Method for predicting primary osteoporosis risk based on bone metabolism markers and clinical indicators

Abstract

This invention relates to the fields of medical artificial intelligence and disease risk prediction technology, and discloses a method for predicting the risk of primary osteoporosis based on bone metabolism biomarkers and clinical indicators. This invention aims to address the problems of existing prediction methods' inability to capture the pathogenic interaction between macroscopic clinical and microscopic biochemical indicators, and the low sensitivity to identifying hidden severe cases due to extreme sample imbalance. This invention constructs a core feature matrix by cross-extracting raw clinical indicators and bone metabolism biomarkers, utilizes a physically isolated dual-channel parallel mapping network and a cross-channel self-attention mechanism to achieve high-dimensional weighted fusion of heterogeneous features, and combines a classification mechanism with imbalanced bias weights to correct the initial risk probability.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and disease risk prediction technology, specifically, a method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators. Background Technology

[0002] Osteoporosis, a systemic degenerative disease characterized by low bone mass and damage to bone microstructure, requires early and accurate warning and long-term intervention for significant public health value in reducing the incidence of fragility fractures. Currently, risk prediction for primary osteoporosis mainly relies on traditional clinical assessment tools (such as the FRAX scale) or bone mineral density screening based on dual-energy X-ray absorptiometry (DXA). With the development of medical digitalization and artificial intelligence technologies, existing technologies are gradually beginning to incorporate machine learning algorithms to comprehensively assess patients by capturing clinical signs and some blood biochemistry test data. In conventional data processing workflows, existing systems typically perform simple linear concatenation of multi-source, multi-dimensional medical data, or directly input the concatenated multi-dimensional data as a whole into a conventional prediction model to output risk probabilities.

[0003] However, in real-world clinical epidemiological scenarios, the pathogenic mechanisms of bone loss are extremely complex, often resulting from a deep cross-coupling between macroscopic physical conditions (such as age, weight, and daily weight-bearing habits) and microscopic metabolic abnormalities (such as bone metabolism markers and calcium-phosphorus metabolism imbalances). Furthermore, real-world clinical sample data typically exhibits an extremely unbalanced distribution, with the vast majority being healthy individuals or those experiencing mild bone loss, while patients with subtle, high-risk characteristics constitute an absolute minority. Existing risk prediction methods struggle to accurately capture and quantify the complex nonlinear pathogenic interactions between macroscopic clinical signs and microscopic biochemical markers. Moreover, when faced with severely imbalanced medical samples, conventional prediction models are easily dominated by the large majority of healthy samples, leading the system to prioritize overall apparent accuracy at the expense of significantly reducing its sensitivity to identifying the minority of patients with subtle, severe bone loss. The lack of heterogeneous feature fusion capability and the resulting high-risk missed diagnosis problem severely restrict the clinical reliability of prediction results, making it difficult for existing technologies to truly support a closed loop of high-precision automated medical decision-making throughout the entire cycle, from in-hospital emergency intervention to early risk prevention and control outside the hospital. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators, which solves the problems of existing prediction methods being unable to capture the pathogenic interaction between macroscopic clinical and microscopic biochemical indicators, and the low sensitivity to identifying hidden severe cases due to extreme sample imbalance.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators, which includes the following steps: S1. Obtain the patient's original clinical indicators and original bone metabolism markers, fill in missing values ​​and standardize the data scale to obtain a standard dataset; S2. Perform cross-operation between the clinical indicators and bone metabolism markers in the standard dataset to extract associated features, and remove redundant data based on correlation analysis to generate a core feature matrix. S3. Perform dual-channel parallel mapping on the core feature matrix to extract nonlinear representations of clinical features and bone metabolism features respectively, and evaluate the interaction pathogenicity of the clinical features and bone metabolism features through a cross-channel self-attention mechanism. Dynamically allocate weights according to the interaction pathogenicity to complete feature fusion and generate a high-order fused feature vector. S4. Perform classification mapping on the high-order fusion feature vector, and output the corresponding risk assessment result based on the mapping result.

[0007] As a preferred embodiment of the method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to the present invention, wherein: the original clinical indicators obtained in S1 include basic human physiological characteristic data and lifestyle data related to bone load, and the original bone metabolism markers include bone formation markers and calcium and phosphorus metabolism-related markers. The missing value imputation step in S1 specifically involves using the nearest neighbor interpolation algorithm to imput missing values. The missing value imputation step using the nearest neighbor interpolation algorithm specifically includes: extracting various known indicators of patients with missing data, matching a preset number of reference patients whose values ​​of various known indicators are closest in the historical database, calculating the average value of the missing indicators corresponding to the preset number of reference patients, and filling the missing values ​​into the missing positions. The specific steps for unifying the data scale include: taking each of the original clinical indicators and the original bone metabolism markers after filling in the missing values ​​as the current processing item, subtracting the corresponding average value in all samples from the value of the current processing item, and dividing the resulting difference by the corresponding standard deviation to convert it into a standard value.

