A method and system for constructing an osteoporosis bone density value prediction model based on clinical data

Abstract

The present application relates to the technical field of machine learning, in particular to a method and system for constructing an osteoporosis bone density value prediction model based on clinical data, in which, through convolutional neural network weight collation of adjacent variables, dot product and calculation are performed and combined with an activation threshold screening mechanism, so that weakly correlated disturbance signals are compressed in the feature space, thereby improving the proportion of locally strong correlation patterns in the overall features, enhancing the model's ability to identify key metabolic indicator linkage relationships, through long short-term memory network sequence modeling of the multi semantic feature body after deep stacking, the time recursive association between different categories of clinical variables is continuously accumulated and expressed, improving the weight distribution stability of global dependence information in feature representation, through hidden vector driven channel weight proportion dynamic adjustment, improving the model's ability to adapt to population heterogeneity, and improving information utilization efficiency and generalization stability under the condition of limited variable quantity.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for constructing a model for predicting osteoporosis bone mineral density based on clinical data. Background Technology

[0002] In the field of machine learning technology, medical data modeling and regression prediction specifically fall under the category of supervised learning model construction technology that uses multivariate clinical data as input and continuous medical indicators as output. This technology is based on statistical learning theory and machine learning algorithms. By constructing a nonlinear mapping relationship between input feature vectors and target variables, it achieves quantitative prediction of continuous medical variables. Core technologies include sample feature space construction, model parameter learning, loss function optimization, and generalization performance evaluation. Commonly used model structures include regression neural networks and gradient boosting regression trees. Unlike traditional rule-based and threshold-based methods, this technology establishes a stable prediction function by learning statistical correlations in historical samples. This allows it to output bone mineral density values ​​with high consistency with actual measurements even without incorporating imaging data, making it suitable for medical risk assessment and assisted diagnosis scenarios.

[0003] A method for constructing a bone mineral density (BMD) prediction model for osteoporosis based on clinical data is a model construction method that establishes a mapping relationship between multiple routine clinical variables and quantitative BMD values ​​through machine learning. The method uses patient demographic information, lifestyle indicators, medical history, medication history, and routine biochemical test indicators as model inputs, and uses BMD values ​​obtained from BMD testing equipment as supervision labels. Through model training, a mathematical model that can directly output predicted BMD values ​​is obtained. The purpose of this method is to eliminate reliance on specialized equipment such as dual-energy X-ray and quantitative CT, enabling primary healthcare institutions to obtain continuous BMD prediction results using only existing clinical data. The model output is no longer a discrete classification of normal, osteopenia, and osteoporosis, thereby improving the value of the prediction results in risk stratification, follow-up assessment, and individualized intervention, and achieving a simultaneous improvement in the accuracy and accessibility of early osteoporosis screening.

[0004] Existing technologies employ unified feature vector input regression neural networks and gradient boosting regression trees to establish the relationship between clinical variables and bone mineral density values ​​through an overall nonlinear mapping function. The model structure typically optimizes for minimizing global error, lacking a specific characterization of local synergistic structures between variables. This leads to the averaging of strongly and weakly correlated indicators within the same mapping space, and the dilution of local risk linkages during the overall function approximation process. When bone metabolism indicators and demographic indicators exhibit nonlinear interactions within specific intervals, the unified mapping function often outputs smooth predictions, failing to accurately reflect local peak changes. Furthermore, traditional regression models maintain a fixed weight distribution after training, failing to dynamically reflect changes in the proportion of variable contributions across different sample groups. The fixed weight structure struggles to reflect changes in real time, resulting in systematic biases in predictions for specific populations. In primary care screening scenarios, this may reduce the accuracy of continuous bone mineral density predictions, impacting the accuracy of risk stratification assessments and follow-up intervention decisions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for constructing a prediction model for osteoporosis bone mineral density based on clinical data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing an osteoporosis bone mineral density prediction model based on clinical data, comprising the following steps: S1: Extract age, body mass index, alkaline phosphatase and osteocalcin scalars from raw clinical data, perform mapping operations to convert them into high-dimensional dense vectors, perform channel splicing based on demographics, bone metabolism and hormone categories to generate a multi-channel clinical semantic embedding matrix. S2: Based on the multi-channel clinical semantic embedding matrix, perform grouped sliding window operation, call the convolutional neural network weight kernel to calculate the sum of the dot products of variables within the channel, compare the collaborative response value with the activation threshold, retain the signal exceeding the threshold and reset the low threshold value to zero, and output the local strong correlation feature map within the group; S3: Based on the local strongly correlated feature maps within the group, construct a multi-semantic feature body by stacking along the depth, perform linear combination operations between channels, drive the long short-term memory network to output context-aware latent vectors, and dynamically adjust the feature channel weight ratios according to the vectors to obtain a cross-domain global collaborative semantic tensor. S4: Based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor, the aggregated semantics are aggregated by superimposing the tensor, and global average pooling is performed on the aggregated feature body. The channel calibration coefficient is calculated and the coefficient is multiplied element-wise with the aggregated feature body to construct an adaptive weighted representation of the key factors of bone density. S5: Based on the adaptive weighted representation of the key bone mineral density factors, the feature vector is flattened into a one-dimensional regression feature vector. The feature vector elements and regression weights are linearly weighted and summed in the input mapping layer. The bias term is superimposed to correct the benchmark and output a continuous scalar to obtain the target osteoporosis bone mineral density quantitative prediction value.

