Breast cancer postoperative bleeding risk prediction method and system based on machine learning

Abstract

The invention relates to the technical field of postoperative risk prediction, in particular to a breast cancer postoperative bleeding risk prediction method and system based on machine learning, and the method comprises the steps: extracting feature variables for breast cancer postoperative bleeding risk prediction from personal data and clinical data through feature importance analysis; utilizing a plurality of machine learning models to construct a plurality of prediction models based on the characteristic variables, and performing performance evaluation to select a prediction model with optimal overall performance; and carrying out local index optimization on the prediction model with the optimal global performance by utilizing reinforcement learning to obtain an optimal prediction model. According to the invention, a plurality of prediction models for breast cancer postoperative bleeding risk prediction are constructed by using a plurality of machine learning models, performance evaluation is carried out on the plurality of prediction models to screen out an optimal prediction model, and then index optimization with insufficient precision is carried out on the optimal prediction model one by one in combination with reinforcement learning. And obtaining a prediction model which is optimized in performance and helps to identify high-risk patients.

Description

Technical Field

[0001] The present invention relates to the technical field of postoperative risk prediction, and in particular to a method and system for predicting postoperative bleeding risk in breast cancer based on machine learning. Background Art

[0002] Breast cancer is a malignant tumor that occurs in the epithelial tissue of the breast, accounting for approximately 99% of breast cancer incidence in women. Modified radical mastectomy is a commonly used surgical procedure for the clinical treatment of breast cancer. It can directly remove tumor lesions and locally invasive tumor tissue, thereby prolonging the patient's survival. However, while surgical treatment is effective, it may also cause a variety of complications. Among them, postoperative bleeding is one of the common complications after radical mastectomy for breast cancer, with an incidence of approximately 5%-10%. Postoperative bleeding may not only lead to prolonged hospitalization and increased medical costs, but may also have a certain impact on the patient's physical and psychological health, hindering postoperative functional recovery and comprehensive anti-tumor treatment. Therefore, early evaluation and intervention of postoperative bleeding risks are very important for reducing adverse clinical outcomes and the management of modified radical mastectomy.

[0003] Currently, risk prediction methods for postoperative bleeding after radical mastectomy for breast cancer are usually based on artificial algorithms to construct risk prediction models for postoperative bleeding and perform intelligent predictions. For example, the LR model is commonly used in binary classification research. However, this single-type binary classification model has limited predictive ability and is difficult to achieve accurate prediction of postoperative bleeding risk. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for predicting postoperative bleeding risk of breast cancer based on machine learning, so as to solve the technical problem that the single-type binary classification model used in the existing technology for postoperative bleeding risk prediction has limited prediction ability and is difficult to achieve accurate prediction of postoperative bleeding risk.

[0005] In order to solve the above technical problems, the present invention specifically provides the following technical solutions:

[0006] A method for predicting bleeding risk after breast cancer surgery based on machine learning, comprising the following steps:

[0007] Access to patients’ personal and clinical data;

[0008] Feature importance analysis was used to extract characteristic variables for predicting bleeding risk after breast cancer surgery from personal data and clinical data;

[0009] Using multiple machine learning models to construct multiple prediction models for predicting the risk of postoperative bleeding after breast cancer surgery based on the characteristic variables and the outcome variable of whether bleeding occurs or not, and performing performance evaluation on the multiple prediction models to select the prediction model with the best global performance;

[0010] Use reinforcement learning to optimize local indicators of the prediction model with the best global performance to obtain the optimal prediction model;

[0011] The SHAP algorithm is used to perform interpretability analysis on the optimized prediction model.

[0012] As a preferred embodiment of the present invention, the personal data includes: age, body mass index (BMI), whether or not there is diabetes, hypertension, or coronary heart disease;

[0013] The clinical data includes: tumor characteristics and treatment information, wherein the tumor characteristics include: tumor size, lesion location, tumor pathological stage, pathological type, and lymph node metastasis;

[0014] Treatment information included: preoperative neoadjuvant chemotherapy, preoperative anticoagulant use, and surgery time.

