A method for predicting the risk of interbody cage subsidence after lumbar interbody fusion

Abstract

The application discloses a risk prediction method for subsidence of an intervertebral fusion cage after lumbar interbody fusion, and comprises the following steps: acquiring preoperative image data and postoperative image data within one week after operation of a target patient, wherein the preoperative image data and the postoperative image data comprise multiple image features reflecting local anatomical morphology, sagittal balance and bone conditions of the spine; inputting the multiple image features into a pre-trained subsidence risk prediction model to obtain a risk probability of the target patient suffering from the subsidence of the intervertebral fusion cage; wherein the subsidence risk prediction model is obtained by screening a core prediction feature set through a "funnel type" feature screening process of fusing preoperative native anatomical parameters and postoperative early image parameters, and by training the core prediction feature set by using a machine learning algorithm based on a historical patient data set containing subsidence samples and non-subsidence samples; and the application reduces the occurrence of postoperative complications.

Description

Technical Field

[0001] This invention relates to the fields of medical data science and orthopedic clinical decision support technology, specifically to a method for predicting the risk of interbody fusion cage subsidence after lower lumbar interbody fusion surgery. Background Technology

[0002] Lumbar fusion surgery is an important treatment for degenerative lumbar spine diseases such as lumbar disc herniation, lumbar spondylolisthesis, and spinal stenosis. With an aging population and the widespread availability of surgical techniques, the number of lumbar fusion surgeries continues to increase. However, postoperative complications—especially cage subsidence (a loss of intervertebral height exceeding 2 mm)—seriously affect the surgical outcome, with an incidence rate as high as 10%-30%. Subsidence not only leads to recurrence of neurological symptoms in 28.7% of patients but is also closely related to serious complications such as pseudoarthrosis and adjacent segment degeneration, sometimes requiring revision surgery.

[0003] Limitations of existing technologies: Existing models primarily focus on the patient's overall bone mass (such as DEXA bone mass or its alternatives CTHU and VBQ values), neglecting the synergistic effect of sagittal alignment and the local biomechanical environment. Endplate length, lumbosacral angle, and other local anatomical factors indirectly reflect the sagittal alignment and stress distribution of the lower lumbar vertebrae. Furthermore, surgical endplate damage or excessive stretching of the intervertebral space can increase endplate stress, a significant factor inducing cage subsidence. These indicators are often overlooked in existing predictive tools. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method for predicting the risk of lumbar interbody fusion cage subsidence after lumbar interbody fusion surgery. It integrates preoperative native anatomical imaging parameters and DR imaging manifestations within one week after surgery, and uses a machine learning-based "funnel-shaped" feature screening algorithm to construct a highly robust and clinically interpretable lumbar interbody fusion cage subsidence (CS) risk prediction model, so as to achieve early and precise intervention for postoperative subsidence in patients.

[0005] Technical solution: The present invention provides a method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery, comprising the following steps:

[0006] (1) Obtain preoperative imaging data and postoperative imaging data within one week after surgery for the target patients, including multi-dimensional imaging features reflecting the local anatomical morphology of the spine, sagittal balance and bone condition.

[0007] (2) Input the multivariate image features into the pre-trained subsidence risk prediction model to obtain the risk probability of interbody fusion cage subsidence in the target patient; wherein, the subsidence risk prediction model is based on a historical patient dataset containing subsidence samples and non-subsidence samples, and the core prediction feature set is selected through a "funnel-shaped" feature screening process that integrates preoperative original anatomical parameters and postoperative early image parameters, and is trained using machine learning algorithms.

[0008] Furthermore, in step (2), the "funnel-shaped" feature screening process is as follows: the initial multivariate image features are screened using a random forest-based algorithm to identify a subset of features that are related to the settlement outcome; the feature subset is further screened using a stability selection method based on regularized regression, and the core features selected by high frequency are retained through multiple iterations; the core features selected are verified for multicollinearity and independence to form the final core prediction feature set used for model training.

[0009] Furthermore, the core predictive feature set includes a combination of at least several features, including postoperative intervertebral height, lumbosacral angle, subcutaneous fat thickness, length of the superior endplate of the vertebra below the responsible segment, concavity angle of the superior endplate of the vertebra below the responsible segment, and bone mineral density related values ​​of a specific lumbar vertebra.