[0007] As a further aspect of the present invention, the multi-channel clinical semantic embedding matrix includes an age mapping vector, a body mass index mapping vector, an alkaline phosphatase mapping vector, and an osteocalcin mapping vector; the intra-group local strong correlation feature map includes a set of over-threshold response values ​​and a set of zero-value response values; the cross-domain global collaborative semantic tensor includes a context-aware latent vector and a set of feature values ​​adjusted by weights; the adaptive weighted representation of key bone mineral density factors includes global average pooling results, channel calibration coefficients, and element-wise multiplication results; and the one-dimensional regression feature vector includes a flattened and expanded sequence of feature values.

[0008] As a further aspect of the present invention, the specific steps for generating the multi-channel clinical semantic embedding matrix are as follows: Based on the subjects' original clinical data, age, body mass index, alkaline phosphatase and osteocalcin scalars are directly extracted, and each scalar is projected into a high-dimensional continuous space by inputting a preset weight matrix. The values ​​are then expanded into multi-dimensional feature vectors to generate independent clinical indicator feature vector groups. Based on the independent clinical indicator feature vector group, each vector is assigned to a preset demographic, bone metabolism and hormone regulation group. Vectors in the same group are concatenated to construct a local semantic channel. The channels are stacked and spliced ​​to form a three-dimensional feature tensor, generating a multi-channel clinical semantic embedding matrix.

[0009] As a further aspect of the present invention, the specific steps for outputting the intra-group local strong correlation feature map are as follows: Based on the multi-channel clinical semantic embedding matrix, a convolutional neural network weight kernel matrix is ​​configured to cover adjacent feature vectors, the kernel parameter is multiplied by the corresponding vector element, the full length of the channel is traversed to obtain local interaction values, and local interaction slices of adjacent variables within the channel are generated. Based on the local interactive slices of adjacent variables within the channel, all product values ​​within the slice are accumulated, the local cooperative response intensity at the window position is calculated, the continuous response sequence of each slice position is obtained by aggregating the response sequences of each channel, and a multidimensional response tensor is constructed by recombining the response sequences of each channel to generate a multidimensional local cooperative response value set. Based on the multi-dimensional local collaborative response value set, a preset activation threshold bias parameter is introduced. Each response value is compared with the threshold bias, positive response values ​​greater than the threshold are retained, and negative response values ​​less than the threshold are reset to zero, thereby generating a local strong correlation feature map within the group.

[0010] As a further aspect of the present invention, the convolutional neural network constructs a network weight kernel matrix with a preset kernel width and input channel depth, maps the weight kernel matrix to the feature dimension of the multi-channel clinical semantic embedding matrix, adjusts the spatial position of the weight kernel matrix to match the current processing step size, and makes the parameter window of the weight kernel matrix precisely aligned in dimension and covers a set of consecutive adjacent feature vectors in the channel.

[0011] As a further aspect of the present invention, the specific steps for obtaining the cross-domain global collaborative semantic tensor are as follows: Based on the local strongly correlated feature maps within the group, a feature tensor is constructed by stacking the maps along the depth. The channel data is linearly weighted and summed using a unit-size kernel. The heterogeneous semantic data stream is fused, and the independent channels are reorganized into high-dimensional entities. The numerical mapping between semantic groups is established, and a multi-semantic fusion linear feature body is generated. Based on the multi-semantic fusion linear feature body, the spatial dimension is compressed to obtain the channel global descriptor, driving the long short-term memory network to evolve the memory unit, using the gating mechanism to model the long-range dependency between channels, outputting the context-aware latent vector, normalizing the values ​​to construct attention weights, and generating the global context dependency weight vector. Based on the global context-dependent weight vector, the weight dimension is expanded along the multi-semantic fusion linear feature volume space, the weight and feature values ​​are expanded by element-wise multiplication, the channel response intensity is recalibrated, the heterogeneous feature distribution is calibrated and the feature response amplitude is adjusted, and a cross-domain global collaborative semantic tensor is generated.

[0012] As a further aspect of the present invention, the Long Short-Term Memory network uses the channel global descriptor obtained by compressing the spatial dimension as the input of the current time step and maps it to the recursive unit of the network. It performs gating operation by combining the hidden state vector of the previous time step with the cell state vector. The retention ratio of historical cell states is determined by the forget gate parameter. The writing weight of the current channel descriptor is controlled by the input gate parameter to update the current cell state. The nonlinear mapping is performed based on the updated cell state by the output gate parameter to generate the context-aware hidden vector of the current time step.

[0013] As a further aspect of the present invention, the specific steps for constructing the adaptive weighted representation of the key bone mineral density factors are as follows: Based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor, align the tensor space dimension and channel number, calculate the sum of each element at the corresponding position, superimpose local details and global context signals, integrate heterogeneous information streams into a unified data structure, construct a full-scale feature container, and generate a dual-path aggregated semantic feature body. Based on the dual-path aggregated semantic feature body, the global average value of the spatial axis is calculated, the feature dimension is compressed to obtain the channel descriptor, the descriptor is mapped to the nonlinear activation interval, the channel calibration coefficient is output, the coefficient is multiplied element by element with the feature body value, the channel response intensity distribution is recalibrated, and an adaptive weighted representation of the key bone density factor is generated.

