A multi-source bone density data fusion method, system and storage medium

Abstract

The application discloses a multi-source bone density data fusion method, system and storage medium, and belongs to the field of medical big data and intelligent diagnosis. The method aims at the systematic deviation between DXA devices of different manufacturers and the high detection uncertainty problem under the primary medical environment. Firstly, the multi-source bone density data and the corresponding physical calibration parameters are projected to a unified measurement space through standardized preprocessing and device-related mapping functions, and the device barriers are eliminated by using distribution alignment constraints. Then, Gaussian uncertainty modeling is performed on the projected data to quantify the influence of specific clinical factors on measurement reliability. Based on the uncertainty propagation mechanism, the multi-source discrete evidence is adaptively compensated and fused by weight, and multi-dimensional features including physiology, medical history and anatomical topology are introduced for joint modeling. Finally, end-to-end optimization is performed through a joint function of fusion uncertainty constraint and task loss, and high-robustness osteoporosis risk intelligent evaluation is realized. The application significantly improves the consistency of cross-device T value prediction and the accuracy of missed diagnosis / misdiagnosis review, and is suitable for multi-center bone density mutual recognition and primary medical bone health intelligent evaluation scene.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to a method, system, and storage medium for multi-source bone mineral density data fusion, applicable to data fusion of multi-source bone mineral density detection equipment and intelligent assessment of osteoporosis risk. Background Technology

[0002] In clinical bone mineral density (BMD) testing, the inherent differences in imaging principles, calibration standards, and reference databases among different manufacturers and models of equipment lead to systematic biases in measurement results for the same subject on different devices. This bias manifests not only as differences in the mean values ​​of measurements but also as an overall shift in data distribution and statistical characteristics, making it difficult for traditional fusion methods to achieve unified analysis and consistent assessment across devices. Furthermore, primary healthcare settings commonly face challenges such as limited equipment accuracy, missing clinical data, and significant environmental interference, resulting in significant instability and uncertainty in BMD assessment results. This can easily lead to misdiagnosis, missed diagnosis, and overtreatment, reducing the reliability of diagnosis and treatment.

[0003] Existing multi-source data fusion methods do not specifically align the measurement space for cross-device differences in bone density testing, nor do they establish a reasonable uncertainty propagation and weighting mechanism. They cannot quantify and suppress data uncertainty while eliminating device bias, making it difficult to meet the clinical demand for highly robust and universal bone health assessment. Summary of the Invention

[0004] Purpose of the invention: The present invention aims to provide a method, system and storage medium for multi-source bone mineral density data fusion. The present invention includes a unified measurement space and considers the uncertainty propagation of bone mineral density evidence across devices. It can solve the problems of inconsistent measurement standards, significant differences in data distribution and difficulty in characterizing uncertainty in the process of multi-source information fusion in the prior art.

[0005] Technical solution: A multi-source bone mineral density data fusion method, which includes unifying the metric space and considering the uncertainty propagation of bone mineral density evidence across devices, comprising the following steps: S1. Data Acquisition and Preprocessing: Obtain bone mineral density measurement data and related clinical information of subjects from multi-source bone mineral density detection devices, fill missing values ​​and detect and remove outliers in the raw data, and standardize the bone mineral density measurement data output from different devices to obtain standardized bone mineral density data. S2. Construct a unified metric space across devices: Define a device-related mapping function to map the standardized bone density data of different devices in step S1 to a unified feature space. Introduce distribution alignment constraints and complete unified modeling by minimizing the differences in data representation distributions between different devices, thereby eliminating systematic biases between devices. S3. Uncertainty Modeling of Bone Density Data: Represent bone density data in a unified metric space as random variables following a Gaussian distribution, construct a mapping function to output the predicted bone density value and the corresponding uncertainty, introduce an uncertainty constraint loss function, and jointly constrain the prediction error and uncertainty magnitude. S4. Weighted fusion of multi-source data based on uncertainty propagation: Calculate the corresponding data source weights based on the uncertainty of bone density data of each device obtained in step S3, and perform weighted fusion of the predicted bone density values ​​of each device based on these weights to obtain a fused bone density representation; perform joint feature modeling with clinical structured information to obtain the final feature representation; S5. Osteoporosis Risk Assessment: Based on the final feature representation of step S4, a prediction model is constructed. The osteoporosis risk prediction result is output through linear mapping and activation function. The prediction model is trained using the cross-entropy loss function. S6. Joint Model Optimization: Construct a joint loss function that includes task loss, cross-device alignment loss, and uncertainty modeling loss. Perform joint optimization on the prediction model and the mapping function of the unified metric space to obtain the optimal model and complete the fusion of bone mineral density evidence and osteoporosis risk assessment.

