A multi-modal osteoporosis stratification early warning system

Abstract

The present application relates to the technical field of health care informatics, and discloses a multi-modal osteoporosis stratification early warning system, comprising a data interface module, a prescription gradient monitoring module, a semantic feature mapping module, a time sequence feature coding module, a cross-attention fusion module and a risk calculation module, wherein the prescription gradient monitoring module is used for real-time calculation of the change gradient of analgesic drug dosage; when the physical image data is missing and the drug dosage gradient is abnormal, a high-risk proxy vector is forcibly called as a morphological feature input, so that the image diagnosis blind area is filled with treatment behavior data, the technical problem that the traditional system cannot effectively warn the risk of occult fracture when lacking image is solved, and the logical continuity and acute risk capturing ability of the system under the condition of incomplete multi-modal data are ensured.

Description

Technical Field

[0001] This invention relates to a multimodal osteoporosis stratified early warning system, belonging to the field of healthcare informatics technology. Background Technology

[0002] Osteoporosis is currently an insidious, progressive, systemic metabolic bone disease characterized by low bone mass and bone microstructure destruction. Dual-energy X-ray absorptiometry (DXA) measures bone mineral density at specific sites, providing quantitative indicators and serving as a basis for screening and diagnosing osteoporosis. Serum biochemical indicators further reflect the body's current bone metabolic transformation status. However, in healthcare informatics applications, conventional methods based on single-point discrete physical detection have technical limitations. Osteoporosis has a continuous and insidious course. DXA is limited by equipment costs and radiation exposure, resulting in low sampling frequencies. Low-frequency sampling cannot capture high-frequency metabolic fluctuations caused by drug intervention or secondary factors, leading to delayed risk warnings. Hospital heterogeneous information systems store a large amount of underutilized data. Radiology information systems accumulate massive amounts of unstructured morphological image report texts, and hospital information systems record continuous high-frequency biochemical test values ​​and drug prescription information. Existing technologies lack heterogeneous logical coupling mechanisms, making it difficult to achieve semantic alignment and fusion of high-frequency biochemical data and morphological text, resulting in fragmented modal data.

[0003] Existing technologies utilize algorithmic models to integrate clinical data and imaging features to improve recognition accuracy. However, most solutions are based on the assumption of data completeness, making it difficult to adapt to the common clinical situation of missing data. Chinese invention patent CN111755129B discloses a multimodal osteoporosis stratified early warning method and system, which constructs a three-layer feature system including individual information, biochemical indicators, and lumbar CT images. It uses a genetic algorithm to optimize the SVM classifier for risk stratification. However, the technical logic heavily relies on the acquisition of explicit CT image data and the manual image segmentation and extraction process. Patients may not undergo timely imaging examinations due to mobility issues or pain, resulting in missing data or outdated image data. The model cannot establish a mapping relationship due to the lack of key feature dimensions. It cannot infer potential skeletal structural damage using alternative behaviors data of analgesic drug use during the silent period when morphological evidence is lacking. The reliance on explicit imaging evidence leads to the interruption and failure of the early warning logic in the normalized medical scenario where multimodal data is incomplete.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a system that deeply integrates unstructured morphological text and structured time-series metabolic data to address data gaps, drug masking, and time misalignment in the calculation of working conditions, and to achieve continuous dynamic hierarchical early warning of osteoporosis risk. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A multimodal osteoporosis stratified early warning system, comprising:

[0006] The data interface module is used to connect the hospital information system and the radiology information system to collect structured time-series data and unstructured image report text of the target object, respectively. The structured time-series data includes the time series of biochemical test indicators and drug prescription records.

[0007] The prescription gradient monitoring module, connected to the data interface module, is used to extract the prescription dosage of analgesics from the drug prescription record and calculate the first derivative of the prescription dosage over time to generate prescription gradient values.

[0008] The semantic feature mapping module is connected to the data interface module and the prescription gradient monitoring module respectively, and is used to execute the feature vector generation logic: when the data interface module collects unstructured image report text, it performs vectorization encoding on the unstructured image report text based on the preset skeletal morphology terminology library to generate morphological feature vectors; when the data interface module does not collect unstructured image report text and the prescription gradient value exceeds the preset threshold, it calls the preset high-risk proxy vector as the morphological feature vector.

[0009] The time-series feature encoding module, connected to the data interface module, is used to perform standardization processing on biochemical test indicators and input the processed data into a pre-set recurrent neural network unit to generate a hidden state vector representing metabolic trends.

[0010] The cross-attention fusion module is connected to the semantic feature mapping module and the temporal feature encoding module respectively. It is used to use the morphological feature vector as the query vector and the hidden state vector as the key vector and value vector to perform cross-attention calculation. It adjusts the weights of each temporal feature dimension in the hidden state vector according to the semantic information in the morphological feature vector and generates a multimodal fusion feature vector.

[0011] The risk calculation module, connected to the cross-attention fusion module, is used to map the multimodal fusion feature vector to a preset risk probability space and output the risk level to which the target object belongs.

