A multi-model fusion bone density value prediction method and system

Abstract

The present disclosure relates to the technical field of bone density value prediction, and proposes a multi-model fusion bone density value prediction method and system, including the following steps: acquiring clinical feature data for bone density prediction; constructing a multi-model fusion prediction model including sequentially connected base learners and meta learners, inputting the acquired features into multiple base learners for prediction and recognition to obtain primary prediction results; adaptively calculating fusion weights according to the prediction accuracy of the base learners, and fusing the primary prediction results of each base learner based on the calculated fusion weights; and transmitting the fused primary prediction results to the meta learners for prediction to obtain the prediction results of the bone density. The present disclosure designs multiple machine learning models as a multi-stage prediction architecture, and each stage of the model learns and optimizes the prediction results of the previous stage and the original input features, thereby gradually improving the accuracy of the prediction.

Description

Technical Field

[0001] This disclosure relates to the technical field of bone mineral density prediction, specifically to a method and system for predicting bone mineral density using a multi-model fusion approach. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Bone mineral density (BMD) refers to the amount of bone contained per unit area or unit volume. Currently, the academically recognized method for BMD testing is dual-energy X-ray absorptiometry (DXA), and its measured value is considered the gold standard for diagnosing osteoporosis.

[0004] In recent years, machine learning-based methods have been applied in the field of bone mineral density (BMD) prediction. These methods collect multimodal data from patients and utilize machine learning models for prediction, aiming to provide a convenient and low-cost solution. However, most current machine learning-based BMD prediction methods rely on a single model, commonly including linear regression, decision trees, support vector machines (SVMs), or simple neural networks. These models lack sufficient ability to characterize complex factors influencing BMD (such as age, gender, and lifestyle habits), resulting in unsatisfactory prediction accuracy. A single model struggles to adapt to the diverse feature distributions in the training data; it may perform well for some groups but poorly for others. Furthermore, the performance of a single model can degrade significantly when faced with noisy data or incomplete features. Summary of the Invention

[0005] To address the aforementioned issues, this disclosure proposes a multi-model fusion method and system for predicting bone mineral density values. It designs multiple machine learning models as a multi-level prediction architecture, with each level of the model learning and optimizing the prediction results and original input features from the previous stage, thereby progressively improving the accuracy of the prediction.

[0006] To achieve the above objectives, the present disclosure adopts the following technical solution:

[0007] One or more embodiments provide a method for predicting bone mineral density values ​​through multi-model fusion, comprising the following steps:

[0008] Obtain clinical characteristic data for bone mineral density prediction;

[0009] A multi-model fusion prediction model is constructed, consisting of base learners and meta-learners connected in sequence. The acquired features are input into multiple base learners for prediction and recognition to obtain a preliminary prediction result.

[0010] The fusion weights are adaptively calculated based on the prediction accuracy of the base learners, and the initial prediction results of each base learner are fused based on the calculated fusion weights.

[0011] The fused primary prediction results are transmitted to the meta-learner for prediction, resulting in a bone density prediction.

[0012] One or more embodiments provide a multi-model fusion bone mineral density prediction system, comprising:

[0013] The acquisition module is configured to acquire clinical feature data for bone mineral density prediction;

[0014] The primary prediction module is configured to construct a multi-model fusion prediction model that includes base learners and meta learners connected in sequence. The acquired features are input into multiple base learners for prediction and recognition to obtain the primary prediction result.

[0015] The fusion module is configured to adaptively calculate fusion weights based on the prediction accuracy of the base learners, and then fuse the initial prediction results of each base learner based on the calculated fusion weights.

[0016] The secondary prediction module is configured to transmit the fused primary prediction results to the meta-learner for prediction, thereby obtaining the bone density prediction result.

[0017] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the aforementioned multi-model fusion bone mineral density prediction method.

[0018] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the aforementioned multi-model fusion method for predicting bone mineral density values.

[0019] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0020] This disclosure utilizes a multi-model fusion method to fully leverage the strengths of different models, enhancing the robustness of predictions. Adaptive weight calculation reduces the impact of low-precision models on the final prediction results. A meta-learner further optimizes the fusion results, eliminating biases in the base learner's predictions and ensuring stable model performance on unknown data, ultimately providing accurate bone mineral density predictions. This method can be extended to other medical prediction tasks simply by changing the input features and the type of base learner.

[0021] The advantages of this disclosure, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0022] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute a limitation thereof.

