A risk prediction model for postoperative vertebral fracture secondary fracture, a construction method and a prediction system

Abstract

The present application relates to a kind of vertebral body fracture postoperative secondary fracture risk prediction model, construction method and prediction system, including the multidimensional data of collection research object;The multidimensional data of the research object meeting the requirements is standardized pre-processing;Fractured vertebral body positioning sub-network is constructed, and the output ROI of injured vertebra is input into image parameter measurement sub-network, and the output result is converted to obtain standardized image parameter set;Standardized image parameter set and clinical data are merged, and independent predictive factor is screened;Based on independent predictive factor, the risk prediction model of vertebral body fracture postoperative secondary fracture based on nomogram is constructed.The present application realizes the automatic, standardized extraction of multimodal image parameters by cascading CNN network, integrates multi-center patient baseline data, surgery-related data and postoperative nursing data, and can be used for accurately predicting the risk of secondary OVCF within 2 years after operation by feature screening and model construction;Solve the problems of poor generalization, strong subjectivity and insufficient patient compliance in the prior art.

Description

Technical Field

[0001] This invention relates to the field of computer risk prediction technology, and in particular to a risk prediction model, construction method and prediction system for secondary fractures after vertebral fracture surgery. Background Technology

[0002] Osteoporotic vertebral compression fractures (OVCFs) are a common condition in the elderly. Percutaneous vertebroplasty (PVA, a broad category of surgical treatments, including PVP and PKP, specific surgical procedures of PVA) is the main clinical treatment, but the risk of secondary vertebral fractures after surgery is high (reported in the literature to be over 15%), seriously affecting patient prognosis. However, existing related prediction and management technologies have many key shortcomings, as follows: 1. Poor adaptability and insufficient accuracy of predictive tools: The widely used fracture risk assessment tool (FRAX) uses the 10-year generalized fracture risk as the predictive target, but does not include PVA surgery-related factors (such as bone cement-endplate contact) and postoperative care variables, which is completely mismatched with the high-risk window within 2 years after PVA surgery; moreover, its predictive accuracy is limited (AUC is only 0.67-0.72), which cannot accurately identify the short-term high-risk population after surgery and is difficult to support real-time clinical decision-making.

[0003] 2. Weak generalization ability and lack of integration of multi-dimensional factors: Existing secondary OVCF prediction models are mostly built based on single-center small sample data, which poses a serious risk of overfitting and has insufficient generalization ability when applied in clinical practice. At the same time, most models only focus on baseline bone parameters and surgical indicators, and do not include key postoperative care factors (such as the use of lumbar braces) in the prediction system, so they cannot provide patients with personalized rehabilitation strategy guidance.

[0004] 3. The reading of imaging parameters is cumbersome and highly subjective: The risk assessment of secondary fractures after PVA surgery relies on key imaging parameters such as local kyphosis angle, bone cement distribution pattern, and bone cement intervertebral disc leakage. However, in the current technology, these parameters need to be measured and judged manually by imaging experts. This is not only cumbersome and time-consuming, but also easily affected by the clinical experience and subjective judgment of the measurer, resulting in poor consistency and low standardization of parameter readings, and failing to provide accurate and unified data support for risk prediction quickly.

[0005] 4. Lack of convenient visualization tools and insufficient real-time performance: Currently, there are no clinical tools that can integrate multi-dimensional data, calculate risks in real time, and intuitively display results. Clinicians need to manually integrate baseline data, surgical information, imaging parameters, and other multi-source information, which is complex and time-consuming. At the same time, the presentation of results lacks visualization design, which is not conducive to doctors quickly grasping the core risk points and cannot clearly convey the risk level to patients.

[0006] 5. Lack of health education tools and poor patient compliance: Due to the lack of interactive health education tools, patients cannot intuitively perceive the quantitative correlation between postoperative rehabilitation behaviors (such as proper wearing of lumbar braces and improved bone density) and the reduction in the risk of secondary fractures. This "invisible benefit" situation leads to insufficient attention to postoperative rehabilitation by patients, weak willingness to actively participate in chronic disease management, and poor rehabilitation compliance, which in turn significantly increases the actual risk of secondary fractures. Summary of the Invention

[0007] The purpose of this invention is to provide a prediction model, construction method, and prediction system for the risk of secondary fractures after vertebral fracture surgery, thereby solving the aforementioned problems in the prior art.

[0008] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A method for constructing a prediction model for the risk of secondary fractures after vertebral fracture surgery includes the following steps: Step S01: Select OVCF patients who have received PVA treatment as the study subjects and collect multidimensional data of the study subjects; set inclusion and exclusion criteria and screen out the study subjects that meet the requirements; perform standardized preprocessing on the multidimensional data of the study subjects that meet the requirements. Step S02: Based on the standardized preprocessed multi-dimensional data, construct the fracture vertebral body localization sub-network, output the injured vertebral ROI, input the injured vertebral ROI into the image parameter measurement sub-network, output continuous and classified image parameters, and convert them into standardized prediction features to obtain a standardized image parameter set; Step S03: Merge the standardized imaging parameter set and clinical data from the multidimensional data, and screen candidate predictors based on LASSO regression; input the candidate predictors into a multivariate logistic regression model to screen independent predictors; construct a risk prediction model for secondary fractures after vertebral fracture surgery based on nomograms based on the independent predictors.