[0012] Preferably, the temporal feature encoding module further includes a residual calculation unit, which stores a pre-set pharmacodynamic mapping table. The pharmacodynamic mapping table defines the correspondence between the dose of anti-osteoporosis drugs and the theoretical changes in biochemical indicators. The residual calculation unit is used to calculate the theoretically expected biochemical values ​​based on the drug prescription record and the pharmacodynamic mapping table, and to calculate the difference between the actual detected values ​​of biochemical test indicators and the theoretically expected biochemical values ​​to generate a metabolic residual vector. The temporal feature encoding module is used to transmit the metabolic residual vector as input data to the recurrent neural network unit to extract the hidden state vector characterizing the body's drug response capability.

[0013] Preferably, the semantic feature mapping module also stores a pre-set basic risk prior vector and is used to execute time decay calibration logic: obtain the generation time of the unstructured image report text, and calculate the time interval between the generation time and the current system time; use a pre-set nonlinear decay function to map the time interval to a normalized semantic confidence coefficient. And generate time-calibrated morphological feature vectors according to the following formula. , where V s The morphological feature vector V generated by the semantic feature mapping module base The basic risk prior vector is used; the cross-attention fusion module is used to utilize the time-calibrated morphological feature vector as the query vector.

[0014] Preferably, the analgesics monitored by the prescription gradient monitoring module include nonsteroidal anti-inflammatory drugs and opioids; the preset threshold is set as the prescription frequency mutation threshold characterizing acute pain events; and the high-risk proxy vector is a pre-trained feature embedding vector characterizing the semantics of vertebral compression fractures.

[0015] Preferably, the recurrent neural network unit is a gated recurrent unit, which is used to process the biochemical test index sequence after interpolation and capture the dynamic dependence of the biochemical test index over time.

[0016] Preferably, the cross-attention fusion module includes a static cross-attention layer, which is used to calculate the attention score between the morphological feature vector and the hidden state vector, and to use the attention score to perform a weighted summation of the hidden state vector; when the morphological feature vector is a high-risk surrogate vector, the attention score forcibly increases the weight of the biochemical indicator dimension representing the bone metabolic stress response in the hidden state vector.

[0017] Preferably, the data interface module is also used to perform data cleaning logic, perform linear interpolation on missing data points in biochemical test indicators, and perform Z-score standardization on all numerical data to eliminate dimensional differences.

[0018] Preferably, the semantic feature mapping module does not contain convolutional neural network units for processing pixel-level image data; the morphological feature vectors only represent the pathological semantic logical relationships in the unstructured image report text or the specific risk state indicated by the high-risk proxy vector.

[0019] Preferably, the risk calculation module includes a fully connected classification layer and a Softmax activation function, which are used to output the probability distribution of the target object belonging to low-risk, medium-risk and high-risk levels respectively, and take the level with the highest probability value as the final warning result.

[0020] Preferably, the skeletal morphology terminology database contains feature roots describing trabecular bone structure, vertebral body edge morphology, and bone density; the vectorization encoding uses a pre-trained word embedding model to convert the matched feature roots into high-dimensional sparse vectors.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. Based on prescription behavior gradient, the system completes hidden risk data, monitors prescription gradient, establishes a mapping between the first derivative of analgesic drug dosage changes over time and skeletal structural events, and responds to drug prescription gradient mutation signals when unstructured image report data is missing in the radiology information system. The system generates a pre-set high-risk proxy vector and injects it into the cross-attention fusion module, transforming the hospital information system's treatment behavior data into input features representing potential morphological risks. The system maintains the logical continuity of the multimodal computation model under the condition of interrupted physical examination data flow, enabling the system to capture potential acute skeletal structural injury risks based on rapid changes in drug use behavior during the silent period when explicit imaging evidence is lacking.

[0023] 2. Metabolic state denoising calibration based on pharmacodynamic residual analysis: The theoretical expected values ​​of biochemical indicators are calculated using a pre-set pharmacodynamic mapping table. A metabolic residual vector based on the difference between observed and expected values ​​is constructed as input data for the time-series coding module. The physical dimension of data features is converted from absolute biochemical concentration to residuals of the body's response to drug intervention. The numerical masking effect of external drug intake on the body's true metabolic level is eliminated. The residual features are processed to identify samples that present falsely normal values ​​under drug maintenance but have low actual metabolic response capabilities. Attention is paid to the underestimation of risk level due to the absolute value meeting the target, thereby improving the accuracy of stratified calculation under complex drug interference environments.

[0024] 3. Dynamic regulation of semantic confidence based on time decay: In the semantic feature mapping stage, a nonlinear decay function based on time interval is integrated. According to the difference between the image report generation time and the current system time, the weight coefficient of morphological feature vector in cross-attention calculation is dynamically adjusted. As the timeliness of data decreases, the dependence on historical morphological description is forcibly reduced, and the weight of attention to current real-time biochemical indicators or basic statistical priors is smoothly increased. Time dimension weighted calibration solves the problem of time misalignment between low-frequency image data and high-frequency biochemical data, prevents old historical normal reports from erroneously suppressing the expression of current abnormal metabolic trend features in fusion calculation, and ensures that the risk warning results meet the data time validity boundary. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the logical architecture and data processing of the multimodal osteoporosis stratified early warning system of the present invention.

