Model and system for predicting lifetime of non-small cell lung cancer patient after radiotherapy

Abstract

The invention provides a model and a system for predicting the lifetime of a non-small cell lung cancer patient after radiotherapy. The system comprises an image processing module; a feature extraction module; the tumor radiomics model is trained on the basis of important tumor radiomics characteristics; the lung radiomics model is trained based on the important lung radiomics features; integrating a radiomics model, fusing the important tumor radiomics features and the important lung radiomics features, and screening out the most critical features; and finally fusing the model, and constructing and training to obtain a lifetime prediction model based on the most critical features and the clinical features. Tumor and lung radiomics characteristics and clinical parameters are integrated into the machine learning model and can be used for accurately predicting the overall lifetime of a lung cancer patient receiving radical radiotherapy, and the model is high in prediction precision, high in flooding ability and high in application value.

Description

[0001] This application claims the priority of the Chinese patent application filed with the China Patent Office on November 8, 2024, with application number 202411588428.2 and invention name “A survival prediction model and system for patients with non-small cell lung cancer after radiotherapy”, the entire contents of which are incorporated by reference in this application. Technical Field

[0002] The present invention relates to the technical field of computer-aided medicine, and in particular to a survival prediction model and system for patients with non-small cell lung cancer after radiotherapy. Background Art

[0003] Lung cancer is the leading cause of cancer-related death worldwide. Unresectable locally advanced lung cancer (LALC) includes stage III non-small cell lung cancer (NSCLC) and some limited-stage small cell lung cancer (SCLC). The current standard of care is concurrent / sequential chemotherapy-radiotherapy (RT) followed by immunotherapy for selected NSCLC patients. Although immunotherapy has shown the potential to improve prognosis, the prognosis of LALC varies due to the heterogeneity of LALC itself. However, since most patients still die from lung cancer, accurate prediction of overall survival (OS) will facilitate personalized treatment and shared decision-making.

[0004] Currently, many studies predict the overall survival (OS) of lung cancer patients by combining radiomics features with Cox proportional hazards model or logistic regression model. Although these traditional statistical models are easy to implement and have certain interpretability, their linear assumptions limit their ability to process complex data. The relationship between radiomics features is usually nonlinear, and the Cox model cannot capture this nonlinear pattern and the interaction effects between features. In addition, due to the high dimensionality and complexity of radiomics data, the Cox model is prone to overfitting when dealing with a large number of features, reducing the accuracy of prediction. At the same time, this type of model has high requirements on data quality and is sensitive to missing data. Once the data is incomplete, the model performance will drop significantly.

[0005] Deep learning models (such as convolutional neural networks, CNNs) have been widely used in medical image analysis in recent years. They can automatically extract high-level features from images, avoid the limitations of manually selected features, and have strong capabilities in processing high-dimensional data and capturing complex nonlinear relationships. This makes deep learning a powerful tool in survival prediction. However, such models also face obvious challenges. Existing deep learning models often have multiple defects or deficiencies: Insufficient model interpretability: Although deep learning models have advantages in modeling nonlinear relationships, their "black box" characteristics make the model prediction process difficult to explain, which is not conducive to clinicians' understanding and trust. Limitations of a single feature source: Most existing models only use the radiomics features of the tumor area, while ignoring the impact of the patient's overall lung condition. Lung features, such as abnormal lung tissue structure and functional damage, are closely related to the patient's prognosis, but have not been fully used in current survival prediction models. The ability of traditional models to capture complex relationships is limited: Traditional statistical methods such as the Cox proportional hazard model have difficulty capturing the nonlinear relationships and potential interaction effects between radiomics features, limiting the accuracy and application of prediction models. Data heterogeneity and lack of generalization: Many existing studies only train and validate models on a single dataset and lack validation across centers and countries, resulting in insufficient generalization capabilities of the models.

[0006] The above-mentioned record of background technical knowledge is intended to help ordinary technicians in this field understand the existing technologies that are relatively close to the present invention, and at the same time facilitate the understanding of the inventive concept and technical solution of the present invention. It should be clear that in the absence of clear evidence that the above-mentioned content has been disclosed before the filing date of this patent application, the above-mentioned background technology should not be used to evaluate the novelty of the technical solution of this application. Summary of the invention

[0007] Technical issues In response to the defects and shortcomings of the existing technology, the present invention proposes a comprehensive machine learning model that combines tumor and lung imaging features and clinical parameters to improve the accuracy and clinical practicality of overall survival (OS) prediction of lung cancer patients. Compared with most studies that only analyze tumor characteristics, the present invention takes into account the overall health status of patients more comprehensively by integrating lung characteristics with tumor characteristics. Existing studies often ignore the important impact of underlying lung diseases (such as chronic obstructive pulmonary disease and interstitial lung disease) on treatment effects and survival prognosis, and the model of the present invention can capture the complex interactions between these factors and tumor characteristics, providing more accurate support for clinical decision-making.