[0014] As a further aspect of the present invention, the specific steps for obtaining the quantitative prediction value of the target osteoporosis bone mineral density are as follows: Based on the adaptive weighted representation of the key bone density factors, the multidimensional tensor structure is reshaped into a single linear dimension. Feature values ​​are concatenated in channel order to construct a continuous one-dimensional numerical array. The vector length is aligned to the number of fully connected nodes and the geometric constraints of the tensor space are removed to generate a globally flattened regression feature vector. Based on the global flattened regression feature vector, the product of the vector elements and the regression weights is calculated, the weighted values ​​are accumulated to obtain the sum of the linear responses, the scalar bias term is superimposed to correct the prediction benchmark, the feature values ​​are mapped to the continuous physical real number domain, the bone density scalar is output, and the target osteoporosis bone density quantitative prediction value is generated.

[0015] A system for constructing an osteoporosis bone mineral density (BMD) prediction model based on clinical data, wherein the system is used to execute the aforementioned method for constructing an osteoporosis BMD prediction model based on clinical data, and the system includes: Semantic construction module: Based on raw clinical data, read age, body mass index, alkaline phosphatase and osteocalcin scalars and classify them, perform vector expansion and channel splicing and arrangement, establish a three-dimensional continuous numerical structure, and generate a multi-channel clinical semantic embedding matrix. Local modeling module: Based on the multi-channel clinical semantic embedding matrix, channel window arrangement and response calculation are performed, thresholds are compared and values ​​are filtered, responses exceeding the threshold are retained and low-value data are set to zero, a multi-channel response structure is constructed, and a local strong correlation feature map within the group is obtained; Global Collaboration Module: Based on the local strongly correlated feature maps within the group, deep overlay and channel integration processing are performed, the recursive structure update state is input, the channel weight ratio is adjusted according to the latent vector, a recalibrated feature structure is established, and a cross-domain global collaborative semantic tensor is obtained. The calibration weighting module performs dimensional alignment and numerical superposition based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor to generate an aggregated feature body, calculates the channel calibration coefficients and multiplies them element by element to construct an adaptive weighted representation of key bone density factors. Regression output module: Based on the adaptive weighted representation of the key bone mineral density factors, calculate the linear expansion and weight product, accumulate the values ​​and superimpose the bias term, map to the continuous numerical domain, and output a single scalar to obtain the target osteoporosis bone mineral density quantitative prediction value.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the weight kernel of a convolutional neural network performs dot product calculation on adjacent variables and combines it with an activation threshold screening mechanism to compress weakly correlated perturbation signals in the feature space, thereby increasing the proportion of local strong correlation patterns in the overall features and enhancing the model's ability to identify the linkage relationship of key metabolic indicators. In this invention, a long short-term memory network is used to perform sequence modeling on the multi-semantic feature body after deep stacking, so that the temporal recursive association between different categories of clinical variables can be continuously accumulated and expressed, thereby improving the stability of the weight distribution of global dependency information in feature representation. In this invention, the proportion of channel weights is dynamically adjusted by driving the latent vector, so that the contribution ratio of variables changes adaptively under different sample conditions, thereby improving the model's ability to adapt to population heterogeneity, enhancing the consistency and numerical resolution of continuous bone density numerical prediction, and improving information utilization efficiency and generalization stability under the condition of a limited number of variables. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 This invention provides a technical solution: a method for constructing a predictive model for osteoporosis bone mineral density based on clinical data, comprising the following steps: S1: Extract age, body mass index, alkaline phosphatase and osteocalcin scalars from raw clinical data, perform mapping operations to convert them into high-dimensional dense vectors, perform channel splicing based on demographics, bone metabolism and hormone categories to generate a multi-channel clinical semantic embedding matrix. S2: Based on the multi-channel clinical semantic embedding matrix, perform grouped sliding window operation, call the convolutional neural network weight kernel to calculate the sum of the dot products of variables within the channel, compare the collaborative response value with the activation threshold, retain the signal exceeding the threshold and reset the low threshold value to zero, and output the local strong correlation feature map within the group. S3: Based on the strong local correlation feature maps within the group, multi-semantic feature bodies are constructed by stacking along the depth, and linear combination operations between channels are performed to drive the long short-term memory network to output context-aware latent vectors. The feature channel weight ratios are dynamically adjusted according to the vectors to obtain cross-domain global collaborative semantic tensors. S4: Based on the intra-group local strong correlation feature map and cross-domain global collaborative semantic tensor, the semantics are aggregated by tensor superposition, global average pooling is performed on the aggregated feature body, the channel calibration coefficient is calculated and the coefficient is multiplied element by element with the aggregated feature body to construct an adaptive weighted representation of key bone density factors. S5: Based on the adaptive weighted representation of key bone mineral density factors, the feature vector is flattened into a one-dimensional regression feature vector. The feature vector elements are input to the mapping layer to calculate the linear weighted sum of the regression weights. The bias term is superimposed to correct the benchmark and output a continuous scalar to obtain the target osteoporosis bone mineral density quantitative prediction value.

[0020] This solution specifically establishes a correlation mapping between clinical scalars and the imaging numerical domain. It extracts the volumetric bone mineral density value obtained by QCT detection and sets it as the target true value, with the quantification unit being milligrams per cubic centimeter. It adaptively weights and maps the key factors of bone mineral density to the imaging numerical space. It uses a bias convergence mechanism to calibrate the logical distance between the distribution of clinical features and the distribution of imaging features. Then, by inputting the patient's clinical scalar, the corresponding imaging-level bone mineral density value can be directly inferred, realizing the acquisition of imaging-level parameters without hardware scanning.