[0015] As a preferred embodiment of the present invention, the criteria for determining whether bleeding occurs include:

[0016] Clinical observation: obvious blood oozing from the incision, requiring frequent dressing changes; the drainage fluid is bright red or bloody and the amount is large; swelling, lumps, or hematomas appear in the surgical area;

[0017] Objective indicators: Postoperative hemoglobin decreases by more than the preset threshold compared with preoperative level; drainage volume exceeds the preset threshold within 24 hours after surgery; ultrasound or CT shows effusion or hematoma in the surgical area.

[0018] As a preferred solution of the present invention, the method for extracting characteristic variables includes:

[0019] The importance of each variable in personal data and clinical data was ranked using the XGBoost model;

[0020] According to the importance ranking results of each variable in the personal data and clinical data, the top 7 variables were selected as the characteristic variables.

[0021] As a preferred embodiment of the present invention, the method for constructing multiple prediction models includes:

[0022] The dataset consisting of multiple patients is divided into training set and test set in a ratio of 7:3;

[0023] On the training set, the XGBoost model, LR model, DT model, SVM model, and RF model were used respectively, with feature variables as input and outcome variables as output, to conduct model training, and construct prediction models based on the XGBoost model, the LR model, the DT model, the SVM model, and the RF model for predicting the risk of postoperative bleeding in breast cancer.

[0024] As a preferred solution of the present invention, the method for determining the prediction model with the best global performance includes:

[0025] On the test set, the prediction performance of the XGBoost model, the LR model, the DT model, the SVM model, and the RF model was evaluated for postoperative bleeding risk prediction after breast cancer surgery.

[0026] Among the prediction models based on the XGBoost model, the prediction model based on the LR model, the prediction model based on the DT model, and the prediction model based on the SVM model, the prediction model with the highest AUC value is selected as the prediction model with the best global performance.

[0027] As a preferred embodiment of the present invention, the evaluation indicators include sensitivity, specificity, F1 score, AUC value, AUC 95% CI, accuracy and precision.

[0028] As a preferred embodiment of the present invention, the method for optimizing the prediction model includes:

[0029] Among the prediction models based on the XGBoost model, the prediction model based on the LR model, the prediction model based on the DT model, and the prediction model based on the SVM model, all prediction models whose evaluation index values ​​are higher than the prediction model with the best global performance are selected as reinforcement learning models;

[0030] Through mutual reinforcement training between each reinforcement learning model and the prediction model with the best global performance, the prediction model with the best global performance after training is used as the optimized prediction model;

[0031] Among them, the reinforcement training loss of the prediction model with the best global performance is:

[0032]

[0033] in,

[0034] Where, L s is the reinforcement training loss of the prediction model with the best global performance, Y s Y is the prediction result of the outcome variable output by the prediction model with the best global performance, t,i is the predicted result of the outcome variable output by the i-th reinforcement learning model, Y r is the true result of the outcome variable, H CE (·) is the cross entropy loss function, γ i H CE (Y s ,Yt,i ), k(i) is the evaluation index item of the reinforcement learning model that is higher than the prediction model with the best global performance, P t,i,k(i) is the value of the evaluation index k(i) in the i-th reinforcement learning model, P s,k(i) is the value of the evaluation index k(i) in the prediction model with the best global performance, and n is the total number of reinforcement learning models;

[0035] The reinforcement training loss of the reinforcement learning model is:

[0036]

[0037] in,

[0038] Where, L t is the reinforcement training loss of the reinforcement learning model, ω i H CE (Y t,i ,Y r )’s weight item.

[0039] As a preferred solution of the present invention, the performance of the optimized prediction model in predicting postoperative bleeding risk after breast cancer surgery is evaluated on a test set.

[0040] As a preferred embodiment of the present invention, the present invention provides a machine learning-based breast cancer postoperative bleeding risk prediction system, which is applied to a machine learning-based breast cancer postoperative bleeding risk prediction method. The system includes:

[0041] Data collection module, used to obtain patients' personal data and clinical data;

[0042] A data processing module is configured to extract characteristic variables for predicting the risk of postoperative bleeding after breast cancer surgery from personal data and clinical data using feature importance analysis; construct multiple prediction models for predicting the risk of postoperative bleeding after breast cancer surgery based on the characteristic variables and outcome variables indicating whether bleeding occurs or not using multiple machine learning models; and perform performance evaluation on the multiple prediction models to select the prediction model with the best global performance; and perform local indicator optimization on the prediction model with the best global performance using reinforcement learning to obtain the optimal prediction model.