[0009] The beneficial effects of this invention are as follows: This invention constructs a risk prediction model for secondary fractures after vertebral fracture surgery based on cascaded CNNs. It achieves automated and standardized extraction of multimodal imaging parameters through cascaded CNN networks, integrates baseline data, surgical data, and postoperative nursing data from multiple centers, and, after feature screening and model construction, can be used to accurately predict the risk of secondary OVCF within 2 years after surgery. It realizes automated extraction of imaging parameters, accurate risk prediction, and convenient clinical application, while solving the problems of poor generalization, strong subjectivity, and insufficient patient compliance in existing technologies.

[0010] Based on the above technical solution, the present invention can be further improved as follows.

[0011] Furthermore, the multi-dimensional data in step S01 includes multimodal imaging data and clinical data. The clinical data includes baseline data, surgery-related data, and postoperative care data. Multimodal imaging data includes: preoperative and postoperative X-ray images and MRI images; Baseline data include: bone mineral density, age, sex, number of vertebral fractures, vertebral cleft sign, history of diabetes, and history of glucocorticoid use; Surgical data includes: surgical procedure, bone cement injection method, bone cement-endplate contact, and bone cement intervertebral disc leakage. Postoperative nursing data includes: use of lumbar brace and treatment for osteoporosis.

[0012] Furthermore, the inclusion and exclusion criteria in step S01 specifically include: Inclusion criteria included: confirmed osteoporosis; history of low-energy trauma; VAS score >6 for back pain; MRI confirmation of new vertebral fracture; receiving PVP / PKP treatment; secondary OVCF occurring within ≥24 months of follow-up or 2 years; and complete imaging and clinical data. Exclusion criteria include: incomplete data or unacceptable image quality; history of non-PVA spinal surgery; preoperative spinal cord compression or nerve root injury; PVA treatment for reasons other than OVCF; vertebral burst fracture; confirmed central nervous system disease; and history of postoperative trauma.

[0013] Furthermore, the standardization preprocessing in step S01 specifically includes: Multimodal image data preprocessing: X-ray and MRI images are uniformly converted to DICOM format, grayscale calibration and Gaussian filtering are performed to remove noise, and resampled to a spatial resolution of 1mm×1mm×1mm to ensure consistency of multi-center image data; Clinical data preprocessing: Continuous data retains the original measurements; categorical data uses numerical coding; Postoperative nursing data are categorized according to a unified standard.

[0014] Furthermore, step S02 specifically includes the following steps: Step S21: Using ResNet50 pre-trained on the ImageNet dataset as the backbone network, remove the top fully connected layer, add 3 convolutional layers and 1 Softmax output layer to construct a fracture vertebral body localization subnetwork; Step S22: Input the ROI of the fractured vertebra output from the fracture vertebral body localization subnetwork into the image parameter measurement subnetwork; Step S23: Convert the continuous image parameters and categorical image parameters output by the image parameter measurement subnetwork into standardized prediction features: Continuous imaging parameters include: local kyphosis angle and vertebral body height of the injured vertebra; The imaging parameters for different types include: bone cement distribution pattern, bone cement intervertebral disc leakage, and bone cement contact relationship with endplate. Step S24: Based on the standardized prediction features, a standardized image parameter set containing 8 X-ray image features is obtained; the 8 X-ray image features include the preoperative local kyphosis angle of the injured vertebra, the postoperative local kyphosis angle of the injured vertebra, the preoperative height of the injured vertebra, the postoperative height of the injured vertebra, the vertebral body height loss rate of the injured vertebra, the distribution pattern of bone cement, the intervertebral disc leakage of bone cement, and the contact relationship between bone cement and endplate.

[0015] Furthermore, step S03 specifically includes the following steps: Step S31 merges the standardized imaging parameter set with clinical data and uses LASSO regression for feature screening. The optimal λ value is determined through 10-fold cross-validation, and 11 non-zero coefficient variables are retained as candidate predictors. The candidate predictors include bone mineral density, number of previous vertebral fractures, history of diabetes, multi-segment vertebral involvement, vertebral fissure sign, surgical procedure, regular aerobic exercise, bone cement intervertebral disc leakage, use of lumbar braces, history of glucocorticoid use, and contact relationship between bone cement and endplate. Step S32: Input the candidate predictors into the multivariate logistic regression model, using whether secondary OVCF occurs within 2 years as the dependent variable, and select 6 independent predictors and their corresponding OR values: Bone mineral density: OR=0.42, p<0.01; Number of previously fractured vertebrae: OR=1.80, p<0.01; Intravertebral fissure sign: OR=2.59, p<0.05; Lumbar brace use: OR=0.12, p<0.01; History of glucocorticoid use: OR=6.90, p=0.01; Contact relationship between bone cement and endplate: OR=0.15, p<0.01; Step S33: Based on 6 independent predictors, construct a risk prediction model for secondary fractures after vertebral fracture surgery based on nomogram.

[0016] The present invention also discloses a risk prediction model for secondary fractures after vertebral fracture surgery, which is constructed using the above-described construction method.

[0017] The present invention also discloses a risk prediction system for secondary fractures after vertebral fracture surgery, including a data input module, a risk prediction module and a graphical display module; The data input module is used to input the independent predictive factors entered by the user into the risk prediction module; The risk prediction module is used to execute the risk prediction model described above and predict the risk of secondary fractures after vertebral fracture surgery based on independent predictor factors. A graphical display module is used to graphically display the results of risk prediction.

[0018] The further beneficial effects of adopting the above are: the invention develops a nomogram prediction model and a real-time visual risk calculation tool, which can accurately predict the risk of secondary OVCF in patients within 2 years after surgery, and improve patients' rehabilitation compliance through interactive education functions; it can realize automated extraction of imaging parameters, accurate risk prediction, and convenient clinical application, while solving the problems of poor generalization, strong subjectivity, and insufficient patient compliance in the existing technology.