[0021] The multi-channel clinical semantic embedding matrix includes age mapping vector, body mass index mapping vector, alkaline phosphatase mapping vector, and osteocalcin mapping vector. The intra-group local strong correlation feature map includes the set of over-threshold response values ​​and the set of zero-value response values. The cross-domain global collaborative semantic tensor includes context-aware latent vectors and the set of feature values ​​after weight adjustment. The adaptive weighted representation of key bone mineral density factors includes global average pooling results, channel calibration coefficients, and element-wise multiplication results. The one-dimensional regression feature vector includes the flattened and expanded feature value sequence.

[0022] The specific steps for generating a multi-channel clinical semantic embedding matrix are as follows: Based on the subjects' original clinical data, age, body mass index, alkaline phosphatase and osteocalcin scalars are directly extracted, and each scalar is projected into a high-dimensional continuous space by inputting a preset weight matrix. The values ​​are then expanded into multi-dimensional feature vectors to generate independent clinical indicator feature vector groups. Based on the feature vector group of independent clinical indicators, each vector is assigned to a preset demographic, bone metabolism and hormone regulation group. Vectors in the same group are concatenated to construct a local semantic channel. The channels are stacked and spliced ​​to form a three-dimensional feature tensor, generating a multi-channel clinical semantic embedding matrix. Based on the subjects' original clinical data, a fixed-parameter linear mapping method was used. The input values ​​were set as follows: age 65 years, body mass index 23.5 kg / m², alkaline phosphatase 120 units / liter, and osteocalcin 18 ng / ml. Linear multiplication and addition were performed according to the preset weight matrix parameters. The weight matrix had 4 rows and 128 columns, with each column corresponding to an output dimension. Each weight element was set to a specific floating-point value between 0.010 and 0.030, such as 0.012, 0.018, 0.021, and 0.027. The bias parameter was uniformly set to 0.005. Multiplication and addition were performed item by item according to the input variable order, and the bias value was superimposed to expand each scalar into a 128-dimensional continuous numerical sequence. The precision of each dimension's output value was retained to 4 decimal places, forming four groups of 128-dimensional numerical vectors. These vectors were independently numbered according to the variable category to generate independent clinical indicator feature vector groups. Based on independent clinical indicator feature vector groups, a channel grouping and stacking method was adopted to classify age vectors and body mass index vectors into the population group, alkaline phosphatase vectors into the bone metabolism group, and osteocalcin vectors into the hormone regulation group. The population group was sequentially concatenated to form a 2-row, 128-column numerical structure, while the bone metabolism group and the hormone regulation group maintained a 1-row, 128-column structure. The three groups of data were stacked in the order of population group first, bone metabolism group in the middle, and hormone regulation group last, constructing a 4-row, 128-column data structure with a depth of 1. The number of channels was set to 3, the feature length was fixed at 128, and the numerical type was uniformly set to 32-bit floating-point format, generating a multi-channel clinical semantic embedding matrix.

[0023] The specific steps for outputting the local strongly correlated feature map within the group are as follows: Based on the multi-channel clinical semantic embedding matrix, the weight kernel matrix of the convolutional neural network is configured to cover the adjacent feature vectors. The product of the kernel parameter and the corresponding vector element is calculated. The local interaction values ​​are obtained by traversing the entire length of the channel and generating local interaction slices of adjacent variables within the channel. Based on the local interactive slices of adjacent variables within the channel, the product values ​​within the slice are accumulated to calculate the local cooperative response intensity at the window position. The continuous response sequence of each slice position is obtained by aggregating the response sequences of each channel to construct a multidimensional response tensor and generate a multidimensional local cooperative response numerical set. Based on a multi-dimensional local collaborative response numerical set, a preset activation threshold bias parameter is introduced. Each response value is compared with the threshold bias. Positive response values ​​greater than the threshold are retained, and negative response values ​​less than the threshold are reset to zero, generating a local strong correlation feature map within the group. Based on a multi-channel clinical semantic embedding matrix, a one-dimensional convolution calculation method is adopted. The input tensor size is set to 4 rows, 128 columns, and 1 depth. The number of convolution kernels is set to 8, the kernel width is 3, the stride is set to 1, and the padding method is set to zero padding in 1 column. The convolution kernel parameters are initialized with a fixed numerical matrix. Each convolution kernel contains 3 consecutive positional parameters with values ​​set to 0.015, 0.020, and 0.025 respectively. The bias parameter is uniformly set to 0.010. A sliding cover operation is performed on each channel, moving sequentially from column 1 to column 126. For each window position, 3 consecutive columns of values ​​are extracted and multiplied element-wise with the corresponding convolution kernel parameters. The 3 multiplied values ​​are accumulated and the bias value is superimposed. The calculation results of each window position are recorded and arranged in the order of window movement to form a response sequence of 126 columns. The same steps are performed on all 4 channels to generate local interaction slices of adjacent variables within the channel. Based on local interactive slicing of adjacent variables within a channel, a window accumulation aggregation method is adopted. The 126 response values ​​in each channel are read item by item, and the summation operation is performed on the three product results corresponding to each window position to form a local collaborative response intensity value at a single window position. The 126 window position values ​​are stored sequentially according to the channel order to form a continuous response sequence. After generating the corresponding response sequences for each of the four channels, a 4-row 126-column structure is constructed according to the channel order. The structure is then expanded into a 4-row 126-column 1-depth data form to generate a multi-dimensional local collaborative response value set. Based on a multi-dimensional local collaborative response numerical set, a threshold comparison screening method is adopted. The preset activation threshold bias parameter is set to 0.5. The threshold value is obtained by averaging all response values ​​in the training samples to 0.48 and adding a bias of 0.02. Element-by-element comparison operation is performed on all response values ​​in 4 rows and 126 columns. Values ​​greater than 0.5 are retained, and values ​​less than or equal to 0.5 are directly assigned the value 0. The assignment precision is uniformly set to three decimal places. The same processing is performed on all channels to keep the matrix dimension unchanged and generate a local strong correlation feature map within the group.