[0043] The model analysis unit is used to perform interpretability analysis on the optimized prediction model using the SHAP algorithm.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This paper uses feature importance analysis to explore the key risk factors for postoperative bleeding, and uses multiple machine learning models to construct multiple prediction models for predicting the risk of postoperative bleeding in breast cancer. The performance of multiple prediction models is evaluated to screen out the best prediction model. Reinforcement learning is then used to optimize the indicators of the best prediction model one by one that lack precision, and a prediction model with optimized performance is obtained to help identify high-risk patients, providing a scientific basis for formulating scientific and effective interventions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0047] Figure 1 A flow chart of a method for predicting bleeding risk after breast cancer surgery based on machine learning provided in an embodiment of the present invention;

[0048] Figure 2 A block diagram of a system for predicting postoperative bleeding risk for breast cancer based on machine learning provided by an embodiment of the present invention;

[0049] Figure 3 ROC curves of the five prediction models provided in the embodiments of the present invention;

[0050] Figure 4 A feature importance analysis diagram provided for an embodiment of the present invention;

[0051] Figure 5 SHAP feature analysis diagram of the XGBoost model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] like Figure 1 As shown, the present invention provides a method for predicting bleeding risk after breast cancer surgery based on machine learning, comprising the following steps:

[0054] Access to patients’ personal and clinical data;

[0055] Feature importance analysis was used to extract characteristic variables for predicting bleeding risk after breast cancer surgery from personal data and clinical data;

[0056] Multiple machine learning models were used to construct multiple prediction models for postoperative bleeding risk prediction after breast cancer surgery based on feature variables and outcome variables indicating whether bleeding occurred or not. The performance of multiple prediction models was evaluated to select the prediction model with the best global performance.

[0057] Use reinforcement learning to optimize local indicators of the prediction model with the best global performance to obtain the optimal prediction model;

[0058] The SHAP algorithm is used to perform interpretability analysis on the optimized prediction model, such as Figure 5 shown.

[0059] The present invention first collects the patient's personal data and clinical data, and analyzes them to mine characteristic variables related to the risk of postoperative bleeding after breast cancer surgery, so as to predict the risk of postoperative bleeding after breast cancer surgery.

[0060] After completing feature extraction, the present invention uses the extracted feature variables as a machine learning model for predicting the risk of bleeding after breast cancer surgery. Given that the results of bleeding risk prediction are: bleeding occurs and bleeding does not occur, which belongs to the binary classification category, the machine learning model for predicting the risk of bleeding after breast cancer surgery is constructed using a binary classification method.

[0061] The present invention selects multiple machine learning models with good performance in binary classification problems, namely, XGBoost model (Extreme Gradient Boosting), logistic regression LR model (Logistic Regression), random forest RF model (Random Forest), support vector machine SVM model, and decision tree DT model (Decision Tree). These binary classification models are trained as basic models for postoperative bleeding risk prediction to select binary classification models with performance advantages in postoperative bleeding risk prediction, and then carry out targeted intensive training on them to strengthen all-round performance indicators, construct an optimized postoperative bleeding risk prediction model, and ensure high-performance classification prediction of postoperative bleeding risk.

[0062] In the present invention, the above five machine learning models are trained one by one using the training set to construct five postoperative bleeding risk prediction models, and the performance of these five postoperative bleeding risk prediction models is evaluated one by one using the test set to obtain their performance index values. The AUC index in the performance index is a global index or a comprehensive index. It is generally believed that the higher the AUC index value, the better the classification performance of the model. Therefore, the present invention uses the AUC value as a screening index to screen out the binary classification model with performance advantages in postoperative bleeding risk prediction. The model with the best performance of the five machine learning models in a single training, that is, the model with the highest AUC value, is screened as the binary classification model with performance advantages in postoperative bleeding risk prediction, and subsequent intensive training is carried out.