[0008] As a preferred embodiment of the method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to the present invention, wherein: the step of performing cross-operation between the clinical indicators and bone metabolism markers in the standard dataset in S2 to extract associated features specifically includes: performing cross-multiplication of each clinical indicator in the standard dataset with each bone metabolism marker one by one to generate a new feature containing all cross-category combinations, and merging the new feature with the original indicators in the standard dataset to construct a set of candidate features.

[0009] As a preferred embodiment of the method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to the present invention, the specific steps in S2 for generating a core feature matrix by removing redundant data based on correlation analysis include: The Pearson correlation coefficient between the first feature and the second feature is obtained by calculating the sum of the products of the differences between any first feature and the second feature in the candidate feature set and their respective deviations from the mean of the corresponding samples, and by dividing the sum of the products of the differences by the product of the square roots of the sum of the squares of the deviations of the first feature and the second feature from the mean of the corresponding samples. If the Pearson correlation coefficient between the first feature and the second feature is determined to be greater than a preset correlation threshold, then the feature with the larger variance between the first feature and the second feature is retained, and the feature with the smaller variance is removed; the removal step is repeated until the Pearson correlation coefficient between the retained features is not greater than the preset correlation threshold, and all the retained features are combined to generate the core feature matrix.

[0010] As a preferred embodiment of the method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to the present invention, the step of performing dual-channel parallel mapping on the core feature matrix in S3 to extract the nonlinear representations of clinical features and bone metabolism features respectively includes: decomposing the core feature matrix into mutually independent clinical feature sub-matrices and bone metabolism feature sub-matrices. The clinical feature sub-matrix is ​​input into the first processing channel and nonlinearly superimposed and mapped through multiple layers of neuronal nodes to extract deep clinical signals that characterize the external physical condition while avoiding interference from biochemical data. Simultaneously, the bone metabolism feature sub-matrix is ​​input into a second processing channel that is physically isolated from the first processing channel. After nonlinear superposition mapping through multiple layers of neuronal nodes, deep biochemical signals characterizing the internal metabolic state are extracted under the condition of avoiding the suppression of macroscopic signs.

[0011] As a preferred embodiment of the primary osteoporosis risk prediction method based on bone metabolism markers and clinical indicators according to the present invention, the step of S3, which assesses the pathogenicity of the interaction between the clinical feature and the bone metabolism feature through a cross-channel self-attention mechanism and dynamically assigns weights according to the pathogenicity of the interaction to complete feature fusion, specifically includes: The deep clinical signals are mapped to a query matrix, and the deep biochemical signals are mapped to a key matrix and a value matrix, respectively. Calculate the dot product of the query matrix and the transpose of the key matrix, and divide by the scaling factor of the key vector dimension to obtain the cross-channel association matching score; Signal combinations with abnormal pathogenicity risk are extracted, and computational weights higher than the conventional standard are assigned to these signal combinations. The computational weights are generated by processing the cross-channel association matching scores through a normalized exponential function. Based on the assigned computational weights, the computational weights are multiplied by the value matrix containing biochemical features, and the weighted fusion is used to generate the higher-order fusion feature vector.

[0012] As a preferred embodiment of the primary osteoporosis risk prediction method based on bone metabolism markers and clinical indicators described in this invention, the step of classifying and mapping the higher-order fusion feature vector in S4 specifically includes: The high-order fused feature vector is nonlinearly classified using a classifier with an internal imbalanced bias weight. The imbalanced bias weight is achieved by introducing a dynamic adjustment factor based on the initial classification probability into the cross-entropy loss function of the classifier. If the initial classification probability of the classifier for highly concealed severe osteoporosis samples decreases, the dynamic adjustment factor is amplified exponentially, increasing the loss gradient weight of the classifier for the highly concealed severe osteoporosis samples during backpropagation. This amplifies the recognition sensitivity of the high-order fusion feature vector reflecting the concealed features of severe osteoporosis, corrects the initial classification probability obtained from the classification mapping, and outputs the initial risk value after probability correction.

[0013] As a preferred embodiment of the method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to the present invention, the step of outputting the corresponding risk assessment result according to the mapping result in S4 specifically includes: mapping the initial risk value into a predicted score within a preset score range. A warning judgment threshold is set, which includes a first risk threshold and a second risk threshold, wherein the first risk threshold is less than the second risk threshold; If the predicted score is less than the first risk threshold, it is determined to be a low-risk level; if the predicted score is not less than the first risk threshold and not greater than the second risk threshold, it is determined to be a medium-risk level; if the predicted score is greater than the second risk threshold, it is determined to be a high-risk level, and the corresponding risk level label is output.

[0014] As a preferred embodiment of the primary osteoporosis risk prediction method based on bone metabolism markers and clinical indicators described in this invention, the step of outputting the corresponding risk assessment result based on the mapping result further includes: If the risk level label is high risk, a data packet containing a bone condition specialist follow-up appointment request will be automatically triggered and generated. At the same time, a standardized intervention report containing targeted bone metabolism intervention recommendations will be generated. The data packet and the intervention report will then be merged and pushed to an external application terminal to execute medical intervention instructions.