[0024] The convolutional neural network constructs a network weight kernel matrix with a preset kernel width and input channel depth. The weight kernel matrix is ​​mapped to the feature dimension of the multi-channel clinical semantic embedding matrix. The spatial position of the weight kernel matrix is ​​adjusted to match the current processing step size, so that the parameter window of the weight kernel matrix is ​​precisely aligned in dimension and covers a set of continuous adjacent feature vectors in the channel.

[0025] The specific steps to obtain the cross-domain global collaborative semantic tensor are as follows: Based on the local strong correlation feature map within the group, a feature tensor is constructed by stacking the map along the depth. The channel data is linearly weighted and summed using a unit-size kernel. The heterogeneous semantic data stream is fused, and the independent channels are reorganized into high-dimensional entities. The numerical mapping between semantic groups is established, and a multi-semantic fusion linear feature body is generated. Based on multi-semantic fusion linear feature volume, the spatial dimension is compressed to obtain the channel global descriptor, which drives the long short-term memory network to evolve memory units. The gating mechanism is used to model long-range dependencies between channels, outputting context-aware latent vectors, normalizing values ​​to construct attention weights, and generating global context dependency weight vectors. Based on the global context-dependent weight vector, the weight dimension is expanded along the multi-semantic fusion linear feature volume space, the weight and feature values ​​are expanded by element-wise multiplication, the channel response intensity is recalibrated, the heterogeneous feature distribution is calibrated and the feature response amplitude is adjusted, and a cross-domain global collaborative semantic tensor is generated. Based on the intra-group local strong correlation feature map, a tensor stacking and unit convolution kernel linear weighting method is adopted. The input data size is set to 4 rows and 126 columns with 1 depth. The depth direction stacking operation is performed according to the channel order to form a 4-row, 126-column, 4-depth structure. The unit size convolution kernel parameter is set to 1 row and 1 column with 4 depths. The convolution kernel parameter values ​​are set to 0.250, 0.250, 0.250, and 0.250 respectively, and the bias parameter is set to 0.000. For each spatial position, a channel-wise multiplication and summation operation is performed. The 4 depth values ​​are multiplied element-wise with the corresponding convolution kernel parameters and summed. The same steps are performed sequentially for all positions of 126 columns to form a 4-row, 126-column single-depth numerical structure. The data arrangement is reorganized according to the channel order to establish a unified high-dimensional entity structure and generate a multi-semantic fusion linear feature body. Based on multi-semantic fusion linear feature bodies, a Long Short-Term Memory (LSTM) network computation method is adopted. The input sequence length is set to 126, the input dimension to 4, and the number of hidden units to 32. The initial values ​​of the forget gate weight matrix are set to 0.010 to 0.030, the input gate weight matrix to 0.015 to 0.035, and the output gate weight matrix to 0.020 to 0.040. The initial cell state value is set to 0.000. The global descriptor values ​​of the input channels are input in the order of time steps 1 to 126. At each time step, the forget gate is linearly multiplied and added with a bias of 0.005, the input gate is linearly multiplied and added with a bias of 0.006, the candidate state is updated with a bias of 0.004, the output gate is linearly multiplied and added with a bias of 0.003, the cell state and hidden state vectors are updated, all values ​​of the 32-dimensional hidden state vector are summed and divided by 32 to obtain a single-channel descriptor, the four-channel descriptors are normalized, and the values ​​of each channel are divided by the sum of all channel values ​​to generate a global context-dependent weight vector. Based on the global context-dependent weight vector, a weight expansion and element-wise multiplication method is adopted. The weight vector length is set to 4. The four weight values ​​are copied 126 times to form a 4-row, 126-column weight matrix. The element-wise multiplication operation is performed on the corresponding 4-row, 126-column values ​​in the multi-semantic fusion linear feature body. The multiplication precision is retained to 4 decimal places. The value replacement operation is performed on all positions to keep the original row and column structure unchanged. The parameter 1 is re-inserted into the depth dimension to form a 4-row, 126-column, 1-depth data structure, generating a cross-domain global collaborative semantic tensor.

[0026] The Long Short-Term Memory (LSTM) network uses the channel global descriptor obtained by compressing the spatial dimension as the input of the current time step and maps it to the recursive unit of the network. It performs gating operations by combining the hidden state vector of the previous time step with the cell state vector. The retention ratio of historical cell states is determined by the forget gate parameter. The input gate parameter controls the writing weight of the current channel descriptor to update the current cell state. The output gate parameter performs nonlinear mapping based on the updated cell state to generate the context-aware hidden vector of the current time step.