[0063] The binary classification model with performance advantages in predicting postoperative bleeding risk screened out by the present invention has the highest AUC indicator, but its performance in other local indicators such as sensitivity, specificity, F1 score, accuracy and precision is not necessarily excellent. Therefore, the present invention adopts an intensive training method for the binary classification model selected by the AUC indicator to enhance its expressiveness in other local indicators, thereby ultimately obtaining a postoperative bleeding risk prediction model with optimized performance.

[0064] In the intensive training of the binary classification model selected by the AUC indicator, the present invention selects five postoperative bleeding risk prediction models that perform better than the binary classification model selected by the AUC indicator (marked as the prediction model with the best global performance) in sensitivity, specificity, F1 score, accuracy and precision (marked as the reinforcement learning model) to guide the training of the binary classification model selected by the AUC indicator and enhance its performance.

[0065] The present invention controls the prediction model with the best global performance to learn the prediction output of the reinforcement learning model, so that the output of the prediction model with the best global performance that satisfies the global advantage can be close to the output of the local advantage, so that after the training is completed, the output of the prediction model with the best global performance can be as consistent as possible with the local advantage while obtaining the global advantage. In other words, an optimal effect can be achieved in all evaluation indicators.

[0066] In order to enable the prediction model with the best global performance to learn the prediction output of the reinforcement learning model, the present invention sets the reinforcement training loss of the prediction model with the best global performance to two parts. One part is the cross entropy loss between the output of the prediction model with the best global performance and the true value, which is used to ensure that in the process of the prediction model with the best global performance learning to the reinforcement learning model, the output of the prediction model with the best global performance is kept close to the true value, and the original global advantage is maintained. It is equivalent to the constraint item of the global advantage, and avoids only focusing on the output close to the reinforcement learning model in the process of learning to the reinforcement learning model, resulting in overfitting learning of the local advantage and loss of the original global performance advantage.

[0067] The other part is the cross-entropy loss between the output of the prediction model with the best global performance and the output of the reinforcement learning model. This is used to ensure that during the learning process of the prediction model with the best global performance from the reinforcement learning model, the output of the prediction model with the best global performance is close to the output of the reinforcement learning model, thereby learning local advantages and achieving the goal of reinforcement learning. Moreover, since there are more than one reinforcement learning models, the prediction model with the best global performance must learn from multiple reinforcement learning models during the reinforcement training process. To this end, the present invention sets a reinforcement learning weight, and the prediction model with the best global performance determines the priority of learning from the reinforcement learning model based on the weight.

[0068] The setting of the reinforcement learning weight in the present invention is determined according to the degree of advantage of the evaluation indicators (sensitivity, specificity, F1 score, accuracy and precision) of the reinforcement learning model over the prediction model with the best global performance. The higher the evaluation indicators of the reinforcement learning model are than the prediction model with the best global performance, the greater the reinforcement learning weight of the reinforcement learning model, the higher the priority of reinforcement learning from it, and the more emphasis is placed on learning from the reinforcement learning model, so that the performance shortcomings of the prediction model with the best global performance can be compensated, and better local performance learning effects can be obtained.

[0069] In the process of reinforcement learning of the prediction model with the best global performance, the present invention also performs reinforcement learning on each reinforcement learning model, so that when the prediction model with the best global performance has a deep learning process of a certain reinforcement learning model, or is close to mastering the local performance indicator advantage of a certain reinforcement learning model, it is expected to further learn the local performance. At this time, the reinforcement learning model will be controlled to be further trained, and the local performance advantage of the reinforcement learning model itself will be further explored and strengthened, and it is expected that it will be improved. If there is improvement, the prediction model with the best global performance will continue to follow the reinforcement learning model to learn, and then achieve improvement in local performance advantages.

[0070] During the reinforcement learning process of multiple reinforcement learning models, the reinforcement learning process is also controlled by controlling their priority. The deeper the learning process of a reinforcement learning model by the prediction model with the best global performance, or the closer it is to mastering the local performance indicator advantages of a reinforcement learning model, the more urgent the self-performance reinforcement learning of the reinforcement learning model is, and the higher the priority needs to be set for the reinforcement learning model.