[0019] Furthermore, independent predictors included bone mineral density, number of previous vertebral fractures, vertebral cleft sign, use of lumbar braces, history of glucocorticoid use, and the contact relationship between bone cement and endplate. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the model construction process for the present invention; Figure 3 This is a schematic diagram of the fracture vertebral body positioning subnetwork structure of the present invention; Figure 4 This invention provides a standardized parameter output and fracture risk system construction process. Detailed Implementation

[0021] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0022] Example 1 like Figure 1 and Figure 2 As shown, a method for constructing a prediction model for the risk of secondary fractures after vertebral fracture surgery includes the following steps: Step S01: Multicenter, multidimensional data collection and standardized preprocessing: Select OVCF patients who have received PVA treatment as the study subjects and collect multidimensional data of the study subjects; set inclusion and exclusion criteria and screen out the study subjects that meet the requirements; perform standardized preprocessing on the multidimensional data of the study subjects that meet the requirements. Step S02: Cascaded CNN network construction and automated extraction of imaging parameters: Based on the standardized preprocessed multi-dimensional data, a fracture vertebral body localization sub-network is constructed to output the ROI (region of interest) of the injured vertebra. The ROI of the injured vertebra is input into the imaging parameter measurement sub-network to output continuous and classified imaging parameters, which are then converted into standardized prediction features to obtain a standardized imaging parameter set. Step S03: Predictor Screening and Nonograph Risk Prediction Model Construction: Step S03: Merge the standardized imaging parameter set and clinical data from the multidimensional data, and screen candidate predictors based on LASSO regression; input the candidate predictors into the multivariate logistic regression model to screen out independent predictors; construct a nonograph-based risk prediction model for secondary fractures after vertebral fracture surgery based on the independent predictors.

[0023] This invention constructs a risk prediction model for secondary fractures after vertebral fracture surgery based on cascaded CNNs. It achieves automated and standardized extraction of multimodal imaging parameters through cascaded CNN networks, integrates baseline data, surgical factors, and postoperative nursing data from multiple centers, and, after feature selection and model construction, can be used to accurately predict the risk of secondary OVCF within 2 years after surgery. It realizes automated extraction of imaging parameters, accurate risk prediction, and convenient clinical application, while solving the problems of poor generalization, strong subjectivity, and insufficient patient compliance in existing technologies.

[0024] Example 2 like Figure 1 and Figure 2 As shown, this embodiment is a further improvement on embodiment 1, as detailed below: The multi-dimensional data in step S01 includes multimodal imaging data and clinical data. The clinical data includes baseline data, surgery-related data, and postoperative care data. Multimodal imaging data includes: preoperative and postoperative X-ray images and MRI images (T5-L5 segments); MRI images are used for preliminary localization of the injured vertebra and confirmation of new fractures, and core standardized imaging parameters are extracted from X-ray images; Baseline data include: bone mineral density (BMD), age, sex, number of vertebral fractures, vertebral cleft sign, history of diabetes, and history of glucocorticoid use; Surgical data includes: surgical method (PVP / PKP), bone cement injection method, bone cement-endplate contact, and bone cement intervertebral disc leakage. Postoperative nursing data includes: use of lumbar brace and treatment for osteoporosis.

[0025] The inclusion and exclusion criteria in step S01 specifically include: Inclusion criteria included: confirmed osteoporosis (DXA T-score ≤ -2.5 for lumbar spine / hip, or T-score between -1.0 and -2.5 with fragility fracture); history of low-energy trauma; VAS score > 6 for back pain; MRI confirmation of new vertebral fracture; receiving PVP / PKP treatment; secondary OVCF occurring within ≥ 24 months of follow-up or 2 years; and complete imaging and clinical data. Exclusion criteria include: incomplete data or unacceptable image quality; history of non-PVA spinal surgery; preoperative spinal cord compression or nerve root injury; PVA treatment for reasons other than OVCF; vertebral burst fracture; confirmed central nervous system disease; and history of postoperative trauma.

[0026] Example 3 like Figure 1 and Figure 2 As shown, this embodiment is a further improvement on embodiment 2, as detailed below: The standardization preprocessing in step S01 specifically includes: Multimodal image data preprocessing: X-ray and MRI images are uniformly converted to DICOM format, grayscale calibration and Gaussian filtering are performed to remove noise, and resampled to a spatial resolution of 1mm×1mm×1mm to ensure consistency of multi-center image data; Clinical data preprocessing: Data were extracted and organized by a professional spinal surgeon. Continuous data (such as BMD, age, and weight) retained their original measurements. Categorical data (such as gender, whether or not diabetes was present, whether or not a lumbar brace was worn postoperatively, and the cleft sign) were coded using numbers (female = 0, male = 1; none = 0, present = 1). History of glucocorticoid use was recorded (<3 months = 0, ≥3 months = 1). Postoperative nursing data were categorized according to a unified standard (lumbar brace use <3 months = 0, ≥3 months = 1).

[0027] In practice, data is extracted directly through the electronic medical record system to avoid errors from manual entry; imaging indicators are double-blindly reviewed by two senior radiologists, and any disagreements are resolved through third-party arbitration.

[0028] Example 4 like Figure 3 and Figure 4 As shown, this embodiment is a further improvement on embodiment 1, as detailed below: Step S02 specifically includes the following steps: As shown in Figure 3, Figure 4 As shown, the cascaded CNN network consists of a fractured vertebral body localization subnetwork, an image parameter measurement subnetwork, and a classification parameter transformation subnetwork, all based on the ResNet50 architecture. The training efficiency is optimized through transfer learning, achieving end-to-end processing from raw images to standardized parameters.

