Pathological stage analysis method, device and equipment for skeleton diseases and storage medium
By constructing machine learning models and analytical methods based on bone muscle characteristics, the limitations of early diagnosis and individualized treatment of osteoporosis are solved, and more comprehensive and accurate analysis and prediction of bone diseases are achieved.
Patent Information
- Application Number
- CN202510126101.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-27
- Publication Date
- 2025-05-13
AI Technical Summary
The prior art has limitations in the early diagnosis and individualized treatment strategies of osteoporosis, and relying solely on bone density assessment cannot fully capture the complex interactions between bone and muscle.
By constructing a machine learning model based on bone muscle characteristics, combining SHAP value analysis and comprehensive bone muscle interaction network, multi-dimensional analysis and prediction of the development stage of bone disease are achieved.
This method can more accurately predict the early stages of osteoporosis and provide individualized treatment strategies to improve diagnosis accuracy and treatment targeting.
Smart Images
Figure CN119990212A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of machine learning technology, and in particular to a method, device, equipment and storage medium for analyzing the pathological stages of bone diseases. Background Art
[0002] Osteoporosis and its most serious complication, osteoporotic fractures, have become a global public health challenge, especially in the context of an aging global population. Global statistics show that about one-third of women over 50 and one-fifth of men over 50 will suffer osteoporotic fractures in their lifetime. These fractures not only lead to long-term disability and loss of self-care ability, but also significantly reduce the quality of life of patients. At the same time, the high costs of acute fracture treatment, long-term care and rehabilitation also impose a heavy economic burden on patients' families and the medical system. Among them, the core pathological characteristics of osteoporosis are the reduction of bone mineral density (BMD) and the degeneration of bone microarchitecture. These degenerations weaken the mechanical strength of bones, making them more susceptible to fractures. Although the decline in BMD is an important indicator for predicting fracture risk, more and more studies have found that the complex interaction between bones and muscles is also crucial in the occurrence and development of the disease. In recent years, researchers have gradually realized that muscle atrophy, fat infiltration, and the coordinated degeneration of bones and muscles are particularly significant in the late stages of the disease. Therefore, relying solely on BMD for risk assessment has limitations, and ignoring changes in the muscle system may lead to insufficient risk prediction. Therefore, how to conduct a more comprehensive pathological analysis of bone diseases needs to be addressed. Summary of the invention
[0003] In view of this, the purpose of the present invention is to provide a method, device, equipment and storage medium for analyzing the pathological stage of bone diseases, which can provide important theoretical support for the early diagnosis and individualized treatment strategy of osteoporosis. The specific scheme is as follows:
[0004] In a first aspect, the present application discloses a method for analyzing the pathological stage of a bone disease, comprising:
[0005] Based on each bone disease development stage, a corresponding binary classification task is constructed, and based on the bone muscle characteristics to be detected and the preset model combination, a bone disease prediction model corresponding to each of the binary classification tasks is determined, so as to predict the bone health of the target patient based on the bone disease prediction model; the preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset bone muscle image interactive analysis framework;
[0006] The gradient boosting machine model in the preset bone muscle image interactive analysis framework is trained using the bone muscle features to be detected to obtain a target machine learning model, and the SHAP value of each of the bone muscle features to be detected is determined using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model, so as to determine the importance of each of the bone muscle features to be detected to bone diseases based on the SHAP value;
[0007] Based on the gradient enhanced decision tree model in the preset bone-muscle image interaction analysis framework, the SHAP interaction value between each of the bone-muscle features to be detected is determined, and a comprehensive bone-muscle interaction network is constructed based on the SHAP interaction value, so as to determine the interaction intensity between the bone-muscle features to be detected in each stage of bone disease development based on the comprehensive bone-muscle interaction network.
[0008] Optionally, the method further includes:
[0009] Obtain bone and muscle imaging examination data of a number of patients, and screen the bone and muscle imaging examination data based on preset data screening rules to obtain bone and muscle features to be detected; the bone and muscle features to be detected include vertebral feature values and muscle feature values;
[0010] Based on each of the bone disease development stages, the bone muscle features to be detected are partitioned to obtain partitioned training data corresponding to each of the bone disease development stages.
[0011] Optionally, constructing a corresponding binary classification task based on each bone disease development stage, and determining a bone disease prediction model corresponding to each binary classification task based on the bone muscle characteristics to be detected and a preset model combination, includes:
[0012] Combining any two of the bone disease development stages in each of the bone disease development stages to construct a corresponding binary classification task, and determining the target bone muscle feature training data corresponding to the binary classification task based on the partitioned training data;
[0013] Using a preset recursive feature elimination method to perform a data dimension reduction operation on the target bone muscle feature training data to obtain a first target training feature;
[0014] Based on the Stacking model in the preset musculoskeletal image interactive analysis framework, a plurality of integrated learning models are integrated to obtain an integrated learning model combination, and based on the preset model combination rule, each integrated learning model in the integrated learning model combination is combined to obtain each training model combination;
[0015] Inputting the first target training features into each of the training model combinations for training, and using a preset five-fold cross-validation algorithm to determine the bone disease prediction sub-model corresponding to the binary classification task from each of the training model combinations;
[0016] The bone disease prediction sub-models corresponding to all the binary classification tasks are determined as the target bone disease prediction model.
[0017] Optionally, the step of using the bone muscle features to be detected to train the gradient boosting machine model in the preset bone muscle image interactive analysis framework to obtain a target machine learning model, and using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model to determine the SHAP value of each of the bone muscle features to be detected, includes:
[0018] Using the gradient boosting decision tree model in the preset bone muscle image interactive analysis framework to perform data dimension reduction processing on the partitioned training data to obtain the second target training features corresponding to each of the bone disease development stages;
[0019] Using the second target training feature, the gradient boosting machine model in the preset skeletal muscle image interactive analysis framework is trained to obtain a target machine learning model;
[0020] The SHAP method in the preset bone muscle image interactive analysis framework is used to analyze the performance of the target machine learning model in the bone muscle features to be detected to obtain the SHAP value of each bone muscle feature to be detected.