[0071] The present invention sets ω in the loss function of the reinforcement learning model i To control the self-reinforcement priority of each reinforcement learning model, ω i The cross entropy loss between the output of the prediction model with the best global performance and the output of the reinforcement learning model is set. The smaller the cross entropy loss, the deeper the learning process of the prediction model with the best global performance for a reinforcement learning model, or the closer it is to mastering the local performance indicator advantage of a reinforcement learning model. i The larger it is, the higher the training priority of the reinforcement learning model is, and the more emphasis is placed on the self-reinforcement of the reinforcement learning model, hoping to further explore and strengthen the local performance advantages.

[0072] Therefore, the prediction model with the best global performance and the reinforcement learning model reinforce and promote each other, ensuring that the risk bleeding prediction model finally integrates the prediction performance advantages of multiple binary classification models to achieve prediction optimization.

[0073] Personal data include: age, body mass index (BMI), whether there is diabetes, hypertension, or coronary heart disease;

[0074] Clinical data include: tumor characteristics and treatment information, where tumor characteristics include: tumor size, lesion location, tumor pathological stage, pathological type, and lymph node metastasis;

[0075] Treatment information included: preoperative neoadjuvant chemotherapy, preoperative anticoagulant use, and surgery time.

[0076] The criteria for determining whether bleeding occurs include:

[0077] Clinical observation: obvious blood oozing from the incision, requiring frequent dressing changes; the drainage fluid is bright red or bloody and the amount is large; swelling, lumps, or hematomas appear in the surgical area;

[0078] Objective indicators: Postoperative hemoglobin decreases by more than a preset threshold compared to preoperative levels (e.g., a decrease of >2 g / dL); drainage volume exceeds a preset threshold within 24 hours after surgery (e.g., >200 mL); ultrasound or CT shows effusion or hematoma in the surgical area.

[0079] The extraction methods of feature variables include:

[0080] The importance of each variable in personal data and clinical data is ranked by the XGBoost model, such as Figure 4 As shown, they are preoperative neoadjuvant chemotherapy, preoperative use of anticoagulants, tumor size, age, body mass index, operation time, and hypertension;

[0081] Based on the importance ranking results of each variable in personal data and clinical data, the top 7 variables were selected as feature variables.

[0082] Methods for building multiple prediction models include:

[0083] The dataset consisting of multiple patients is divided into training set and test set in a ratio of 7:3;

[0084] On the training set, the XGBoost model, LR model, DT model, SVM model, and RF model were used respectively, with feature variables as input and outcome variables as output, to conduct model training, and construct prediction models based on the XGBoost model, the LR model, the DT model, the SVM model, and the RF model for predicting the risk of postoperative bleeding in breast cancer.

[0085] Methods for determining the prediction model with the best global performance include:

[0086] On the test set, the prediction performance of the XGBoost model, the LR model, the DT model, the SVM model, and the RF model was evaluated for postoperative bleeding risk prediction after breast cancer surgery.

[0087] Among the prediction models based on the XGBoost model, the prediction model based on the LR model, the prediction model based on the DT model, and the prediction model based on the SVM model, the prediction model with the highest AUC value is selected as the prediction model with the best global performance.

[0088] Evaluation indicators include sensitivity, specificity, F1 score, AUC value, AUC 95% CI, accuracy and precision.

[0089] In the present invention, the five machine learning models were trained one by one using the training set to construct five postoperative bleeding risk prediction models, and the performance of the five postoperative bleeding risk prediction models was evaluated one by one using the test set to obtain their performance index values, as shown in Table 1 and Figure 3As shown in the figure, the XGBoost model has the best AUC indicator. Therefore, the present invention uses the XGBoost model as the prediction model with the best global performance for subsequent reinforcement training. It also shows that the LR model and the DT model are superior to the XGBoost model in accuracy and precision, and can be used as reinforcement learning models to guide the XGBoost model for reinforcement training.

[0090] Table 1 Performance of five prediction models on the test set

[0091]

[0092]

[0093] Correspondingly, in the reinforcement learning of the XGboost model, the performance index of the LR model is higher than that of the XGboost model by (0.942-0.937)+(0.926-0.727)=0.204, and the performance index of the DT model is higher than that of the XGboost model by (0.767-0.727)=0.04. After normalization, the reinforcement learning weight γ of the LR model is obtained. LR =0.836, reinforcement learning weight γ of DT model DT =0.164, substitute the value L s The XGBoost model is trained intensively in the process, and the XGBoost model after training is used as the optimized prediction model.

