XAI-based young triple negative breast cancer prognosis prediction system
By applying machine learning models based on interpretable artificial intelligence in the prognosis prediction of young triple-negative breast cancer, the problems of inaccurate and opaque prediction in the prior art are solved, and higher prediction accuracy and model interpretability are achieved, and personalized diagnosis and treatment decisions are supported.
Patent Information
- Application Number
- CN202411756207.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-03-18
AI Technical Summary
In the prior art, the inaccurate, opaque and lack of interpretability of young triple-negative breast cancer, which limits its wide application in the medical field.
Using machine learning models based on interpretability artificial intelligence (XAI), multimodal data is obtained through SEER database, feature screening and model construction are carried out, and interpretability analysis technologies such as Permutation Importance and LIME are combined to provide explanatory support for the model.
It improves the accuracy and transparency of the prognosis prediction of young triple-negative breast cancer, makes the medical decision-making process more credible and explainable, and provides a scientific basis for the formulation of personalized diagnosis and treatment plans.
Smart Images
Figure BDA0005166351910000051 
Figure HDA0005166351920000011 
Figure HDA0005166351920000012
Abstract
Description
[0001] The present invention relates to a prognosis prediction system for young triple-negative breast cancer based on explainable artificial intelligence (XAI), which combines artificial intelligence with medical health technology and focuses on intelligent medical applications for prognosis prediction and personalized diagnosis and treatment of young triple-negative breast cancer. Background Art
[0002] Triple-negative breast cancer is a highly heterogeneous and aggressive subtype of breast cancer, characterized by the lack of estrogen receptors, progesterone receptors, and overexpression of the HER2 gene. Traditional treatments usually include chemotherapy, but due to the different biological characteristics and disease manifestations of young patients, the clinical outcome is highly uncertain. Especially in young patients, the prognosis of triple-negative breast cancer is usually worse and the risk of recurrence is higher. How to effectively predict the prognosis and conduct personalized intervention is an important topic in current medical research.
[0003] With the continuous integration of computer science and medical technology, machine learning technology has been widely used in cancer diagnosis and treatment. By analyzing a large amount of clinical data and molecular biology data, machine learning models can mine hidden patterns from data and make prognostic predictions. However, traditional machine learning models are considered to be "black boxes" and cannot provide clear explanations for prediction results, which limits their widespread application in the medical field, especially in breast cancer prognosis assessment.
[0004] In recent years, Explainable AI (XAI), as an emerging research direction, has begun to introduce transparency and credibility into intelligent diagnosis and treatment systems. The goal of XAI is to develop technologies that can explain the prediction mechanism of the model so that doctors can understand and trust the model output. In the prognosis prediction of triple-negative breast cancer, the introduction of an explainable machine learning model can not only improve the accuracy of the prediction, but also make the medical decision-making process more transparent. For example, by explaining how the model uses tumor location, tumor grade, and TNM staging to make predictions, doctors can optimize treatment strategies and reasonably evaluate the patient's prognosis. Summary of the invention
[0005] The purpose of the present invention is to provide a prognosis prediction system for young triple-negative breast cancer based on XAI to solve the problems of inaccurate, opaque and lack of explainability in clinical prognosis prediction in the prior art.
[0006] To achieve the above objectives, the present invention proposes a prognosis prediction system for young triple-negative breast cancer based on XAI. The working process of the model and system includes the following steps:
[0007] S1) Data acquisition and preprocessing
[0008] Multimodal data related to young triple-negative breast cancer patients were obtained from the public SEER (Surveillance, Epidemiology, and End Results) database, including basic demographic information (such as gender, age, income, marital status), treatment information (such as whether radiotherapy, chemotherapy, surgery), TNM stage, tumor grade, and other important clinical characteristics. At the same time, the collected raw data were preprocessed, including missing value filling, outlier detection, and data standardization to ensure data quality and consistency.
[0009] S2) Feature Screening
[0010] A combination of univariate and multivariate COX regression analysis was used to screen out features significantly associated with prognosis from a variety of clinical features. Univariate analysis preliminarily screened potential important variables through statistical tests, while multivariate analysis further corrected the interaction between covariates. The main prognostic features finally screened out included chemoradiotherapy status, surgery, TNM stage, tumor grade, etc.