[0026] Figure 2 This is a comparison chart of the risk identification accuracy of the present invention and the control group under the drug efficacy masking environment.

[0027] Figure 3 This is an interaction timing diagram of the prescription gradient monitoring and high-risk proxy vector injection logic of the present invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] This invention provides a multimodal osteoporosis stratified early warning system. The system is constructed as a healthcare informatics data processing platform running in a server cluster or cloud computing environment. It includes a data interface module, a prescription gradient monitoring module, a semantic feature mapping module, a temporal feature encoding module, a cross-attention fusion module, and a risk calculation module. These modules interact and collaborate logically via a high-speed data bus. The system establishes encrypted communication links with the hospital information system and the radiology information system based on the HL7 or FHIR standard protocol through the data interface module. This is used to collect structured time-series data and unstructured image report text of the target object in real time. The structured time-series data includes time-stamped biochemical test indicators, specifically including serum calcium concentration, serum phosphorus concentration, etc. The data includes alkaline phosphatase concentration, serum creatinine concentration, and bone turnover markers, as well as the subject's medication prescription records over a historical period, particularly prescription details for anti-osteoporosis drugs, nonsteroidal anti-inflammatory drugs, and opioids. The unstructured image report text is a natural language description written by radiologists of X-ray, CT, or MRI examination results, including morphological evaluations of trabecular bone structure, vertebral body margin morphology, and bone mineral density. The data interface module's built-in data cleaning subroutine performs linear interpolation on the collected biochemical test indicators, fills in time series gaps caused by irregular visits, and performs Z-score standardization on all numerical data using preset mean and standard deviation parameters, mapping physiological indicators of different dimensions to a unified dimensionless numerical range.

[0030] To address the technical challenge of identifying occult acute fracture risks in medical settings due to missing morphological data caused by patients not undergoing imaging examinations, the system utilizes a prescription gradient monitoring module to execute risk sentinel logic based on treatment behavior. This module scans drug prescription records in real time, extracting prescription frequency and cumulative dosage data for nonsteroidal anti-inflammatory drugs (NSAIDs) and opioid analgesics. The module's internal discrete differential calculation unit sets a 14-day time sliding window to calculate the first derivative of the analgesic drug dosage over time, generating a prescription gradient value reflecting the changing trend of pain management intensity. The system has a pre-set prescription frequency mutation threshold characterizing acute pain events. When the calculated prescription gradient value exceeds this threshold, and the data interface module does not detect any new unstructured image report text within the same time window, the prescription gradient monitoring module generates a high-risk state trigger signal. This signal is transmitted to the semantic feature mapping module. The semantic feature mapping module forcibly activates the hidden event proxy logic, which calls a high-risk proxy vector pre-trained based on a labeled dataset to represent the semantics of vertebral compression fractures or bone microstructural damage in a high-dimensional space. This high-risk proxy vector is then directly injected as a morphological feature vector into the subsequent attention calculation process. This procedure translates drastically changing analgesic drug use behavior into potential skeletal morphological damage signals, maintaining the system's logical perception of acute skeletal risks when physical imaging evidence is lacking. To address the issue of misalignment between low-frequency image reports and high-frequency biochemical data in the temporal dimension, leading to outdated images misleading current risk assessments, the semantic feature mapping module runs a time-decay-based semantic confidence calibration logic. This module utilizes a pre-built skeletal morphology terminology library and employs a word embedding model to perform keyword matching and vectorization encoding on the collected unstructured image report text, generating an initial morphological feature vector V. s The time difference calculation unit within the module obtains the generation timestamp of the image report, calculates the time interval Δt between it and the current system analysis time, and uses a preset nonlinear decay function to map this time interval into a normalized semantic confidence coefficient. The coefficient decreases monotonically with increasing Δt. The module synchronously calls the basic risk prior vector V, which is constructed based on the statistical distribution characteristics of age, gender, and BMI, stored in the database. base The module is based on the formula Perform a weighted fusion operation to generate a time-calibrated morphological feature vector. , where V s V is the initial morphological feature vector. base Based on the fundamental risk prior vector, The semantic confidence coefficient is calculated to ensure that the system automatically and smoothly transitions to a priori patterns that rely on basic demographic features when image data becomes outdated.