[0094] The performance of the optimized prediction model is then evaluated. As shown in Table 2, the enhanced XGBoost model maintains the optimal AUC indicator while improving local performance indicators compared to the XGBoost model before enhancement. Similarly, the LR model and DT model, which are reinforcement learning models, also have certain performance improvements after self-reinforcement learning.

[0095] Table 2 Performance of five prediction models on the test set (results on the test set after enhancement)

[0096]

[0097] Methods for optimizing forecasting models include:

[0098] Among the prediction models based on the XGBoost model, the prediction model based on the LR model, the prediction model based on the DT model, and the prediction model based on the SVM model, all prediction models whose evaluation index values ​​are higher than the prediction model with the best global performance are selected as reinforcement learning models;

[0099] Through mutual reinforcement training between each reinforcement learning model and the prediction model with the best global performance, the prediction model with the best global performance after training is used as the optimized prediction model;

[0100] Among them, the reinforcement training loss of the prediction model with the best global performance is:

[0101]

[0102] in,

[0103] Where, L s is the reinforcement training loss of the prediction model with the best global performance, Y s Y is the prediction result of the outcome variable output by the prediction model with the best global performance, t,i is the predicted result of the outcome variable output by the i-th reinforcement learning model, Y r is the true result of the outcome variable, H CE (·) is the cross entropy loss function, γ i H CE (Y s ,Y t,i ), k(i) is the evaluation index item of the reinforcement learning model that is higher than the prediction model with the best global performance, P t,i,k(i) is the value of the evaluation index k(i) in the i-th reinforcement learning model, P s,k(i) is the value of the evaluation index k(i) in the prediction model with the best global performance, and n is the total number of reinforcement learning models;

[0104] The reinforcement training loss of the reinforcement learning model is:

[0105]

[0106] in,

[0107] Where, L t is the reinforcement training loss of the reinforcement learning model, ω i H CE (Y t,i ,Y r )’s weight item.

[0108] The performance of the optimized prediction model in predicting bleeding risk after breast cancer surgery was evaluated on the test set.

[0109] The method for collecting personal data and clinical data in the present invention is as follows: 2,000 patients who underwent radical mastectomy in the Department of Thyroid and Breast Surgery of the hospital from January 2023 to October 2024 were retrospectively collected. Inclusion criteria: (1) All patients' "breast cancer" diagnosis met the diagnostic criteria for breast cancer in the "Guidelines and Standards for the Diagnosis and Treatment of Breast Cancer of the Chinese Anti-Cancer Association (2024 Edition)" and were confirmed to be breast cancer by pathological examination; (2) received modified radical mastectomy; (3) had complete clinical data; (4) were aged ≥18 years; (5) underwent breast cancer surgery for the first time; Exclusion criteria: (1) patients with combined immune dysfunction and coagulation dysfunction; (2) patients with combined blood system diseases or other malignant tumors; (3) patients who only received chemotherapy, radiotherapy or targeted therapy without surgery; (4) patients with combined cognitive dysfunction and communication disorders; (5) patients with incomplete clinical data; The sample size of this study was divided into training set and test set with a ratio of 7:3. After theoretical research and integration, 13 possible risk factors were found, and the incidence of bleeding in patients after radical mastectomy was 5%-10%. The logistic regression model requires that the number of positive events is more than 10 times the independent variable, and the decision tree CART and random forest models require the sample size to be more than 2 times the explanatory variable. Therefore, this study adopted the sample size estimation method of the logistic regression model, including 1400 cases in the training set and 600 cases in the test set.

[0110] Through the hospital's medical information system and electronic medical record system, patients' personal and clinical data were collected for data analysis and model construction. To ensure predictive effectiveness, as many clinical factors as possible were included as model predictor variables during data collection. Based on the research objectives, data were collected by reviewing relevant literature and referring to relevant clinical experience. Duplicate features and completely collinear features were removed from the collected data. Variables or subjects with a missing rate exceeding 20% ​​were deleted, and variables with a missing rate less than 20% were imputed. The data were normalized using a linear function normalization method. Because the bleeding group accounted for a relatively low proportion of subjects (approximately 5% to 10%), the data were unbalanced. To address this data imbalance, this study processed the training set using random oversampling and synthetic minority over-sampling techniques (SMOTE). Random oversampling is a basic method for addressing class imbalance in datasets, primarily by replicating minority class samples to balance the class distribution.