[0011] S3) Build machine learning models and evaluate performance
[0012] Based on the processed feature data, multiple machine learning models (such as Gradient Boosting Machine (GBM), XGBoost, Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR) and KNN) are constructed. The screened features are input into each model for training and testing, and the prediction ability of different models is evaluated by performance indicators (such as AUC, sensitivity, specificity, etc.).
[0013] S4) Optimal model selection and optimization
[0014] According to the model evaluation results, the machine learning model with the best performance is selected as the basic model. Subsequently, the basic model is hyperparameter tuned, and the model parameters are further optimized using the random search method, combined with the cross-validation technology to ensure the robustness and generalization ability of the model. After multiple rounds of iterative training, the best machine learning model is finally obtained.
[0015] S5) Model interpretability analysis
[0016] We use two mainstream interpretable analysis techniques, PI (Permutation Importance) and LIME (Local Interpretable Model-agnostic Explanations), to deeply analyze the decision-making process of the best machine learning model. We use Permutation Importance to explain the importance of features globally and identify the features that have the most significant impact on patient prognosis. At the same time, we use LIME to provide local explanations, so that clinicians can understand the individualized prediction results of each patient and the model logic behind them.
[0017] S6) Model application and feedback
[0018] The trained model is deployed as a clinical decision support system to help doctors accurately predict the prognosis of young triple-negative breast cancer patients and provide a scientific basis for formulating personalized treatment plans (such as whether to intensify treatment and follow-up).
[0019] The interpretable machine learning model constructed in the present invention based on the SEER database has a wide range of data sources, high authenticity and authority, and scientifically screens out prognostic features with statistical significance and clinical relevance by combining univariate and multivariate analysis methods, and ensures that the final model has higher prediction accuracy and robustness through comparison and optimization of multiple models. The model introduces interpretable analysis technologies such as Permutation Importance and LIME to achieve global and local explanations, enhance the transparency of the model and the trust of clinicians, and can effectively integrate clinical information, providing reliable support for the prognosis assessment and formulation of personalized treatment plans for young patients with triple-negative breast cancer, with significant clinical application value and broad promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 : Flow chart of the triple-negative breast cancer prognosis prediction model of the present invention.
[0021] Figure 2 : Receiver operating characteristic curve (ROC) diagram of all prediction models in the test set in the embodiments of the present invention.
[0022] Figure 3 : Receiver operating characteristic curve (ROC) diagram of the best prediction model (GBM) in the embodiment of the present invention.
[0023] Figure 4 : Calibration curve of the best prediction model (GBM) in an embodiment of the present invention.
[0024] Figure 5: Permutation Importance feature importance diagram of the best prediction model (GBM) in an embodiment of the present invention.
[0025] Figure 6 : Schematic diagram of LIME local interpretation of the best prediction model (GBM) in an embodiment of the present invention.
[0026] Figure 7 : Web page prediction system diagram of the present invention. DETAILED DESCRIPTION
[0027] 1. Data acquisition and preprocessing
[0028] like Figure 1 As shown, the present invention first collects multimodal data of young triple-negative breast cancer patients based on the SEER database, including demographic information (such as age and gender), tumor characteristics (TNM stage, tumor size, tumor grade) and treatment-related information (whether surgery, chemotherapy and radiotherapy were received). As an internationally renowned cancer epidemiology database, the SEER database is derived from real clinical cases, covers a wide range, is highly authoritative and reliable, and provides a solid data foundation for model construction.
[0029] In data preprocessing, a series of rigorous processing methods are used to address possible problems with the original data: Missing value filling: Use the KNN-based multiple interpolation method to fill in the missing values of key variables, thereby avoiding the impact of incomplete data on model performance. Outlier processing: Detect and eliminate outliers through statistical methods such as box plots and Z-score to ensure the rationality of data distribution. Data standardization: Normalize numerical variables to eliminate the influence of different dimensions for easy model training and prediction. Categorical variable encoding: For categorical variables (such as whether surgery, chemotherapy, and radiotherapy are received), one-hot encoding or factorization is used to meet the input requirements of machine learning models.