[0015] As a preferred embodiment of the method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to the present invention, the step of outputting the corresponding risk assessment result based on the mapping result further includes generating a long-term health management strategy. The step of generating the long-term health management strategy specifically includes: Obtain the predicted scores over multiple consecutive historical assessment periods, and fit them to generate a sequence map that reflects the dynamic trend of risk changes; Based on the risk level labels, a daily care plan containing lifestyle intervention parameters and non-pharmacological management strategies is generated. The sequence map is combined with the daily care plan and then pushed to an external application terminal to provide chronic disease intervention prompts.

[0016] The beneficial effects of this invention are as follows: This invention effectively improves the ability of risk prediction models to quantify complex pathogenic factors and the sensitivity to identify a small number of hidden high-risk samples, reduces the clinical missed diagnosis rate, and realizes an automated hierarchical intervention and execution closed loop from in-hospital clinical auxiliary decision-making to out-of-hospital long-term health management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The flowchart shows the overall process for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators. Figure 2This is a schematic diagram of a dual-channel parallel and cross-channel feature fusion architecture provided in an embodiment of the present invention; Figure 3 A flowchart illustrating the automated execution of high-risk clinical auxiliary decision-making within a hospital, as provided in this embodiment of the invention. Figure 4 A flowchart illustrating the tiered implementation of long-term non-drug management for the entire outpatient population, provided for embodiments of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0022] Example 1 Reference Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators, including the following steps: S1. Obtain the patient's original clinical indicators and original bone metabolism markers, fill in missing values ​​and standardize the data scale to obtain a standard dataset; The raw clinical indicators obtained in S1 include basic human physiological characteristics data and lifestyle data related to bone load-bearing, and raw bone metabolism markers include bone formation markers and calcium and phosphorus metabolism-related markers. The missing value imputation step in S1 is specifically to use the nearest neighbor interpolation algorithm to imput missing values. The steps of using the nearest neighbor interpolation algorithm to imput missing values ​​specifically include: extracting the known indicators of patients with missing data, matching the preset number of reference patients whose values ​​of the known indicators are closest in the historical database, calculating the average value of the missing indicators corresponding to the preset number of reference patients and filling them into the missing positions. The specific steps for standardizing data scales include: taking each original clinical indicator and original bone metabolism marker after filling in missing values ​​as the current treatment item, subtracting the corresponding mean value in all samples from the value of the current treatment item, and dividing the resulting difference by the corresponding standard deviation to convert it into a standard value.

[0023] In real clinical settings, due to differences in patient compliance or varying hospital testing priorities, the acquired medical data often contains a large number of missing values. In this embodiment, the original clinical indicators may specifically include macroscopic physical data such as age, height, body mass index, years since menopause (for female patients), and frequency of daily weight-bearing exercise; the original bone metabolism markers may specifically include microscopic biochemical indicators such as total type I collagen N-terminal elongated peptide (t1NP), N-MID osteocalcin, and 25-hydroxyvitamin D.

[0024] To avoid a sharp decrease in data volume due to directly deleting samples containing missing values, this system uses the nearest neighbor interpolation algorithm for intelligent imputation. In practice, a preset number K=5 is set. Taking a patient missing N-MID osteocalcin value as an example, the system calculates the Euclidean distance between this patient and existing patients in the historical database for all known indicators such as age and body mass index. It then selects the five closest reference patients, extracts the N-MID osteocalcin values ​​from these five patients, calculates the arithmetic mean, and uses this estimated value to fill the missing position.

[0025] After imputation, because the physical dimensions of each indicator are completely different (e.g., age is measured in years, and BMI is measured in micrograms per liter, unlike vitamin D), directly inputting them into the neural network could lead to gradient explosion or model bias. Therefore, the Z-Score standardization formula is used to unify the data scale. Specifically, for any current processing item... Its corresponding standardized value The calculation formula is: in, This is the arithmetic mean of the indicator across the entire historical sample set. This represents the corresponding standard deviation. Through this step, all macroscopic clinical data and microscopic biochemical data are mapped to a standard normal distribution interval with a mean of 0 and a variance of 1, thus constructing a high-quality standard dataset.

[0026] S2. Cross-operate with clinical indicators and bone metabolism markers in the standard dataset to extract associated features, and remove redundant data based on correlation analysis to generate a core feature matrix; The steps in S2 that involve cross-operating clinical indicators and bone metabolism biomarkers in the standard dataset to extract associated features specifically include: performing cross-multiplication on each clinical indicator in the standard dataset with each bone metabolism biomarker to generate a new feature containing all cross-category combinations, and merging the new feature with the original indicators in the standard dataset to construct a set of candidate features.

[0027] The specific steps in S2 for removing redundant data based on correlation analysis and generating the core feature matrix include: The Pearson correlation coefficient between the first and second features is obtained by calculating the sum of the products of the differences between any first feature and second feature in the candidate feature set and their respective deviations from the mean of the corresponding samples, and then dividing the sum of the products of the differences by the product of the square roots of the sum of the squares of the deviations of the first feature and second feature from the mean of the corresponding samples. If the Pearson correlation coefficient between the first feature and the second feature is greater than the preset correlation threshold, then the feature with the larger variance between the first feature and the second feature is retained, and the feature with the smaller variance is removed; the removal step is repeated until the Pearson correlation coefficient between the retained features is not greater than the preset correlation threshold, and all the retained features are combined to generate the core feature matrix.