[0021] Optionally, after analyzing the performance of the target machine learning model in the bone and muscle features to be detected by using the SHAP method in the preset bone and muscle image interactive analysis framework to obtain the SHAP value of each of the bone and muscle features to be detected, the method further includes:
[0022] A feature importance ranking diagram and a SHAP value distribution diagram of all the bone and muscle features to be detected are generated using the SHAP value of each of the bone and muscle features to be detected, and a feature importance ranking diagram and a SHAP value distribution diagram of all the patients are generated.
[0023] Optionally, determining the SHAP interaction value between each of the bone-muscle features to be detected based on the gradient enhanced decision tree model in the preset bone-muscle image interaction analysis framework, and constructing a comprehensive bone-muscle interaction network based on the SHAP interaction value, includes:
[0024] Determine the independent prediction result contribution value of each of the bone and muscle features to be detected by using the gradient enhancement decision tree model in the preset bone and muscle image interaction analysis framework, and determine the SHAP interaction value between each of the bone and muscle features to be detected based on the independent prediction result contribution value;
[0025] The SHAP interaction values that are greater than a preset threshold value among the SHAP interaction values are determined as target interaction network edges, and a comprehensive bone-muscle interaction network is constructed based on the target interaction network edges and each of the bone-muscle features to be detected.
[0026] Optionally, after determining the SHAP interaction values greater than a preset threshold among the SHAP interaction values as target interaction network edges, and constructing a comprehensive bone-muscle interaction network based on the target interaction network edges and each of the bone-muscle features to be detected, the method further includes:
[0027] Calculating the centrality parameters corresponding to each characteristic node in the comprehensive bone-muscle interaction network through the preset bone-muscle image interaction analysis framework, and determining the community center node in the comprehensive bone-muscle interaction network based on the centrality parameters;
[0028] The Louvain community discovery algorithm and the community center nodes are used to divide the comprehensive bone-muscle interaction network into communities to obtain each node community, so as to perform an interactivity analysis of each bone-muscle feature to be detected based on each node community.
[0029] In a second aspect, the present application discloses a device for analyzing the pathological stage of a bone disease, comprising:
[0030] A model prediction module is used to construct a corresponding binary classification task based on each bone disease development stage, and determine the bone disease prediction model corresponding to each of the binary classification tasks based on the bone muscle characteristics to be detected and a preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; the preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset bone muscle image interactive analysis framework;
[0031] A feature importance analysis module, used to train the gradient boosting machine model in the preset bone muscle image interactive analysis framework using the bone muscle features to be detected to obtain a target machine learning model, and to determine the SHAP value of each of the bone muscle features to be detected using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model, so as to determine the importance of each of the bone muscle features to be detected to bone diseases based on the SHAP value;
[0032] A feature interaction analysis module is used to determine the SHAP interaction value between each of the bone and muscle features to be detected based on the gradient enhanced decision tree model in the preset bone and muscle image interaction analysis framework, and to construct a comprehensive bone and muscle interaction network based on the SHAP interaction value, so as to determine the interaction intensity between the bone and muscle features to be detected in each stage of bone disease development based on the comprehensive bone and muscle interaction network.
[0033] In a third aspect, the present application discloses an electronic device, comprising:
[0034] Memory, used to store computer programs;
[0035] A processor is used to execute the computer program to implement the aforementioned method for analyzing the pathological stage of bone diseases.
[0036] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the aforementioned method for analyzing the pathological stage of bone diseases.
[0037] It can be seen that in the present application, a corresponding binary classification task is constructed based on each stage of bone disease development, and a bone disease prediction model corresponding to each of the binary classification tasks is determined based on the bone and muscle features to be detected and the preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; the preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset bone and muscle image interaction analysis framework; the gradient boosting machine model in the preset bone and muscle image interaction analysis framework is trained using the bone and muscle features to be detected to obtain a target machine learning model, and the SHAP value of each of the bone and muscle features to be detected is determined using the SHAP method in the preset bone and muscle image interaction analysis framework and the target machine learning model, so as to determine the importance of each of the bone and muscle features to be detected to the bone disease based on the SHAP value; the SHAP interaction value between each of the bone and muscle features to be detected is determined based on the gradient enhanced decision tree model in the preset bone and muscle image interaction analysis framework, and a comprehensive bone and muscle interaction network is constructed based on the SHAP interaction value, so as to determine the interaction intensity between the bone and muscle features to be detected in each stage of bone disease development based on the comprehensive bone and muscle interaction network. That is, by predicting bone health through constructing a bone disease prediction model, and conducting a multi-dimensional analysis of various bone-muscle characteristics in bone diseases through the SHAP method and the construction of a comprehensive bone-muscle interaction network, it can provide important theoretical support for the early diagnosis and personalized treatment strategies of osteoporosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0039] Figure 1 A flow chart of a method for analyzing the pathological stage of a bone disease disclosed in the present application;
[0040] Figure 2 A flowchart of a specific method for analyzing the pathological stage of a bone disease disclosed in the present application;
[0041] Figure 3 A schematic diagram of a preset skeletal muscle image interactive analysis framework disclosed in this application;
[0042] Figure 4 A schematic diagram of a zoning processing result of a bone and muscle feature to be detected disclosed in this application;
[0043] Figure 5 A schematic diagram of a bone muscle characteristic analysis result to be detected disclosed in this application;
[0044] Figure 6 A schematic diagram of the results of model interpretation using SHAP disclosed in this application;
[0045] Figure 7 A schematic diagram of a comprehensive bone-muscle interaction network analysis result disclosed in this application;
[0046] Figure 8 This is a schematic diagram of the structure of a pathological stage analysis device for bone diseases disclosed in the present application;
[0047] Fig. 9 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION
[0048] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0049] Although existing studies have revealed the close relationship between bones and muscles, current clinical practice still lacks efficient and systematic tools to analyze bone-muscle interactions. This deficiency limits the comprehensive understanding of osteoporosis and leads to a significant gap in fracture risk assessment. Therefore, this application will specifically introduce a pathological stage analysis method for bone diseases, which can effectively integrate the multidimensional characteristics of bones and muscles into a unified analysis framework.