[0008] Technical Solution In order to achieve the above objectives, the inventors of the present application conducted in-depth research and found that the integration of tumor and lung imaging features and clinical parameters into a machine learning model can be used to accurately predict the overall survival (OS) of lung cancer patients receiving radical radiotherapy. The model not only captures complex nonlinear relationships and interaction effects, but also achieves interpretability through the SHAP framework, which helps clinicians understand the prediction process of the model and supports the formulation of personalized treatment plans.

[0009] That is, the present invention is: The first aspect is a survival prediction system for patients with non-small cell lung cancer after radiotherapy, including: Image processing module, feature extraction module, tumor imaging omics model, lung imaging omics model, comprehensive imaging omics model and final fusion model; The image processing module first pre-processes the CT image before radiotherapy, and then divides the CT image into a lung area image and a tumor area image; The feature extraction module extracts lung imaging features from lung region images, tumor imaging features from tumor region images, and clinical features from clinical parameters. The tumor imaging omics model (Model-T) is trained based on the selected important tumor imaging omics features; the important tumor imaging omics features are preferably 8-15; The lung imaging omics model (Model-L) is trained based on the selected important lung imaging omics features; the important lung imaging omics features are preferably 8-15; The comprehensive radiomics model (Model-LT) combines important tumor radiomics features with important lung radiomics features to screen out 5-15 most critical features; The final fusion model (Model-LTC) is constructed and trained to obtain a survival prediction model for patients with non-small cell lung cancer after radiotherapy based on the most critical features and clinical features screened out.

[0010] Preferably, the survival prediction system for non-small cell lung cancer patients after radiotherapy does not involve the diagnosis and treatment of any disease.

[0011] Preferably, the pathological type of the non-small cell lung cancer is non-small cell lung cancer.

[0012] Preferably, the radiotherapy is radical radiotherapy (RT) treatment.

[0013] Preferably, the tumor imaging omics model, lung imaging omics model, comprehensive imaging omics model and final fusion model are constructed using one of the XGBoost algorithm, random forest (RF), LightGBM (LGBM), support vector machine (SVM) or convolutional neural network (CNN), preferably the XGBoost algorithm. Random Forest (RF) is suitable for processing high-dimensional data and complex interactions between features, and achieves a certain degree of interpretability through feature importance analysis. LightGBM (LGBM) is similar to XGBoost and is also an algorithm based on gradient boosting. It is faster and more suitable for large-scale data, can capture nonlinear relationships between features, and provide feature importance analysis. Support vector machine (SVM) is suitable for high-dimensional data, especially in small data sets. Deep learning models (such as convolutional neural networks, CNN) can automatically extract image features and reduce the workload of manual selection.

[0014] Preferably, the preprocessing of the CT image before radiotherapy includes uniformly resampling the voxel size of the CT image to 1×1×1 mm, eliminating the influence of the difference in image resolution of different centers on feature extraction.

[0015] Preferably, segmenting the CT image into a lung region image and a tumor region image comprises: A deep learning algorithm was used to perform preliminary segmentation of the lung region, followed by manual correction by experienced radiation oncologists to ensure the accuracy of the segmentation; The tumor area is automatically extracted from the RT structure file and reviewed and confirmed by the doctor. The present invention uses a deep learning algorithm combined with manual correction to segment the lung and tumor areas. It can also be replaced by the following methods: fully automatic segmentation tools (such as 3D Slicer or ITK-SNAP), fully automated processing, reducing human intervention and improving segmentation efficiency. Traditional image processing algorithms (such as region growing, threshold segmentation, etc.) are suitable for the segmentation of simple structures and consume less computing resources.

[0016] Preferably, the lung imaging features include morphological features, intensity features, and texture features; more preferably, they include: 14 morphological features, 17 intensity features, and 73 texture features.

[0017] Preferably, the tumor imaging features include morphological features, intensity features, and texture features; more preferably, they include: 14 morphological features, 17 intensity features, and 73 texture features.

[0018] Preferably, the important tumor imaging features are screened by the following steps: removing highly correlated features (absolute value of Spearman correlation coefficient ≥ 0.8) to reduce redundancy; screening out the top 8-15, more preferably 10-15, and most preferably 10 most important features based on the SHAP value for model training.

[0019] Preferably, the important lung imaging omics features are screened by the following steps: removing highly correlated features (absolute value of Spearman correlation coefficient ≥ 0.8) to reduce redundancy; screening out the top 8-15, more preferably 10-15, and most preferably 10 most important features based on the SHAP value for model training.

[0020] Preferably, the clinical parameters include: smoking status, KPS score, and tumor stage.

[0021] Preferably, in the final fusion model, patients are divided into a high-risk group and a low-risk group based on the median risk score of the training set, and a Kaplan-Meier curve is used for survival analysis, and the high-risk group has a short survival period after radiotherapy.

[0022] Preferably, when each model is trained, the patient's prognosis survival label is also included.