[0023] Figure 1 This is a schematic diagram of the structure of the multi-model fusion prediction model of Embodiment 1 of this disclosure;

[0024] Figure 2 This is a schematic diagram of the prediction process of Embodiment 1 of this disclosure;

[0025] Figure 3 This is a simulation experiment result diagram of the multi-model fusion prediction model in Embodiment 1 of this disclosure; Detailed Implementation

[0026] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0028] It should be noted that the terminology used herein is for descriptive purposes only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0029] Example 1

[0030] In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 3 As shown, a multi-model fusion method for predicting bone mineral density values ​​includes the following steps:

[0031] Step 1: Obtain clinical characteristic data for bone mineral density prediction;

[0032] Step 2: Construct a multi-model fusion prediction model consisting of base learners and meta-learners connected in sequence. Input the acquired features into multiple base learners for prediction and recognition to obtain the initial prediction results.

[0033] Step 3: Adaptively calculate the fusion weights based on the prediction accuracy of the base learners, and fuse the initial prediction results of each base learner based on the calculated fusion weights;

[0034] Step 4: Transmit the fused primary prediction results to the meta-learner for prediction to obtain the bone density prediction results;

[0035] In this embodiment, leveraging the advantages of multi-model fusion technology, multiple base learners can analyze clinical feature data from different perspectives to generate preliminary prediction results. Through adaptive weight calculation, the weights of high-accuracy models are appropriately amplified to improve the overall reliability of predictions. The meta-learner further optimizes the fusion results, eliminating biases in the base learner predictions, and ultimately providing accurate bone mineral density predictions.

[0036] In step 1, data acquisition and preprocessing are performed.

[0037] The bone mineral density prediction method in this embodiment is based on multimodal clinical characteristic data. It forms a data input set by collecting individual characteristics and indicators related to bone mineral density. The specific steps are as follows:

[0038] Specifically, by using scale data and physical examination data, the clinical characteristics required for bone density prediction are obtained, including but not limited to demographic data, behavioral habit data and other clinical indicators.

[0039] Optional demographic data may include age, sex, height, weight, occupation, etc. Behavioral habits may include whether the individual smokes or drinks alcohol. Other clinical indicators may include bone mineral density-related characteristics from medical examinations, such as calcium intake and exercise frequency.

[0040] Furthermore, the acquired dataset undergoes preprocessing, including data formatting, missing value handling, and encoding; the preprocessing includes the following:

[0041] Step 1.1: Format the data. Organize the above features into a unified CSV format data file to facilitate subsequent data processing and model training. Each record includes 28 features for the individual and the target variable (T-Value of bone mineral density). The pandas library in Python can be used to import and process the CSV file.

[0042] Step 1.2: Perform missing value processing. Detect missing values ​​and determine the proportion of missing values ​​for each feature. Fill in missing values ​​with the mean or median for features with few missing values. Remove features with many missing values ​​and low information value to ensure data quality.

[0043] Step 1.3: Perform feature encoding. For the collected multimodal features, use appropriate encoding strategies to process non-numerical features. Binary feature encoding and multi-class feature encoding can be used.

[0044] Binary feature encoding uses 0 and 1 values ​​to represent binary features such as gender, smoking status, and alcohol consumption, for example:

[0045] Gender: 1 for male, 0 for female.

[0046] Does the user smoke? 1 for yes, 0 for no.

[0047] Whether you have consumed alcohol: 1 for yes, 0 for no.

[0048] Multi-class feature encoding is used for features such as occupation that have multiple categories. It employs an integer encoding method from 0 to N, where N is the number of categories.

[0049] In some embodiments, the multi-model fusion prediction model constructed in step 2 can be as follows: Figure 1 As shown, it includes a two-layer model. The first layer includes multiple base learners, and the second layer includes a meta-learner.

[0050] The first layer model setup includes a support vector machine (SVM) model, an XGBoost model, and a GBDT model set up side by side; each base learner is trained and predicted independently, and the prediction results of the first layer are shown.

[0051] The meta-learner uses a random forest model. It combines the prediction results of the first layer into a new feature vector and makes predictions to obtain the final prediction result.

[0052] A feasible technical solution is proposed to construct a support vector machine (SVR) model, optimize the objective function through kernel function and Lagrange multiplier method, and use GridSearchCV to fine-tune the kernel function type and regularization parameters.