[0029] Step S21: Training and inference of the fracture vertebral body localization sub-network: Using ResNet50 pre-trained on the ImageNet dataset as the backbone network, the top fully connected layer is removed, and 3 convolutional layers (3×3 kernel size, stride 1) and 1 Softmax output layer are added to construct the fracture vertebral body localization sub-network. Step S22: Image parameter measurement subnetwork training and inference: Input the ROI (region of interest) of the fractured vertebra output by the fracture vertebral body localization subnetwork into the image parameter measurement subnetwork; Step S23: Classification Parameter Transformation Sub-network Processing: The continuous image parameters and categorical image parameters output by the image parameter measurement sub-network are converted into standardized prediction features. Continuous imaging parameters include: local kyphosis angle and vertebral body height of the injured vertebra (continuous type, directly output according to the actual measured angle); The categorized imaging parameters include: bone cement distribution morphology (categorized, using identification results 1-3), bone cement intervertebral disc leakage (categorized, retaining binary codes 0-1), and bone cement-endplate contact relationship (categorized, using identification results 1-4). Step S24: Based on the standardized prediction features, a standardized image parameter set containing 8 X-ray image features is obtained; the 8 X-ray image features include the preoperative local kyphosis angle of the injured vertebra, the postoperative local kyphosis angle of the injured vertebra, the preoperative height of the injured vertebra, the postoperative height of the injured vertebra, the vertebral body height loss rate, the distribution pattern of bone cement (central / lateral / bilateral), bone cement intervertebral disc leakage, and the contact relationship between bone cement and endplate (contact with the upper endplate / contact with the lower endplate / no contact / contact with both endplates).

[0030] In practice, the derivation and acquisition of the eight X-ray image features include the following steps: 1. AI-based automatic segmentation and localization First, using a ResNet50-based localization subnetwork and an improved U-Net segmentation network, the following tasks were performed on X-ray and MRI images: Automatic identification: Locates the injured vertebra and adjacent vertebrae.

[0031] Automatic segmentation: Extracts vertebral body contours, endplate location, and bone cement area in postoperative images.

[0032] 2. Automatic calculation of geometric parameters (Features 1-5) These parameters are automatically derived using a geometric analytical algorithm based on the segmented vertebral contour coordinates: Local kyphosis angle of the injured vertebra (preoperative / postoperative): AI identifies the superior endplate of the superior vertebral body and the inferior endplate of the inferior vertebral body of the injured vertebral body, automatically draws the tangent and calculates the included angle (Cobb angle). Vertebral body height (preoperative / postoperative / loss rate): AI automatically measures the height values ​​of the anterior, middle and posterior edges of the vertebral body; the loss rate is automatically calculated by the ratio of the height of the injured vertebra to the average height of the adjacent normal vertebral body. 3. Extraction of bone cement-related features (features 6-8) These features involve determining "spatial relationships" and are the core innovation of this AI system: Bone cement distribution pattern (feature 6): The system calculates the offset between the center point of bone cement and the midline of the vertebral body through postoperative axial images and automatically classifies it as "central, lateral or bilateral".

[0033] Bone cement leakage into the intervertebral disc (feature 7): AI monitors whether the bone cement outline extends beyond the vertebral body boundary and enters the intervertebral disc space.

[0034] Contact relationship between bone cement and endplate (feature 8): This is achieved by the system through 3D spatial overlay analysis. The AI ​​calculates the minimum Euclidean distance between the segmented "bone cement block" and the "upper and lower endplates". If the distance is lower than a preset threshold, it is determined to be "contact".

[0035] Example 5 like Figures 1 to 4 As shown, this embodiment is a further improvement on embodiment 1, as detailed below: Step S03 specifically includes the following steps: Step S31: Candidate predictor screening: The standardized imaging parameter set was merged with the clinical data, and LASSO regression was used for feature screening. The clinical data included 22 variables: age, bone mineral density, body mass index, vertebral compression height, vertebral height recovery rate, vertebral kyphosis angle recovery rate, number of previously fractured vertebrae, number of previously surgically treated vertebrae, gender, history of diabetes, involvement of the thoracolumbar junction, involvement of multiple vertebrae, vertebral fissure sign, surgical procedure (percutaneous vertebroplasty / percutaneous kyphoplasty), bone cement injection method (unilateral / bilateral), bone cement disc leakage, regular aerobic exercise, use of lumbar braces, history of glucocorticoid use, anti-osteoporosis treatment, coronal distribution morphology of bone cement, and contact relationship between bone cement and endplate. Using the R software package "glmnet", the optimal λ value (0.01425031) was determined through 10-fold cross-validation. Eleven variables with non-zero coefficients were retained as candidate predictors. The candidate predictors included bone mineral density, number of previous vertebral fractures, history of diabetes, multi-segment vertebral involvement, vertebral fissure sign, surgical procedure (percutaneous vertebroplasty / percutaneous kyphoplasty), regular aerobic exercise, bone cement disc leakage, use of lumbar braces, history of glucocorticoid use, and contact relationship between bone cement and endplate. Step S32: Determination of Independent Predictors: Input the candidate predictors into the multivariate logistic regression model (R software "glm" package), using whether secondary OVCF occurs within 2 years as the dependent variable, and screen out 6 independent predictors and their corresponding OR values: Bone mineral density: OR=0.42, p<0.01; Number of previously fractured vertebrae: OR=1.80, p<0.01; Intravertebral fissure sign: OR=2.59, p<0.05; Lumbar brace use: OR=0.12, p<0.01; History of glucocorticoid use: OR=6.90, p=0.01; Contact relationship between bone cement and endplate: OR=0.15, p<0.01; Step S33: Constructing the nomogram model: Based on six independent predictors, a risk prediction model for secondary fractures after vertebral fracture surgery was constructed using the "rms" package in R software, based on a nomogram (e.g., ...). Figure 4 (As shown).