[0028] Osteoporosis often results from the coupling of macroscopic physical condition and microscopic metabolic abnormalities. Relying solely on single-dimensional indicators is insufficient to capture hidden pathogenic factors. This embodiment uses a Cartesian product-like cross-multiplication operation to forcibly construct cross-boundary associations. For example, the system multiplies the standardized age by each biochemical indicator such as t1NP and osteocalcin, generating new features such as the age_t1NP cross term. These new features, together with the original features, constitute a highly dimensional set of candidate features.

[0029] However, the dramatic increase in the number of features not only significantly increases the computational load of the system, but also introduces a large amount of redundant data carrying overlapping information, which can easily interfere with the subsequent prediction accuracy. Therefore, this system introduces the Pearson correlation coefficient matrix for dimensionality reduction and cleaning. In this embodiment, a preset correlation threshold of 0.85 is set. The system iterates through and calculates the Pearson correlation coefficient for all feature pairs. Specifically, for any two feature vectors in the candidate feature set... and (For example, feature X is the age_t1NP interaction term, and feature Y is BMI), its Pearson correlation coefficient The calculation formula is: In the formula, The total number of samples in the historical database; and They represent the first Each sample in features and characteristics The value on; and Features and characteristics The arithmetic mean of all samples.

[0030] In the overall calculation process, the formula uses the sample mean of the feature as a benchmark. By calculating the product of the deviations of the feature variables from the mean and standardizing and normalizing it, it accurately measures the degree of linear covariance between macroscopic clinical indicators and microscopic biochemical indicators, and outputs a coefficient value in the range of [-1,1]. This provides a solid mathematical basis for subsequent screening and elimination of highly overlapping redundant information.

[0031] When the correlation coefficient between feature X and feature Y reaches 0.88 (greater than the threshold of 0.85), it indicates that these two features provide statistically similar pathogenic information. At this point, the system calculates the data variance for features X and Y separately. A larger variance indicates more significant numerical fluctuations of the feature across different patient groups, and thus higher classification information entropy and stronger discriminative power. Therefore, the system automatically retains features with larger variances and ruthlessly eliminates redundant features with smaller variances. Through this iterative feature selection process, key difference data are retained while redundant interference is eliminated, ultimately constructing a low-dimensional core feature matrix containing effective classification information.

[0032] S3. Perform dual-channel parallel mapping on the core feature matrix to extract the nonlinear representations of clinical features and bone metabolism features respectively. Then, evaluate the interaction pathogenicity of clinical features and bone metabolism features through a cross-channel self-attention mechanism. Dynamically allocate weights according to the interaction pathogenicity to complete feature fusion and generate a high-order fused feature vector. The steps in S3 to perform dual-channel parallel mapping on the core feature matrix and extract nonlinear representations of clinical features and bone metabolism features respectively include: decomposing the core feature matrix into mutually independent clinical feature sub-matrices and bone metabolism feature sub-matrices; The clinical feature submatrix is ​​input into the first processing channel and nonlinearly superimposed and mapped through multiple layers of neuronal nodes to extract deep clinical signals that characterize the external physical condition while avoiding interference from biochemical data. Simultaneously, the bone metabolism feature sub-matrix is ​​input into a second processing channel that is physically isolated from the first processing channel. After nonlinear superposition mapping through multiple layers of neuronal nodes, deep biochemical signals characterizing the internal metabolic state are extracted under the condition of avoiding the suppression of macroscopic signs.

[0033] In S3, the interaction pathogenicity of clinical and bone metabolic features is assessed through a cross-channel self-attention mechanism. The specific steps for feature fusion, including dynamically assigning weights based on the interaction pathogenicity, include: Deep clinical signals are mapped to query matrices, and deep biochemical signals are mapped to key matrices and value matrices, respectively. Calculate the dot product of the query matrix and the transpose of the key matrix, and divide by the scaling factor of the key vector dimension to obtain the cross-channel association matching score; Signal combinations with abnormal pathogenicity risk are extracted, and computational weights higher than the conventional standard are assigned to these combinations. The computational weights are generated by processing the cross-channel correlation matching scores with a normalized exponential function. Based on the assigned computational weights, the computational weights are multiplied by the value matrix containing biochemical features, and a weighted fusion is performed to generate a higher-order fusion feature vector.

[0034] This step is the core of the entire prediction algorithm architecture. Traditional single-channel networks, when processing mixed data, are prone to swallowing weak bone metabolism abnormality signals during backpropagation due to the excessively strong gradient signals of macroscopic clinical features such as age and years since menopause. To address this issue, this system designs an innovative physically isolated dual-channel network. During the network input stage, the system splits the core feature matrix into two independent sub-matrices according to feature attributes. The first processing channel (clinical feature network) focuses on processing macroscopic signs, extracting pure deep clinical signals through three layers of fully connected neurons (with hidden layer dimensions set to, for example, 128-64-32) containing the ReLU nonlinear activation function. Simultaneously, the second processing channel (biochemical feature network), in a physically isolated state completely unaffected by macroscopic data interference, performs nonlinear extraction of the bone metabolism feature sub-matrix at the same depth, outputting deep biochemical signals.