[0006] On the other hand, the present invention provides a multi-source bone mineral density data fusion system for performing the above-described method, and the system is deployed in a hospital information system or a bone mineral density testing device terminal, including sequentially communicatively connected systems: The data access module is used to acquire the subject's bone mineral density measurement data and related clinical information from the multi-source bone mineral density detection device and transmit them to the data preprocessing module; The data preprocessing module is used to impute missing values, detect and remove outliers, and standardize bone mineral density measurement data and clinical information to obtain standardized bone mineral density data, which is then transmitted to the unified metric space modeling module. The unified metric space modeling module is used to construct device-related mapping functions, map standardized bone density data to a unified feature space, complete cross-device unified modeling through distribution alignment constraints, and transmit bone density data from the unified metric space to the uncertainty calculation module. The uncertainty calculation module is used to perform Gaussian probability modeling on bone density data in a unified metric space, calculate the predicted bone density value and the corresponding uncertainty, and transmit it to the multi-source fusion module. The multi-source fusion module is used to calculate the weight of each data source based on uncertainty, complete the weighted fusion of bone density data from multiple devices, and perform joint feature modeling with clinical structured information to obtain the final feature representation before transmitting it to the risk assessment module. The risk assessment module is used to build and train an osteoporosis risk prediction model based on the final feature representation and output osteoporosis risk assessment results. The model optimization module is used to construct a joint loss function, jointly optimize the mapping function of the unified metric space modeling module and the prediction model of the risk assessment module, and realize iterative updates of model parameters. The results display module receives the assessment results from the risk assessment module and outputs standardized bone mineral density values, uncertainty indices, and osteoporosis risk levels in the form of a visual report.

[0007] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multi-source bone density data fusion method.

[0008] Beneficial effects: Compared with the prior art, the substantive features and significant effects of the present invention include: (1) This invention proposes for the first time a device-specific mapping function and distribution alignment constraint to map the measurement data of different bone density devices to a unified feature space, thereby fundamentally eliminating systematic deviations between devices and realizing consistent expression of multi-source data.

[0009] (2) Gaussian probability modeling is performed on bone density data in a unified space to quantify the uncertainty caused by measurement noise and model error. The uncertainty constraint loss function is used to achieve joint control of prediction error and uncertainty amplitude, providing a quantitative basis for data credibility.

[0010] (3) The inverse variance weighting mechanism is adopted to dynamically allocate weights based on the uncertainty of each data source. The lower the uncertainty of the device data, the higher the weight, so as to realize the reliable fusion of multi-source bone density evidence and suppress the interference of low-quality data on the results.

[0011] (4) The fused bone density features are combined with clinical structured information (age, gender, medical history, etc.) to form a joint model, introducing medical prior knowledge and improving the feature expression ability of osteoporosis risk assessment.

[0012] (5) Construct a joint loss function that includes task loss, cross-device alignment loss and uncertainty modeling loss, and perform end-to-end joint optimization of the mapping function and the prediction model to achieve synergistic improvement in evaluation accuracy, cross-device consistency and model robustness. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is an architecture diagram of the system described in this invention. Detailed Implementation

[0014] To illustrate the technical solution provided by this invention in detail, further description is provided below with reference to the accompanying drawings.