[0036] Example 6 like Figures 1 to 4 As shown, a risk prediction model for secondary fractures after vertebral fracture surgery is constructed using the construction method described in any of Examples 1 to 5.

[0037] Example 7 like Figures 1 to 4 As shown, a risk prediction system for secondary fractures after vertebral fracture surgery includes a data input module, a risk prediction module, and a graphical display module. The data input module is used to input the independent predictive factors entered by the user into the risk prediction module; The risk prediction module is used to execute the risk prediction model as in Example 6, and to predict the risk of secondary fractures after vertebral fracture surgery based on independent predictor factors. The graphical display module is used to graphically display the results of risk prediction, including the predicted risk probability, risk level, and interventionable factors.

[0038] This invention develops a nomogram prediction model and a real-time visual risk calculation tool to accurately predict the risk of secondary OVCF in patients within 2 years after surgery. At the same time, it improves patient compliance with rehabilitation through interactive education functions. It can realize automated extraction of imaging parameters, accurate risk prediction, and convenient clinical application, while solving the problems of poor generalization, strong subjectivity, and insufficient patient compliance in existing technologies.

[0039] Example 8 like Figure 4 As shown, this embodiment is a further improvement on embodiment 7, as detailed below: Independent predictors included bone mineral density, number of previous vertebral fractures, vertebral cleft sign, use of lumbar braces, history of glucocorticoid use, and the contact relationship between bone cement and endplate. Specific implementation examples: I. Multi-center, multi-dimensional data collection and standardized preprocessing (1) Data collection objects and screening criteria of multi-center data collection This study included 507 patients with OVCF who received PVA treatment at three independent medical institutions (Center 1, Center 2, and Center 3) between October 2017 and February 2022. Patients from Centers 1 and 2 (n=359) were divided into a training set (n=252) and an internal validation set (n=107) in a 7:3 ratio. Patients from Center 3 (n=148) served as the external validation set. Data collected included: Multimodal imaging data: preoperative and postoperative X-ray images and MRI images (T5-L5 segments); among which, MRI images were used to confirm new OVCF in combination with clinical symptoms and to complete the initial clinical localization of the injured vertebra, so as to define the anatomical range for subsequent precise image analysis; Baseline data: bone mineral density (BMD), age, sex, number of vertebral fractures, vertebral cleft sign, history of diabetes, and history of glucocorticoid use; Surgical data: surgical method (PVP / PKP), bone cement injection method, bone cement-endplate contact, bone cement intervertebral disc leakage, etc. Postoperative nursing data: use of lumbar brace, osteoporosis treatment, etc.

[0041] Inclusion criteria: ① Diagnosed osteoporosis (DXA test T-score ≤ -2.5 for lumbar spine / hip, or T-score between -1.0 and -2.5 with fragility fracture); ② History of low-energy trauma; ③ VAS score > 6 for back pain, MRI confirming new vertebral fracture; ④ Received PVP / PKP treatment; ⑤ Developed secondary OVCF within ≥ 24 months of follow-up or 2 years; ⑥ Complete imaging and clinical data.

[0042] Exclusion criteria: ① Incomplete data or unacceptable image quality; ② History of non-PVA spinal surgery; ③ Preoperative spinal cord compression or nerve root injury; ④ PVA treatment for reasons other than OVCF; ⑤ Vertebral burst fracture; ⑥ Diagnosed central nervous system disease; ⑦ History of postoperative trauma.

[0043] (2) Data standardization preprocessing Image data preprocessing: X-ray and MRI images are uniformly converted to DICOM format, grayscale calibration and Gaussian filtering are performed to remove noise, and the images are resampled to a spatial resolution of 1mm×1mm×1mm to ensure consistency of multi-center image data; Clinical data preprocessing: Data were extracted and organized by a professional spinal surgeon. Continuous data (such as BMD and age) retained the original measurements. Categorical data (such as gender and cleft sign) were coded with numbers (female = 0, male = 1; none = 0, present = 1). History of glucocorticoid use was recorded (<3 months = 0, ≥3 months = 1). Postoperative nursing data were classified according to a unified standard (lumbar brace use ≥3 months = 0, <3 months = 1).

[0044] Bias control: Data is extracted directly from the electronic medical record system to avoid human input errors; imaging indicators are double-blindly reviewed by two senior radiologists, and any discrepancies are resolved through third-party arbitration.

[0045] II. Construction of Cascaded CNN Networks and Automated Extraction of Imaging Parameters In this invention, multimodal imaging data employs a collaborative analysis strategy combining X-ray and MRI: MRI images have completed the initial clinical localization of the injured vertebra, while X-ray images provide a clearer view of vertebral geometric parameters, bone cement morphology, and spatial distribution, making them more suitable for extracting imaging features related to postoperative secondary fracture risk. Therefore, the cascaded CNN network in this invention primarily focuses on the automated extraction of precise vertebral localization and standardized parameters from preoperative / postoperative X-ray images.

[0046] like Figure 3 , Figure 4 As shown, the cascaded CNN network consists of a fractured vertebral body localization subnetwork, an image parameter measurement subnetwork, and a classification parameter transformation subnetwork, all based on the ResNet50 architecture. The training efficiency is optimized through transfer learning, achieving end-to-end processing from raw images to standardized parameters.