[0053] This embodiment uses the SVM (Support Vector Machine) SVR (Support Vector Regression) model to predict bone mineral density values.

[0054] Specifically, for a dataset consisting of n samples containing m features...

[0055]

[0056] SVM uses a linear or nonlinear function to predict the output. The goal of the SVR model is to minimize the following objective function:

[0057]

[0058] Where ω is the weight vector, C is the regularization parameter used to control the model's complexity and the degree of penalty for error, and ξ...i These are slack variables, used to allow for a certain degree of prediction error. The first term of the objective function is the L2 norm of the weight vector, used to control the complexity of the model; the second term is the error term, used to measure the difference between the predicted value and the target value.

[0059] In the SVR model of this embodiment, an insensitive loss function is introduced, meaning that when the difference between the predicted value and the target value is less than or equal to ε, the prediction is considered correct, and no loss is incurred. Therefore, the optimization problem can be further transformed into:

[0060]

[0061] The constraints are:

[0062]

[0063] Where, ξ i and These represent the degree to which the deviation of the sample point above and below the hyperplane exceeds ε, respectively.

[0064] To solve this optimization problem, this embodiment further employs the Lagrange multiplier method and kernel function. Specifically, the original problem is transformed into a dual problem using the Lagrange function, and the input data is mapped to a high-dimensional space using the kernel function in order to find a better hyperplane to fit the data in the new space.

[0065] During training, the SVR model iterates through all sample points, finding support vectors (i.e., sample points located near the hyperplane or violating constraints), and uses these support vectors to determine the position and orientation of the hyperplane. Meanwhile, the regularization parameter C and the threshold ε of the insensitive loss function also significantly impact the model's performance.

[0066] To prevent overfitting, cross-validation, feature scaling, and kernel function parameter tuning can be applied to the SVR model. Finally, the trained SVR model can be used to predict new input data to obtain predicted bone density values. To evaluate the model's performance, various metrics are used to compare the differences between predicted and true values.

[0067] Furthermore, this embodiment uses the GridSearchCV class from the scikit-learn library to automatically search for the optimal parameter combination for the Support Vector Regression (SVR) model. Specifically, it includes the following steps:

[0068] Step 21.1: Define the parameter grid (parameters) to store the two key parameters of the SVR model: kernel function type and regularization parameter;

[0069] The kernel function type can include linear kernel, radial basis function kernel (rbf), and polynomial kernel (poly); the regularization parameter C varies between 0.01 and 100 and can be transformed with different step sizes.

[0070] Step 21.2: Then, a GridSearchCV object, grid_search_svr, is created. This object will train the SVR model using the specified parameter grid and evaluate the performance of each parameter combination using 10-fold cross-validation.

[0071] During training, the verbose=1 parameter is used to output information about the search process. Finally, by calling the grid_search_svr.fit method, the code trains the model using the training dataset (reg_train as features and target_reg_train as the target variable) and finds the optimal combination of parameters.

[0072] An feasible technical solution for constructing an XGBoost model involves using a greedy algorithm to optimize the objective function of the XGBoost model. During training, overfitting is prevented through strategies such as shrinkage and feature bucketing. The construction process includes the following:

[0073] First, a greedy algorithm is used to optimize the objective function of the XGBoost model;

[0074] Specifically, the XGBRegressor regression model of the XGBoost model is used for model training, for a dataset consisting of n samples with m features.

[0075]

[0076] The XGBoost model uses an additive function to predict the output:

[0077]

[0078] Where l represents the loss function, used to measure the prediction. and target y i The difference between them. The second term Ω penalizes the complexity of the model, which helps to smooth the final learned weights to avoid overfitting. Equation (5) includes the function as a parameter and is trained additively. Formally, if This is the prediction for the i-th instance at iteration t, and f needs to be added. t To minimize the following objectives:

[0079]

[0080] This means greedily adding f according to equation (5). t To optimize the model to the greatest extent possible. Second-order approximations can be used to quickly optimize the objective in general settings:

[0081]

[0082] Among them, g i and h i It relates to the first and second gradients of the loss function.

[0083] In the t-th iteration, the constant term can be removed to obtain the following simplified objective function:

[0084]

[0085] XGBoost uses a greedy algorithm to traverse all feature split points across all features, dividing the resulting objective function value by the gain of the objective function of a single leaf node. To limit tree growth, a threshold is added; splits only occur when the gain exceeds this threshold. Overfitting can be prevented by setting a maximum tree depth and stopping growth when the sum of sample weights falls below the set threshold.