[0010] Furthermore, in step (2), the settlement risk prediction model is a logistic regression model, and its judgment threshold has been optimized. The optimization is achieved by iteratively finding the maximum value of the F1 score within a preset probability range to determine the optimal probability cut-off point for distinguishing settlement and non-settlement risks.

[0011] Furthermore, the method also includes: constructing a visual clinical assessment tool based on the core predictive feature set and its regression coefficients in the settlement risk prediction model; the visual clinical assessment tool is used to receive the actual feature values ​​of the target patient, calculate the corresponding scores by accumulating them, and map the total score to an intuitive settlement risk probability value for clinical decision-making reference.

[0012] Furthermore, the visualization tool for clinical assessment is the nomogram, which assigns a scoring interval to each core predictive feature based on its regression coefficient, and achieves a visual assessment of risk through the mapping relationship between the cumulative total score and the risk probability.

[0013] The present invention discloses a risk prediction system for interbody fusion cage subsidence after lumbar interbody fusion surgery, comprising:

[0014] Acquisition module: used to acquire preoperative imaging data and postoperative imaging data within one week after surgery for the target patient, including multi-dimensional imaging features reflecting local spinal anatomy, sagittal balance and bone condition;

[0015] The subsidence risk prediction module is used to input multivariate imaging features into a pre-trained subsidence risk prediction model to obtain the risk probability of interbody fusion cage subsidence in the target patient. The subsidence risk prediction model is based on a historical patient dataset containing both subsidence and non-subsidence samples. It uses a funnel-shaped feature selection process that integrates preoperative native anatomical parameters and early postoperative imaging parameters to select the core prediction feature set, and then trains the model using machine learning algorithms.

[0016] The computer device of the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the program, it implements any of the methods described herein.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described herein.

[0018] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention only requires routine preoperative imaging examinations and routine postoperative DR images, without the need for additional feature measurements, making it economical and practical; the optimized model of this invention achieves strong generalization ability while maintaining low complexity. The AUC on the training set is 0.767, and the AUC on the test set is 0.762 (0.609-0.900). This shows that the model has high stability and reliability in sedimentation prediction across different patient groups. A visual nomogram tool has been developed, enabling clinicians to easily calculate the probability of postoperative cage sedimentation through simple and efficient cumulative calculations, without limitations on equipment or location. This helps doctors select surgical plans preoperatively or adjust postoperative prophylactic medications and rehabilitation measures, ultimately reducing the occurrence of postoperative complications. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention;

[0020] Figure 2 This is the SFTT measurement diagram of the present invention;

[0021] Figure 3 This is a graph showing the CTHU values ​​of each vertebra in this invention;

[0022] Figure 4 This is a schematic diagram of the preoperative DR height measurement at various angles according to the present invention;

[0023] Figure 5 This is a schematic diagram of the postoperative DR height measurement at various angles according to the present invention;

[0024] Figure 6 This is a Boruta feature screening importance distribution diagram of the present invention;

[0025] Figure 7 This is a heatmap showing the correlation between the feature independence and the outcome of the present invention;

[0026] Figure 8 This is the ROC curve of the multi-model comparison training set of the present invention;

[0027] Figure 9 The ROC curve of the multi-model comparison test set of this invention

[0028] Figure 10 This is the receiver operating characteristic (ROC) curve of the present invention.

[0029] Figure 11 This is the calibration curve of the present invention.

[0030] Figure 12 This is the decision curve DCA of the present invention;

[0031] Figure 13 This is a global contribution analysis of the present invention;

[0032] Figure 14 This is the Summary Plot of the present invention;

[0033] Figure 15 This is a dependency graph of the continuous variables in this invention;

[0034] Figure 16 This is the Decision Plot of the present invention;

[0035] Figure 17 This is a heatmap of model feature contribution based on SHAP values ​​according to the present invention;

[0036] Figure 18 This is the nomogram of the present invention.