[0031] To address the pharmacodynamic masking effect, which masks the body's true metabolic level and leads to falsely normal indicators, the temporal feature encoding module utilizes a residual calculation unit and a pharmacodynamic mapping table for feature reconstruction. The pharmacodynamic mapping table defines the theoretical correction magnitude and half-life of various anti-osteoporosis drugs for specific biochemical indicators at standard unit doses. The residual calculation unit retrieves this mapping table based on the patient's medication prescription records and administration time points, and combines it with a pharmacodynamic superposition model to calculate the theoretically expected biochemical value of the target at the current sampling time. The unit performs a difference operation, subtracting the theoretically expected biochemical value from the actual detection value input from the data interface module, generating a metabolic residual vector after removing the drug masking effect. This metabolic residual vector is transmitted as input data to the gated loop unit of the temporal feature encoding module. This gated loop unit processes the interpolated and completed metabolic residual vector sequence, capturing the dynamic dependence of the residual signal over time, and outputting a hidden state vector containing information about the body's true intrinsic metabolic response. The cross-attention fusion module serves as the logical coupling mechanism for heterogeneous data features. The system receives morphological feature vectors or high-risk proxy vectors from the semantic feature mapping module as query vectors, and latent state vectors from the temporal feature encoding module as key and value vectors. The static cross-attention layer within the module performs dot product attention calculations to measure the correlation between the query vector and the key vector in the feature space, generating an attention score matrix. This attention score dynamically adjusts the weights of each temporal feature dimension in the latent state vector and aggregates the weighted features to generate a multimodal fusion feature vector. When the input query vector is the aforementioned high-risk proxy vector, the attention mechanism increases the weight contribution of the biochemical indicator dimension related to bone metabolic stress response in the latent state vector based on the pathological semantic features contained in the vector. The risk calculation module receives the multimodal fusion feature vector, maps it to a preset risk probability space through a fully connected neural network layer, and uses the Softmax activation function to calculate the probability distribution of the target object belonging to low-risk, medium-risk, and high-risk levels, respectively. The level with the highest probability value is selected as the final stratified warning result output.

[0032] Example 1: In a community health management application scenario targeting elderly individuals at high risk of osteoporosis, the target patient is a 75-year-old female patient with a history of osteoporosis. Her most recent radiological imaging record was generated 24 months ago, showing that her bone mineral density was borderline but her vertebral body morphology was intact. In the past 14 days, she has frequently visited the hospital due to lower back pain and increased the dosage of her nonsteroidal anti-inflammatory drug (NSAID). However, due to mobility limitations, she has not undergone any new imaging examinations, resulting in the radiology information system being in a data-silent state within the current time window. At this time, the conventional logic based on multimodal alignment of image and biochemistry faces insufficient judgment criteria due to the lack of real-time morphological feature input. In response to this situation, the system collects the patient's structured time-series data and medication prescription records through the data interface module. The prescription gradient monitoring module immediately performs discretization on the NSAID prescription data. Differential operations are performed to calculate the first derivative of the prescription dose change of the object within the current 14-day sliding window, i.e., the prescription gradient value. Based on the fact that the calculated prescription gradient value has exceeded the preset mutation threshold and no new unstructured image report text has been collected within the same time window, the system determines that it is currently in a hidden risk state and triggers the proxy mechanism. The semantic feature mapping module responds to the trigger signal by calling the high-risk proxy vector that represents the semantics of vertebral compression fracture in high-dimensional space to replace the missing real-time morphological feature vector injected into the calculation process. At the same time, the residual calculation unit in the temporal feature encoding module calculates the theoretical expected biochemical value of the object under the current high-dose drug maintenance according to the pharmacodynamic mapping table. The metabolic residual vector is generated by subtracting the theoretical expected value from the actual detection value. After being processed by the gated loop unit, this vector generates a hidden state vector that represents the body's true metabolic response.

[0033] The cross-attention fusion module receives a high-risk proxy vector as a query vector and a latent state vector as a key vector and a value vector. During the attention calculation process, the pathological semantic features in the high-risk proxy vector are coupled with the metabolic residual features in the latent state vector, which increases the weight contribution of the dimensions of alkaline phosphatase concentration fluctuation and abnormal trend of bone turnover markers related to bone metabolic stress response in the latent state vector. This reconstructs the logical relationship between bone structure destruction and metabolic stress in the feature space. The risk calculation module receives a weighted aggregated multimodal fusion feature vector and maps it to the risk probability space through a fully connected neural network layer. Finally, it outputs a probability distribution indicating that the object is currently in a high-risk level. In the absence of real-time imaging evidence, this result identifies the potential risk of occult vertebral compression fractures through the synergistic determination of abrupt changes in treatment behavior gradient and metabolic residual abnormalities.

[0034] Example 2: This example is based on a medical data simulation verification platform built in an isolated server environment. It objectively verifies the engineering effectiveness of the present invention's technical solution under non-ideal conditions such as data loss and drug efficacy masking. This verification platform integrates data interfaces from a simulated Hospital Information System (HIS) and Radiology Information System (RIS) conforming to the HL7 / FHIR standard. It includes a pre-set test dataset containing 500 simulated cases. The test dataset is generated from real clinical desensitized data through parameter reconstruction, covering the full spectrum of pathological states from normal bone mass and osteopenia to severe osteoporosis and occult fractures. It introduces a normally distributed random biochemical detection error (±5%) and prescription record noise to simulate data disturbances in a real medical environment. This verifies the effectiveness of the prescription gradient monitoring module and... The synergistic effect of the hidden event proxy logic in the semantic feature mapping module was investigated. The experiment set up a sample group of the present invention and a partially missing control group. The sample group of the present invention fully enabled the prescription gradient monitoring function and the high-risk proxy vector injection mechanism; the partially missing control group physically disabled the prescription gradient monitoring module, so that the system relied on biochemical indicators and historical images for inference. The experiment screened 150 specific samples marked as the silent period. The characteristics of this type of sample were defined as: vertebral compression microfractures (true label is high risk) that occurred in the past 14 days, but no new imaging examinations were performed (RIS data missing), and the cumulative prescription dose of nonsteroidal anti-inflammatory drugs (NSAIDs) showed a rapid upward trend. Table 1 shows the comparative test data when risk stratification was performed for this type of specific sample.