[0086]

[0087] Furthermore, during the XGBoost model training process, overfitting prevention strategies are implemented, including:

[0088] 1) In each iteration of XGBoost model training, the leaf node score of each tree is multiplied by the reduction weight so that the subsequently generated trees can further optimize the model; before splitting the nodes at each level, some features are randomly selected to determine the optimal split point.

[0089] Specifically, to further prevent overfitting, the Shrinkage method is used. In each iteration, the score of each leaf node of the tree is multiplied by a reduction weight, so that the influence of each tree is not too large, leaving more room for subsequent generated trees to optimize the model. Column Subsampling is also used for layer-by-layer random sampling. Before splitting each node in the same layer, a subset of features is randomly selected, and then only this subset of features needs to be traversed to determine the optimal split point.

[0090] 2) During XGBoost model training, continuous input features are divided into multiple buckets according to preset quantiles. Samples between adjacent quantiles are in the same bucket. The optimal division is calculated by traversing each quantile, which reduces the range of feature values ​​and optimizes computational efficiency.

[0091] Specifically, the feature is binned, meaning L split points are found, and samples located between adjacent split points are grouped into one bin. When traversing the feature, only each split point needs to be traversed to calculate the optimal split, further preventing overfitting of the model.

[0092] 3) Tree growth limit: Set the maximum tree depth max_depth and sample weight threshold to prevent the tree from growing without limit. The splitting stops when the gain is less than the set threshold.

[0093] Furthermore, during the XGBoost model training process, the optimal parameters of the XGBoost model are selected through grid search and cross-validation. The parameter tuning and grid search process is as follows:

[0094] By defining a dictionary of parameters, the candidate value ranges for the three key parameters of the XGBoost model are determined: maximum tree depth (max_depth), number of trees (n_estimators), and learning rate (eta). A grid search for cross-validation is performed using GridSearchCV to find the optimal parameter combination. GridSearchCV tries all parameter combinations and evaluates the performance of each combination using cross-validation (set to 10-fold cross-validation in this example). The verbose=1 parameter is used to output information during the search process. The best parameter combination found in the grid search can be obtained using grid_search_xgb_reg.best_params_, while grid_search_xgb_reg.best_score_ displays the corresponding best cross-validation score. Using the obtained optimal parameters, a new XGBoost regressor is initialized and trained on the training set reg_train and target_reg_train using the fit method. Finally, the feature importance graph is plotted using the xgb.plot_importance function. This uses a pre-trained regressor model, specifies the ax parameter as the ax in fig,ax, and sets max_num_features=25 to display the top 25 most important features.

[0095] An feasible technical solution is proposed: the GBDT model is constructed by using a gradient boosting decision tree, and the trained GBDT model is obtained by optimizing the model by reducing the residuals.

[0096] When performing bone mineral density (BMD) regression prediction, GBDT (Gradient Boosting Decision Tree) is based on the Boosting concept in ensemble learning. It obtains the final prediction by stacking the predictions of multiple decision trees, with each decision tree being optimized based on the predictions of the previous trees. The goal of the GBDT regression model is to achieve regression prediction by continuously reducing the residuals (i.e., the difference between the predicted and actual values) accumulated during training.

[0097] Specifically, the GradientBoostingRegressor class from the scikit-learn library was chosen as the implementation of the GBDT model. Based on the characteristics of the dataset and model requirements, the parameters of the GBDT model were set, including key parameters such as n_estimators (the number of weak learners), learning_rate (the learning rate), and max_depth (the maximum depth of the decision tree). The GBDT model was trained using the training set data. During training, multiple decision trees were gradually built, and model performance was optimized through gradient boosting. After training, the GBDT model was evaluated using the test set data. The predictive performance of the model was evaluated by calculating the error between the predicted and actual values, such as mean absolute error (MAE) and root mean square error (RMSE). For the trained GBDT model, feature importance analysis was performed to understand which features have the greatest impact on bone density prediction.

[0098] An feasible technical solution is proposed: the construction of the RandomForest model is based on ensemble learning of multiple decision trees. The depth and number of trees are optimized through grid search, and the average prediction result of multiple decision trees is used as the output of the model.

[0099] In this embodiment, an ensemble learning method is used for model training. By constructing multiple decision trees and taking their average prediction results, the stability and accuracy of the model are improved.