[0035] After independently refining the signals from both channels, the system needs to facilitate a high-dimensional medical convergence of these two purified signals. This is where a cross-channel self-attention mechanism is introduced. The system maps deep clinical signals to query vectors and deep biochemical signals to key and value vectors, respectively. Furthermore, the system evaluates the cross-channel association matching score for each collision between clinical and biochemical representations using a self-attention calculation formula and performs feature aggregation. The specific calculation formula is as follows: In the formula, A query matrix generated for mapping deep clinical signals; The bond matrix generated for mapping deep biochemical signals. It is its transpose matrix; The value matrix generated for mapping deep biochemical signals; To determine the dimension of the key vector, we introduce... As a scaling factor to prevent the model gradient from vanishing due to excessively large inner product values; This is a normalized exponential function used to convert the matching score into attention weight coefficients that sum to 1.

[0036] In the overall calculation process, the system first goes through Matrix and The matrix dot product operation is used to calculate the original cross-channel correlation between macroscopic clinical features and microscopic biochemical features; subsequently, scaling factors and... The activation function outputs a dynamically assigned weight probability distribution; finally, this weight probability matrix is ​​compared with the value matrix containing specific biochemical features. Matrix multiplication is performed to complete high-dimensional weighted fusion operations between signals of different modes.

[0037] For example, when the algorithm discovers in a large amount of training data that the clinical signal of advanced menopause and the biochemical signal of abnormal free osteocalcin coexist, the corresponding high-risk label has an extremely high incidence rate. The system will then determine that this combination is a signal combination with abnormal pathogenic risk and assign it a very high (close to 1) attention weight using the Softmax function. Conversely, for combinations without a clear pathogenic association, a very low weight is assigned. Finally, based on these highly sensitive weight coefficients, the algorithm performs weighted summation and aggregation on all underlying signals, outputting a high-order fused feature vector that combines macroscopic vision with microscopic insight.

[0038] S4. Perform classification mapping on the high-order fusion feature vectors, and output the corresponding risk assessment results based on the mapping results.

[0039] The specific steps for classifying and mapping high-order fused feature vectors in S4 include: A classifier with an internal imbalanced bias weight is used to perform nonlinear classification mapping on the high-order fused feature vector. The imbalanced bias weight is achieved by introducing a dynamic adjustment factor based on the initial classification probability into the cross-entropy loss function of the classifier. If the initial classification probability of the classifier for highly concealed severe osteoporosis samples decreases, the dynamic adjustment factor is amplified exponentially, increasing the loss gradient weight of the classifier for highly concealed severe osteoporosis samples during backpropagation. This amplifies the recognition sensitivity of the high-order fused feature vector reflecting the concealed features of severe osteoporosis, corrects the initial classification probability obtained from the classification mapping, and outputs the initial risk value after probability correction.

[0040] The steps in S4 that output the corresponding risk assessment results based on the mapping results specifically include: mapping the initial risk values ​​into predicted scores within a preset score range; Set early warning judgment thresholds, which include a first risk threshold and a second risk threshold, wherein the first risk threshold is less than the second risk threshold; If the predicted score is less than the first risk threshold, it is determined to be a low-risk level; if the predicted score is not less than the first risk threshold and not greater than the second risk threshold, it is determined to be a medium-risk level; if the predicted score is greater than the second risk threshold, it is determined to be a high-risk level, and the corresponding risk level label is output.

[0041] In real-world clinical epidemiological statistics, the vast majority of people are healthy or have mild bone loss, while patients with severe primary osteoporosis constitute an absolute minority. If a conventional classifier is used, the algorithm, in pursuit of a superficially high overall accuracy, tends to misclassify most hidden severe cases as healthy (i.e., a very high false negative rate). This embodiment creatively introduces a classifier with imbalanced bias weights (e.g., the Focal Loss loss function). This classifier uses the following imbalanced bias loss function when training and updating network parameters. Model optimization: In the formula, This is the initial probability (between 0 and 1) that the model currently predicts a patient to be at high risk of severe osteoporosis. These are class weight parameters used to balance the significant disparity in the total number of positive and negative samples; To adjust the focusing parameters for the weights of difficult samples, this embodiment preferably sets them as follows: .

[0042] In the overall calculation mechanism of this formula, These factors constitute the core dynamic adjustment factors: when the model faces highly concealed severe osteoporosis samples (i.e., the model's prediction accuracy is extremely low, predicting the probability of developing the disease), (when close to 0) The calculated value will be drastically amplified, forcing the network to incur a huge gradient penalty cost on these difficult samples; conversely, for easily classified samples with obvious features, this penalty factor will decay exponentially. Through the above-mentioned probability-based nonlinear error constraint mechanism, the system effectively amplifies the weight contribution of minority difficult samples in the model gradient update, significantly improving the sensitivity to the identification of severe and hidden pathogenic features, thereby correcting the model output to an initial risk value that is closer to the actual clinical severity rate.