[0111] The present invention uses SPSS26.0 software for statistical analysis, and the enumeration data are expressed as number of cases (percentage) [n (%)], and the comparison between groups is expressed as X 2Test; the Shapiro-Wilk method was used to test the normality of the measurement data. The data that met the normal distribution were expressed as mean ± standard deviation (x ± s), and the independent sample t test was used for inter-group comparison; the data that did not meet the normal distribution were expressed as median (four-digit) [M(P25, P75)], and the Mann-Whitney U test was used for inter-group comparison. The difference was considered statistically significant when P < 0.05.

[0112] A total of 2,000 patients undergoing modified radical mastectomy for breast cancer were enrolled. Of these, 186 experienced postoperative bleeding (the bleeding group), with a bleeding rate of 9.3%. The remaining 1,814 patients did not bleed (the non-bleeding group). A comparison of general characteristics between the bleeding and non-bleeding groups is shown in Table 3.

[0113] Table 3 Comparison of general data between the bleeding group and the non-bleeding group

[0114]

[0115]

[0116] like Figure 2 As shown, the present invention provides a machine learning-based breast cancer postoperative bleeding risk prediction system, which is applied to a machine learning-based breast cancer postoperative bleeding risk prediction method. The system includes:

[0117] Data collection module, used to obtain patients' personal data and clinical data;

[0118] The data processing module is used to extract characteristic variables for predicting postoperative bleeding risk after breast cancer surgery from personal and clinical data using feature importance analysis; multiple machine learning models are used to construct multiple prediction models for predicting postoperative bleeding risk after breast cancer surgery based on the characteristic variables and outcome variables indicating whether bleeding occurs or not; the performance of multiple prediction models is evaluated to select the prediction model with the best global performance; reinforcement learning is used to optimize the local indicators of the prediction model with the best global performance to obtain the optimal prediction model;

[0119] The model analysis unit is used to perform interpretability analysis on the optimized prediction model using the SHAP algorithm.

[0120] This paper uses feature importance analysis to explore the key risk factors for postoperative bleeding, and uses multiple machine learning models to construct multiple prediction models for predicting the risk of postoperative bleeding in breast cancer. The performance of multiple prediction models is evaluated to screen out the best prediction model. Reinforcement learning is then used to optimize the indicators of the best prediction model one by one that lack precision, and a prediction model with optimized performance is obtained to help identify high-risk patients, providing a scientific basis for formulating scientific and effective interventions.

[0121] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present application.

Claims

1. A method for predicting bleeding risk after breast cancer surgery based on machine learning, characterized in that: The following steps are involved: Access to patients’ personal and clinical data; Feature importance analysis was used to extract characteristic variables for predicting bleeding risk after breast cancer surgery from personal data and clinical data; Using multiple machine learning models to construct multiple prediction models for predicting the risk of postoperative bleeding after breast cancer surgery based on the characteristic variables and the outcome variable of whether bleeding occurs or not, and performing performance evaluation on the multiple prediction models to select the prediction model with the best global performance; Use reinforcement learning to optimize local indicators of the prediction model with the best global performance to obtain the optimal prediction model; The SHAP algorithm is used to perform interpretability analysis on the optimized prediction model.

2. The method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning according to claim 1, characterized in that: The personal data include: age, body mass index (BMI), whether there is diabetes, hypertension, or coronary heart disease; The clinical data includes: tumor characteristics and treatment information, wherein the tumor characteristics include: tumor size, lesion location, tumor pathological stage, pathological type, and lymph node metastasis; Treatment information included: preoperative neoadjuvant chemotherapy, preoperative anticoagulant use, and surgery time.

3. The method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning according to claim 1, characterized in that: The criteria for determining whether bleeding occurs include: Clinical observation: obvious blood oozing from the incision, requiring frequent dressing changes; the drainage fluid is bright red or bloody and the amount is large; swelling, lumps, or hematomas appear in the surgical area; Objective indicators: Postoperative hemoglobin decreases by more than the preset threshold compared with preoperative level; drainage volume exceeds the preset threshold within 24 hours after surgery; ultrasound or CT shows effusion or hematoma in the surgical area.