[0100] For a dataset consisting of n samples with m features:

[0101]

[0102] The output is predicted by constructing multiple decision trees. Each decision tree is trained based on a different subset of features and training samples, and outputs a prediction. The final prediction is the average of the predictions from all decision trees.

[0103] During training, model performance can be optimized by adjusting the parameters of the ensemble learning. The `n_estimators` parameter can be adjusted to specify the number of decision trees, the `max_depth` parameter can be adjusted to limit the maximum depth of the decision trees, and other related parameters can be adjusted to further optimize the model.

[0104] After model training is complete, predictions are performed using the test set. First, a parameter dictionary is defined, specifying the range of parameter values ​​for the RandomForestRegressor model (n_estimators, max_depth, etc.). Then, GridSearchCV is used for grid search with cross-validation to find the optimal parameter combination. GridSearchCV tries all parameter combinations and evaluates the performance of each combination using cross-validation (here set to 10-fold cross-validation).

[0105] The best parameter combination found in the grid search can be obtained using `grid_search_rf_reg.best_params_`, while `grid_search_rf_reg.best_score_` displays the corresponding best cross-validation score. Using the obtained best parameters, a new `RandomForestRegressor` is initialized and trained on the training set using the `fit` method.

[0106] Finally, the trained regressor is used to make predictions on the test set, and the model's performance is evaluated. The model's accuracy can be assessed by calculating the error between the predicted results and the true values ​​(such as mean absolute error (MAE), root mean square error (RMSE), etc.). Furthermore, the `feature_importances_` attribute in the scikit-learn library can be used to view the importance of features, providing a better understanding of how the model works.

[0107] The above technical solution provides methods for constructing and improving the training process of each model of the base learner and meta learner. By optimizing the training process, the obtained model can make the prediction of bone density more accurate.

[0108] Furthermore, the training process for the multi-model fusion prediction model is as follows:

[0109] Step S1: Training Base Learners: First, train multiple different models, which are called base learners or individual learners. Each base learner is trained independently and can use different algorithms or parameters.

[0110] Step S2: Generate first-level prediction results: Use these base learners to make predictions on the training set or another dataset to obtain the prediction results for each model.

[0111] Step S3: Construct a meta-learner: Input the prediction results of the base learner as new features into a meta-learner for training to obtain the meta-learner;

[0112] The goal of a meta-learner is to obtain the final prediction result by combining the predictions of the base learners. During training, the data is first input into the base learners for training. After training, each algorithm in the base learners outputs its corresponding prediction result. These prediction results are then combined into a new feature matrix, which is then input into the secondary learner for training.

[0113] Step S4: Use the meta-learner to make a secondary prediction on the prediction result of the base learner to obtain the final prediction result; iterate the training until the set accuracy is met to obtain the trained multi-model fusion prediction model.

[0114] Optionally, the accuracy of each model's predictions can be evaluated based on the MAE (mean absolute error) and RMSE (root mean square error) metrics.

[0115] The following specific training example illustrates the above training process;

[0116] (1) Data preparation: Obtain the training set and test set. Divide the training set into 10 parts: train1, train2, train3, train4, train5, train6, train7, train8, train9, and train10 for subsequent 10-fold cross-validation.

[0117] (2) Selecting base models: Select SVM, XGBoost, and GBDT as the base models for the Stacking regression model, and then train and predict the base models (taking XGBoost as an example).

[0118] (3) Training and prediction of the base model: Train1, Train2, Train3, Train4, Train5, Train6, Train7, Train8, Train9, and Train10 are used as the validation set, and the remaining 9 are used as the training set. In each fold, the XGBoost regression model is trained using the training set, and predictions are made on the validation set to obtain the regression prediction value for that fold. After 10-fold cross-validation, the prediction values ​​of the 10 validation sets are merged vertically to obtain the prediction value matrix A1 generated by the XGBoost model on the training set (each row corresponds to an original training sample, and each column corresponds to the prediction value of a cross-validation fold. However, in regression tasks, the average or median of the predictions for each fold is usually taken as the final prediction value for that sample to reduce the number of features and avoid overfitting, and to obtain a prediction value vector A1' with the same number of samples as the original training set). A final XGBoost regression model is trained using the entire training set (or using the average / ensemble of the 10-fold model), and predictions are made on the test set to obtain the prediction value B1 on the test set.