[0037] Figure 19 This is the SHAP Force Plot of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the risk of interbody fusion cage subsidence after lower lumbar interbody fusion surgery, including the following steps:

[0040] Step 1: Data Collection and Image Processing. Study Subjects: 171 patients (205 segments) who underwent lumbar interbody fusion, vertebral fusion, or internal fixation of interbody fusion in the spine at the Department of Orthopedics, Subei People's Hospital from January 2020 to June 2021. The mean intervertebral height of the surgical segment measured on DR at 3 months postoperatively and at the last follow-up was compared with the mean intervertebral height within one week postoperatively. A decrease in intervertebral height exceeding 2 mm was defined as subsidence.

[0041] Inclusion criteria: age ≥ 45 years; patients with indications for lumbar fusion surgery such as lumbar disc herniation, lumbar spondylolisthesis, spinal stenosis and discogenic low back pain; patients whose MRI and CT scans are related to their clinical symptoms; good compliance with regular follow-up for at least 3 months after surgery; complete medical records and complete preoperative general information.

[0042] Exclusion criteria: history of spinal tumors, infections, etc.; severe scoliosis and coronal / sagittal imbalance that makes surgery unsuitable; incomplete medical records, failure to have regular check-ups, or failure to complete follow-up.

[0043] Relevant data:

[0044] It mainly includes three aspects: patient baseline information, surgical information, and imaging examination information and measured imaging indicators.

[0045] Baseline information: gender, age, height, weight, BMI, duration of clinical symptoms / month, hypertension, diabetes, heart disease, history of lumbar spine surgery, smoking history.

[0046] Surgical data: Surgical name, vertebral location, number of surgical segments, bone guidance, fusion device brand, surgical time / minute, intraoperative blood loss.

[0047] Imaging data: Preoperative: Subcutaneous fat thickness SFTT (as shown in the image) Figure 2 The SFTT measurement diagram, the Pfirrmann classification of intervertebral disc degeneration, and the HU values ​​of each vertebral body from L1 to L5 on CT scans are shown below. Figure 3 The CTHU values ​​of each vertebral body are shown in the diagram, along with the following measurements: scoliosis, lumbar lordosis angle (LL), lumbosacral angle (SS), lordosis angle (LSA) of the responsible segment, intervertebral height (DH) of the responsible segment, concavity angle of the inferior endplate of the superior vertebral body of the responsible segment (IECA), and concavity angle of the superior endplate of the inferior vertebral body of the responsible segment (SECA). Figure 4 The preoperative DR height measurement diagram is shown below; DR within one week postoperatively: lumbar lordosis angle LL-0, lordosis angle LSA-0 of the responsible segment, intervertebral height DH-0, anteroposterior ratio of CAGE, CAGE inclination angle, and length of superior endplate of the vertebral body below the responsible segment SEL are shown below. Figure 5 The postoperative DR height measurements at various angles are shown in the diagram. The difference and ratio of the postoperative and preoperative intervertebral heights were calculated. During follow-up DR examinations, the intervertebral height of the surgical segment was measured and analyzed using the Subei People's Hospital PACS system and 3D Slice 5.6.2.

[0048] Step 2: Dataset Splitting and Evaluation: There are 205 samples in total, with 45 positive samples (21.95%). The training and test sets are split into training and test sets using a stratified splitting method with a ratio of 7:3.

[0049] Training set: 143 samples (70%), with a positive rate of 21.68%;

[0050] Test set: Sample size 62 (30%), with a positive rate of 22.58%;

[0051] Categorical variables are expressed as frequencies and percentages (n, %), and comparisons between groups are performed using the chi-square test or Fisher's exact test. Continuous variables are first tested for normality using the Shapiro-Wilk test. Normally distributed continuous variables are expressed as mean ± standard deviation (Mean ± SD), and comparisons between two groups are performed using Student's t-test; variables that are not normally distributed are expressed as median and interquartile range [Median (IQR)], and comparisons between groups are performed using the Mann-Whitney U test.

[0052] As shown in Tables 1 and 2, the differences between the training and test sets in demographic characteristics (such as age, sex, and BMI), surgery-related indicators (such as operation time and blood loss), and imaging parameters (such as mean DH and CTHU values) were mostly not statistically significant (P>0.05). This indicates that the basic data of the training and test sets are evenly distributed and have good comparability.

[0053] Table 1. Demographic characteristics and surgery-related indicators for the training and test sets.