[0030] 2. Feature Screening
[0031] In the feature screening stage, the present invention uses a combination of univariate Cox regression analysis and multivariate Cox regression analysis to screen out key features that are significantly associated with the prognosis of triple-negative breast cancer from a variety of clinical features. The specific operation is as follows:
[0032] Univariate Cox regression analysis: The correlation between each feature and the patient's survival outcome was evaluated by performing a Cox proportional hazard model analysis on each candidate feature one by one. This process calculates the hazard ratio (HR) and its 95% confidence interval for each feature, and screens out candidate features that are potentially related to prognosis based on the statistical significance level (P<0.05). For the candidate features screened out by univariate Cox regression, a multivariate Cox proportional hazard model was further used to perform a joint analysis to correct the covariate effects between features. This process eliminates possible multicollinearity problems by evaluating the independent contribution of each feature after correcting for the effects of other variables, thereby screening out key features that have independent predictive power for prognosis. For example, features such as TNM stage, tumor grade, tumor location, radiotherapy and chemotherapy status, and whether surgery was performed were identified as major prognostic features in multivariate analysis. Criteria for feature screening: Throughout the screening process, the present invention strictly controls the statistical significance (P<0.05) and clinical relevance of the variables to ensure that the selected features are not only statistically significant, but also of great value in clinical practice. After the screening was completed, the key features that were finally retained were integrated into a standardized data set and divided into a training set and a test set in a ratio of 7:3 to ensure the scientificity and fairness of subsequent model training and testing. At the same time, in the feature screening stage, the imbalance problem of the data set has been handled through appropriate oversampling techniques to further improve model performance. Through the combination of univariate and multivariate Cox regression analysis, not only can the features that have a significant impact on the prognosis of triple-negative breast cancer be screened out, but the reliability and interpretability of the screening results can also be ensured, thereby providing a high-quality data foundation for the construction of subsequent machine learning models.
[0033] 3. Construction and performance evaluation of prediction models
[0034] The present invention uses a variety of machine learning algorithms, including Gradient Boosting Machine (GBM), XGBoost, Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR) and KNN, etc., and constructs different prediction models by inputting processed feature data and label data. The model performance is comprehensively evaluated by cross-validation technology.
[0035] Evaluation indicators include: AUC-ROC: measures the model's ability to distinguish between positive and negative samples at different thresholds. The closer the value is to 1, the better the performance. Accuracy: indicates the proportion of samples correctly classified by the model, reflecting the overall prediction accuracy of the model. F1-score: combines the harmonic mean of Precision and Recall, suitable for scenarios with imbalanced positive and negative samples. Sensitivity: indicates the predictive ability of positive samples. Specificity: indicates the predictive ability of negative samples. Kappa: measures the consistency between the predicted results and the actual category. The higher the value, the stronger the consistency. Precision (positive prediction value): reflects the proportion of samples predicted by the model to be positive that are actually positive. As shown in Table 1, after comparing the performance of multiple models, the GBM model performed best, with a test set AUC-ROC of 0.830 and an Accuracy of 0.757, which is better than other models in multiple evaluation indicators ( Figure 2 ). This indicates that the GBM model has high accuracy and reliability in predicting the prognosis of triple-negative breast cancer.
[0036] Table 1: Performance indicators of each model
[0037] Models AUC Accuracy Specificity Sensitivity Precision F1 Kappa XGB 0.819 0.733 0.722 0.752 0.607 0.671 0.451 GBM 0.830 0.757 0.759 0.752 0.641 0.692 0.493 RF 0.823 0.739 0.731 0.752 0.615 0.677 0.461 SVM 0.803 0.733 0.755 0.694 0.618 0.654 0.437 KNN 0.704 0.673 0.698 0.628 0.543 0.582 0.316 LR 0.740 0.694 0.736 0.620 0.573 0.595 0.349
[0038] 4. Hyperparameter Optimization and Calibration Analysis
[0039] After GBM is determined to be the best model, the model's hyperparameters are optimized using the random search method.