[0050] See also Figure 1 As shown, the present application embodiment discloses a method for analyzing the pathological stage of a bone disease, comprising:
[0051] Step S11: Construct corresponding binary classification tasks based on each bone disease development stage, and determine the bone disease prediction model corresponding to each binary classification task based on the bone and muscle characteristics to be detected and the preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; the preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset bone and muscle image interactive analysis framework.
[0052] like Figure 2 As shown, this application builds a preset bone muscle image interaction analysis framework (BMINet), which encapsulates basic analysis and drawing codes, including visual display of data, simple data statistics, age-related analysis and drawing, PCA (principal components analysis, i.e. principal component analysis technology) and PLS-DA (Partial Least Squares Discriminant Analysis, i.e. partial least squares discriminant analysis) analysis dimensionality reduction, and single feature ROC curve (Receiver Operating Characteristic, one of the common tools for evaluating the performance of binary classification models) drawing. In addition, the BMINet framework mainly includes two parts: prediction of classification models and analysis of bone muscle interaction networks. First, the multi-classification problem is converted into a binary classification problem of two-to-two combinations, and the interface for model construction and evaluation is defined to support binary and multi-classification predictions. Users can use their own data to build models from scratch, or directly apply BMINet's trained models to analyze osteoporosis and other related disease systems. Through BMINet's bone muscle interaction network analysis framework, users can establish interaction networks for selected important bones or muscles. BMINet also integrates network centrality parameter calculation and community division modules to facilitate users to further analyze the interaction network.
[0053] In this embodiment, it also includes: obtaining bone muscle imaging examination data of several patients, and screening the bone muscle imaging examination data based on preset data screening rules to obtain bone muscle features to be detected; the bone muscle features to be detected include vertebral feature values and muscle feature values; based on each of the bone disease development stages, the bone muscle features to be detected are partitioned to obtain the partitioned training data corresponding to each of the bone disease development stages. That is, patient data examined by dual-energy X-ray absorptiometry (DXA, i.e. Dual-energy X-ray Absorptiometry), computed tomography (CT, i.e. Computed Tomography) and magnetic resonance imaging (MRI, i.e. Magnetic Resonance Imaging) are obtained. Therefore, the bone muscle features to be detected include vertebral feature CT values, MRI signal values and muscle feature CT values, MRI signal values. Further, the patient data to be examined are screened based on preset data screening rules to obtain the bone muscle features to be detected. Finally, the bone muscle features to be detected can be partitioned according to each of the bone disease development stages to obtain the partitioned training data corresponding to each of the bone disease development stages. For example, in actual practice, data on patients who underwent dual-energy X-ray absorptiometry (DXA), computed tomography (CT), and magnetic resonance imaging (MRI) examinations were collected. MRI was used to confirm whether there was a recent fracture. Inclusion criteria were: patients aged 50 years and above; the spine underwent the above three imaging examinations. Exclusion criteria included: incomplete patient information, failure to complete necessary imaging or diagnostic examinations, the presence of other diseases that affect bone density, or previous anti-osteoporosis treatment. There were 1,397 patients in the initial data, of whom 656 were excluded because they had received osteoporosis treatment. After further verification, 111 patients were excluded due to duplicate names or data errors. After a detailed reading of the imaging examination results, 186 patients were excluded again for reasons including the lack of lumbar spine CT examination (for example, only cervical or thoracic spine CT), primary pathological fractures, or image artifacts (such as motion artifacts, foreign body artifacts, etc.) that led to unreliable data measurements. Finally, 444 patients were included, including 315 females (71%) and 129 males (29%). In total, CT values of 6 vertebrae and 30 muscles were measured, and CT value data of 36 bone and muscle features were obtained. Figure 3 As shown, the data can be partitioned according to the development stage of the bone disease: normal → osteopenia → osteoporosis → osteoporotic fracture to obtain the partitioned training data corresponding to each of the development stages of the bone disease.
[0054] In the present embodiment, the corresponding binary classification task is constructed based on each bone disease development stage, and the bone disease prediction model corresponding to each binary classification task is determined based on the bone muscle features to be detected and the preset model combination, including: combining any two bone disease development stages in each of the bone disease development stages to construct the corresponding binary classification task, and determining the target bone muscle feature training data corresponding to the binary classification task based on the partitioned training data; performing data dimension reduction operation on the target bone muscle feature training data using a preset recursive feature elimination method to obtain a first target training feature; fusing several integrated learning models based on the Stacking model in the preset bone muscle image interactive analysis framework to obtain an integrated learning model combination, and combining each integrated learning model in the integrated learning model combination based on the preset model combination rule to obtain each training model combination; inputting the first target training feature into each of the training model combinations for training, and determining the bone disease prediction sub-model corresponding to the binary classification task from each of the training model combinations using a preset five-fold cross-validation algorithm; and determining the bone disease prediction sub-model corresponding to all the binary classification tasks as the target bone disease prediction model. Specifically, BMINet uses the standard dataframe (a two-dimensional tabular data structure) data structure format of Pandas (a tool for solving data analysis tasks) in Python (a widely used high-level programming language), requiring the first column of the data to represent the stage of the disease, arranged from mild to severe, such as A, B, C, D. For a disease classification problem with multiple stages, BMINet decomposes the multi-classification problem into multiple binary classification tasks. For example, for the four-class ABCD, BMINet splits it into AvsB, AvsC, AvsD, BvsC, BvsD, and CvsD. For each binary classification model, BMINet uses the recursive feature elimination (RFE) method with XGBoost (Extreme Gradient Boosting), LightGBM (Light Gradient Boosting Machine) and CatBoost (categorical boosting, i.e., decision tree based on gradient boosting) as the core to reduce the dimensionality of high-dimensional features, which will make the data redundant and interfere with the machine learning model. After the dimensionality is reduced, the features will become fewer, retaining only the important vertebral and muscle features in the development of the disease stage.Next, the selected features are input into a variety of machine learning models. BMINet builds a Stacking model and integrates the default four basic models - RandomForest, XGBoost, LightGBM and CatBoost (users can also customize model combinations) through logistic regression. The system will search these four model combinations under 5-fold cross-validation, and finally select the model combination with the best AUC, so as to find the best parameters and model combination for each binary classification task. It should be noted here that the number of models for model combination is not fixed, that is, RandomForest and XGBoost can form a model combination, or XGBoost, LightGBM and CatBoost can form a model combination, or XGBoost can be a model combination alone. Therefore, the four basic models can form 32 model combinations, and the best model combination is the best model combination determined from these 32 model combinations. In the model verification stage, BMINet provides two methods: binary classification and multi-classification. For binary classification testing, BMINet selects the data corresponding to the two labels, and then predicts and verifies the matching between the predicted value and the actual label. For multi-classification testing, BMINet adopts a stepwise approximation strategy. For data with unknown labels, first classify it into the most easily distinguishable stage, and then gradually approximate it until the most likely stage is predicted. For example, for a data to be predicted, first use the best model combination corresponding to AvsD to classify the data to be predicted as D, then use the best model combination corresponding to BvsD to classify the data to be predicted as B, and finally use the best model combination corresponding to BvsC to classify the data to be predicted as C; then the final classification prediction result of the data to be predicted is C.