[0023] Preferably, when each model is trained, Optuna is used to optimize hyperparameters, with the goal of minimizing the Cox negative log-likelihood (cox-nloglik) value, and 3-fold cross validation is used to ensure the robustness of the model. The present invention uses Optuna to optimize the hyperparameters of the model, and other optimization algorithms can also be used. Alternative optimization algorithms include: Grid Search, which can exhaustively search all possible hyperparameter combinations and is suitable for situations with fewer parameters. Random Search, which randomly selects some parameter combinations for search, is more efficient than Grid Search. Bayesian Optimization, uses Bayesian theory to guide hyperparameter selection and improve search efficiency.

[0024] The present invention proposes a comprehensive machine learning system and model that combines tumor and lung imaging features and clinical parameters to improve the accuracy and clinical practicality of overall survival (OS) prediction in patients with non-small cell lung cancer. Existing studies often only analyze tumor characteristics, while ignoring the important impact of underlying lung diseases (such as chronic obstructive pulmonary disease and interstitial lung disease) on treatment efficacy and survival prognosis. The present invention considers the overall health status of patients more comprehensively by integrating lung characteristics with tumor characteristics. The model of the present invention can capture the complex interaction between the overall lung factors of patients and tumor characteristics, providing more accurate support for clinical decision-making.

[0025] In a second aspect, the method for constructing the aforementioned survival prediction system comprises the following steps: (A) Data preparation: CT images were preprocessed into a uniform voxel size, segmented into lung region images and tumor region images, and clinical, lung, and tumor imaging features were extracted; and divided into training and validation sets; (B) The top 10 SHAP features were selected for model construction, and combined with clinical factors, extreme gradient boosting (XGB) was used for predictive modeling, including: tumor radiomics model, lung radiomics model, comprehensive radiomics model and the final fusion model; the Cox proportional hazard model was established as a baseline control; (C) The training set data were used to train the tumor imaging model, lung imaging model, comprehensive imaging model and the final fusion model, and the validation set data were used for external validation. The model performance was evaluated by consistency index (C index), time-based receiver operating characteristic curve (ROC), Kaplan-Meie curve (KM curve), calibration curve and decision curve analysis.

[0026] The third aspect is a method for predicting the survival of patients with non-small cell lung cancer after radiotherapy, which is not related to the diagnosis and treatment of the disease, including: Based on the survival prediction system for small cell lung cancer patients after radiotherapy described in the first aspect, the pre-chemotherapy CT images and clinical parameters of the non-small cell lung cancer patients to be predicted are input into the system, and the patient's survival prediction results are output. The survival prediction method for non-small cell lung cancer patients after radiotherapy does not involve the diagnosis and treatment of any disease.

[0027] The fourth aspect is a device for predicting the survival of patients with non-small cell lung cancer after radiotherapy, comprising: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method for predicting the survival period of patients with non-small cell lung cancer after radiotherapy as described in the third aspect above.

[0028] In a fifth aspect, a computer-readable storage medium stores a program executable by a processor, wherein the program executable by the processor is used to execute the method for predicting the survival period of patients with non-small cell lung cancer after radiotherapy described in the third aspect.

[0029] Based on the common knowledge in this field, the above-mentioned preferred conditions can be combined with each other to obtain a specific implementation method.

[0030] Beneficial Effects The model of the present invention achieves the following technical effects by combining the radiomic features of tumors and lungs with clinical parameters: Improved prediction accuracy: Compared with the traditional Cox model, the predictive ability of Model-LTC is significantly improved. Enhanced model interpretability: With the help of the SHAP framework, clinicians can understand the predictive logic of the model and get support in formulating personalized treatment plans. Wide applicability: The model has been validated on multi-national data sets and shows good generalization ability and practical application value. Data fusion and feature interaction: The model captures the complex relationship between lung and tumor characteristics, providing new ideas for future model optimization and multimodal data fusion.

[0031] The present invention has been validated on multinational and multicenter datasets, demonstrating its good generalization and adaptability. Due to differences in imaging equipment and data formats used by different countries and institutions, the performance of traditional models is often unstable when applied across centers, while the model of the present invention has achieved good performance on datasets from China, the Netherlands, and the United States, ensuring its wide applicability in different medical environments. This advantage enables the model of the present invention to not only support the formulation of individualized treatment plans, but also to be stably applied in a variety of clinical settings, providing support for long-term follow-up and treatment adjustments for lung cancer patients.

[0032] By fusing multimodal data and providing explanatory analysis, the present invention makes up for the problems of single perspective, limited assumptions and lack of explanatory power in the prior art. The application of the model not only helps doctors understand the health status of patients more comprehensively, but also provides data support for personalized treatment. In this way, doctors can flexibly adjust treatment strategies according to the specific conditions of patients, improve treatment effects, and minimize unnecessary side effects.