[0054]

[0055] Table 2. Information on imaging parameters between the training and test sets.

[0056] Step 3: Data preprocessing and 'funnel' feature selection:

[0057] Data preprocessing: Z-score standardization is performed on continuous variables to eliminate the influence of different units on the model's penalty term. Categorical variables are converted into dummy variables through one-hot encoding. All these steps are performed on the training set.

[0058] 'Funnel-style Feature Screening': First, a full feature screening is performed using the Boruta algorithm. This algorithm, based on the Random Forest algorithm, identifies statistically significant variables through multiple iterations of the Random Forest by constructing "shadow features" as significance benchmarks. Unlike traditional screening methods that only retain the top few features, the Boruta algorithm can uncover factors that are potentially correlated with the outcome variable. The preprocessed training set data, totaling 45 features, was analyzed. After the initial screening using the Boruta algorithm, 10 features were confirmed, and 35 features were rejected. See details... Figure 6 The Boruta feature selection importance distribution plot; refined feature selection based on Lasso stability selection. This study uses a stability selection method under repeated stratified sampling, iterating 50 times, and utilizing the L1 regularization property of Lasso regression to automatically select the most representative indicators from the candidate features. Only features with a selection frequency of 90% or higher are ultimately retained. This method overcomes the sensitivity of a single Lasso selection to the sample distribution. High-frequency iterations can select robust and representative core features. After Lasso stability selection, the 10 Confirmed features obtained from the initial screening were reduced to 6 features. See Table 3 for feature selection transparency report; multicollinearity and independence verification. The variance inflation factor (VIF) was calculated for the 6 final selected features, and features with VIF > 5 were planned to be deleted. The results show that the VIF of all features is less than 1.5, indicating that there is no serious multicollinearity interference among the selected features. See Table 4 for VIF. A plot of the selected features and outcome variables was also created. Figure 7 The feature independence and outcome correlation matrix heatmap visually demonstrates the independence among the selected feature indicators and the contribution of features to outcome prediction.

[0059] Table 3 Feature Filtering Transparency Report

[0060] Table 4 VIF

[0061] Step 4: Predictive Model Selection, Optimization, and Validation: Multiple Algorithm Selection and Initial Optimization: To more comprehensively capture the complex relationship between candidate features and sedimentation outcomes, this approach constructs a comprehensive model selection pool that includes linear baseline models (logistic regression, support vector machine (SVM), ensemble learning models (random forest, XGBoost, LightGBM, etc.), and probabilistic inference models (Gaussian Naive Bayes (GaussianNB)). The cleaned and standardized data, combined with the features selected in the previous step, are input into the model selection pool. Hierarchical ten-fold cross-validation is used for initial screening of the above models. The TEST-AUC is calculated, and Lasso logistic regression wins the initial screening.

[0062] For the champion model in the initial screening, this approach abandons blind automated search and instead adopts a two-layer parameter tuning mechanism based on AUC-Brier synergy, which is more in line with clinical prediction requirements. High-density iterative exploration is conducted on the regularization strength C. While ensuring that the model's discrimination AUC is above 98% of its maximum value, the parameter point with the lowest prediction deviation (BrierScore) is prioritized; in this approach, C is selected as 0.1.

[0063] Subsequently, to further explore the model's predictive performance on the independent test set, a non-parametric bootstrap method was introduced. This involves performing 1000 identically repeated samplings with replacement on the test set, calculating the predictive power of each subsample, and finally determining the AUC performance within the 95% confidence interval, further demonstrating the model's stability. See details... Figure 8 , Figure 9 AUC curves for each model on the training and test sets.

[0064] Screening for clinical decision-making thresholds:

[0065] Traditional machine learning models default to a decision threshold of 0.5 for sample outcomes, but in clinical practice, the incidence and probability of complications vary across diseases. Therefore, this study optimized the F1 score (a balancing metric between precision and recall) through high-density iterative exploration within the range of [0.05-0.95], finding the optimal F1 value. The optimal probability cutoff point for this study was calculated to be P=0.28. The TEST F1 score improved from 0.4889 to 0.5714. This optimization not only improved the model's sensitivity to predicting early or subtle sedimentation but also significantly reduced the risk of missed diagnoses.