[0055] Step S12: Use the bone and muscle features to be detected to train the gradient boosting machine model in the preset bone and muscle image interactive analysis framework to obtain a target machine learning model, and use the SHAP method in the preset bone and muscle image interactive analysis framework and the target machine learning model to determine the SHAP value of each of the bone and muscle features to be detected, so as to determine the importance of each of the bone and muscle features to be detected to bone diseases based on the SHAP value.
[0056] In this embodiment, for example Figure 4 When using the BMINet package, each CT feature is first quantified using the receiver operating characteristic curve (ROC) and the area under the curve (AUC) to evaluate its classification performance ( Figure 4a). For example, in the early stage, the L5 vertebra had the highest classification performance in the classification of group A (normal) and group D (osteoporotic fracture), with an AUC value of 0.921, indicating its importance in fracture risk prediction. L4 also showed high AUC values (>0.7) in the classification of A vs. B and A vs. C, so L5 and L4 features may be key biomarkers in the early stage of the disease. However, as the disease enters the late stage, the classification performance based solely on vertebral CT values decreases. At this time, some muscle features (such as L1-L2_6) performed better in distinguishing the late stage groups, with an AUC value of 0.743, indicating that muscle atrophy and changes play a more important role in the classification of the late stage of the disease.
[0057] In order to further understand the characteristic distribution of vertebral bodies and muscles, the BMINet package provides two dimensionality reduction methods: principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA) ( Figure 4 b, 4c). First, PCA unsupervised analysis was performed to explore the intrinsic structure of the data. The results showed that in the dimensionality reduction space, group A (normal group), group B (osteopenia group) and group C (osteoporosis group) had a large overlap in the PC1 direction, indicating that the difference in characteristics was small in the early stage of the disease. However, as the disease progressed, group C (osteoporosis group) and group D (fracture group) gradually separated in the PC1 direction, suggesting that the CT features related to bone and muscle degeneration gradually showed differences in the late stage.
[0058] The BMINet package provides PLS-DA as a supervised dimensionality reduction method to verify the differences in feature distribution between different disease stages. PLS-DA further reveals the changes in feature distribution at each stage by maximizing the differences between groups. The results show that PLS-DA more clearly separates the features of group C and group D in the PC1 and PC2 directions than PCA, verifying that when the disease progresses to the fracture stage, the changes in CT features of different groups are more obvious.
[0059] Through the hierarchical clustering of the maximum distance method, five significant categories can be obtained, namely, one category consisting of all vertebrae, one category consisting of all muscles No. 2 and No. 5, one category consisting of all muscles No. 3 and No. 6, one category consisting of all muscles No. 1 and No. 4 except L5_S1_1 and L5-S1_4, and one category consisting of L5_S1_1 and L5-S1_4 alone. This indicates that in the entire process of vertebral muscle degeneration, the degeneration between vertebrae has similar characteristics, and the spatially symmetrical muscle degeneration also has certain similar characteristics. In particular, the degeneration characteristics of the two symmetrical muscles L5-S1_1 and L5-S1_4 are relatively unique compared to other muscles No. 1 and No. 4. Combined with statistics, it can be found that the degree of degeneration of these two muscles is low during the development of osteoporosis. Through the above steps, the BMINet package can systematically analyze the classification performance, distribution rules and common degeneration modules of vertebral and muscle CT features at various stages of the disease, providing a powerful tool for early identification of key biomarkers and fracture risk prediction.
[0060] Then, the gradient boosting machine model in the preset bone muscle image interactive analysis framework is trained using the bone muscle features to be detected to obtain a target machine learning model, and the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model are used to determine the SHAP value of each of the bone muscle features to be detected, including: using the gradient boosting decision tree model in the preset bone muscle image interactive analysis framework to perform data dimensionality reduction processing on the partitioned training data to obtain the second target training features corresponding to each stage of bone disease development; using the second target training features to train the gradient boosting machine model in the preset bone muscle image interactive analysis framework to obtain a target machine learning model; using the SHAP method in the preset bone muscle image interactive analysis framework to analyze the performance of the target machine learning model in the bone muscle features to be detected to obtain the SHAP value of each of the bone muscle features to be detected. It should be noted here that Figure 4 In the figure, (a) is a single-feature ROC curve between each disease group to quantitatively evaluate the classification performance of the CT value feature. (b) The principal component analysis (PCA) reveals the distribution trend of the CT value features between the four groups. There is overlap between group A and group B, but as the disease progresses, the separation between group C and group D becomes increasingly obvious. (c) Partial least squares discriminant analysis (PLS-DA) further shows the degree of separation of the four groups. (d) The correlation heat map shows the significant correlation between vertebral and muscle features. The high correlation reflects the coordinated degeneration of bones and muscles during disease progression.