[0033] The present invention adopts the above technical solution to achieve the above purpose, which makes up for the shortcomings of the prior art and has reasonable design and convenient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to make the above and / or other purposes, features, advantages and examples of the present invention more obvious and easy to understand, the drawings required for use in the specific implementation of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 An overall flow chart showing a method for predicting the survival period of patients with non-small cell lung cancer after radiotherapy in some embodiments; Figure 2 Indicates the impact of SHAP value on the output of model-L model; Figure 3 Indicates the impact of SHAP value on the output of model-T model; Figure 4 Indicates the impact of SHAP value on the output of MODEL-LT model; Figure 5 Indicates the impact of SHAP value on the output of MODEL-LTC model; Figure 6 Represents the ROC curves of Model-L and Model-T in the training set and validation set (where A1 represents Model-L in the training set; A2 represents Model-L in validation set 1; A3 represents Model-L in validation set 2; B1 represents Model-T in the training set; B2 represents Model-T in validation set 1; B3 represents Model-T in validation set 2); Figure 7 Represents the ROC curves of Model-LT and Model-LTC in the training set and validation set (where C1 represents Model-LT in the training set; C2 represents Model-LT in validation set 1; C3 represents Model-LT in validation set 2; D1 represents Model-LTC in the training set; D2 represents Model-LTC in validation set 2); Figure 8 KM analysis curve graph representing the training set; Fig. 9 KM analysis curve diagram representing validation set 2; Fig.10 A structural block diagram showing a system for predicting survival period of non-small cell lung cancer patients after radiotherapy provided by an embodiment of the present invention; Fig.11 A structural block diagram showing a device for predicting survival period of non-small cell lung cancer patients after radiotherapy provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] Those skilled in the art can refer to the content of this article and appropriately replace and / or modify the process parameters to achieve the same. However, it should be particularly noted that all similar replacements and / or modifications are obvious to those skilled in the art and are considered to be included in the present invention. The products and preparation methods described in the present invention have been described through preferred examples. It is obvious that relevant personnel can modify or appropriately change and combine the products and preparation methods described herein without departing from the content, spirit and scope of the present invention to implement and apply the technology of the present invention.

[0037] Unless otherwise defined, technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention belongs. The present invention uses the methods and materials described herein; however, other suitable methods and materials known in the art may also be used. The materials, methods, and examples described herein are illustrative only and are not intended to be limiting. All publications, patent applications, patents, provisional applications, database entries, and other references mentioned herein are incorporated herein by reference in their entirety. In the event of a conflict, the present specification, including definitions, shall prevail.

[0038] Unless specifically stated, the materials, methods, and examples described herein are illustrative only and not limiting. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, suitable methods and materials are still described herein.

[0039] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. At the same time, the embodiments in this application and the features in the embodiments can be combined with each other without conflict.

[0040] It should be understood that any technical solution claimed for protection in the present invention does not involve the diagnosis and treatment of diseases.

[0041] In order to facilitate understanding of the embodiments of the present invention, the abbreviations and key terms that may be involved in the embodiments of the present invention are first explained or defined. For undefined abbreviations or key terms, they are commonly understood by those skilled in the art.

[0042] In addition, the test methods used in the examples are conventional methods unless otherwise specified; the materials, reagents, etc. used are all commercially available unless otherwise specified. The reagents or instruments used without indicating the manufacturer are all conventional products that can be obtained commercially. All the publications and other references mentioned in this article are incorporated herein by reference in their entirety.

[0043] The present invention is described in detail below Embodiment 1: Reference Figure 1 As shown in the overall flow chart of the project, this embodiment provides a method for predicting the survival period of patients with non-small cell lung cancer after radiotherapy without involving the diagnosis and treatment of the disease, which specifically includes the following steps.

[0044] A. Data Preparation The three datasets were integrated; images were normalized to a uniform voxel size; lung and tumor regions were delineated to define regions of interest (ROIs); and clinical, lung, and tumor radiomic features were extracted. The overall patient characteristics are shown in Table 1.

[0045] Table 1. Overall patient characteristics

[0046] *: The differences in features among the training set, validation set 1, and validation set 2 were assessed using Fisher's exact test (categorical variables) and Kruskal-Wallis H test (continuous variables).

[0047] a.1. Data source: Three datasets from China, the Netherlands, and the United States, containing 661 lung cancer patients who received radical radiotherapy (RT). The patients’ imaging data included CT images before radiotherapy and were segmented according to ROIs (lung and tumor regions).

[0048] a.2. Image preprocessing: The voxel size of the CT image is uniformly resampled to 1×1×1mm to eliminate the impact of the difference in image resolution at different centers on feature extraction.

[0049] a.3. ROI segmentation: A deep learning algorithm is used to perform preliminary segmentation of the lung area, followed by manual correction by experienced radiation oncologists to ensure the accuracy of the segmentation. The tumor area is automatically extracted from the RT structure file and reviewed and confirmed by the doctor.

[0050] a.4. Radiomics feature extraction: 104 radiomics features were extracted from the lung and tumor regions respectively using the PyRadiomics Python package. These features included 14 morphological features, 17 intensity features, and 73 texture features.

[0051] B. Model construction and training process The top 10 SHAP features were selected for lung and / or tumor models and combined with clinical factors to perform predictive modeling using eXtreme Gradient Boosting (XGB). A Cox proportional hazard model was established as a baseline control.