[0119] (4) Repeat the training and prediction of the base model: Repeat the above steps for the support vector machine (SVM) and GBDT models to obtain the predicted value vectors A2' and A3' on the training set, and the predicted values ​​B2 and B3 on the test set, respectively.

[0120] (5) Constructing the meta-model: The random forest regression model is used as the meta-model. The predicted value vectors A1', A2', and A3' of the three base models on the training set are used as new features. Together with the target variable (continuous numerical value) of the original training set, they are used to train the random forest regression model as the meta-model. The feature vector of each training sample consists of the predicted values ​​of the three base models, and the target variable is the original continuous numerical value.

[0121] (6) Meta-model prediction: Using the trained random forest regression meta-model, prediction is made on the new feature vector constructed from the prediction values ​​B1, B2, and B3 of the three base models on the test set, and the final prediction value (continuous numerical value) is obtained.

[0122] In step 3, the fusion method of the primary prediction results of the base learner includes the following steps:

[0123] Step 31: Obtain the prediction accuracy of the base learner;

[0124] Specifically, the prediction accuracy can be statistically analyzed based on the test set. During the use of the model, the test set data is updated according to the set time. Based on the test results of the test set and the actual values, the prediction accuracy of each base learner is calculated.

[0125] Step 32: Use the accuracy as the fusion weight, multiply it by the primary prediction result of the corresponding base learner, and use it as the fused primary prediction result;

[0126] In this step, the initial prediction results of the base learner are adjusted according to their accuracy. This adjusts the features input to the meta learner based on the accuracy, which can significantly improve the overall prediction accuracy and robustness of the model.

[0127] Simulation experiments verified that the evaluation metrics of the fusion model in this embodiment are improved compared to the individual prediction models. For example... Figure 3 As shown, Figure 3 This is a statistical chart comparing the true and predicted values ​​of bone mineral density. The chart shows the comparison between the true and predicted values. It can be seen that the overall predicted value line of the model follows the true value line, and the prediction effect is good. However, there are some differences at certain sample points.

[0128] Table 1. Prediction Comparison Results

[0129]

[0130]

[0131] In the table, FTTransformer: a deep learning model based on Transformer, suitable for tabular data;

[0132] KNN (K-Nearest Neighbors): The K-nearest neighbor algorithm is used for classification and regression.

[0133] ResNet (Residual Network): A residual neural network is an architecture used in deep learning.

[0134] The prediction results compared with existing single models are shown in Table 1. The mean absolute error (MAE) of the fusion model is 0.80 and the root mean square error (RMSE) is 1.02. The Stacking fusion model, which uses SVM regression prediction model, XGBoost prediction model, and GBDT prediction model as primary learners and RandomForest prediction model as secondary learner, has better prediction performance than the single model.

[0135] Example 2

[0136] Based on Example 1, this example provides a multi-model fusion bone mineral density prediction system, including:

[0137] The acquisition module is configured to acquire clinical feature data for bone mineral density prediction;

[0138] The primary prediction module is configured to construct a multi-model fusion prediction model that includes base learners and meta learners connected in sequence. The acquired features are input into multiple base learners for prediction and recognition to obtain the primary prediction result.

[0139] The fusion module is configured to adaptively calculate fusion weights based on the prediction accuracy of the base learners, and then fuse the initial prediction results of each base learner based on the calculated fusion weights.

[0140] The secondary prediction module is configured to transmit the fused primary prediction results to the meta-learner for prediction, thereby obtaining the bone density prediction result.

[0141] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0142] Example 3

[0143] This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the multi-model fusion bone mineral density prediction method of Embodiment 1.

[0144] Example 4

[0145] This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the multi-model fusion bone mineral density prediction method of Embodiment 1.