[0061] Specifically, BMINet uses the SHAP (SHapley Additive exPlanations) method to interpret the machine learning model to reveal the key features and medical mechanisms in the classification and prediction of osteoporosis and its related diseases. Its advantage is that it can not only generate a global feature importance ranking, but also provide an explanation for the prediction of a single sample. BMINet first uses the machine learning algorithm with GBM (boosting tree model) as the core to train the corresponding data of the best feature combination in each disease stage comparison, using 80% of the data as the training set and 20% of the data as the validation set to automatically adjust the hyperparameters to obtain a trained machine learning model. That is, the gradient boosting decision tree model in the preset bone and muscle image interactive analysis framework is used to perform data dimensionality reduction processing on the partitioned training data to obtain the best feature combination corresponding to each of the bone disease development stages to obtain the second target training features, and then the obtained second target training features are used to train the gradient boosting machine model in the preset bone and muscle image interactive analysis framework to obtain the target machine learning model. The feature importance ranking graph and SHAP value distribution graph of all the bone and muscle features to be detected are generated by the SHAP value of each of the bone and muscle features to be detected, and the feature importance ranking graph and SHAP value distribution graph of all the patients are generated. Specifically, BMINet uses the SHAP value to globally interpret the performance of the model in all samples, and generates a feature importance ranking graph and a SHAP value distribution graph. Through these graphs, you can intuitively see which vertebral and muscle CT values have an important impact on the classification prediction of osteoporosis and its related diseases. Next, BMINet uses the SHAP value to generate an explanation graph for each individual to further explore the decision-making process of the model in a single patient sample. This method can decompose the prediction of each sample into the contribution value of each feature, thereby clearly showing which features have the greatest impact in the sample. The SHAP value can accurately identify which vertebral and muscle CT values play a key role in distinguishing disease states in normal, osteopenia, osteoporosis and osteoporotic fracture stages.
[0062] In actual operation, Figure 5 As shown in the figure, BMINet first screened features through the LightGBM model and gradually eliminated variables with lower importance, thereby determining the optimal feature set for the vertebral and muscle combination. Based on these features, BMINet used the Stacking model to classify each disease group. The results showed that the distinction between group A and group D (normal and osteoporotic fracture groups) was the most significant, with an AUROC of 0.942, indicating the efficiency of feature selection and the accuracy of the classification model. At the same time, the model calibration curve ( Figure 5 b) shows that the calibration effect between group A and group D is the best, further verifying the consistency of the model in the prediction of severe diseases. ROC curve ( Figure 5c) shows the sensitivity and specificity of each classification model. AvsD and A vs C have the best classification effects, with AUROC of 0.942 and 0.909, respectively, while BvsC has a lower classification performance (AUROC = 0.708), showing the challenge of distinguishing osteopenia from osteoporosis. In addition, the histogram of the predicted scores ( Figure 5 d) reveals the differences in the distribution of prediction scores between different groups, especially in group A vs D. BMINet also uses the Precision-Recall curve ( Figure 5 e) The performance of the model in an unbalanced dataset was evaluated. It can be found that the cross-validation AUC obtained by BMINet using the Stacking model to integrate multiple features is significantly higher than the single-feature AUC. It should be noted here that Figure 5 In the figure, a) is the RFE screening process, where the red dots are the best feature combinations, b) is the calibration graph of model prediction, c) is the model ROC graph, d) is the distribution graph of cross-validation scores, and e) is the accuracy-recall graph. Then, as Figure 6 As shown, BMINet uses SHAP for model interpretation and uses LightGBM as the core algorithm for interpretation. BMINet encapsulates the model interpretation drawing method, including global interpretation methods and single sample interpretation, including feature importance in classification and the contribution of each feature in a single prediction of a sample. The SHAP value heat map results show that vertebral bodies L4 and L5 have higher importance in the comparison between the normal group and the disease group, and their downward trend drives the model to judge the sample as a disease state ( Figure 6 ac). As the disease progresses (osteopenia to osteoporotic fractures), the importance of S1 and L5 increases ( Figure 6 de). In the late stage of the disease (CvsD group), the importance of vertebral features decreased, and the model relied more on muscle features such as L1-L2_2 and L1-L2_3 to make judgments ( Figure 6 f), indicating that early vertebral changes are critical, while late muscle features are more significant. In addition to the explanation of feature importance and feature contribution, BMINet also provides a display of feature interaction patterns. The SHAP interaction effect diagram shows the relationship between features and their impact on disease prediction. For example, the interaction diagram of L5 and L4 ( Figure 6 h) shows that when the L5 value decreases, the model is more inclined to predict a severe disease state, and the decrease in the L4 value will further strengthen this trend. This phenomenon reveals the coordinated degeneration of bones and muscles during disease progression and provides a basis for disease prediction models.
[0063] Step S13: Determine the SHAP interaction value between each of the bone-muscle features to be detected based on the gradient enhanced decision tree model in the preset bone-muscle image interaction analysis framework, and construct a comprehensive bone-muscle interaction network based on the SHAP interaction value, so as to determine the interaction intensity between the bone-muscle features to be detected in each stage of bone disease development based on the comprehensive bone-muscle interaction network.
[0064] In this embodiment, BMINet can use SHAP interaction effect diagrams to quantify the mutual influence of vertebral and muscle features in different disease stages, and calculate SHAP interaction values to capture significant bone and muscle feature interaction patterns. These interaction effect diagrams show the relationship between features and their interaction variables through color coding, further revealing the dynamic changes of bone and muscle features during disease progression. Among them, the SHAP interaction value between each of the bone and muscle features to be detected is determined based on the gradient enhancement decision tree model in the preset bone and muscle image interaction analysis framework, and a comprehensive bone and muscle interaction network is constructed based on the SHAP interaction value, including: determining the independent prediction result contribution value of each of the bone and muscle features to be detected through the gradient enhancement decision tree model in the preset bone and muscle image interaction analysis framework, and determining the SHAP interaction value between each of the bone and muscle features to be detected based on the independent prediction result contribution value; determining the SHAP interaction value greater than the preset threshold in the SHAP interaction value as the target interaction network edge, and constructing a comprehensive bone and muscle interaction network based on the target interaction network edge and each of the bone and muscle features to be detected. Specifically, in order to deeply detect the complex interactions between bones and muscles, BMINet uses the boosting tree model in the machine learning model to analyze the synergy between bone and muscle CT value features. The boosting tree model can effectively capture the interaction effect between features when processing high-dimensional data. At the same time, BMINet introduces the concept of SHAP interaction value to reveal the strength of the interaction between each pair of features in model prediction. SHAP interaction value explains the role of feature combination in two steps: first, evaluate the contribution of each feature to the prediction result in an independent case, and then calculate the interaction effect between it and other features, that is, the impact of the two features on the prediction result under the joint action of the two features. BMINet provides the core of multiple models (XGBoost, LightGBM, CatBoost) to calculate the SHAP interaction value between each vertebral and muscle CT value feature, thereby determining which features have significant interactions in disease classification. By analyzing these interaction effects, BMINet can identify the synergistic pattern of vertebral and muscle features at different stages of disease development. In order to reveal the interaction between bones and muscles in the progression of osteoporosis, BMINet constructs an interaction network based on SHAP interaction value. First, by calculating the SHAP interaction value between each pair of features and setting the cutoff value, the interactions above this value are included in the edges of the interaction network. The nodes are composed of significant bone and muscle features screened by the LightGBM model.BMINet integrates the sub-networks constructed in each pair of disease group comparisons into a "Combined-Bone-Muscle Interaction Network" (Combined-BMI) to globally evaluate the role of bone-muscle interaction in disease development.