[0035] Table 1: Comparison of Sample Identification Performance for Hidden Risks During the Quiet Period

[0036]

[0037] Referring to Table 1, the partial missing control group, lacking real-time morphological input and with simple biochemical fluctuations not yet reaching abnormal thresholds in the acute phase, had a recall rate of only 18.7% for occult fracture samples. Many actual fracture cases were misclassified as low-risk or medium-risk. In contrast, in the sample group of this invention, the prescription gradient monitoring module successfully captured drug use mutation signals in 94.0% of the samples and triggered 141 high-risk surrogate vector injection operations. This mechanism improved the recall rate of the high-risk level to 89.3% without imaging data. Data shows that although the introduction of surrogate vectors slightly increased the false alarm rate to 6.8%, the improvement in risk capture capability was orders of magnitude, confirming the engineering effectiveness of the logic of filling imaging blind spots through gradient changes in treatment behavior data. Further verification of the pharmacodynamic metabolism residual calibration mechanism is needed. To assess robustness in addressing drug masking effects, the experiment introduced a problem intensity gradient control system. Another 150 osteoporosis samples were selected from patients receiving long-term combined bisphosphonate and calcium therapy. These samples had serum calcium and alkaline phosphatase (ALP) levels within the normal reference range, but their bone microstructure was physically fragile (true label: high risk). The experiment set three levels of drug maintenance intensity: low, medium, and high, corresponding to 0.5, 1.0, and 1.5 times the standard dose, respectively, to simulate different degrees of drug efficacy masking. The control group directly used Z-score-normalized raw biochemical test values ​​as input to the temporal feature encoding module, while the sample group of this invention used the metabolic residual vector processed by the residual calculation unit as input. Table 2 records the accuracy of the two systems in identifying potential metabolic failure risk under different drug maintenance intensities.

[0038] Table 2: Comparison of Risk Identification Accuracy under Different Drug Masking Intensities

[0039]

[0040] Referring to Table 2, the recognition accuracy of the control group showed a non-linear decreasing trend with the increase of drug maintenance intensity. Under high intensity (1.5 times the dose) conditions, the accuracy dropped to 22.1%, indicating that the original biochemical indicators almost completely lost their physical significance in representing the body's true metabolic capacity under the strong masking of high-concentration drugs. In contrast, the recognition accuracy of the sample group of this invention remained stable at over 85% at all three gradient levels, with fluctuations of less than 2%, and did not show a trend of deterioration with increasing drug dose. This data objectively confirms that the residual calculation logic based on the pharmacodynamic mapping table can effectively isolate the superimposed components of external drug intervention mathematically and restore the body's true internal response to drug stimulation.

[0041] Example 3: This example combines Figures 1 to 3 This section describes a multimodal osteoporosis stratified early warning system, such as... Figure 1As shown, the overall logical architecture of the multimodal osteoporosis stratified early warning system mainly consists of a data input end, a core processing module group, and a result output end. The data input end includes the Hospital Information System (HIS) and the Radiology Information System (RIS), which are used to provide biochemical test indicators, drug prescription records, and unstructured image report text, respectively. The data interface module receives the above data and performs data cleaning, Z-score standardization, and linear interpolation processing. The data flow is divided into two channels: one channel enters the prescription gradient monitoring module to calculate the first derivative of analgesic drug dosage and monitor acute pain event mutations. When the trigger threshold is reached, this module sends a signal to the semantic feature mapping module, which is responsible for text vectorization, time decay calibration, and injection of high-risk surrogate vectors to generate morphological feature vectors. The other channel enters the temporal feature encoding module, which calculates the pharmacodynamic residual and extracts the latent state vector using GRU. The above morphological feature vectors and latent state vectors converge in the cross-attention fusion module to generate multimodal fusion features. Finally, after processing by the fully connected classification layer and Softmax activation function of the risk calculation module, the system outputs stratified early warning results containing low, medium, and high risk levels.

[0042] like Figure 2 As shown, the horizontal axis represents drug maintenance intensity, covering three gradient levels: low intensity (0.5×), medium intensity (1.0×), and high intensity (1.5×). The vertical axis represents recognition accuracy (%). The legend distinguishes the recognition accuracy of the control group from that of the present invention's sample group. The data shows that as the drug maintenance intensity increases from low to high, the height of the bars in the control group decreases, reaching extremely low values ​​under high intensity conditions. In contrast, the height of the bars in the present invention's sample group remains at a high and stable level across all three gradients. Figure 3 As shown in the sequence diagram, this diagram details the interaction logic and data flow process between the prescription gradient monitoring module and related modules. The participating entities include the data interface module, the prescription gradient monitoring module, the discrete differential calculation unit, the adaptive threshold mechanism, and the semantic feature mapping module. The process begins with the data interface module sending drug prescription records. The prescription gradient monitoring module extracts the analgesic drug dosage data and then calls the discrete differential calculation unit. This unit sets a 7-day sliding window and calculates the first derivative of the dosage change, returning the generated prescription gradient value. The prescription gradient monitoring module then passes the gradient value to the adaptive threshold mechanism. This mechanism calculates the cumulative sum based on the CUSUM algorithm and compares it with the individual's historical baseline. The logic is divided into two paths: if the gradient value is normal, a normal state signal is sent; if the gradient value exceeds the mutation threshold, an abnormal signal is triggered, and a high-risk state signal is sent to the semantic feature mapping module. This activates the hidden event proxy logic and calls the high-risk proxy vector, completing the closed-loop control from prescription behavior monitoring to feature vector injection.