[0015] This invention addresses two core issues: cross-device data consistency modeling and uncertainty propagation mechanisms. It systematically models the bone mineral density (BMD) assessment process to achieve unified representation and reliable fusion of multi-source BMD data. In its implementation, it first acquires subject BMD measurement data and related clinical information from multiple devices and preprocesses the raw data. The preprocessing includes imputing missing values, detecting and removing outliers, and standardizing data from different devices to ensure analysis on a unified scale. (The last sentence appears to be incomplete and possibly refers to a separate, unrelated point about data sources.) Bone density measurement data from each device are as follows: Obtained through standardization This provides a foundation for subsequent unified modeling. Considering the implicit differences in metric space between different devices—that is, the same physical meaning of bone mineral density index corresponds to different numerical ranges and statistical distributions on different devices—this invention constructs device-related mapping functions to map the original data to a unified metric space, thereby achieving consistent expression of cross-device data.

[0016] Specifically, by defining a mapping function Mapping data from different devices into a unified representation:

[0017] in, This represents the parameters of the mapping function. To ensure the consistency of data from different devices in a unified space, a distribution alignment constraint is introduced. Unified modeling is achieved by minimizing the differences in the representation distributions between different devices. The objective function is:

[0018] in, This represents the distribution distance function.

[0019] As a preferred implementation scheme, the distribution distance function The choice can be flexibly made based on the distribution density of clinical bone mineral density data. For example, when processing small sample DXA data, the maximum mean difference (MMD) can be preferred; when processing subject groups with multimodal distribution characteristics, an alignment mechanism based on contrastive learning can be used.

[0020] Through the above process, data from different devices have a consistent form of expression in a unified space, thereby effectively eliminating systematic biases caused by device differences.

[0021] Building upon the unified metric space, this invention further introduces a probabilistic modeling mechanism to characterize the uncertainties in bone density measurement and modeling, extending the unified representation into a random variable form. Specifically: Will Let it be a random variable that follows a Gaussian distribution:

[0022] in, This represents the predicted bone mineral density value. This represents the corresponding uncertainty, used to reflect the reliability of the data source. This is achieved by constructing a mapping function. ,get:

[0023] And the following uncertainty constraint loss function is introduced:

[0024] This loss function simultaneously constrains the magnitude of prediction error and uncertainty during the optimization process, enabling the model to adaptively assign different confidence levels to different samples, thereby effectively characterizing the impact of factors such as equipment noise, missing data, and model uncertainty on the prediction results.

[0025] Furthermore, in the process of multi-source evidence fusion, this invention designs a weighted fusion mechanism based on uncertainty propagation, enabling different data sources to dynamically adjust their weights according to their uncertainties during the fusion process. Specifically, the weights of each data source are calculated as follows:

[0026] The bone mineral density estimates from each device were then weighted and fused.

[0027] This weighting strategy can effectively suppress the impact of highly uncertain data on the fusion results, thereby improving the stability of the overall results.

[0028] Based on this, the fused bone mineral density representation is combined with clinical structured information. Joint modeling is performed to obtain the final feature representation:

[0029] in, This represents the feature fusion function. This process not only achieves the fusion of data from multiple devices but also incorporates clinical information to assist in the prediction results, thereby improving the model's expressive power.

[0030] As a preferred implementation method, the feature fusion function An attention-based gating fusion structure is adopted, which specifically includes three parts: a clinical feature gating network, a dynamic weight generation unit, and a bias compensation branch. The first step is the construction of a clinical feature-gated network. The gated network uses structured clinical information... As input, the vector is mapped to the hidden feature space through the embedding layer and the fully connected layer, and a gating coefficient vector is generated. The gating coefficients correspond one-to-one with the bone density feature dimensions, and are used to achieve dimensional feature modulation.

[0031] Secondly, dynamic weight adaptive generation is used. This is based on the output of the gated network. , for fusion bone mineral density characteristics Perform dimensional weighting:

[0032] in Represents the activation function (e.g.) Sigmoid ), This indicates element-wise multiplication. Through this operation, the model can dynamically enhance or suppress the response of specific dimensions of bone mineral density features based on clinical conditions (such as age group, comorbidities, etc.).