[0066] Multi-dimensional performance verification of the optimal model:

[0067] The ROC curve, calibration curve, and decision curve of the model were plotted to demonstrate the discrimination, calibration, and clinical benefits of the optimal model.

[0068] Discrimination test, such as Figure 10 The receiver operating characteristic (ROC) curve is shown below.

[0069] Training set AUC: 0.767; Independent test set AUC: 0.762 (95% confidence interval: 0.609–0.900).

[0070] The model demonstrates robustness across different patients and does not exhibit overfitting or underfitting.

[0071] Calibration and reliability verification, such as Figure 11 As shown in the calibration curve,

[0072] The calibration curves, which also employed Brier Score and spline smoothing, show that the observed curves closely approximate the ideal diagonal. The Brier Score is 0.1386 on the training set and 0.1418 on the test set.

[0073] This indicates that the model's predictive efficacy matches the actual sedimentation occurrence during follow-up, demonstrating high clinical value.

[0074] Clinical benefit validation, such as Figure 12 The decision curve DCA is shown.

[0075] Compared to the two extreme cases of full intervention and no intervention, which treat all patients, the net benefit obtained by this model is higher than the two cases within a wider threshold range.

[0076] Step 5: Development of Feature SHAP Analysis and Visualization Nodal Chart Tools:

[0077] The SHAP game theory method is used to analyze the trained and optimized lasso-LR;

[0078] like Figure 13 Global contribution analysis was performed, and the mean absolute value of the SHAP of the test set samples was calculated to show the contribution of each feature to the settlement outcome.

[0079] like Figure 14 Summary plots, combined with scatter plot distributions, revealed the positive and negative relationships between feature intensity and outcome risk. DH-0 and SFTT were positively correlated with outcomes, while inferior vertebral superior endplate length, SS, inferior vertebral superior endplate concavity angle, and HUL3 were negatively correlated with outcomes.

[0080] like Figure 15Dependency graph analysis of each continuous variable is used to extract the independent influence curves of a single feature on the settlement probability.

[0081] like Figure 16 The decision plot backtracks the decision path, showing the decision trajectory from the baseline probability to the final predicted probability, demonstrating the process of multi-feature synergy accumulation until settlement prediction.

[0082] Figure 17 This is a heatmap of model feature contributions based on SHAP values. The horizontal axis represents test set sample instances, the vertical axis selects features in order of importance, and the color represents the degree of influence of that feature on the result trend. The f(x) curve above reflects the output trend after feature aggregation. This figure shows that the model can effectively identify key features and their influence on the prediction results, overcome the black box limitation of computer models, and demonstrate the transparent process of the model's logical decision-making.

[0083] Construction of clinical visualization nomograms:

[0084] To transform complex computer mathematical models into visual, clinically applicable tools, this study used the R programming language to construct nomograms. For example... Figure 18 nomogram.

[0085] By using the logistic regression transformation function, the linear predictions output by the machine learning model are mapped to a settlement risk percentage ranging from 0 to 1.

[0086] The upper part of the nomogram divides the regression coefficients of selected features such as DH-0, SS, and SEL into scores from 0 to 100. Doctors calculate the total score based on the patient's actual indicators, which is then mapped to the substitution risk probability axis at the bottom. This allows for the prediction of subsidence risk and further guides clinical practice.

[0087] To verify the interpretability and demonstrate the practicality of this model, I selected a critical sample in the test set that was near the optimal threshold for individual analysis. The predicted probability of this sample, after rounding, was 0.28, which was exactly at the optimal threshold selected after model tuning.

[0088] like Figure 19 As shown in the SHAP Force Plot, the model quantifies the antagonistic process of each feature in predicting the risk of this case through SHAP contribution values. The values ​​of each feature are standardized; DH-0 and SECA are positively correlated factors, increasing sedimentation risk; while SFTT, SS, etc., are negatively correlated with the outcome, offsetting the risk value. The resulting risk probability explains the synergistic coupling of multiple features.