[0146] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

[0147] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for predicting bone mineral density values ​​using multi-model fusion, characterized in that, Includes the following steps: Step 1: Obtain clinical characteristic data for bone mineral density prediction; clinical characteristic data includes demographic data, behavioral data, and other clinical indicators; demographic data includes age, sex, height, weight, and occupation; behavioral habits include whether one smokes and drinks alcohol; other clinical indicators include characteristics related to bone mineral density found in medical examinations. Step 2: Construct a multi-model fusion prediction model consisting of sequentially connected base learners and meta-learners. The acquired features are input into multiple base learners for prediction and recognition, yielding initial prediction results. The constructed multi-model fusion prediction model includes two layers: the first layer comprises multiple base learners, and the second layer comprises a meta-learner. The first layer model uses a combination of Support Vector Machine (SVM), XGBoost, and GBDT models. Each base learner is trained and predicts independently, yielding initial prediction results. The meta-learner uses a random forest model, combining the prediction results from the first layer to form a new feature vector for prediction, resulting in the final prediction result. The support vector machine model is constructed by optimizing the objective function through kernel functions and the Lagrange multiplier method, and the kernel function type and regularization parameters are tuned using GridSearchCV. Step 3: Adaptively calculate the fusion weights based on the prediction accuracy of the base learners, and fuse the initial prediction results of each base learner based on the calculated fusion weights. The fusion method for the initial prediction results of the base learners includes the following steps: obtaining the prediction accuracy of the base learners; the prediction accuracy is statistically analyzed based on the test set, the test set data is updated according to a set time, and the prediction accuracy of each base learner is calculated based on the test results and actual values ​​of the test set; the accuracy is used as the fusion weight, multiplied by the initial prediction result of the corresponding base learner, to obtain the fused initial prediction result. Step 4: Transmit the fused primary prediction results to the meta-learner for prediction to obtain the bone density prediction result.

2. The bone mineral density prediction method based on multi-model fusion as described in claim 1, characterized in that: The construction of the support vector machine model optimizes the objective function through kernel functions and the Lagrange multiplier method. Specifically, the Lagrange function is used to transform the original problem into a dual problem, and the kernel function is used to map the input data to a high-dimensional space in order to find a better hyperplane to fit the data in the new space.

3. The bone mineral density prediction method based on multi-model fusion as described in claim 1, characterized in that: The XGBoost model is constructed by using a greedy algorithm to optimize the objective function of the XGBoost model.

4. The bone mineral density prediction method based on multi-model fusion as described in claim 1, characterized in that: During XGBoost model training, strategies to prevent overfitting are implemented, including: In each iteration of XGBoost model training, the leaf node score of each tree is multiplied by the reduction weight, so that the subsequently generated trees can further optimize the model; before splitting the nodes at each level, some features are randomly selected to determine the optimal split point. During XGBoost model training, continuous input features are divided into multiple buckets according to preset quantiles. Samples between adjacent quantiles are grouped into the same bucket, and the optimal division is calculated by traversing each quantile. Set a maximum tree depth and a sample weight threshold to prevent the tree from growing indefinitely; stop splitting when the gain is less than the set threshold.

5. A multi-model fusion system for predicting bone mineral density values, characterized in that, include: The acquisition module is configured to acquire clinical characteristic data for bone mineral density prediction. The clinical characteristic data includes demographic data, behavioral data, and other clinical indicators. The demographic data includes age, sex, height, weight, and occupation. The behavioral data includes whether the person smokes and drinks alcohol. Other clinical indicators include features related to bone mineral density from medical examinations. The primary prediction module is configured to construct a multi-model fusion prediction model consisting of sequentially connected base learners and meta-learners. Acquired features are input into multiple base learners for prediction and recognition, yielding primary prediction results. The constructed multi-model fusion prediction model comprises two layers: the first layer includes multiple base learners, and the second layer includes a meta-learner. The first layer model uses a combination of Support Vector Machine (SVM), XGBoost, and GBDT models. Each base learner is trained and predicts independently, yielding primary prediction results. The meta-learner employs a random forest model, combining the prediction results from the first layer to form a new feature vector for prediction, resulting in the final prediction result. The support vector machine model is constructed by optimizing the objective function through kernel functions and the Lagrange multiplier method, and the kernel function type and regularization parameters are tuned using GridSearchCV. The fusion module is configured to adaptively calculate fusion weights based on the prediction accuracy of the base learners, and fuse the initial prediction results of each base learner based on the calculated fusion weights. The fusion method for the initial prediction results of the base learners includes the following steps: obtaining the prediction accuracy of the base learners; the prediction accuracy is statistically analyzed based on the test set, the test set data is updated according to a set time, and the prediction accuracy of each base learner is calculated based on the test results and actual values ​​of the test set; the accuracy is used as the fusion weight, multiplied by the initial prediction result of the corresponding base learner, to obtain the fused initial prediction result. The secondary prediction module is configured to transmit the fused primary prediction results to the meta-learner for prediction, thereby obtaining the bone density prediction result.

6. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the multi-model fusion bone mineral density prediction method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the multi-model fusion bone mineral density prediction method according to any one of claims 1-4.

Images