[0033] Finally, there is the bias-guided compensation mechanism. This further integrates clinical information. Introduced as a bias term into the fusion process, a bias vector is generated through linear transformation. And superimposed on the gated features:

[0034] It is generated independently from clinical information and can effectively compensate for the apparent deviation of bone mineral density caused by differences in physiological state among different individuals. For example, it can enhance the contribution weight of key vertebral features for the elderly or reduce feature fluctuations caused by soft tissue interference for the younger population.

[0035] In the risk assessment phase, a predictive model is constructed based on the final feature representation to quantitatively assess the risk of osteoporosis. Specifically, the prediction results are obtained through linear mapping and activation functions:

[0036] Model training is achieved by minimizing the cross-entropy loss function:

[0037] This enables accurate prediction of osteoporosis risk.

[0038] To further improve the model's generalization ability and stability, this invention employs a joint loss function during the overall optimization process, jointly optimizing the cross-device alignment loss, uncertainty modeling loss, and task loss:

[0039] in, and These are weighting coefficients used to balance the influence of different loss terms. Through a joint optimization strategy, the model ensures prediction accuracy while also considering cross-device consistency and uncertainty modeling capabilities.

[0040] In terms of system implementation, the present invention further provides a bone density intelligent assessment software system, which can be deployed in hospital information systems or bone density testing equipment terminals to realize the full-process processing of automatic access to data from multiple devices, unified modeling, and risk assessment.

[0041] The system achieves data exchange with different devices through standardized interfaces and enables collaborative work between various functional modules through modular design. These modules include a data access module, a unified metric space modeling module, an uncertainty calculation module, a multi-source fusion module, and a risk assessment and result display module. The system can output unified bone mineral density values, uncertainty indices, and risk levels, and present them in the form of visual reports to provide doctors with auxiliary decision-making support.

[0042] In practical applications, when the same patient undergoes bone density testing at different times or in different medical institutions using different devices, the method of this invention can uniformly model and fuse multiple measurement results, eliminating the influence of device differences and outputting consistent and stable assessment results. In situations where primary healthcare data is scarce or of low quality, the uncertainty propagation mechanism automatically reduces the impact of low-quality data on the results, improving the robustness of the overall assessment. In multi-center clinical research scenarios, this invention can achieve unified analysis and modeling of data from different sources, providing reliable technical support for large-scale bone health data analysis.

[0043] Furthermore, the unified metric space and uncertainty propagation mechanism proposed in this invention possess excellent versatility. It is not only applicable to bone density data analysis but can also be extended to other multi-device, multi-modal data fusion scenarios in healthcare, such as the fusion of imaging and clinical data, and collaborative analysis of data from different hospitals. This provides a universal data fusion framework for intelligent medical systems. Through the above technical solutions, this invention achieves a transformation from traditional empirical correction methods to unified modeling and reliable fusion methods, significantly improving cross-device data consistency, uncertainty modeling, and multi-source fusion capabilities. It demonstrates high innovation and practical application value.

[0044] In this embodiment of the invention, three different models of DXA cameras were used to collect data, totaling 1,420 samples.

[0045] Equipment distribution: GE Lunar Prodigy (equipment A, containing 620 samples), Hologic Horizon (equipment B, containing 550 samples), and a domestically produced DXA model (equipment C, containing 250 samples), with equipment B serving as a unified comparison object.

[0046] This experiment used the average BMD measurement of the lumbar spine (L1-L4) of the same subject at a specific hospital as the standard. Evaluation indicators: Root mean square error (RMSE), mean relative bias (MARD), and coefficient of determination were used. To assess numerical consistency, AUROC was used to evaluate the performance of risk classification.

[0047] The proposed "device-related mapping function" is compared with the industry-recognized linear transformation formula for sBMD (standardized bone mineral density).