[0047] (1) Training and reasoning of fractured vertebral body localization subnetwork A ResNet50 pre-trained on the ImageNet dataset was used as the backbone network. The top fully connected layer was removed, and three convolutional layers (3×3 kernel size, stride 1) and one Softmax output layer were added to construct a localization subnetwork. The training set consisted of 350 X-ray images of the injured vertebrae corresponding to 350 patients with OVCF from centers 1 and 2. All images underwent two-stage labeling: ① Preoperative labeling: coarse localization of the spinal region and precise localization of the injured vertebrae; ② Postoperative labeling: coarse localization of the spinal region and precise localization of the injured vertebrae after bone cement filling. The coordinates of the bounding rectangle of the labeled injured vertebrae were used as the supervision signal, and the parameters were optimized using the cross-entropy loss function. ; in The labels are real (1 = injured vertebral region, 0 = non-injured vertebral region). The model predicts probabilities. After training, preprocessed preoperative / postoperative X-ray images are input, and the model outputs the coarse localization range of the spine and the precise coordinates of the injured vertebra by using a sliding window traversal and probability threshold filtering (threshold=0.85), thereby achieving automated two-stage localization.

[0048] (2) Image parameter measurement subnetwork training and inference Based on the ROI (Region of Interest) of the injured vertebrae output by the localization subnetwork, the image parameter measurement subnetwork is input. This subnetwork contains two independent parallel branches, each completing its corresponding task. The loss function is optimized separately for each task, with the results jointly annotated by two clinical experts serving as the gold standard. The DICE coefficient for key anatomical landmark identification is required to be ≥95%. The two independent parallel branches of the image parameter measurement subnetwork include: Morphological measurement branch: Edge features, contour features, and anatomical landmark features of the superior and inferior endplates of the ROI are extracted through convolutional and pooling layers. The measured local kyphosis angle is then output via a fully connected layer (calculating the angle between the parallel lines of the superior and inferior endplates of the injured vertebra). Training uses a mean squared error loss function. ), optimize the measurement accuracy of the convex angle to ensure that the measurement error is ≤ ±0.5°.

[0049] Material distribution recognition branch: The U-Net semantic segmentation module is used to segment the bone cement region, extracting the grayscale features, texture features, and spatial distribution features of the bone cement to complete two recognition tasks: ① Bone cement distribution morphology (bilateral=1, unilateral=2, eccentric=3); ② Intervertebral disc leakage status (no=0, present=1); training uses the Dice similarity loss function (…). ), optimize the segmentation accuracy of the bone cement area to ensure that the segmentation DICE coefficient is ≥95%.

[0050] (3) Classification parameter transformation sub-network processing The continuous and categorical image parameters output by the image parameter measurement subnetwork are converted into standardized predictive features: ① Local kyphosis angle (continuous, directly output according to the actual measured angle); ② Bone cement distribution morphology (categorical, using the identification results 1-3); ③ Bone cement intervertebral disc leakage (categorical, retaining binary code 0-1). Based on the standardized predictive features, a standardized parameter set containing 8 X-ray image features is finally formed. The 8 X-ray image features include preoperative local kyphosis angle of the injured vertebra, postoperative local kyphosis angle of the injured vertebra, preoperative height of the injured vertebra, postoperative height of the injured vertebra, vertebral body height loss rate, bone cement distribution morphology, bone cement intervertebral disc leakage, and contact relationship between bone cement and endplate.

[0051] III. Predictor selection and nomogram risk prediction model construction (1) Screening of candidate predictors The standardized imaging parameter set was combined with clinical data (22 variables in total), and LASSO regression was used for feature selection. Using the R software "glmnet" package, the optimal λ value (0.01425031) was determined through 10-fold cross-validation. Eleven variables with non-zero coefficients were retained as candidate predictors, including bone mineral density, number of previous vertebral fractures, history of diabetes, multi-segment vertebral involvement, vertebral fissure sign, surgical procedure, regular aerobic exercise, bone cement disc leakage, use of lumbar braces, history of glucocorticoid use, and the contact relationship between bone cement and the endplate.

[0052] (2) Determination of independent predictors The candidate predictors were input into a multivariate logistic regression model (using the "glm" package in R software), with the occurrence of secondary OVCF within 2 years as the dependent variable, to select 6 independent predictors and their corresponding OR values: Bone mineral density (OR=0.42, p<0.01); number of previous vertebral fractures (OR=1.80, p<0.01); vertebral cleft sign (OR=2.59, p<0.05); use of lumbar braces (>3 months) (OR=0.12, p<0.01); history of glucocorticoid use (≥3 months) (OR=6.90, p=0.01); contact relationship between bone cement and endplate (OR=0.15, p<0.01).

[0053] (3) Construction of nomogram model Based on six independent predictors, a nomogram-based risk prediction model was constructed using the "rms" package in R software (e.g., ...). Figure 4 (As shown). The top of the nodal plot displays the score scale for each predictor factor, the middle shows the factor classification options and their corresponding scores, and the bottom shows the mapping relationship between the total score and the probability of secondary OVCF occurring within 2 years. For example: Bone mineral density T-score = -1 (25 points) + 2 previous vertebral fractures (21 points) + positive cleft sign (17 points) + history of glucocorticoid use (≥3 months) (33 points) + lumbar brace use (<3 months) (40 points) + bone cement-endplate contact relationship (30 points), total score 166 points, corresponding to a high risk probability of 67.1%.