[0043] To make the algorithm results intuitively understandable for medical staff and patients without a computer science background, the system linearly maps the obscure initial risk values ​​(usually decimals between 0 and 1) into predicted scores within a preset range of 0 to 100. In this embodiment, a first risk threshold of 30 points and a second risk threshold of 60 points are set (these specific values ​​are only preferred examples, and the system can be dynamically fine-tuned based on statistical data from specific medical institutions). When a patient's final predicted score is below 30 points, the system outputs a low-risk label, indicating that the current bone condition is good; when the score is between 30 points (inclusive) and 60 points (inclusive), a medium-risk label is output, indicating a potential risk of bone loss that requires attention; when the score exceeds 60 points, a bright red high-risk label is output, triggering a red alert for medical intervention. Thus, the system completes an automated closed loop from raw multidimensional data to precise medical quantitative assessment.

[0044] Example 2 Reference Figure 3 This is the second embodiment of the present invention. This embodiment provides a method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators. Building upon Embodiment 1, this embodiment further provides an automated closed-loop execution system for clinical auxiliary decision-making in high-risk populations, specifically including: The risk assessment results output based on the mapping results also include: If the risk level label is high risk, a data package containing a bone condition specialist follow-up appointment request will be automatically triggered and generated. At the same time, a standardized intervention report containing targeted bone metabolism intervention recommendations will be generated. The data package and the intervention report will then be merged and pushed to an external application terminal to execute medical intervention instructions.

[0045] In real-world hospital clinical decision support system (CDSS) scenarios, simply outputting risk scores is often insufficient to form a complete healthcare service loop. This embodiment focuses on illustrating the automated hardware response and physical workflow when the system triggers a high-risk label.

[0046] Specifically, in this embodiment, the external application terminal is manifested as the doctor's workstation host in the hospital information system (HIS) or medical image archiving and communication system (PACS), or the mobile smart terminal bound to the patient.

[0047] When the prediction system determines and outputs a high-risk level label in S4, it immediately interacts with the hospital's HIS system's registration and examination scheduling module via the application programming interface (API) to automatically generate a data packet containing a bone condition specialist follow-up examination scheduling request. In this embodiment, the bone condition specialist follow-up examination scheduling request data packet generated by the system is encapsulated in structured XML or JSON format. Its core data items include: patient unique identifier (PID), predicted score value, warning level code (red high risk), recommended examination item code (e.g., DXA-001), and preset recommended follow-up examination time window. This standardized data encapsulation ensures that the instructions output by the prediction engine can be directly parsed by the hospital scheduling system and automatically occupy the corresponding examination slots without manual input.

[0048] While generating the scheduled data package, the system uses its built-in medical rule engine to generate a standardized intervention report based on the patient's abnormal deep clinical and biochemical signals. This intervention report includes not only an analysis of the risk score output by the algorithm but also specific medical-grade prescription recommendations. For example, if the patient's abnormal characteristics are primarily advanced age and extremely low t1NP (a bone formation marker), the rule engine will automatically match an adjunctive therapy pathway to promote bone formation; if it shows calcium and phosphorus metabolism disorders, it will match a mineral supplementation and endocrine regulation pathway. The report not only includes recommended drug names (such as alendronate sodium and vitamin D3) but also further refines the recommended dosing frequency and dosage range, and includes dietary and lifestyle restrictions specific to this risk level.

[0049] Finally, the system encrypts and merges the aforementioned scheduling data package with the intervention report, and synchronously pushes it to the attending physician's external application terminal screen via the medical intranet. On the physician's workstation interface, the system displays a warning window via a blocking asynchronous notification mechanism. This window forces the attending physician to read the risk statement and provides a one-click confirmation or modification option. Only after the physician completes secondary authentication and clicks confirmation will the system officially activate the electronic examination form in the HIS system and automatically generate a follow-up appointment entry in the PACS system.

[0050] Thus, this invention transforms discrete risk assessment data into legally binding and clinically executable medical instructions through standardized protocol interaction and closed-loop logical feedback, significantly eliminating the information gap and response delay between risk discovery and clinical action.

[0051] Example 3 Reference Figure 4This is the third embodiment of the present invention. This embodiment provides a method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators. Building upon Embodiment 1, this embodiment further provides a closed-loop process for long-term health and non-pharmacological management of the entire population, specifically including: The risk assessment results output based on the mapping results also include the generation of long-term health management strategies. The specific steps for generating long-term health management strategies include: Obtain the predicted scores over multiple consecutive historical assessment periods, and fit them to generate a sequence map that reflects the dynamic trend of risk changes; Based on the risk level labels, a daily care plan that includes lifestyle intervention parameters and non-pharmacological management strategies is generated. The sequence map is combined with the daily care plan and then pushed to an external application terminal to provide chronic disease intervention prompts.