4. The method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning according to claim 2, characterized in that: The method for extracting the characteristic variables includes: The importance of each variable in personal data and clinical data was ranked using the XGBoost model; According to the importance ranking results of each variable in the personal data and clinical data, the top 7 variables were selected as the characteristic variables.

5. The method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning according to claim 4, characterized in that: Methods for building multiple prediction models include: The dataset consisting of multiple patients is divided into training set and test set in a ratio of 7:3; On the training set, the XGBoost model, LR model, DT model, SVM model, and RF model were used respectively, with feature variables as input and outcome variables as output, to conduct model training, and construct prediction models based on the XGBoost model, the LR model, the DT model, the SVM model, and the RF model for predicting the risk of postoperative bleeding in breast cancer.

6. The method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning according to claim 5, characterized in that: The method for determining the prediction model with the best global performance includes: On the test set, the prediction performance of the XGBoost model, the LR model, the DT model, the SVM model, and the RF model was evaluated for postoperative bleeding risk prediction after breast cancer surgery. Among the prediction models based on the XGBoost model, the prediction model based on the LR model, the prediction model based on the DT model, and the prediction model based on the SVM model, the prediction model with the highest AUC value is selected as the prediction model with the best global performance.

7. The method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning according to claim 6, characterized in that: The evaluation indicators include sensitivity, specificity, F1 score, AUC value, AUC 95% CI, accuracy and precision.

8. The method for predicting bleeding risk after breast cancer surgery based on machine learning according to claim 7, characterized in that: The method for optimizing the prediction model comprises: Among the prediction models based on the XGBoost model, the prediction model based on the LR model, the prediction model based on the DT model, and the prediction model based on the SVM model, all prediction models whose evaluation index values ​​are higher than the prediction model with the best global performance are selected as reinforcement learning models; Through mutual reinforcement training between each reinforcement learning model and the prediction model with the best global performance, the prediction model with the best global performance after training is used as the optimized prediction model; Among them, the reinforcement training loss of the prediction model with the best global performance is: in, Where, L s is the reinforcement training loss of the prediction model with the best global performance, Y s Y is the prediction result of the outcome variable output by the prediction model with the best global performance, t,i Y is the predicted result of the outcome variable output by the i-th reinforcement learning model, r is the true result of the outcome variable, H CE (·) is the cross entropy loss function, γ i H CE (Y s ,Y t,i ), k(i) is the evaluation index item of the reinforcement learning model that is higher than the prediction model with the best global performance, P t,i,k(i) is the value of the evaluation index k(i) in the i-th reinforcement learning model, P s ,k(i) is the value of the evaluation index k(i) in the prediction model with the best global performance, and n is the total number of reinforcement learning models; The reinforcement training loss of the reinforcement learning model is: in, Where, L t is the reinforcement training loss of the reinforcement learning model, ω i H CE (Y t,i ,Y r )’s weight item.

9. The method for predicting bleeding risk after breast cancer surgery based on machine learning according to claim 8, characterized in that: The performance of the optimized prediction model in predicting bleeding risk after breast cancer surgery was evaluated on the test set.

10. A machine learning-based system for predicting postoperative bleeding risk in breast cancer, characterized in that: A method for predicting postoperative bleeding risk after breast cancer surgery based on machine learning, as described in any one of claims 1 to 9, wherein the system comprises: Data collection module, used to obtain patients' personal data and clinical data; A data processing module is configured to extract characteristic variables for predicting the risk of postoperative bleeding after breast cancer surgery from personal data and clinical data using feature importance analysis; construct multiple prediction models for predicting the risk of postoperative bleeding after breast cancer surgery based on the characteristic variables and outcome variables indicating whether bleeding occurs or not using multiple machine learning models; and perform performance evaluation on the multiple prediction models to select the prediction model with the best global performance; and perform local indicator optimization on the prediction model with the best global performance using reinforcement learning to obtain the optimal prediction model. The model analysis unit is used to perform interpretability analysis on the optimized prediction model using the SHAP algorithm.

Images