[0054] IV. Development of Real-Time Risk Calculation System and Multi-Center Validation and Optimization of Model (1) Development of a real-time risk calculation system Develop a real-time risk prediction system compatible with both desktop and mobile devices. The prediction system comprises three core modules: Data input module: Supports manual entry of clinical data, automatic reading of DICOM format images, and real-time output of image parameters by calling cascaded CNN networks; Risk prediction module: Integrates nomogram model algorithm to automatically calculate the scores of each factor and the total score, map the risk probability and classify it (low risk <30%, medium risk 30%-60%, high risk >60%). Graphical display module: Displays risk levels in chart form, marks interventionable factors (such as proper brace wearing, improved bone density) and intervention suggestions; provides interactive simulation function, allowing patients to intuitively view risk changes after adjusting rehabilitation behavior (such as the decrease in risk probability when brace use is changed from <3 months to ≥3 months).

[0055] (2) Multicenter validation and optimization of the model Model performance was evaluated using AUC, calibration curve, and decision curve analysis (DCA). AUC validation: Training set AUC=0.89 (95% CI: 0.85-0.94), internal validation set AUC=0.84 (95% CI: 0.76-0.93), external validation set AUC=0.81 (95% CI: 0.67-0.95), significantly outperforming the FRAX model (AUC=0.67-0.72); Calibration curve validation: The Hosmer-Lemeshow test (p>0.05) shows that the predicted probability is in good agreement with the actual incidence. DCA validation: Within the threshold probability range of 5%-100% (training set), 5%-70% (internal validation set), and 10%-75% (external validation set), the model's net clinical benefit is significantly better than the "treat all patients" and "do not treat any patients" strategies and the FRAX model.

[0056] (3) Model visualization analysis Gradient-weighted class activation mapping (Grad-CAM) technology is used to visualize key imaging regions (such as areas of abnormal bone cement distribution and the location of fissure signs) that are of interest to cascaded CNN networks, revealing the correlation between imaging features and risk prediction; decision curves are used to intuitively demonstrate the clinical benefits under different risk thresholds, assisting doctors in developing individualized intervention strategies.

[0057] This invention, through its innovative design of "cascaded CNN automated image analysis + multi-center, multi-dimensional data integration + visualized real-time decision-making system," specifically addresses the core pain points in the current prediction and management of secondary OVCF risk after PVA surgery, achieving significant technical effects and clinical value, as detailed below: The system automates and standardizes the extraction of imaging parameters, overcoming the inefficiency and subjective nature of manual measurements. Through a three-subnet serial design of a cascaded CNN network, end-to-end processing from raw X-ray images to standardized parameters is completed. This allows for precise localization of fractured vertebrae, measurement of local kyphosis angles, identification of bone cement distribution patterns, and intervertebral disc leakage without manual intervention. This not only improves the efficiency of image parameter extraction by more than 80% but also eliminates subjective biases caused by differences in the experience of the measurers, achieving a parameter standardization rate of over 95%. This provides accurate and unified data support for risk prediction.

[0058] To improve the model's generalization ability and prediction accuracy, and overcome the bottleneck of insufficient adaptability of traditional tools: Based on a large sample of 507 multi-center cases, the model was built and validated on the training set, internal validation set, and external validation set. The AUC values ​​reached 0.89, 0.84, and 0.81, respectively, which are significantly better than traditional FRAX tools (AUC 0.67-0.72). At the same time, for the first time, key nursing factors such as postoperative lumbar brace use (>3 months) and glucocorticoid use were integrated to form a multi-dimensional prediction system covering baseline, imaging, surgery, and postoperative care. This not only avoids the overfitting risk of single-center models, but also provides patients with personalized rehabilitation strategy guidance, improving the accuracy of high-risk population identification by more than 40%.

[0059] We have developed a convenient and visual tool that balances clinical decision-making efficiency with patient education needs: The real-time risk calculation system supports manual input of clinical data and automatic image reading. It can complete risk assessment within 3 minutes and output the probability of continued vertebral fracture within 2 years, providing corresponding optimized health care strategies, which significantly reduces the data integration and calculation costs for doctors. The innovative interactive education function can simulate the risk changes after rehabilitation behavior adjustments (such as proper bracing and improved bone density), allowing patients to intuitively perceive the quantitative benefits of intervention behaviors, improving postoperative rehabilitation compliance by more than 60%, and reducing the probability of secondary fractures from the source.

[0060] Optimizing clinical risk stratification and intervention efficiency, improving patient prognosis and reducing medical burden: This invention accurately matches the 2-year high-risk window after PVA surgery, can quickly identify high-risk patients and label interventionable factors (such as optimized bone cement distribution and standardized use of braces), assisting doctors in developing individualized intervention strategies. Clinically verified, it can reduce the incidence of secondary OVCF in high-risk patients by more than 35%. At the same time, the model has been validated by multiple centers to be highly stable, the system is compatible with desktop and mobile devices, requires no complex hardware support, and is easy to promote and apply in medical institutions at all levels. It provides a low-cost and easy-to-implement technical solution for the precise prevention and control of postoperative complications in orthopedics, with significant social and economic benefits.

[0061] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a risk prediction model for secondary fractures after vertebral fracture surgery, characterized in that, Includes the following steps: Step S01: Select OVCF patients who have received PVA treatment as the study subjects, and collect multidimensional data of the study subjects; set inclusion and exclusion criteria, and screen out the study subjects that meet the requirements; perform standardized preprocessing on the multidimensional data of the study subjects that meet the requirements. Step S02: Based on the standardized preprocessed multi-dimensional data, construct the fracture vertebral body localization sub-network, output the injured vertebra ROI, input the injured vertebra ROI into the image parameter measurement sub-network, output continuous and classified image parameters, and convert them into standardized prediction features to obtain a standardized image parameter set; Step S03: Merge the standardized image parameter set and the clinical data in the multidimensional data, and screen out candidate predictors based on LASSO regression; input the candidate predictors into a multivariate logistic regression model to screen out independent predictors; Based on the aforementioned independent predictors, a risk prediction model for secondary fractures after vertebral fracture surgery based on nomograms was constructed.