[0052] b.1. Feature selection and filtering: Remove highly correlated features (Spearman correlation coefficient absolute value ≥ 0.8) to reduce redundancy. The top 10 most important features were selected based on the SHAP value for model training. After image preprocessing, ROI delineation, and removal of highly collinear features, a total of 25 GTV-based features and 27 lung-based features were retained for analysis. The top ten features of Model-L, Model-T, and Model-LT were distinguished by SHAP values. The results of feature screening by SHAP values ​​are shown in Figure 2. Figure 2-Figure 5 , features are arranged vertically by importance, with the most important features at the top, descending in order to the least important features, and each data point represents a patient. The SHAP value on the X-axis quantifies the impact of the feature on the model output, reflecting the risk of death for each patient, where values ​​greater than zero have a negative impact on survival (poor prognosis) and values ​​less than zero have a positive impact on survival (good prognosis). Feature values ​​are color-coded, transitioning from blue to red, representing the trend of feature values ​​from low to high. By using SHAP, the model can show the specific contribution of different features in the prediction results. The SHAP value quantifies the impact of each feature on the model output and distinguishes the positive or negative effects of the feature in different situations, while revealing the nonlinear relationship between the feature and the predicted results. In addition, SHAP can capture the interactions between features and show how they jointly affect the model's predictions. By Figure 2-Figure 5 It can be seen that Model-L combines 4 morphological features, 3 intensity features, and 3 texture features. Model-T mainly selects texture features and only includes one intensity feature. Model-LT has a bias towards tumor features, with 6 of the 10 features coming from tumors, while lung features occupy a medium importance range. The "Maximum2DDiameterSlice-tumor" feature is related to T stage, ranking fourth in Model-T and seventh in Model-LT, with significant prognostic relevance.

[0053] b.2. Model construction: b.2.1. Tumor imaging genomics model Model-T: Based on the characteristics of tumor imaging genomics, 10 important features are selected for training.

[0054] b.2.2. Lung imaging genomics model Model-L: Based on the characteristics of lung imaging genomics, 10 important features are selected for training.

[0055] b.2.3. Comprehensive imaging omics model Model-LT: Integrate the 20 important features in Model-T and Model-L to select the 10 most critical features.

[0056] b.2.4. Final fusion model Model-LTC: The imaging features of Model-LT and clinical parameters (such as smoking history, KPS score, stage, etc.) are integrated to construct the final model.

[0057] b.3. Model training and optimization: Optuna is used to optimize hyperparameters, with the goal of minimizing the Cox negative log-likelihood (cox-nloglik) value. Three-fold cross validation is used to ensure the robustness of the model.

[0058] C. Model Validation and Interpretive Analysis Lung- and / or tumor-based models were externally validated using two validation sets, and the ensemble model was specifically validated using validation set 2. Model performance was evaluated by consistency index (C-index), time-based receiver operating characteristic curve (ROC curve), Kaplan-Meier curve, calibration curve, and decision curve analysis. Specifically, model validation used validation set 1 and validation set 2 as external datasets, and Model-LTC was validated in validation set 2. Two main performance indicators were evaluated for each model: consistency was evaluated by Harrell consistency index (C-index), and stratification was determined by dividing the dataset into high-risk and low-risk groups. To stratify patients, the median risk score of the XGBoost algorithm in the training set was used as the criterion for dividing the validation set into high-risk and low-risk groups. In addition, we evaluated the time-based receiver operating characteristic curve (ROC curve) by month to determine the performance of the model over time. The survival endpoint was dichotomized into a two-year survival prognosis, indicating whether the patient survived more than two years after radiotherapy.

[0059] The SHAP method was used to interpret the model and analyze feature importance, nonlinear relationships, and interactions between features to reveal the predictive dynamics of the model.

[0060] All statistical analyses were performed using SPSS 27 and R software (version 4.3.1). The construction of the prediction model, calculation of SHAP values, and generation of the corresponding visualizations were all done in Python (version 3.7.0), using the XGBoost package (version 1.7.3) and the SHAP package (version 0.41.0). The exact Fisher test was used to assess differences in categorical variables between patients in the training and validation sets, and the Kruskal-Wallis H test was used to assess differences in continuous variables. Comparisons between validation sets were performed using the chi-square test or the exact Fisher test for categorical variables and the Student's t test or the Mann-Whitney U test for continuous variables, all two-sided tests, and the significance level was set at a p value of <0.05. The significance of clinical factors in the training set was determined by univariate and multivariate Cox regression analysis, and the p value threshold for significance in model construction was <0.05. The discriminatory ability of the model was assessed by the C index. The survival rate was calculated using the Kaplan-Meier (KM) method, and the difference in survival distribution between the high-risk group and the low-risk group was tested by the log-rank test.