[0043] Example 4: This example addresses the potential algorithm path black box and parameter setting uncertainties that may be encountered in the practical engineering deployment of the prescription gradient-based hidden risk sentinel logic and pharmacodynamic metabolism residual calibration mechanism. It provides targeted technical supplements and transparent explanations. Regarding the specific implementation path of discrete differential operations in the prescription gradient monitoring module, the system adopts a numerical calculation method based on finite difference. For a given nonsteroidal anti-inflammatory drug prescription dose-time series D(t), at the sampling time... Prescription gradient value G(t) i Instead of simply calculating the slope between two adjacent points, the system obtains the slope by applying a sliding time window of length N (N=14 days) and performing a first-order linear regression fit within that window. The system constructs a local regression model D(t)=α+βt+ Where β is the desired prescription gradient value G(t) i This algorithm estimates the parameter β using the least squares method, effectively smoothing the random fluctuations in daily prescription behavior and extracting a robust trend signal reflecting changes in pain management intensity. The input to this process is a prescription dose sequence containing timestamps, and the output is a continuous sequence of prescription gradient values, providing statistically significant quantitative evidence for subsequent threshold determination. Secondly, regarding the setting principle and adaptive logic of the prescription frequency mutation threshold in the occultation risk assessment, the system does not use a single fixed threshold but introduces an adaptive threshold mechanism based on the Page Cumulative Summation (CUSUM) algorithm. This mechanism calculates the individual baseline gradient mean μ based on the target object's historical prescription data over the past 12 months. base With standard deviation σ base During real-time monitoring, the system calculates the cumulative sum statistic S. i =max(0,S i-1 +G(t i )-(μ base +kσ base ), where k is the drift coefficient, and when the cumulative sum statistic S i Exceeding the judgment threshold hσ base When a mutation signal is triggered, h is a preset threshold coefficient used to compare it with the individual's historical standard deviation σ. baseBy jointly constructing a dynamic judgment boundary, the mutation threshold is essentially a dynamic deviation from an individual's historical baseline, rather than an absolute value. This design ensures the system can adapt to differences in the baseline dependence levels of analgesics among different patients, avoiding false positives for patients on long-term high-dose medications or false negatives for patients sensitive to low doses. At the algorithm level, personalized and adaptive threshold settings are achieved. Regarding the construction of the pharmacokinetic mapping table and the calculation of theoretical expected values ​​in the pharmacodynamic metabolic residual calibration mechanism, the system establishes a pharmacodynamic superposition algorithm based on a compartment model. The mapping table does not store simple static values, but rather the pharmacodynamic decay function parameters of various drugs after a standard single dose, typically formalized as a double exponential decay model E(t)=Ae. -αt +Be -βt Where E(t) represents the correction of biochemical indicators after drug administration time t, and A, B, α, β are drug specificity constants. The residual calculation unit calculates the current time t. now When calculating the theoretical expected biochemical value, it is not only based on the most recent dosing, but also on a linear summation of all dose events within the action window from the historical dosing records, i.e., the theoretical expected value. , where V base The baseline biochemical values ​​are those of the patient in a non-medication state, where j represents the dosing event number of the target subject before the current sampling time; E j This indicates that the j-th dosing event occurs at the current time t. now The theoretical correction contribution of t to biochemical indicators j For the j-th dosing time, this computational model accurately quantifies the cumulative effect of multiple dosings, resulting in a calculated metabolic residual vector V. res =V obs -V pred (V) obs (The actual measured value) can purely reflect the body's intrinsic metabolic deviation under the complex superposition of drug effects.