[0065] Further, in this embodiment, the centrality parameters corresponding to each characteristic node in the comprehensive bone-muscle interaction network are calculated by the preset bone-muscle image interaction analysis framework, and the community center node in the comprehensive bone-muscle interaction network is determined based on the centrality parameters; the Louvain community discovery algorithm and the community center node are used to divide the comprehensive bone-muscle interaction network into communities to obtain each node community, so as to perform the interaction analysis of each bone-muscle feature to be detected based on each node community. In order to deeply analyze the topological characteristics of the network, BMINet calculates four centrality parameters of each node in the network: betweenness centrality, closeness centrality, degree and clustering coefficient. These parameters are used to measure the importance of nodes in network communication. Betweenness centrality reflects the frequency of nodes on the shortest path, representing its importance as an information hub in the network. In order to further reveal the interaction module between bone-muscle CT value features during the progression of osteoporosis, BMINet adopts the Louvain community discovery algorithm. The algorithm optimizes community division by maximizing modularity, dividing the nodes in the interaction network into multiple closely connected subgraphs. Figure 7 As shown in Figure 2, BMINet calculated the SHAP interaction value. In this application example, the cutoff value was set to 1.5, and a global bone-muscle interaction network was constructed ( Figure 7 a). This network reveals the complex development mechanism of osteoporosis and its related diseases at different stages. The Louvain community discovery algorithm was further used to automatically identify three main communities ( Figure 7 b), each community exhibited unique interaction patterns at specific disease stages ( Figure 7c): Community 1: With L5, L4 and L1 vertebrae as the center, it mainly reflects the interaction pattern between bones. L5 performed outstandingly in the analysis of betweenness, closeness centrality and degree, indicating that it plays a key role in the early stage (the process from normal to disease state). Community 2: The S1 vertebra is the central node of this community, indicating that it plays a key role in the process from osteopenia to more serious stages. At the same time, the interaction between bones and muscles in the network gradually increased, and 8 muscle features such as L1-L2_2 also appeared as bridge nodes connecting different communities. Community 3: It is composed entirely of muscle interactions, among which L1-L2_6 and L2-L3_6 ranked high in the degree analysis, indicating that they played a core role in the process from osteoporosis to osteoporotic fractures. In this way, by calculating the betweenness centrality, closeness centrality and degree of each node ( Figure 6 d), we have further clarified the importance of each node in the network, and the results show that the L5 vertebra ranks first in all parameters, indicating its dominant position in the entire network communication; followed by the S1 vertebra. In the muscle group, L1-L2_3, L2-L3_6, and L1-L2_6 rank highest in betweenness, closeness centrality, and degree, respectively, showing their key roles in local and global interactions. It should be noted here that Figure 7 (a) is the Combined-BMI total interaction network, (b) is the adjacency matrix of the network, (c) is the three communities detected by the Louvain module partitioning method, and (d) is the centrality histogram of the network nodes.
[0066] It can be seen that in this embodiment, a corresponding binary classification task is constructed based on each stage of bone disease development, and a bone disease prediction model corresponding to each binary classification task is determined based on the bone muscle features to be detected and the preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; the preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset bone muscle image interaction analysis framework; the gradient boosting machine model in the preset bone muscle image interaction analysis framework is trained using the bone muscle features to be detected to obtain a target machine learning model, and the SHAP value of each bone muscle feature to be detected is determined using the SHAP method in the preset bone muscle image interaction analysis framework and the target machine learning model, so as to determine the importance of each bone muscle feature to be detected to the bone disease based on the SHAP value; the SHAP interaction value between each bone muscle feature to be detected is determined based on the gradient enhancement decision tree model in the preset bone muscle image interaction analysis framework, and a comprehensive bone muscle interaction network is constructed based on the SHAP interaction value, so as to determine the interaction intensity between the bone muscle features to be detected in each stage of bone disease development based on the comprehensive bone muscle interaction network. That is, by predicting bone health through constructing a bone disease prediction model, and conducting a multi-dimensional analysis of various bone-muscle characteristics in bone diseases through the SHAP method and the construction of a comprehensive bone-muscle interaction network, it can provide important theoretical support for the early diagnosis and personalized treatment strategies of osteoporosis.
[0067] refer to Figure 8 The present application also discloses a pathological stage analysis device for bone diseases, including:
[0068] The model prediction module 11 is used to construct a corresponding binary classification task based on each bone disease development stage, and determine the bone disease prediction model corresponding to each binary classification task based on the bone muscle characteristics to be detected and the preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; the preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset bone muscle image interactive analysis framework;
[0069] A feature importance analysis module 12 is used to train the gradient boosting machine model in the preset bone muscle image interactive analysis framework using the bone muscle features to be detected to obtain a target machine learning model, and to determine the SHAP value of each of the bone muscle features to be detected using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model, so as to determine the importance of each of the bone muscle features to be detected to bone diseases based on the SHAP value;
[0070] The feature interaction analysis module 13 is used to determine the SHAP interaction value between each of the bone and muscle features to be detected based on the gradient enhancement decision tree model in the preset bone and muscle image interaction analysis framework, and to construct a comprehensive bone and muscle interaction network based on the SHAP interaction value, so as to determine the interaction intensity between the bone and muscle features to be detected in each stage of bone disease development based on the comprehensive bone and muscle interaction network.