[0061] c.1. Performance evaluation: The prediction ability of the model was evaluated using Harrell's C index (C index) and ROC curve. The C index of each model in the training set and validation set 1 and validation set 2 is shown in Table 2. The ROC curves of Model-L and Model-T in the training set and validation set are shown in Table 2. Figure 6 As shown in Figure 2, the ROC curves of Model-LT and Model-LTC in the training set and validation set are as follows: Figure 7 As shown in the figure. The performance indicators of Model-L and Model-T are comparable, while Model-LT shows stronger discrimination ability. Model-LTC outperforms other models, with C index and AUC both exceeding 0.75, reflecting its strong ability to distinguish the prognosis of patients after treatment.

[0062] Patients were divided into high-risk and low-risk groups based on the median risk score of the training set, and survival analysis was performed using Kaplan-Meier curves. Figure 8 The KM survival curve of the training set is shown. Fig. 9The KM survival curves of validation set 2 are shown. Each chart divides patients into high-risk groups (red) and low-risk groups (blue). The horizontal axis represents the time after treatment and the vertical axis represents the survival probability. The Kaplan-Meier (KM) survival curve plays an important role in model prediction. It is used to estimate the survival probability at different time points and intuitively display the survival trends and differences of patients in different risk groups. The KM curve can be used to evaluate the model's ability to distinguish between predicting patient survival risks. The significant survival rate differences between risk groups indicate that the model effectively identifies high-risk and low-risk patients.

[0063] Table 2. C index of each model training set and validation set 1 and validation set 2

[0064] c.2. Model interpretability analysis: The SHAP framework is used to provide model interpretability, and the relationship and interaction effects between features are displayed through dependency graphs and interaction graphs. The importance of features such as "Maximum2DDiameterSlice-tumor" and "LargeDependenceLowGrayLevelEmphasis-tumor" in different models is revealed by the SHAP values.

[0065] In the present invention, we developed an interpretable model to predict the overall survival (OS) of lung cancer patients, especially non-small cell lung cancer, after radiotherapy by integrating radiomic features extracted from lung tissues and tumors and clinical parameter features. The SHAP method played a key role in revealing the importance of these features and their interactions in OS prediction. We validated the robustness of the model using two external datasets from different countries and evaluated its performance from multiple dimensions, including discrimination ability, stratification, calibration, and clinical utility. Our results highlight the importance of considering both tumor and lung radiomics in prognostic modeling. Especially in patients receiving radical radiotherapy, the relationship between radiomic features and clinical factors exhibits nonlinearity and complex interaction effects, which challenges traditional linear modeling methods. Through comparative analysis, our proposed model outperforms the traditional Cox regression model in predictive performance.

[0066] To our knowledge, this is the first study to integrate lung and tumor radiomic features using an explainable artificial intelligence framework to predict the prognosis of lung cancer patients receiving radical radiotherapy. In addition, this is the largest cohort of lung cancer patients to date using radiomics for OS prediction. Our analysis showed that the prognostic significance of lung radiomics was slightly lower than that of tumor radiomics, but still valuable. The predictive ability of Model-L was similar to that of Model-T. In addition, some lung-based features were more important than some tumor features. Therefore, combining lung and tumor radiomics is expected to improve prognostic accuracy, support clinical application, and provide valuable directions for future model development.

[0067] The generalization ability of the model is a key element of its practicality. Therefore, this study included two external test sets with significant heterogeneity, not only in terms of differences in clinical parameters (see Table 1), but also in terms of significant racial differences. The training set consisted of only Chinese patients, while the validation set 2 was mainly Caucasian (84.3%), and lung cancer susceptibility genes may differ between Chinese and Caucasian patients. The heterogeneity of the data also extends to the technical level, such as the diversity of CT machines used in different centers. The validation set 2 was derived from a multicenter prospective trial and therefore presented data from different CT platforms. Despite these multi-level differences, the model we developed maintained its effectiveness in prognostic prediction and showed good generalization.

[0068] The present invention integrates tumor and lung imaging features and clinical parameters to construct an explanatory machine learning model, which achieves significant improvements in multiple key indicators compared to existing technologies. These advantages are not only reflected in the prediction performance, but also in the transparency, applicability and generalization ability of the model. The details are as follows.

[0069] 1. Improve prediction accuracy and differentiation ability This paper uses the eXtreme Gradient Boosting (XGBoost) algorithm to capture the complex nonlinear relationships and interaction effects between radiomics features, which effectively improves the prediction accuracy of the model compared to the traditional Cox proportional hazard model. The C index on the training set is 0.87, and the C index on the validation set 2 reaches 0.76, while the C index of the traditional Cox model is usually between 0.6-0.7.

[0070] The XGBoost algorithm can handle nonlinear relationships between features and improve model performance through integrated learning, while the Cox model can only capture linear relationships, limiting its ability to model complex data. The joint analysis of lung features and tumor features provides a more comprehensive assessment of the patient's condition, increasing the accuracy and robustness of the prediction compared to a single tumor feature model.