[0044] Finally, regarding the generation and injection mechanism of the high-risk proxy vector, this vector is not randomly initialized, but rather obtained through feature embedding from a graph neural network (GNN) pre-trained on a large-scale labeled medical atlas. During the offline training phase, the system constructs a heterogeneous information network encompassing symptoms, medications, images, and diagnoses, and learns the low-dimensional vector representation of each node through random walks and node embedding algorithms (such as Node2Vec). The high-risk proxy vector is the feature vector corresponding to the vertebral compression fracture diagnosis node in this vector space. When the system runs online and triggers the hidden event proxy logic, it directly calls this pre-trained vector as the output of the semantic feature mapping module, ensuring that the injected proxy vector is within the feature space. It exhibits geometric consistency with real-world image pathology semantics, enabling the cross-attention fusion module to correctly parse the data and effectively activate the weights of biochemical indicators related to fracture risk in temporal features. This achieves semantic alignment from discrete symbolic logic to a continuous vector space. The data interface module did not collect unstructured image report text, and the prescription gradient value output by the prescription gradient monitoring module was below the mutation threshold. The semantic feature mapping module executed the default baseline filling procedure, calling the baseline stationary vector in the pre-set read-only memory as the morphological feature vector and transmitting it to the cross-attention fusion module. The baseline stationary vector construction follows the zero-bias principle, and its value is determined by performing a full arithmetic mean operation on the semantic features embedded from healthy sample images in the standard reference database. The dimensionality is limited to a fixed value consistent with the output dimension of the word embedding model, such as 512 or 1024 dimensions, to ensure dimensionality compatibility in subsequent matrix multiplication operations. The cross-attention fusion module uses the baseline stationary vector as the query vector to perform attention calculation, generating an attention weight matrix. The numerical representation is uniformly distributed or low-pass filtered across time steps for the latent state vector, ensuring that the final multimodal fusion feature vector retains the original metabolic trend information of biochemical test indicators over time, without adding additional morphological risk semantic weighting. This logic constitutes a degraded operation mode when the system's imaging data is missing and drug prescription behavior is in a stable period, allowing the risk calculation module to output a risk layer reflecting the body's basic bone metabolic state based on the simple biochemical metabolic residual signal. The system determines the level of judgment results, maintains the connectivity and logical loop of the data flow in the calculation graph of the all-time monitoring task, and performs offline statistical calculations for the calibration and update of the baseline stationary vector. It selects a sample set of samples from the historical database whose bone mineral density T-values ​​by dual-energy X-ray absorptiometry are greater than -1.0 and who have not experienced fracture events within the past year. It extracts the feature vectors corresponding to all image reports in the set and calculates the geometric center coordinates to generate an initial baseline stationary vector. The system then saves this vector to the model parameter file. In subsequent runs, only when the prescription gradient monitoring module detects that the gradient values ​​of three consecutive sliding windows have returned to near zero will the system trigger a fine-tuning and update operation of the baseline vector, thus eliminating the statistical noise interference introduced by short-term drug fluctuations and maintaining the long-term stability of the baseline reference standard.

[0045] Example 5: The system executes a standardized offline calibration procedure based on retrospective cohort analysis. This procedure extracts samples from the hospital's historical database that have received a standard single anti-osteoporosis drug intervention and have high-frequency biochemical indicator monitoring records as a reference cohort. For each target drug, a nonlinear least squares algorithm is used to fit the discrete observation points of biochemical indicators changing over time in the reference cohort to a double exponential decay model E(t)=Ae. -αt +Be -βt The system uses an iterative optimization algorithm to minimize the mean square error between the theoretical curve and the actual observed data to calculate the specific values ​​of the drug's pharmacokinetic constants A, B, α, and β. These statistically validated parameters are then stored in a pharmacokinetic mapping table to complete the engineering transformation from raw clinical data to computational model parameters. After initializing and filling the mapping table, the system uses validation set data, which is mutually exclusive with the reference queue, to execute closed-loop validation logic for model validity. The patient's drug prescription records in the validation set are input into the residual calculation unit to generate theoretically expected biochemical values, which are then compared with actual detection values ​​to calculate the predicted residual sequence. The system determines that the current mapping table parameters are qualified and locks them for online operation only when the statistical distribution characteristics of the predicted residual sequence meet the conditions of zero mean and variance less than a preset engineering tolerance threshold. The system also runs a periodic parameter drift monitoring logic. When changes in the basic statistical distribution of biochemical indicators or drug dosage form cause the variance of the predicted residual to continuously exceed the tolerance threshold, the system automatically triggers a calibration procedure to reconstruct and update the mapping table parameters.

[0046] Example 6: To address the potential discrimination bias issues arising from differences in biochemical testing equipment benchmarks and varying local epidemiological characteristics when deploying a multimodal early warning system across institutions, the system incorporates a standardized pre-deployment calibration and parameter adaptation procedure. This procedure is forcibly executed after the system connects to a new hospital information system instance and before real-time monitoring services are initiated. The data interface module performs a baseline statistical distribution alignment operation, constructing a local background noise dataset D by batch retrieving biochemical data from physical examination populations with no records of bone metabolic diseases within the past 90 calendar days of the medical institution. local Furthermore, the local arithmetic mean μ of key indicators such as serum calcium, phosphorus, and alkaline phosphatase was calculated. local and standard deviation σ local The system then updates the preprocessing parameters in the temporal feature encoding module, reconstructing the Z-score normalization formula as z=(x-μ local ) / σ local This eliminates the input feature distribution drift caused by errors in the detection instrument system, where x is the actual raw detection value of the biochemical test index collected by the data interface module, and μ local σ is the local arithmetic mean. localThe local standard deviation is used to ensure that the numerical features input to the neural network are physically isomorphic to the original training data. After data baseline alignment is completed, the risk calculation module executes the on-site optimization logic of the decision boundary. It uses a local validation set (sample size not less than 100 cases) containing confirmed osteoporotic fracture cases and healthy controls for inference calculation to generate the original risk probability set P output by the model. The system traverses the high-risk judgment threshold θ in the continuous interval [0,1] with a step size of 0.01, and calculates the corresponding Youden index J(θ) = Sensitivity(θ) + Specificity(θ) - 1 for each θ value. The system automatically locks the θ that makes J(θ) reach its maximum value. opt This parameter serves as the threshold for determining the high-risk level in this deployment environment and is embedded into the configuration file of the risk calculation module.