[0048] Table 1. Comparison of BMD values ​​across devices using different calibration methods (L1-L4 lumbar spine)

[0049] Experimental results show that while the traditional linear sBMD formula can partially eliminate mean bias, it cannot handle the nonlinear differences in high / low energy spectrum distributions of different detectors. This invention, through a unified metric space mapping, further reduces the RMSE from 0.0749 to 0.0260, improving accuracy by 65.3%, and achieving a high degree of physical alignment.

[0050] Next, we verified the robustness of the uncertainty modeling proposed in this invention to noise samples. To simulate the common "improper placement" or "low signal-to-noise ratio" situations in primary healthcare, 15% artificial synthesized noise (Gaussian noise σ=0.05) was added to the test set.

[0051] Table 2. Comparison of the performance of uncertainty weighting mechanisms under noise interference

[0052] Under noise interference, the error of simple averaging fusion increases dramatically. However, this invention, by constraining the loss function with uncertainty, enables the model to identify high-noise samples and automatically reduce their weights (weight allocation is reduced from 0.5 to 0.12), thereby controlling the error fluctuation to 36.1% and significantly enhancing the robustness of clinical applications.

[0053] Finally, the gain effect of joint modeling based on "bone mineral density evidence" and "clinical structured information" was verified, as shown in Table 3. Relying solely on bone mineral density or clinical information has limitations. This invention utilizes a feature fusion function... Deep interaction between the two was achieved, with an AUROC of 0.928. Compared to simple feature concatenation, the gated fusion mechanism of this invention can "reinterpret" bone mineral density values ​​based on clinical context (such as advanced age and menopause), effectively reducing the misdiagnosis rate of borderline samples.

[0054] Table 3. Gain effect of joint modeling of "bone mineral density evidence" and "clinical structured information"

[0055] Note: Youden Index = Sensitivity + Specificity - 1, is the most commonly used comprehensive index for evaluating the validity of a prediction model.

Claims

1. A method for fusing multi-source bone mineral density data, characterized in that, This method achieves reliable fusion based on a unified metric space across devices and a bone density measurement uncertainty propagation mechanism. The implementation steps are as follows: S1. Data Acquisition and Preprocessing: Obtain bone mineral density measurement data and related clinical information from the subjects using multi-source bone mineral density detection devices. Perform missing value imputation, outlier detection and removal on the raw data, and standardize and unify the measurement range of the bone mineral density measurement data output from different devices to obtain standardized bone mineral density data. S2. Construct a unified metric space across devices: Define a mapping function related to device hardware characteristics, map the standardized bone density data of different devices in step S1 to a unified feature space, introduce distribution alignment constraints, and complete unified modeling by minimizing the differences in data representation distribution between different devices, thereby eliminating systematic biases caused by differences in device principles and calibration specifications. S3. Uncertainty Modeling of Bone Density Data: Represent bone density data in a unified metric space as random variables following a Gaussian distribution, construct a mapping function to output the predicted bone density value and reflect the uncertainty of measurement and environmental noise, introduce an uncertainty constraint loss function to jointly constrain the prediction error and uncertainty magnitude. S4. Weighted fusion of multi-source data based on uncertainty propagation: Calculate the corresponding data source weights based on the uncertainty of bone density data of each device obtained in step S3, and perform adaptive weighted fusion of the predicted bone density values ​​of each device based on the weights to obtain the fused bone density representation; perform joint feature modeling and bias compensation with clinical structured information to obtain the final feature representation; S5. Osteoporosis Risk Assessment: Based on the final feature representation of step S4, a prediction model is constructed. The osteoporosis risk prediction result is output through linear mapping and activation function. The prediction model is trained using the cross-entropy loss function. S6. Joint Model Optimization: Construct a joint loss function that includes task loss, cross-device alignment loss, and uncertainty modeling loss. Perform end-to-end joint optimization on the prediction model and the mapping function of the unified metric space to obtain the optimal model and complete the fusion of bone mineral density evidence and osteoporosis risk assessment.

