greater than the preset value as the parent generation; S24. Perform single-point crossover, double-point crossover, or uniform crossover operations on the selected parent generation to generate offspring. S25. Perform mutation operations on the generated offspring, adopt an elite strategy, retain the multiple offspring with the highest fitness after the mutation operation, and add the offspring to the population to form a new generation of population. S26. If the maximum number of iterations or the fitness value converges, the optimization is complete. Then, select the individual with the highest fitness from the new generation of population as the optimal decision threshold. If the maximum number of iterations or the fitness value does not converge, repeat S22-S25.

4. The method for constructing a breast cancer axillary lymph node status prediction model according to claim 3, characterized in that, The method involves repeated sampling and statistical analysis using a bootstrap approach to calculate the confidence interval of the breast cancer axillary lymph node status prediction model. Based on the calculated confidence interval, the statistical uncertainty range of the model performance indicators in the features is determined, including: Based on features as the dataset, sample data is extracted from the dataset using a bootstrap random sampling method with replacement to generate multiple bootstrap sample sets. Based on each self-service sample set, the target performance evaluation index of the breast cancer axillary lymph node status prediction model is calculated, forming an index distribution based on the self-service sample set; Arrange the indicators of the indicator distribution in ascending order, and determine the lower and upper limits of the confidence interval according to the required confidence level; The calculated confidence intervals are applied to model performance evaluation, providing the statistical uncertainty range of model performance indicators and assisting in model optimization and selection.

5. The method for constructing a breast cancer axillary lymph node status prediction model according to claim 4, characterized in that, It also includes standardization of medical diagnostic datasets, which include both negative and positive medical diagnostic data; Construct a multilayer perceptron generator, with a 100-dimensional random noise vector as input and a synthetic sample with the same dimensions as the real sample as output. Construct a multilayer perceptron discriminator, which takes a real sample or a synthetic sample as input and outputs a probability value representing the likelihood that a sample is a real sample; The generator and discriminator are optimized separately using the Adam optimizer, and the cross-entropy loss function is used for training. During the training process, the parameters of the generator and discriminator are updated alternately to train the GAN model. Based on the feedback from the training of the GAN model, if the recognition rate of the GAN model on the anomaly class does not meet the standard, more high-quality synthetic real samples will be generated. The generated synthetic samples are scored by a discriminator, and samples with scores higher than 0.9 are selected.

6. The method for constructing a breast cancer axillary lymph node status prediction model according to claim 5, characterized in that, It also includes multiple imputation processing for missing values ​​in medical diagnostic datasets; Features are generated using polynomial features, logarithmic transformations, power transformations, and combined features; Calculate the correlation coefficient between each generated feature and whether or not the disease is present, and select features with a correlation greater than 0.3; Principal component analysis was performed on the selected features, and the top 10 features were retained.

7. The breast cancer axillary lymph node status prediction model constructed by the method of constructing the breast cancer axillary lymph node status prediction model according to any one of claims 1-6.

8. A breast cancer axillary lymph node status prediction system, characterized in that, Includes a data acquisition module and a prediction module; The data acquisition module is used to collect the medical diagnostic data of the test subject. The prediction module is used to input the medical diagnostic data into the breast cancer axillary lymph node status prediction model of claim 7, and to predict and evaluate the breast cancer axillary lymph node status of the test subject.

Images