[0071] It can be seen that in this embodiment, bone health is predicted by constructing a bone disease prediction model, and a multi-dimensional analysis of various bone-muscle characteristics in bone diseases is performed through the SHAP method and the construction of a comprehensive bone-muscle interaction network, which can provide important theoretical support for the early diagnosis and personalized treatment strategies of osteoporosis.
[0072] In some specific embodiments, the bone disease pathological stage analysis device may further include:
[0073] A feature collection module is used to obtain bone and muscle imaging examination data of several patients, and to screen the bone and muscle imaging examination data based on preset data screening rules to obtain bone and muscle features to be detected; the bone and muscle features to be detected include vertebral feature values and muscle feature values;
[0074] The feature processing module is used to partition the bone and muscle features to be detected based on the development stages of each bone disease to obtain partitioned training data corresponding to each bone disease development stage.
[0075] In some specific embodiments, the model prediction module 11 may specifically include:
[0076] A binary classification task construction unit, used for combining any two bone disease development stages in each of the bone disease development stages to construct a corresponding binary classification task, and determining target bone muscle feature training data corresponding to the binary classification task based on the partitioned training data;
[0077] A data dimension reduction unit, used for performing a data dimension reduction operation on the target bone muscle feature training data using a preset recursive feature elimination method to obtain a first target training feature;
[0078] A model combination unit, used for fusing a plurality of integrated learning models based on a stacking model in a preset musculoskeletal image interactive analysis framework to obtain an integrated learning model combination, and combining each integrated learning model in the integrated learning model combination based on a preset model combination rule to obtain each training model combination;
[0079] A prediction model training unit, used for inputting the first target training features into each of the training model combinations for training, and determining a bone disease prediction sub-model corresponding to the binary classification task from each of the training model combinations using a preset five-fold cross-validation algorithm;
[0080] A prediction model determination unit is used to determine the bone disease prediction sub-models corresponding to all the binary classification tasks as the target bone disease prediction model.
[0081] In some specific embodiments, the feature importance analysis module 12 may specifically include:
[0082] A training feature determination unit, used for performing data dimension reduction processing on the partitioned training data using the gradient boosting decision tree model in the preset bone muscle image interactive analysis framework to obtain second target training features corresponding to each of the bone disease development stages;
[0083] A gradient boosting machine model training unit, used for training the gradient boosting machine model in the preset skeletal muscle image interaction analysis framework using the second target training feature to obtain a target machine learning model;
[0084] A SHAP value determination unit is used to analyze the performance of the target machine learning model in the bone and muscle features to be detected by using the SHAP method in the preset bone and muscle image interaction analysis framework to obtain the SHAP value of each bone and muscle feature to be detected.
[0085] In some specific embodiments, the bone disease pathological stage analysis device may further include:
[0086] A distribution map generating module is used to generate a feature importance ranking map and a SHAP value distribution map of all the bone and muscle features to be detected through the SHAP value of each of the bone and muscle features to be detected, and to generate a feature importance ranking map and a SHAP value distribution map of all the patients.
[0087] In some specific embodiments, the feature interaction analysis module 13 may specifically include:
[0088] A SHAP interaction value determination unit, used to determine the independent prediction result contribution value of each of the bone and muscle features to be detected through the gradient enhancement decision tree model in the preset bone and muscle image interaction analysis framework, and determine the SHAP interaction value between each of the bone and muscle features to be detected based on the independent prediction result contribution value;
[0089] The interactive network construction unit is used to determine the SHAP interaction values greater than a preset threshold among the SHAP interaction values as target interactive network edges, and to construct a comprehensive bone-muscle interactive network based on the target interactive network edges and each of the bone-muscle features to be detected.
[0090] In some specific embodiments, the bone disease pathological stage analysis device may further include:
[0091] A community center node determination module, used to calculate the centrality parameters corresponding to each characteristic node in the comprehensive bone-muscle interaction network through the preset bone-muscle image interaction analysis framework, and determine the community center node in the comprehensive bone-muscle interaction network based on the centrality parameters;
[0092] The node community construction module is used to divide the comprehensive bone-muscle interaction network into communities using the Louvain community discovery algorithm and the community center node to obtain each node community, so as to perform the interactivity analysis of each bone-muscle feature to be detected based on each node community.
[0093] Furthermore, the present application also discloses an electronic device. Fig. 9 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram cannot be regarded as any limitation on the scope of use of the present application.
[0094] Fig. 9 A schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the pathological stage analysis method of bone diseases disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.
[0095] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol it follows is any communication protocol that can be applied to the technical solution of the present application, and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs and is not specifically limited here.
[0096] In addition, the memory 22, as a carrier for storing resources, can be a read-only memory, a random access memory, a disk or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.
[0097] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program that can be used to complete the pathological stage analysis method of bone disease performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include a computer program that can be used to complete other specific tasks.
[0098] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned disclosed method for analyzing the pathological stage of a bone disease is implemented. The specific steps of the method may refer to the corresponding contents disclosed in the aforementioned embodiments, and will not be described in detail here.
[0099] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0100] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0101] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
[0102] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0103] The technical solution provided by the present application is introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technicians in this field, according to the idea of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
Claims
1. A method for analyzing the pathological stage of a bone disease, characterized in that: include: Constructing corresponding binary classification tasks based on each bone disease development stage, and determining the bone disease prediction model corresponding to each binary classification task based on the bone muscle characteristics to be detected and the preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; The preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset skeletal muscle image interactive analysis framework; The gradient boosting machine model in the preset bone muscle image interactive analysis framework is trained using the bone muscle features to be detected to obtain a target machine learning model, and the SHAP value of each of the bone muscle features to be detected is determined using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model, so as to determine the importance of each of the bone muscle features to be detected to bone diseases based on the SHAP value; Based on the gradient enhanced decision tree model in the preset bone-muscle image interaction analysis framework, the SHAP interaction value between each of the bone-muscle features to be detected is determined, and a comprehensive bone-muscle interaction network is constructed based on the SHAP interaction value, so as to determine the interaction intensity between the bone-muscle features to be detected in each stage of bone disease development based on the comprehensive bone-muscle interaction network.