[0071] 2. Make the model interpretable Compared with the "black box" model that relies on deep learning, the present invention provides interpretability through the SHAP framework, allowing clinicians to intuitively understand the prediction process of the model. The SHAP value assigns a specific contribution value to each feature, reveals which features are most important in the model prediction, and displays the interaction between features. Dependency graphs and interaction graphs reveal the nonlinear effects of "Maximum2DDiameterSlice-tumor" on survival time prediction, helping doctors understand its prediction logic.

[0072] Through the interpretability of the SHAP framework, doctors can make treatment decisions more confidently based on the personalized feature reports of each patient, improving clinical operability. Explanatory analysis not only improves the transparency of the model, but also enhances doctors' trust and acceptance of the model's prediction results.

[0073] 3. Multimodal data fusion to improve the comprehensiveness of prediction The present invention integrates tumor imaging, lung imaging and clinical parameters for analysis. This integration of multimodal data can provide more comprehensive predictions than a single data source model. By comparing the performance of Model-T (tumor feature model), Model-L (lung feature model) and Model-LT (comprehensive imaging feature model), it was found that Model-LTC (comprehensive model) had the best C index (0.76) and AUC (>0.75). Feature interaction analysis showed that certain lung features were more important than certain tumor features in predicting patient survival.

[0074] Analyzing tumor characteristics alone ignores the importance of the overall health of the lungs, and lung characteristics (such as the extent of chronic obstructive pulmonary disease or interstitial lung disease) directly affect the patient's survival time. The fusion of multimodal data allows the model to not only focus on tumor characteristics, but also comprehensively consider the patient's overall health status, providing comprehensive support for personalized treatment.

[0075] 4. The model has strong generalization ability and wide applicability The present invention has been validated using multicenter datasets from China, the Netherlands, and the United States. Compared with models validated only in a single center, the model of the present invention has stronger generalization capabilities. The model performance was validated on datasets from different countries and medical institutions, and the C index and AUC of both the training set and the validation set remained at a high level without obvious overfitting. 84.3% of the patients in validation set 2 were Caucasians, while the training set was Chinese patients, but the model still showed stable predictive ability among different races.

[0076] Multinational data verification ensures that the model is applicable to different races and medical institutions, improving the practical application value of the model. This generalization ability is particularly important for medical AI models, ensuring that the model can work stably in a variety of clinical scenarios.

[0077] 5. Reduce data redundancy and improve computing efficiency The present invention uses SHAP value for feature screening and removes redundant features to make the model more computationally efficient. Specifically, after removing highly correlated features (|Spearman coefficient|≥0.8), the top 10 important features of each model are selected for training. Feature dimensionality reduction reduces the complexity of the model and speeds up the training speed. After removing redundant features, the training speed and reasoning speed of the model are optimized, which is crucial for real-time clinical applications. Feature screening enables the model to retain only the most informative features, improving the accuracy and stability of the prediction.

[0078] 6. Personalized risk reporting supports clinical decision making The model generates a personalized risk report based on the characteristics of each patient, helping doctors develop personalized treatment plans. Patients are divided into high-risk and low-risk groups based on the median risk score predicted by the model. The Kaplan-Meier survival curve shows the difference in survival between different risk groups, providing doctors with a basis for treatment. Personalized reports can help doctors develop more accurate treatment plans based on the specific characteristics of each patient and improve treatment outcomes. This makes the model more than just a tool, but a part of the clinical workflow.

[0079] Embodiment 2: Reference Fig.10 The embodiment of the present invention further provides a system for predicting the survival period of patients with non-small cell lung cancer after radiotherapy, comprising: The image processing module first pre-processes the CT image before radiotherapy, and then divides the CT image into lung area image and tumor area image; The feature extraction module extracts lung imaging features from lung region images, tumor imaging features from tumor region images, and clinical features from clinical parameters; The tumor imaging model (Model-T) was trained based on 10 important tumor imaging features; The lung radiomics model (Model-L) was trained based on 10 important lung radiomics features selected; The comprehensive radiomics model (Model-LT) combines 10 important tumor radiomics features with 10 important lung radiomics features to screen out the 10 most critical features; The final fusion model (Model-LTC) was constructed and trained to obtain a survival prediction model for patients with non-small cell lung cancer after radiotherapy based on the 10 most critical features and clinical characteristics screened out.

[0080] Embodiment 3: Reference Fig.11 The embodiment of the present invention further provides a device for predicting the survival period of patients with non-small cell lung cancer after radiotherapy, comprising: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements at least one step of the method for predicting survival of patients with non-small cell lung cancer after radiotherapy that does not involve diagnosis and treatment of the disease as described in the aforementioned embodiment 1, and can achieve the same technical effect. To avoid repetition, this embodiment will not be repeated.

[0081] Embodiment 4: An embodiment of the present invention further provides a computer-readable storage medium, in which a computer program executable by a processor is stored. When the computer program is executed by the processor, at least one step of the method for predicting survival of patients with non-small cell lung cancer after radiotherapy that does not involve disease diagnosis and treatment as described in the aforementioned embodiment 1 is run, and the same technical effect can be achieved. To avoid repetition, this embodiment will not be described in detail.