[0027] The specific steps for constructing an adaptive weighted representation of key bone mineral density factors are as follows: Based on the strong local correlation feature map within the group and the cross-domain global collaborative semantic tensor, the tensor space dimension and channel number are aligned, the sum of each element of the corresponding position is calculated, local details and global context signals are superimposed, heterogeneous information flow is integrated into a unified data structure, a full-scale feature container is constructed, and a dual-path aggregated semantic feature body is generated. Based on the dual-path aggregated semantic feature body, the global average value of the spatial axis is calculated, the feature dimension is compressed to obtain the channel descriptor, the descriptor is mapped to the nonlinear activation interval, the channel calibration coefficient is output, the coefficient is multiplied element by element with the feature body value, the channel response intensity distribution is recalibrated, and an adaptive weighted representation of the key bone density factor is generated. Based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor, an element-wise aligned accumulation method is adopted. The size of the two types of input tensors is set to 4 rows and 126 columns with a depth of 1. The number of rows and columns of the tensor is checked for consistency to confirm that the number of channels is 4 and the column length is 126. The element-wise reading operation is performed on the position from the 1st column to the 126th column of each row. The two values ​​at the corresponding positions are added and 4 decimal places are retained. The accumulation result is written into a new 4-row and 126-column data structure. The same steps are performed on all 4 rows in sequence, keeping the depth dimension unchanged. The generated data structure is uniformly numbered and stored to generate a dual-path aggregated semantic feature body. Based on the dual-path aggregated semantic feature body, a global averaging and Sigmoid mapping method is adopted. The spatial axial averaging is performed row by row on the 4-row, 126-column data structure. The 126 values ​​in each row are summed and divided by 126 to obtain the values ​​of the four channel descriptors. The slope of the Sigmoid function is set to 1.000 and the bias parameter is set to 0.000. The values ​​between 0 and 1 are calculated by substituting the four channel descriptors into the Sigmoid function to obtain the channel calibration coefficients. The four channel calibration coefficients are copied 126 times to form a 4-row, 126-column weight matrix. Element-wise multiplication is performed on the corresponding values ​​in the dual-path aggregated semantic feature body and four decimal places are retained. The entire 4-row, 126-column data is overwritten and updated, keeping the original depth dimension of 1, to generate an adaptive weighted representation of the key bone density factors.

[0028] The specific steps to obtain the target osteoporosis bone mineral density quantitative prediction value are as follows: Based on the adaptive weighted representation of key bone mineral density factors, the multidimensional tensor structure is reshaped into a single linear dimension. Feature values ​​are concatenated in channel order to construct a continuous one-dimensional numerical array. The vector length is aligned to the number of fully connected nodes and the geometric constraints of the tensor space are removed to generate a globally flattened regression feature vector. Based on the global flattened regression feature vector, the product of the vector elements and the regression weights is calculated, the weighted values ​​are accumulated to obtain the sum of the linear responses, the scalar bias term is superimposed to correct the prediction benchmark, the feature values ​​are mapped to the continuous physical real number domain, the bone mineral density scalar is output, and the target osteoporosis bone mineral density quantitative prediction value is generated. Based on the adaptive weighted representation of key bone density factors, a sequential rearrangement and flattening method is adopted. The input data structure is set to 4 rows, 126 columns, and 1 depth. A deletion operation is performed on the depth dimension. The 4 rows and 126 columns of data are read linearly in channel order. First, the values ​​of the first row and the first column to the 126th column are read and written to the target array in sequence. Then, the values ​​of the second row and the first column to the 126th column are read and written to the target array in sequence. The values ​​of the third and fourth rows are read and written in the same order to construct a one-dimensional numerical array with a length of 504. The number of fully connected nodes is set to 504. A length consistency check is performed to confirm that the vector length is equal to 504. Index numbers 0 to 503 are applied to all 504 positions and a continuous linear storage structure is established to generate a global flattening regression feature vector. Based on the globally flattened regression feature vector, a linear regression calculation method is adopted. The length of the regression weight array is set to 504, the weight values ​​are set to a fixed floating-point number between 0.001 and 0.010, and the bias parameter is set to 0.050. The values ​​at positions 0 to 503 in the vector are multiplied element-wise with the corresponding weight values. The 504 product results are sequentially accumulated to obtain a single linear response value. The accumulated result is superimposed with the bias parameter 0.050 to form the final value. The value is retained to 4 decimal places and the output range is limited to between 0.300 and 1.500. A single continuous real scalar is output to generate the target osteoporosis bone mineral density quantitative prediction value.

[0029] This solution performs cross-domain extrapolation from clinical indicators to the numerical domain of imaging data. Specific steps include: performing numerical calibration of the target domain of imaging; obtaining the raw volumetric bone mineral density (BMD) values ​​of the lumbar spine region from QCT and setting them as the target ground truth label; standardizing the physical quantification to milligrams per cubic centimeter and limiting the value range to 50-250; constructing a cross-modal manifold mapping structure; inputting a 128-dimensional cross-domain global collaborative semantic tensor into the mapping layer; calling a mapping matrix with 128 rows and 1 column to perform a dot product operation and adding a bias parameter of 0.025 for numerical calibration; and implementing image correlation characterization. The model uses mean squared error logic to calculate the deviation between clinically extrapolated scalars and QCT measured scalars. When the deviation value is greater than 0.01, it triggers parameter gradient backpropagation and adjusts the convolution kernel parameters and recursive unit weights. This drives the nonlinear combination of clinical variables within the model to converge to the image domain feature distribution, producing image-level extrapolation results. During the inference stage, multiple scalars such as age, body mass index, and metabolic indicators are input. Through high-dimensional mapping and cross-domain feature alignment processing, it produces a quantitative prediction value of bone mineral density that matches the accuracy of QCT images. Specifically, it infers the patient's image-level bone mineral density data by processing clinical data without starting the hardware scanning program.