[0047] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0048] Finally, 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.

Claims

1. A multimodal osteoporosis stratified early warning system, characterized in that, include: The data interface module is used to connect the hospital information system and the radiology information system to collect structured time-series data and unstructured image report text of the target object, respectively. The structured time-series data includes the time series of biochemical test indicators and drug prescription records. The prescription gradient monitoring module, connected to the data interface module, is used to extract the prescription dosage of analgesics from the drug prescription record and calculate the first derivative of the prescription dosage over time to generate prescription gradient values. The semantic feature mapping module is connected to the data interface module and the prescription gradient monitoring module respectively, and is used to execute the feature vector generation logic: when the data interface module collects unstructured image report text, it performs vectorization encoding on the unstructured image report text based on the preset skeletal morphology terminology library to generate morphological feature vectors; when the data interface module does not collect unstructured image report text and the prescription gradient value exceeds the preset threshold, it calls the preset high-risk proxy vector as the morphological feature vector. The time-series feature encoding module, connected to the data interface module, is used to perform standardization processing on biochemical test indicators and input the processed data into a pre-set recurrent neural network unit to generate a hidden state vector representing metabolic trends. The cross-attention fusion module is connected to the semantic feature mapping module and the temporal feature encoding module respectively. It is used to use the morphological feature vector as the query vector and the hidden state vector as the key vector and value vector to perform cross-attention calculation. It adjusts the weights of each temporal feature dimension in the hidden state vector according to the semantic information in the morphological feature vector and generates a multimodal fusion feature vector. The risk calculation module, connected to the cross-attention fusion module, is used to map the multimodal fusion feature vector to a preset risk probability space and output the risk level to which the target object belongs; through the joint determination of the mutation of treatment behavior gradient and metabolic residual abnormality, potential risks of occult vertebral compression fractures are identified. The time-series feature encoding module also includes a residual calculation unit, which stores a pre-set pharmacodynamic mapping table. The pharmacodynamic mapping table defines the correspondence between the dose of anti-osteoporosis drugs and the theoretical changes in biochemical indicators. The residual calculation unit is used to calculate the theoretical expected biochemical values ​​based on the drug prescription record and the pharmacodynamic mapping table, and to calculate the difference between the actual detected values ​​of biochemical test indicators and the theoretical expected biochemical values ​​to generate a metabolic residual vector. The temporal feature encoding module is used to transmit the metabolic residual vector as input data to the recurrent neural network unit to extract the hidden state vector that characterizes the body's drug response capability; The cross-attention fusion module includes a static cross-attention layer, which is used to calculate the attention score between the morphological feature vector and the hidden state vector, and to use the attention score to perform a weighted summation of the hidden state vector; when the morphological feature vector is a high-risk surrogate vector, the attention score forcibly increases the weight of the biochemical index dimension representing the bone metabolic stress response in the hidden state vector. The skeletal morphology terminology database contains feature roots describing trabecular bone structure, vertebral body edge morphology, and bone density; vectorization encoding uses a pre-trained word embedding model to convert the matched feature roots into high-dimensional sparse vectors.

2. The multimodal osteoporosis stratified early warning system according to claim 1, characterized in that, The semantic feature mapping module also stores a pre-set basic risk prior vector and is used to execute time decay calibration logic: it obtains the generation time of the unstructured image report text and calculates the time interval between the generation time and the current system time; it then uses a pre-set nonlinear decay function to map the time interval to a normalized semantic confidence coefficient. And generate time-calibrated morphological feature vectors according to the following formula. , where V s The morphological feature vector V generated by the semantic feature mapping module base The basic risk prior vector is used; the cross-attention fusion module is used to utilize the time-calibrated morphological feature vector as the query vector.

3. The multimodal osteoporosis stratified early warning system according to claim 1, characterized in that, The prescription gradient monitoring module monitors analgesics including nonsteroidal anti-inflammatory drugs and opioids; the preset threshold is set as the critical value for prescription frequency mutations that characterize acute pain events; the high-risk proxy vector is a pre-trained feature embedding vector that characterizes the semantics of vertebral compression fractures.

4. The multimodal osteoporosis stratified early warning system according to claim 1, characterized in that, The recurrent neural network unit is a gated recurrent unit, which is used to process the biochemical test index sequence after interpolation and capture the dynamic dependence of the biochemical test index over time.

5. The multimodal osteoporosis stratified early warning system according to claim 1, characterized in that, The data interface module is also used to perform data cleaning logic, perform linear interpolation on missing data points in biochemical test indicators, and perform Z-score standardization on all numerical data to eliminate dimensional differences.

6. The multimodal osteoporosis stratified early warning system according to claim 1, characterized in that, The risk calculation module includes a fully connected classification layer and a Softmax activation function, which output the probability distribution of the target object belonging to low-risk, medium-risk, and high-risk levels, and take the level with the highest probability value as the final warning result.

Images