[0082] Computer readable media include permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PR AM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0083] The conventional techniques in the above embodiments are prior arts known to those skilled in the art, and thus will not be described in detail here.

[0084] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0085] Although the present invention has been described in detail and some specific embodiments have been cited, it is obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit and scope of the present invention.

[0086] Although the above-mentioned specific embodiments have been shown, described and pointed out the novel features applied to various embodiments, it should be understood that various omissions, substitutions and changes can be made to the form and details of the described device or method without departing from the spirit of the present disclosure. In addition, the above-mentioned various features and methods can be used independently of each other, or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of the present disclosure. Many of the above-mentioned embodiments include similar components, and therefore, these similar components are interchangeable in different embodiments. Although the present invention has been disclosed in the context of certain embodiments and examples, it should be understood by those skilled in the art that the present invention can extend beyond the specifically disclosed embodiments to other alternative embodiments and / or applications and obvious modifications and equivalents thereof. Therefore, the present invention is not intended to be limited by the specific disclosure of the preferred embodiments herein.

[0087] Matters not covered in the present invention are all known technologies.

Claims

1. A survival prediction system for patients with non-small cell lung cancer after radiotherapy, characterized in that include: Image processing module, feature extraction module, tumor imaging omics model, lung imaging omics model, comprehensive imaging omics model and final fusion model; The image processing module first pre-processes the CT image before radiotherapy, and then divides the CT image into a lung area image and a tumor area image; The feature extraction module extracts lung imaging features from lung region images, tumor imaging features from tumor region images, and clinical features from clinical parameters. The tumor imaging omics model is trained based on the screened important tumor imaging omics features; The lung radiomics model is trained based on the screened important lung radiomics features; The comprehensive radiomics model combines important tumor radiomics features with important lung radiomics features to screen out 5-15 most critical features; The final fusion model is based on the most critical features and clinical features screened out to construct and train a survival prediction model for patients with non-small cell lung cancer after radiotherapy.

2. The system according to claim 1, characterized in that: The tumor imaging omics model, lung imaging omics model, comprehensive imaging omics model and final fusion model are constructed using one of the XGBoost algorithm, random forest, LightGBM, support vector machine or convolutional neural network.

3. The system according to claim 1, characterized in that: The step of segmenting the CT image into a lung region image and a tumor region image comprises: A deep learning algorithm was used to perform preliminary segmentation of the lung region, followed by manual correction by experienced radiation oncologists to ensure the accuracy of the segmentation; The tumor area is automatically extracted from the RT structure file and reviewed and confirmed by the doctor.

4. The system according to claim 1, characterized in that: The lung imaging features include morphological features, intensity features, and texture features; and / or The tumor imaging features include morphological features, intensity features, and texture features.

5. The system according to claim 1, characterized in that: In the final fusion model, patients were divided into high-risk group and low-risk group based on the median risk score of the training set, and survival analysis was performed using the Kaplan-Meier curve. The high-risk group had a short survival period after radiotherapy.

6. The system according to claim 1, characterized in that: When each model is trained, Optuna is used for hyperparameter optimization, with the goal of minimizing the Cox negative log-likelihood value, and 3-fold cross-validation is used to ensure the robustness of the model.

7. The method for constructing a system for predicting survival of patients with non-small cell lung cancer after radiotherapy according to any one of claims 1 to 6, comprising the following steps: (A) Data preparation: CT images were preprocessed into a uniform voxel size, segmented into lung region images and tumor region images, and clinical, lung, and tumor imaging features were extracted; and divided into training and validation sets; (B) The top 10 SHAP features were selected for model construction and combined with clinical factors to perform predictive modeling using XGBoost, including: Tumor radiomics model, lung radiomics model, comprehensive radiomics model and final fusion model; Cox proportional hazards model was established as a baseline control; (C) The training set data were used to train the tumor imaging model, lung imaging model, comprehensive imaging model and the final fusion model, and the validation set data were used for external validation. The model performance was evaluated by consistency index, time-based receiver operating characteristic curve, Kaplan-Meie curve, calibration curve and decision curve analysis.

8. A method for predicting the survival of patients with non-small cell lung cancer after radiotherapy for non-disease diagnosis and treatment, characterized in that include: Based on the survival prediction system for small cell lung cancer patients after radiotherapy as described in any one of claims 1-6, the pre-chemotherapy CT images and clinical parameters of the non-small cell lung cancer patient to be predicted are input into the system, and the patient's survival prediction result is output.

9. A device for predicting the survival of patients with non-small cell lung cancer after radiotherapy, characterized in that include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method for predicting the survival period of non-small cell lung cancer patients after radiotherapy for non-disease diagnosis and treatment as described in claim 8.

10. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor executable program is used to execute the method for predicting the survival period of non-small cell lung cancer patients after radiotherapy for non-disease diagnosis and treatment as claimed in claim 8 when executed by the processor.

Images