[0052] In the broader context of outpatient health and chronic disease management, osteoporosis prevention and treatment emphasizes early preventative intervention and continuous tracking throughout the entire life cycle. This embodiment focuses on how the system transcends the limitations of a single medical diagnosis, transforming complex algorithms and computing power into a digital health management service accessible to ordinary users.

[0053] Specifically, in this embodiment, the physical form of the external application terminal extends from the hospital setting to the outside world, specifically manifested as the patient's own smartphone mobile application, smart wearable devices with health data collection and interaction functions (such as smartwatches and bracelets), and the chronic disease management cloud platform terminal of the community health check center.

[0054] Firstly, regarding the generation and visualization of dynamic sequence maps, the system not only outputs the single prediction score but also retrieves the patient's time-series data from the historical database. The system uses the patient's previous prediction scores as the ordinate and time points as the abscissa to generate an intuitive time-series line chart. Simultaneously, for anomalous features assigned high computational weight by the cross-channel self-attention mechanism in Example 1 (such as the intersection of continuously increasing bone resorption markers and body mass index), the system uses a multi-dimensional radar chart for local highlighting. Through this visualized map mapping, even users without a medical background can clearly understand the long-term evolution trajectory and contributing factors of their own bone loss.

[0055] Secondly, regarding the generation and execution of daily care plans, the system has a built-in tiered lifestyle intervention model, which automatically issues health instructions in a step-by-step manner based on risk level labels (low, medium, and high). When the risk level label is low, the system pushes basic bone health science popularization and prevention programs to external application terminals. The dietary nutrition indicators are specifically expressed as a recommended high-calcium and low-salt diet structure model (such as pushing quantitative intake suggestions for daily dark green vegetables and dairy products).

[0056] When the risk level label is medium risk, it indicates that the patient is in the critical warning zone of bone loss. At this point, the system not only pushes enhanced dietary guidelines but also generates mandatory weight-bearing exercise parameters. For example, it generates a specific digital prescription for resistance training (such as dumbbell exercises and resistance band stretching) three times a week for 30 minutes each time. Furthermore, the system can monitor the patient's daily exercise frequency and duration and assess compliance by calling the inertial measurement unit (IMU) interface of a smart wearable device.

[0057] When the risk level label is high risk, the system triggers the medical-grade response closed loop in Example 2 (such as DXA examination and drug prescription) and simultaneously adds home environment modification suggestions for preventing falls to the daily care plan (such as suggesting the installation of bathroom handrails, increasing nighttime lighting, etc.), thereby cutting off the external physical causes of osteoporosis leading to fragility fractures to the greatest extent.

[0058] Thus, relying on the underlying feature fusion algorithm architecture, this invention has extended its application from in-hospital clinical decision support to out-of-hospital long-term health management, effectively improving the applicability and intervention continuity of the risk prediction model in actual medical environments.

[0059] In summary, this invention effectively improves the ability of risk prediction models to quantify complex pathogenic factors and their sensitivity to identifying a small number of hidden high-risk samples, reduces the rate of missed diagnoses in clinical practice, and realizes an automated hierarchical intervention and execution closed loop from in-hospital clinical decision support to long-term health management outside the hospital.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators, characterized in that, Includes the following steps: S1. Obtain the patient's original clinical indicators and original bone metabolism markers, fill in missing values ​​and standardize the data scale to obtain a standard dataset; S2. Perform cross-operation between the clinical indicators and bone metabolism markers in the standard dataset to extract associated features, and remove redundant data based on correlation analysis to generate a core feature matrix. S3. Perform dual-channel parallel mapping on the core feature matrix to extract nonlinear representations of clinical features and bone metabolism features respectively, and evaluate the interaction pathogenicity of the clinical features and bone metabolism features through a cross-channel self-attention mechanism. Dynamically allocate weights according to the interaction pathogenicity to complete feature fusion and generate a high-order fused feature vector. S4. Perform classification mapping on the high-order fusion feature vector, and output the corresponding risk assessment result based on the mapping result.

2. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 1, characterized in that, The original clinical indicators obtained in S1 include basic human physiological characteristics data and lifestyle data related to bone load-bearing, and the original bone metabolism markers include bone formation markers and calcium and phosphorus metabolism-related markers. The missing value imputation step in S1 specifically involves using the nearest neighbor interpolation algorithm to imput missing values. The missing value imputation step using the nearest neighbor interpolation algorithm specifically includes: extracting various known indicators of patients with missing data, matching a preset number of reference patients whose values ​​of various known indicators are closest in the historical database, calculating the average value of the missing indicators corresponding to the preset number of reference patients, and filling the missing values ​​into the missing positions. The specific steps for unifying the data scale include: taking each of the original clinical indicators and the original bone metabolism markers after filling in the missing values ​​as the current processing item, subtracting the corresponding average value in all samples from the value of the current processing item, and dividing the resulting difference by the corresponding standard deviation to convert it into a standard value.

3. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 1, characterized in that, The step in S2 of performing cross-operation between clinical indicators and bone metabolism biomarkers in the standard dataset to extract associated features specifically includes: performing cross-multiplication of each clinical indicator in the standard dataset with each bone metabolism biomarker to generate a new feature containing all cross-category combinations, and merging the new feature with the original indicators in the standard dataset to construct a set of candidate features.

4. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 3, characterized in that, The specific steps in S2 for generating the core feature matrix by removing redundant data based on correlation analysis include: The Pearson correlation coefficient between the first feature and the second feature is obtained by calculating the sum of the products of the differences between any first feature and the second feature in the candidate feature set and their respective deviations from the mean of the corresponding samples, and by dividing the sum of the products of the differences by the product of the square roots of the sum of the squares of the deviations of the first feature and the second feature from the mean of the corresponding samples. If the Pearson correlation coefficient between the first feature and the second feature is determined to be greater than a preset correlation threshold, then the feature with the larger variance between the first feature and the second feature is retained, and the feature with the smaller variance is removed; the removal step is repeated until the Pearson correlation coefficient between the retained features is not greater than the preset correlation threshold, and all the retained features are combined to generate the core feature matrix.

5. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 1, characterized in that, The step in S3 of performing dual-channel parallel mapping on the core feature matrix to extract nonlinear representations of clinical features and bone metabolism features respectively includes: decomposing the core feature matrix into mutually independent clinical feature sub-matrices and bone metabolism feature sub-matrices; The clinical feature sub-matrix is ​​input into the first processing channel and nonlinearly superimposed and mapped through multiple layers of neuronal nodes to extract deep clinical signals that characterize the external physical condition while avoiding interference from biochemical data. Simultaneously, the bone metabolism feature sub-matrix is ​​input into a second processing channel that is physically isolated from the first processing channel. After nonlinear superposition mapping through multiple layers of neuronal nodes, deep biochemical signals characterizing the internal metabolic state are extracted under the condition of avoiding the suppression of macroscopic signs.

6. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 5, characterized in that, The step in S3, which assesses the pathogenicity of the interaction between the clinical feature and the bone metabolism feature through a cross-channel self-attention mechanism and dynamically assigns weights based on the pathogenicity of the interaction to complete feature fusion, specifically includes: The deep clinical signals are mapped to a query matrix, and the deep biochemical signals are mapped to a key matrix and a value matrix, respectively. Calculate the dot product of the query matrix and the transpose of the key matrix, and divide by the scaling factor of the key vector dimension to obtain the cross-channel association matching score; Signal combinations with abnormal pathogenicity risk are extracted, and computational weights higher than the conventional standard are assigned to these signal combinations. The computational weights are generated by processing the cross-channel association matching scores through a normalized exponential function. Based on the assigned computational weights, the computational weights are multiplied by the value matrix containing biochemical features, and the weighted fusion is used to generate the higher-order fusion feature vector.

7. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 1, characterized in that, The step of classifying and mapping the higher-order fused feature vector in S4 specifically includes: The high-order fused feature vector is nonlinearly classified using a classifier with an internal imbalanced bias weight. The imbalanced bias weight is achieved by introducing a dynamic adjustment factor based on the initial classification probability into the cross-entropy loss function of the classifier. If the initial classification probability of the classifier for highly concealed severe osteoporosis samples decreases, the dynamic adjustment factor is amplified exponentially, increasing the loss gradient weight of the classifier for the highly concealed severe osteoporosis samples during backpropagation. This amplifies the recognition sensitivity of the high-order fusion feature vector reflecting the concealed features of severe osteoporosis, corrects the initial classification probability obtained from the classification mapping, and outputs the initial risk value after probability correction.

8. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 7, characterized in that, The step of outputting the corresponding risk assessment result based on the mapping result in S4 specifically includes: mapping the initial risk value into a predicted score within a preset score range; A warning judgment threshold is set, which includes a first risk threshold and a second risk threshold, wherein the first risk threshold is less than the second risk threshold; If the predicted score is less than the first risk threshold, it is determined to be a low-risk level; if the predicted score is not less than the first risk threshold and not greater than the second risk threshold, it is determined to be a medium-risk level; if the predicted score is greater than the second risk threshold, it is determined to be a high-risk level, and the corresponding risk level label is output.

9. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 8, characterized in that, The step of outputting the corresponding risk assessment result based on the mapping result also includes: If the risk level label is high risk, a data packet containing a bone condition specialist follow-up appointment request will be automatically triggered and generated. At the same time, a standardized intervention report containing targeted bone metabolism intervention recommendations will be generated. The data packet and the intervention report will then be merged and pushed to an external application terminal to execute medical intervention instructions.

10. The method for predicting the risk of primary osteoporosis based on bone metabolism markers and clinical indicators according to claim 8, characterized in that, The step of outputting the corresponding risk assessment result based on the mapping result also includes generating a long-term health management strategy. The specific steps for generating the long-term health management strategy include: Obtain the predicted scores over multiple consecutive historical assessment periods, and fit them to generate a sequence map that reflects the dynamic trend of risk changes; Based on the risk level labels, a daily care plan containing lifestyle intervention parameters and non-pharmacological management strategies is generated. The sequence map is combined with the daily care plan and then pushed to an external application terminal to provide chronic disease intervention prompts.

Images