[0089] like Figure 18 To make nomograms more convenient for clinical use and not limited by location or equipment, this study also added a more intuitive clinical assessment tool by converting the complex computer logic mentioned above. Using the above sample as an example, the original values ​​for each feature are: SFTT 7.5, SS 39.9, SECA 158.1, DH-0 16.355, SEL 44.28, and HUL3 106.11. The corresponding hazard scores above the image for each feature are: SFTT around 7, SS around 31, SECA around 25, DH-0 around 70, SEL around 33, and HUL3 around 10, adding up to a total score of 176. Refer to the Total Points section in the lower half of the table.

[0090] A score of 176 corresponds to a Subsidence Risk of 0.28, which is highly consistent with the SHAP analysis. This example demonstrates that the model not only addresses the lack of interpretability in machine learning models through SHAP weight analysis and nomogram-based score calculation, but also assists doctors in identifying high-risk subsidence patients, thereby enabling them to select appropriate surgical options or provide patients with suitable early medication recommendations, rehabilitation guidance, and health education.

Claims

1. A method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery, characterized in that, Includes the following steps: (1) Obtain preoperative imaging data and postoperative imaging data within one week after surgery for the target patients, including multi-dimensional imaging features reflecting the local anatomical morphology of the spine, sagittal balance and bone condition. (2) Input the multivariate image features into the pre-trained subsidence risk prediction model to obtain the risk probability of interbody fusion cage subsidence in the target patient; wherein, the subsidence risk prediction model is based on a historical patient dataset containing subsidence samples and non-subsidence samples, and the core prediction feature set is selected through a "funnel-shaped" feature screening process that integrates preoperative original anatomical parameters and postoperative early image parameters, and is trained using machine learning algorithms.

2. The method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery according to claim 1, characterized in that, In step (2), the "funnel-shaped" feature screening process is as follows: the initial feature screening of the initial multivariate image features is carried out using a random forest-based algorithm to identify the feature subsets that are related to the settlement outcome. For the feature subset, a stability selection method based on regularized regression is used for fine screening, and the core features that are frequently selected are retained through multiple iterations; The selected core features are subjected to multicollinearity and independence verification to form the final core prediction feature set used for model training.

3. The method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery according to claim 2, characterized in that, The core predictive feature set includes a combination of at least several features, such as postoperative intervertebral height, lumbosacral angle, subcutaneous fat thickness, length of the superior endplate of the vertebral body below the responsible segment, concavity angle of the superior endplate of the vertebral body below the responsible segment, and bone mineral density related values ​​of a specific lumbar vertebral body.

4. The method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery according to claim 1, characterized in that, In step (2), the settlement risk prediction model is a logistic regression model. Its judgment threshold has been optimized. The optimization is achieved by iteratively finding the maximum value of the F1 score within a preset probability range to determine the optimal probability cut-off point for distinguishing settlement and non-settlement risks.

5. The method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery according to claim 1, characterized in that, The method further includes: constructing a visual clinical assessment tool based on the core predictive feature set and its regression coefficients in the settlement risk prediction model; the visual clinical assessment tool is used to receive the actual feature values ​​of the target patient, calculate the corresponding scores by summing them, and map the total score to an intuitive settlement risk probability value for clinical decision-making reference.

6. The method for predicting the risk of interbody fusion cage subsidence after lumbar interbody fusion surgery according to claim 5, characterized in that, The visualization tool for clinical assessment is a nomogram. The nomogram assigns a scoring interval to each core predictive feature based on its regression coefficient, and achieves a visual assessment of risk through the mapping relationship between the cumulative total score and the risk probability.

7. A risk prediction system for interbody fusion cage subsidence after lumbar interbody fusion surgery, characterized in that, include: Acquisition module: used to acquire preoperative imaging data and postoperative imaging data within one week after surgery for the target patient, including multi-dimensional imaging features reflecting local spinal anatomy, sagittal balance and bone condition; The subsidence risk prediction module is used to input multivariate imaging features into a pre-trained subsidence risk prediction model to obtain the risk probability of interbody fusion cage subsidence in the target patient. The subsidence risk prediction model is based on a historical patient dataset containing both subsidence and non-subsidence samples. It uses a "funnel-shaped" feature selection process that integrates preoperative native anatomical parameters and early postoperative imaging parameters to select the core prediction feature set, and then trains the model using machine learning algorithms.

8. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, implements the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Images