[0040] Stochastic Gradient Boosting No pre-processingResampling:Cross-Validated(10fold,repeated 4times)
[0041] Summary of sample sizes:1500,1501,1502,1503,1504,1505,...
[0042] Resampl i ng resul ts across tuni ng parameters:
[0043]
[0044] ROC was used to select the optimal model usi ng a tol erance rul e.
[0045] The final values used for the model were n.trees=228, interaction.depth=4, shrinkage=0.06265188 and n.minobsi nnode=7.
[0046] The optimized GBM model not only improves the prediction accuracy, but also shows higher robustness (such as Figure 3 ). Then, the consistency between the predicted probability of the GBM model and the actual observation results was further verified by the calibration curve (such as Figure 4 ). The part of the curve close to the diagonal line shows that the performance of the model under different prediction probabilities is consistent with the actual situation, and has good reliability and stability.
[0047] 5. Model interpretability analysis
[0048] In order to enhance the clinical applicability of the model, the present invention introduces the Permutation Importance and LIME methods to interpret and analyze the GBM model.
[0049] Permutation Importance Analysis: Figure 5 As shown in the figure, Permutation Importance randomly perturbs the values of specific features, observes the changes in model performance (such as AUC-ROC and Accuracy), and quantitatively analyzes the contribution of each feature to model prediction. The results show that age, tumor site, T stage, and tumor grade contribute the most to model performance. This analysis further verifies the importance of these features, is consistent with actual clinical experience, and improves the credibility of the model.
[0050] LIME analysis: Figure 6 As shown in the figure, the LIME method provides local explanations for individual patients, showing the positive and negative effects of specific features on the prediction results. For example, for a certain patient, tumor partition and N stage may play a major supporting role in their prognosis, while tumor M stage and histological grade may play an opposing role in their prognosis prediction. This personalized explanation not only improves the transparency of the model, but also supports doctors in formulating precise treatment plans.
[0051] 6. Interpretable Machine Learning Triple Negative Breast Cancer Prognosis Prediction System
[0052] The triple-negative breast cancer auxiliary prognosis prediction system of the present invention is implemented in the form of a web application, such as Figure 7As shown in Figure 1, users can access and use the system through the Internet. The system consists of a front-end user interface, a back-end data processing module, a prediction model engine, and an interpretable analysis module, providing clinicians with a convenient operating experience and accurate prediction results. The following is a detailed description of the system.
[0053] The system architecture is divided into four main parts. First, the front-end user interface is implemented through web technologies (such as HTML, CSS, and JavaScript), and the front-end provides users with a friendly interactive interface. Users can upload patient data on the web page to view prediction results, calibration curves, and explanatory analysis charts. The front-end design focuses on simplicity and ease of use to help doctors quickly understand the patient's prognosis. Second, the back-end data processing module uses a high-performance server (such as based on Python Flask or Django framework) to manage the data processing process. The back-end module is responsible for receiving the data uploaded by the user and performing the following preprocessing tasks: data cleaning (detecting and removing duplicate values and outliers), missing value filling (using multiple interpolation methods such as KNN-based interpolation to fill missing values of key features), and feature engineering (unique hot encoding of categorical variables and standardization of numerical variables to adapt to machine learning models). Third, the prediction model engine integrates the trained GBM (Gradient Boosting Machine) model as the core prediction engine. The model is developed based on the SEER database and has high accuracy and robustness. The preprocessed patient data is input into the prediction model, and the system generates prognosis prediction results in real time, including key information such as survival probability and recurrence risk. Fourth, in order to improve the transparency and credibility of the prediction results, the interpretable analysis module integrates a variety of interpretable analysis techniques, including Permutation Importance, LIME analysis, and calibration curves.
[0054] In terms of system functions, the first is data input and verification. Users upload patients' multimodal data (such as demographic information, treatment information, and staging data) through the web interface. The system detects the integrity and accuracy of the data in real time, and provides error prompts or correction suggestions. The second is real-time prediction and feedback. The system backend calls the GBM model to predict the input data and generate results within seconds. The results include survival probability and corresponding confidence intervals. In addition, the system provides visual display, including receiver operating characteristic curves, calibration curves, feature importance graphs, and local interpretation graphs.