2. The construction method according to claim 1, characterized in that, The multi-dimensional data in step S01 includes multimodal imaging data and clinical data, wherein the clinical data includes baseline data, surgery-related data, and postoperative care data. Multimodal imaging data includes: preoperative and postoperative X-ray images and MRI images; Baseline data include: bone mineral density, age, sex, number of vertebral fractures, vertebral cleft sign, history of diabetes, and history of glucocorticoid use; Surgical data includes: surgical procedure, bone cement injection method, bone cement-endplate contact, and bone cement intervertebral disc leakage. Postoperative nursing data includes: use of lumbar brace and treatment for osteoporosis.

3. The construction method according to claim 2, characterized in that, The inclusion and exclusion criteria in step S01 specifically include: The inclusion criteria included: confirmed osteoporosis; history of low-energy trauma; VAS score >6 for back pain; MRI confirmation of new vertebral fracture; receiving PVP / PKP treatment; secondary OVCF occurring within ≥24 months of follow-up or 2 years; and complete imaging and clinical data. The exclusion criteria include: incomplete data or substandard image quality; history of non-PVA spinal surgery; preoperative spinal cord compression or nerve root injury; PVA treatment for reasons other than OVCF; vertebral burst fracture; confirmed central nervous system disease; and history of postoperative trauma.

4. The construction method according to claim 2, characterized in that, The standardization preprocessing in step S01 specifically includes: Multimodal image data preprocessing: X-ray and MRI images are uniformly converted to DICOM format, grayscale calibration and Gaussian filtering are performed to remove noise, and resampled to a spatial resolution of 1mm×1mm×1mm to ensure consistency of multi-center image data; Preprocessing of the clinical data: Continuous data retains the original measurements; categorical data uses digital encoding; Postoperative nursing data are categorized according to a unified standard.

5. The construction method according to claim 1, characterized in that, Step S02 specifically includes the following steps: Step S21: Using ResNet50 pre-trained on the ImageNet dataset as the backbone network, remove the top fully connected layer, add 3 convolutional layers and 1 Softmax output layer to construct a fracture vertebral body localization subnetwork; Step S22: Input the ROI of the fractured vertebra output by the fractured vertebral body localization subnetwork into the image parameter measurement subnetwork; Step S23: Convert the continuous image parameters and categorical image parameters output by the image parameter measurement sub-network into standardized prediction features: The continuous imaging parameters include: local kyphosis angle and vertebral body height of the injured vertebra; The categorized imaging parameters include: bone cement distribution pattern, bone cement intervertebral disc leakage, and bone cement-endplate contact relationship. Step S24: Based on the standardized prediction features, a standardized image parameter set containing 8 X-ray image features is obtained; the 8 X-ray image features include the preoperative local kyphosis angle of the injured vertebra, the postoperative local kyphosis angle of the injured vertebra, the preoperative height of the injured vertebra, the postoperative height of the injured vertebra, the vertebral body height loss rate of the injured vertebra, the distribution pattern of bone cement, the intervertebral disc leakage of bone cement, and the contact relationship between bone cement and endplate.

6. The construction method according to claim 1, characterized in that, Step S03 specifically includes the following steps: Step S31 merges the standardized image parameter set with the clinical data and uses LASSO regression for feature screening; the optimal λ value is determined by 10-fold cross-validation, and 11 non-zero coefficient variables are retained as candidate predictors. The candidate predictors include bone mineral density, number of previous vertebral fractures, history of diabetes, multi-segment vertebral involvement, vertebral fissure sign, surgical procedure, regular aerobic exercise, bone cement intervertebral disc leakage, use of lumbar braces, history of glucocorticoid use, and contact relationship between bone cement and endplate. Step S32: Input the candidate predictors into a multivariate logistic regression model, using whether secondary OVCF occurs within 2 years as the dependent variable, and select 6 independent predictors and their corresponding OR values: Bone mineral density: OR=0.42, p<0.01; Number of previously fractured vertebrae: OR=1.80, p<0.01; Intravertebral fissure sign: OR=2.59, p<0.05; Lumbar brace use: OR=0.12, p<0.01; History of glucocorticoid use: OR=6.90, p=0.01; Contact relationship between bone cement and endplate: OR=0.15, p<0.01; Step S33: Based on the six independent predictors, construct a risk prediction model for secondary fractures after vertebral fracture surgery based on a nomogram.

7. A risk prediction model for secondary fractures after vertebral fracture surgery, characterized in that, It is constructed using the construction method described in any one of claims 1 to 6.

8. A risk prediction system for secondary fractures after vertebral fracture surgery, characterized in that, It includes a data input module, a risk prediction module, and a graphical display module; The data input module is used to input the independent predictive factors input by the user into the risk prediction module; The risk prediction module is used to execute the risk prediction model as described in claim 7 and predict the risk of secondary fractures after vertebral fracture surgery based on the independent predictor factors. The graphical display module is used to graphically display the results of risk prediction.

9. The risk prediction system according to claim 8, characterized in that, The independent predictors include bone mineral density, number of previous vertebral fractures, vertebral cleft sign, use of lumbar braces, history of glucocorticoid use, and the contact relationship between bone cement and endplate.

Images