2. The method for analyzing the pathological stage of bone disease according to claim 1, characterized in that: Also includes: Obtain bone and muscle imaging examination data of a number of patients, and screen the bone and muscle imaging examination data based on preset data screening rules to obtain bone and muscle features to be detected; The bone and muscle features to be detected include vertebral feature values and muscle feature values; Based on each of the bone disease development stages, the bone muscle features to be detected are partitioned to obtain partitioned training data corresponding to each of the bone disease development stages.
3. The method for analyzing the pathological stage of bone disease according to claim 2, characterized in that: The binary classification tasks corresponding to the development stages of the bone diseases are constructed, and the bone disease prediction models corresponding to the binary classification tasks are determined based on the bone and muscle features to be detected and the preset model combination, including: Combining any two of the bone disease development stages in each of the bone disease development stages to construct a corresponding binary classification task, and determining the target bone muscle feature training data corresponding to the binary classification task based on the partitioned training data; Using a preset recursive feature elimination method to perform a data dimension reduction operation on the target bone muscle feature training data to obtain a first target training feature; Based on the Stacking model in the preset musculoskeletal image interactive analysis framework, a plurality of integrated learning models are integrated to obtain an integrated learning model combination, and based on the preset model combination rule, each integrated learning model in the integrated learning model combination is combined to obtain each training model combination; Inputting the first target training features into each of the training model combinations for training, and using a preset five-fold cross-validation algorithm to determine the bone disease prediction sub-model corresponding to the binary classification task from each of the training model combinations; The bone disease prediction sub-models corresponding to all the binary classification tasks are determined as the target bone disease prediction model.
4. The method for analyzing the pathological stage of bone disease according to claim 2, characterized in that: The method of using the bone muscle features to be detected to train the gradient boosting machine model in the preset bone muscle image interactive analysis framework to obtain a target machine learning model, and using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model to determine the SHAP value of each of the bone muscle features to be detected, includes: Using the gradient boosting decision tree model in the preset bone muscle image interactive analysis framework to perform data dimension reduction processing on the partitioned training data to obtain the second target training features corresponding to each of the bone disease development stages; Using the second target training feature, the gradient boosting machine model in the preset skeletal muscle image interactive analysis framework is trained to obtain a target machine learning model; The SHAP method in the preset bone muscle image interactive analysis framework is used to analyze the performance of the target machine learning model in the bone muscle features to be detected to obtain the SHAP value of each bone muscle feature to be detected.
5. The method for analyzing the pathological stage of bone disease according to claim 4, characterized in that: After analyzing the performance of the target machine learning model in the bone and muscle features to be detected by using the SHAP method in the preset bone and muscle image interactive analysis framework to obtain the SHAP value of each of the bone and muscle features to be detected, the method further includes: A feature importance ranking diagram and a SHAP value distribution diagram of all the bone and muscle features to be detected are generated using the SHAP value of each of the bone and muscle features to be detected, and a feature importance ranking diagram and a SHAP value distribution diagram of all the patients are generated.
6. The method for analyzing the pathological stage of bone disease according to claim 2, characterized in that: The step of determining the SHAP interaction value between the bone-muscle features to be detected based on the gradient enhanced decision tree model in the preset bone-muscle image interaction analysis framework, and constructing a comprehensive bone-muscle interaction network based on the SHAP interaction value, includes: Determine the independent prediction result contribution value of each of the bone and muscle features to be detected by using the gradient enhancement decision tree model in the preset bone and muscle image interaction analysis framework, and determine the SHAP interaction value between each of the bone and muscle features to be detected based on the independent prediction result contribution value; The SHAP interaction values that are greater than a preset threshold value among the SHAP interaction values are determined as target interaction network edges, and a comprehensive bone-muscle interaction network is constructed based on the target interaction network edges and each of the bone-muscle features to be detected.
7. The method for analyzing the pathological stage of bone disease according to claim 6, characterized in that: After determining the SHAP interaction values greater than a preset threshold value among the SHAP interaction values as target interaction network edges, and constructing a comprehensive bone-muscle interaction network based on the target interaction network edges and each of the bone-muscle features to be detected, the method further includes: Calculating the centrality parameters corresponding to each characteristic node in the comprehensive bone-muscle interaction network through the preset bone-muscle image interaction analysis framework, and determining the community center node in the comprehensive bone-muscle interaction network based on the centrality parameters; The Louvain community discovery algorithm and the community center nodes are used to divide the comprehensive bone-muscle interaction network into communities to obtain each node community, so as to perform an interactivity analysis of each bone-muscle feature to be detected based on each node community.
8. A device for analyzing the pathological stage of bone disease, characterized in that: include: A model prediction module is used to construct a corresponding binary classification task based on each bone disease development stage, and determine the bone disease prediction model corresponding to each binary classification task based on the bone muscle characteristics to be detected and the preset model combination, so as to predict the bone health of the target patient based on the bone disease prediction model; The preset model combination is an integrated learning model combination obtained by fusing several integrated learning models based on the Stacking model in the preset skeletal muscle image interactive analysis framework; A feature importance analysis module, used to train the gradient boosting machine model in the preset bone muscle image interactive analysis framework using the bone muscle features to be detected to obtain a target machine learning model, and to determine the SHAP value of each of the bone muscle features to be detected using the SHAP method in the preset bone muscle image interactive analysis framework and the target machine learning model, so as to determine the importance of each of the bone muscle features to be detected to bone diseases based on the SHAP value; A feature interaction analysis module is used to determine the SHAP interaction value between each of the bone and muscle features to be detected based on the gradient enhanced decision tree model in the preset bone and muscle image interaction analysis framework, and to construct a comprehensive bone and muscle interaction network based on the SHAP interaction value, so as to determine the interaction intensity between the bone and muscle features to be detected in each stage of bone disease development based on the comprehensive bone and muscle interaction network.
9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the method for analyzing the pathological stage of a bone disease as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: Used to store a computer program, which, when executed by a processor, implements the pathological stage analysis method of a bone disease as described in any one of claims 1 to 7.