[0055] Through a web-based interpretable machine learning system, the present invention provides an efficient, transparent and easy-to-use tool for predicting the prognosis of young triple-negative breast cancer patients. The system can not only help doctors quickly obtain accurate prediction results, but also provide additional decision support information through a variety of interpretable analysis methods. The following are the main clinical values. First, improve decision-making efficiency. Doctors can complete data input, model prediction and result interpretation in a few minutes, significantly improving diagnosis and treatment efficiency. Secondly, optimize treatment plans. Based on the patient's personalized prediction results, doctors can formulate more accurate treatment plans. In addition, enhance patient trust, and visually explain the impact of individual characteristics of patients on prognosis, help patients understand their health status more intuitively, and improve patients' compliance with treatment.
[0056] The above embodiments are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above contents. Any obvious changes, adjustments or equivalent replacements made by any person skilled in the art within the technical scope disclosed in the present invention according to the technical solution of the present invention shall be deemed to fall within the protection scope of the present invention.
Claims
1. A prognosis prediction system for young triple-negative breast cancer based on XAI, characterized by: The following steps are involved: S1. Collect clinical information and treatment-related data of triple-negative breast cancer patients based on the SEER database, construct a data set, and preprocess the data set; S2. Based on the pre-processed data, univariate and multivariate COX regression analysis was performed to select the key features that significantly affect the prognosis of patients with triple-negative breast cancer; S3. Use the processed data to build multiple machine learning models and evaluate the performance of each model; S4. Select the best model based on the performance evaluation results, further optimize the hyperparameters, and obtain the best model through cross-validation and iterative training; S5. Perform interpretive analysis on the best model, combining global feature importance with individual feature interpretation to identify the main factors affecting prognosis; S6. Build a system platform on the web page, import patient information through the data upload module, and the system automatically processes data, generates predictions and interpretation results, and displays the prediction and analysis results on the web page for users to use and refer to.
2. The XAI-based prognosis prediction system for young triple-negative breast cancer according to claim 1, characterized in that: The data preprocessing in step S1 includes: removing duplicate values, filling missing values, processing outliers and normalizing data, wherein missing value filling adopts a multiple interpolation method.
3. The XAI-based prognosis prediction system for young triple-negative breast cancer according to claim 1, characterized in that: The model performance evaluation indicators in step S3 include: AUC-ROC, Accuracy, F1-score, Sensitivity, Specificity, Kappa and Precision. AUC-ROC is used to measure the ability of the model to distinguish between positive and negative samples, and the other indicators reflect the accuracy and precision of the classification respectively.
4. The XAI-based prognosis prediction system for young triple-negative breast cancer according to claim 1, characterized in that: The machine learning algorithms in step S3 include: Gradient Boosting Machine (GBM), XGBoost, Random Forest, Support Vector Machine (SVM), Logistic Regression, and K Nearest Neighbor (KNN).
5. The XAI-based prognosis prediction system for young triple-negative breast cancer according to claim 1, characterized in that: Further optimizing the hyperparameters in step S4 includes: adjusting the hyperparameters of the optimal model by a random search method, and evaluating the model performance by cross-validation.
6. The XAI-based prognosis prediction system for young triple-negative breast cancer according to claim 1, characterized in that: Providing explanatory information by combining global feature importance and local feature contribution in step S5 refers to performing explanatory analysis on the model by using PermutationImportance and LIME methods.
7. The XAI-based prognosis prediction system for young triple-negative breast cancer according to claim 1, characterized in that: The step S6 integrates data uploading, processing, analysis and result display functions on the web page, specifically including: a) Data upload module, which allows users to upload patients’ clinical and treatment data through the web page; b) Model prediction module, which performs preprocessing, feature extraction and model prediction based on input data; c) Explanatory analysis module, which provides interpretable results through methods such as Permutation Importance and LIME; d) The result display module displays the prediction results and explanatory information in a visual form on the web page interface for user reference.