[0030] The steps to obtain the target quantitative prediction value are also related to the image numerical extrapolation process. Specifically, the volumetric bone mineral density value detected by QCT is obtained and set as the target true value label. The key factors of bone mineral density are adaptively weighted and projected onto the image numerical domain. The deviation between the extrapolated scalar and the QCT measured scalar is calculated. The feature distribution is calibrated through the gradient backpropagation algorithm until the deviation is less than the preset threshold. Thus, the corresponding image-level bone mineral density value is inferred through clinical data input.

[0031] Please see Figure 2 A system for constructing a prediction model for osteoporosis bone mineral density based on clinical data, the system comprising: Semantic construction module: Based on raw clinical data, read age, body mass index, alkaline phosphatase and osteocalcin scalars and classify them, perform vector expansion and channel splicing and arrangement, establish a three-dimensional continuous numerical structure, and generate a multi-channel clinical semantic embedding matrix. Local modeling module: Based on the multi-channel clinical semantic embedding matrix, channel window arrangement and response calculation are performed, thresholds are compared and values ​​are filtered, responses exceeding the threshold are retained and low-value data are set to zero, a multi-channel response structure is constructed, and a local strong correlation feature map within the group is obtained; Global Collaboration Module: Based on the strong local correlation feature map within the group, deep overlay and channel integration processing are performed. The recursive structure is input to update the state. The channel weight ratio is adjusted according to the latent vector to establish a recalibrated feature structure and obtain a cross-domain global collaborative semantic tensor. The calibration weighting module: Based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor, it performs dimensional alignment and numerical superposition to generate an aggregated feature body, calculates the channel calibration coefficients and multiplies them element by element, and constructs an adaptive weighted representation of key bone density factors. Regression output module: Based on the adaptive weighted representation of key bone mineral density factors, it calculates the linear expansion and weight product, accumulates the values ​​and superimposes the bias term, maps to the continuous numerical domain, and outputs a single scalar to obtain the quantitative prediction value of target osteoporosis bone mineral density.

[0032] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing a model to predict osteoporosis bone mineral density based on clinical data, characterized in that, Includes the following steps: S1: Extract age, body mass index, alkaline phosphatase and osteocalcin scalars from raw clinical data, perform mapping operations to convert them into high-dimensional dense vectors, perform channel splicing based on demographics, bone metabolism and hormone categories to generate a multi-channel clinical semantic embedding matrix. S2: Based on the multi-channel clinical semantic embedding matrix, perform grouped sliding window operation, call the convolutional neural network weight kernel to calculate the sum of the dot products of variables within the channel, compare the collaborative response value with the activation threshold, retain the signal exceeding the threshold and reset the low threshold value to zero, and output the local strong correlation feature map within the group. S3: Based on the local strongly correlated feature maps within the group, construct a multi-semantic feature body by stacking along the depth, perform linear combination operations between channels, drive the long short-term memory network to output context-aware latent vectors, and dynamically adjust the weight ratio of feature channels according to the vectors to obtain a cross-domain global collaborative semantic tensor. S4: Based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor, the aggregated semantics are aggregated by superimposing the tensor, and global average pooling is performed on the aggregated feature body. The channel calibration coefficient is calculated and the coefficient is multiplied element-wise with the aggregated feature body to construct an adaptive weighted representation of the key factors of bone density. S5: Based on the adaptive weighted representation of the key bone mineral density factors, the feature vector is flattened into a one-dimensional regression feature vector. The feature vector elements and regression weights are linearly weighted and summed in the input mapping layer. The bias term is superimposed to correct the benchmark and output a continuous scalar to obtain the target osteoporosis bone mineral density quantitative prediction value.

2. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The multi-channel clinical semantic embedding matrix includes an age mapping vector, a body mass index mapping vector, an alkaline phosphatase mapping vector, and an osteocalcin mapping vector. The intra-group local strong correlation feature map includes a set of over-threshold response values ​​and a set of zero-value response values. The cross-domain global collaborative semantic tensor includes a context-aware latent vector and a set of feature values ​​adjusted by weights. The adaptive weighted representation of key bone mineral density factors includes global average pooling results, channel calibration coefficients, and element-wise multiplication results. The one-dimensional regression feature vector includes a flattened and expanded sequence of feature values.

3. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The specific steps for generating the multi-channel clinical semantic embedding matrix are as follows: Based on the subjects' original clinical data, age, body mass index, alkaline phosphatase and osteocalcin scalars are directly extracted, and each scalar is projected into a high-dimensional continuous space by inputting a preset weight matrix. The values ​​are then expanded into multi-dimensional feature vectors to generate independent clinical indicator feature vector groups. Based on the independent clinical indicator feature vector group, each vector is assigned to a preset demographic, bone metabolism and hormone regulation group. Vectors in the same group are concatenated to construct a local semantic channel. The channels are stacked and spliced ​​to form a three-dimensional feature tensor, generating a multi-channel clinical semantic embedding matrix.

4. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The specific steps for outputting the local strongly correlated feature map within the group are as follows: Based on the multi-channel clinical semantic embedding matrix, a convolutional neural network weight kernel matrix is ​​configured to cover adjacent feature vectors, the kernel parameter is multiplied by the corresponding vector element, the full length of the channel is traversed to obtain local interaction values, and local interaction slices of adjacent variables within the channel are generated. Based on the local interactive slices of adjacent variables within the channel, all product values ​​within the slice are accumulated, the local cooperative response intensity at the window position is calculated, the continuous response sequence of each slice position is obtained by aggregating the response sequences of each channel, and a multidimensional response tensor is constructed by recombining the response sequences of each channel to generate a multidimensional local cooperative response value set. Based on the multi-dimensional local collaborative response value set, a preset activation threshold bias parameter is introduced. Each response value is compared with the threshold bias, positive response values ​​greater than the threshold are retained, and negative response values ​​less than the threshold are reset to zero, thereby generating a local strong correlation feature map within the group.

5. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The convolutional neural network constructs a network weight kernel matrix with a preset kernel width and input channel depth. The weight kernel matrix is ​​mapped to the feature dimension of the multi-channel clinical semantic embedding matrix. The spatial position of the weight kernel matrix is ​​adjusted to match the current processing step size, so that the parameter window of the weight kernel matrix is ​​precisely aligned in dimension and covers a set of continuous adjacent feature vectors in the channel.

6. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The specific steps to obtain the cross-domain global collaborative semantic tensor are as follows: Based on the local strongly correlated feature maps within the group, a feature tensor is constructed by stacking the maps along the depth. The channel data is linearly weighted and summed using a unit-size kernel. The heterogeneous semantic data stream is fused, and the independent channels are reorganized into high-dimensional entities. The numerical mapping between semantic groups is established, and a multi-semantic fusion linear feature body is generated. Based on the multi-semantic fusion linear feature body, the spatial dimension is compressed to obtain the channel global descriptor, driving the long short-term memory network to evolve the memory unit, using the gating mechanism to model the long-range dependency between channels, outputting the context-aware latent vector, normalizing the values ​​to construct attention weights, and generating the global context dependency weight vector. Based on the global context-dependent weight vector, the weight dimension is expanded along the multi-semantic fusion linear feature volume space, the weight and feature values ​​are expanded by element-wise multiplication, the channel response intensity is recalibrated, the heterogeneous feature distribution is calibrated and the feature response amplitude is adjusted, and a cross-domain global collaborative semantic tensor is generated.

7. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network uses the channel global descriptor obtained by compressing the spatial dimension as the input of the current time step and maps it to the recursive unit of the network. It performs gating operations by combining the hidden state vector of the previous time step with the cell state vector. The retention ratio of historical cell states is determined by the forget gate parameter. The writing weight of the current channel descriptor is controlled by the input gate parameter to update the current cell state. The nonlinear mapping is performed based on the updated cell state by the output gate parameter to generate the context-aware hidden vector of the current time step.

8. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The specific steps for constructing the adaptive weighted representation of the key bone mineral density factors are as follows: Based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor, align the tensor space dimension and channel number, calculate the sum of each element at the corresponding position, superimpose local details and global context signals, integrate heterogeneous information streams into a unified data structure, construct a full-scale feature container, and generate a dual-path aggregated semantic feature body. Based on the dual-path aggregated semantic feature body, the global average value of the spatial axis is calculated, the feature dimension is compressed to obtain the channel descriptor, the descriptor is mapped to the nonlinear activation interval, the channel calibration coefficient is output, the coefficient is multiplied element by element with the feature body value, the channel response intensity distribution is recalibrated, and an adaptive weighted representation of the key bone density factor is generated.

9. The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to claim 1, characterized in that, The specific steps for obtaining the target osteoporosis bone mineral density quantitative prediction value are as follows: Based on the adaptive weighted representation of the key bone density factors, the multidimensional tensor structure is reshaped into a single linear dimension. Feature values ​​are concatenated in channel order to construct a continuous one-dimensional numerical array. The vector length is aligned to the number of fully connected nodes and the geometric constraints of the tensor space are removed to generate a globally flattened regression feature vector. Based on the global flattened regression feature vector, the product of the vector elements and the regression weights is calculated, the weighted values ​​are accumulated to obtain the sum of the linear responses, the scalar bias term is superimposed to correct the prediction benchmark, the feature values ​​are mapped to the continuous physical real number domain, the bone density scalar is output, and the target osteoporosis bone density quantitative prediction value is generated.

10. A system for constructing a predictive model for osteoporosis bone mineral density based on clinical data, characterized in that, The method for constructing an osteoporosis bone mineral density prediction model based on clinical data according to any one of claims 1-9, wherein the system comprises: Semantic construction module: Based on raw clinical data, read age, body mass index, alkaline phosphatase and osteocalcin scalars and classify them, perform vector expansion and channel splicing and arrangement, establish a three-dimensional continuous numerical structure, and generate a multi-channel clinical semantic embedding matrix. Local modeling module: Based on the multi-channel clinical semantic embedding matrix, channel window arrangement and response calculation are performed, thresholds are compared and values ​​are filtered, responses exceeding the threshold are retained and low-value data are set to zero, a multi-channel response structure is constructed, and a local strong correlation feature map within the group is obtained; Global Collaboration Module: Based on the local strongly correlated feature maps within the group, deep overlay and channel integration processing are performed, the recursive structure update state is input, the channel weight ratio is adjusted according to the latent vector, a recalibrated feature structure is established, and a cross-domain global collaborative semantic tensor is obtained. The calibration weighting module performs dimensional alignment and numerical superposition based on the intra-group local strong correlation feature map and the cross-domain global collaborative semantic tensor to generate an aggregated feature body, calculates the channel calibration coefficients and multiplies them element by element to construct an adaptive weighted representation of key bone density factors. Regression output module: Based on the adaptive weighted representation of the key bone mineral density factors, calculate the linear expansion and weight product, accumulate the values ​​and superimpose the bias term, map to the continuous numerical domain, and output a single scalar to obtain the target osteoporosis bone mineral density quantitative prediction value.

Images