2. The multi-source bone mineral density data fusion method according to claim 1, characterized in that, In step S2, let's assume it comes from the first... Bone density measurement data from each device are as follows: Obtained through standardization By defining a mapping function Will Mapping is represented as a unified representation ,in, Indicates data dimension, Represents the set of real numbers. These are the parameters of the mapping function; The objective function of the distribution alignment constraint is , This represents the distribution distance function.

3. The multi-source bone mineral density data fusion method according to claim 1, characterized in that, Step S3 includes processing bone mineral density data in a unified metric space. Let represent a random variable that follows a Gaussian distribution. ,in This represents the predicted bone mineral density value. This indicates the corresponding uncertainty, used to reflect the reliability of the data source; By constructing mapping functions get ; An uncertainty-constrained loss function is introduced to constrain the magnitude of prediction error and uncertainty during the optimization process, enabling adaptive assignment of different confidence levels to different samples. The uncertainty-constrained loss function is as follows: In the formula, For the sample size, For the sample true value, Indicates the first Predicted bone mineral density values ​​for each sample.

4. The multi-source bone mineral density data fusion method according to claim 1, characterized in that, The formula for calculating the weights of each data source in step S4 is as follows: The weighted fusion bone mineral density is expressed as: Through feature fusion function Joint feature modeling is achieved, and the final feature representation is as follows: In the above formula, and They represent the first The data and the first The uncertainty of the data, Indicates the first Predicted bone mineral density values ​​for each data point. Indicates the first The weights of the predicted bone mineral density values ​​for each data point For structured clinical information.

5. The multi-source bone mineral density data fusion method according to claim 1, characterized in that, The formula for calculating the osteoporosis risk prediction result in step S5 is as follows: In the formula, For activation function, This is the weight matrix. For bias terms; The cross-entropy loss function is: In the formula, Indicates the first The true label of each sample, taking the value 0 or 1. Indicates the first The predicted probability of a sample.

6. The multi-source bone mineral density data fusion method according to claim 1, characterized in that, Step S6 involves jointly optimizing the cross-device alignment loss, uncertainty modeling loss, and task loss: in, and These are weighting coefficients used to balance the influence between different loss terms. Let cross-entropy be the loss function. The objective function of the distribution alignment constraint, This is the uncertainty constraint loss function.

7. A multi-source bone mineral density data fusion system, characterized in that, The system is used to perform the method as described in any one of claims 1-6, and the system is deployed in a hospital information system or a bone density testing device terminal, including sequentially communicatively connected components: The data access module is used to acquire the subject's bone mineral density measurement data and related clinical information from the multi-source bone mineral density detection device and transmit them to the data preprocessing module; The data preprocessing module is used to fill in missing values, detect and remove outliers, and standardize bone mineral density measurement data and clinical information to obtain standardized bone mineral density data, which is then transmitted to the unified metric space modeling module. The unified metric space modeling module is used to construct device-related mapping functions, map standardized bone density data to a unified feature space, complete cross-device unified modeling through distribution alignment constraints, and transmit bone density data from the unified metric space to the uncertainty calculation module. The uncertainty calculation module is used to perform Gaussian probability modeling on bone density data in a unified metric space, calculate the predicted bone density value and the corresponding uncertainty, and transmit it to the multi-source fusion module. The multi-source fusion module is used to calculate the weight of each data source based on uncertainty, complete the weighted fusion of bone density data from multiple devices, and perform joint feature modeling with clinical structured information to obtain the final feature representation before transmitting it to the risk assessment module. The risk assessment module is used to build and train an osteoporosis risk prediction model based on the final feature representation and output osteoporosis risk assessment results. The model optimization module is used to construct a joint loss function, jointly optimize the mapping function of the unified metric space modeling module and the prediction model of the risk assessment module, and realize iterative updates of model parameters. The results display module receives the assessment results from the risk assessment module and outputs standardized bone mineral density values, uncertainty indices, and osteoporosis risk levels in the form of a visual report.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the cross-device bone density evidence fusion method based on a unified metric space and uncertainty propagation as described in any one of claims 1-6.

Images