A post-knee surgery distal deep vein thrombosis prediction system

Abstract

The application belongs to the technical field of medical decision support, and discloses a knee joint postoperative distal deep vein thrombosis prediction system; by acquiring multi-dimensional parameters such as coagulation function, metabolic function, inflammation ratio index and intra-articular injection history, the sodium hyaluronate injection history is converted into an articular lesion severity score, an inflammation-coagulation coupling characteristic vector and a metabolic disturbance correction model are innovatively constructed, and the synergistic effect among multiple systems is captured. A penalty regression model with cross-validation is used for sparse prediction factor screening, redundant variables are screened out through a regularization path shrinkage technology, and key prediction factors are reserved. Finally, an individualized thrombosis risk probability nomogram is constructed, and the prediction accuracy is verified through resampling calibration. The application realizes the multi-system integration of inflammation, coagulation, metabolism and joint degeneration factors, converts the complex statistical model into a clinically friendly tool, and optimizes the postoperative thrombosis prevention and control strategy.

Description

Technical Field

[0001] This invention relates to the field of medical decision support technology, and more specifically, to a system for predicting distal deep vein thrombosis after knee surgery. Background Technology

[0002] Currently, widely used clinical methods for predicting the risk of distal deep vein thrombosis (DVT) after knee surgery suffer from several drawbacks. Traditional assessment models rely excessively on simplified indicators or coarse clinical scoring systems, failing to comprehensively capture the complex pathological process of thrombosis involving coagulation cascades, inflammatory responses, metabolic stress, and local tissue damage. In clinical practice, physicians often face patients with complex medical histories and multiple underlying diseases, yet they can only rely on experience or simplified scoring to make risk assessments. This leads to low-risk patients receiving unnecessary preventative measures, while truly high-risk patients are not identified in a timely manner. Existing joint disease assessments rely excessively on costly, time-consuming, and subjectively interpretable imaging examinations, making it difficult to standardize and quantify the degree of joint degeneration and directly link it to thrombosis risk. This results in complex preoperative assessment procedures and inconsistent results. Data from routinely tested biomarkers in clinical laboratories, such as D-dimer and inflammatory markers, are not fully utilized. In particular, the non-linear distribution characteristics and synergistic effects of these indicators are generally overlooked, causing predictive models to fail to accurately reflect the actual thrombosis mechanism. Crucially, existing methods treat inflammation, coagulation function, metabolic disorders, and joint lesions as independent risk factors. However, clinicians frequently observe a significantly increased risk of thrombosis in patients with hyperglycemia and concurrent inflammation, indicating that the complex mechanisms by which multiple systems synergistically promote thrombosis are not effectively captured by the models. Traditional feature selection lacks rigorous statistical support, often relying on experience or simple correlation analysis, resulting in models containing a large amount of redundant information or omitting key predictive factors, thus reducing predictive stability and accuracy. Complex statistical predictions are difficult to translate into intuitive and practical clinical tools, and the consistency between preoperative assessment results and actual thrombosis incidence is insufficient, making it difficult for surgeons and anesthesiologists to develop individualized prevention strategies based on the predictions.

[0003] In view of this, the present invention proposes a system for predicting distal deep vein thrombosis after knee surgery to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a system for predicting distal deep vein thrombosis after knee surgery, comprising: The perioperative data acquisition module is used to collect patients' coagulation function indicators, metabolic function indicators, inflammation ratio indicators, anesthesia duration records, and intra-articular injection history records. Coagulation function indicators include D-dimer levels, metabolic function indicators include fasting blood glucose levels, and inflammation ratio indicators include the ratio of neutrophils to lymphocytes. The joint degeneration replacement marker quantification module is used to extract hyaluronic acid injection exposure characteristics based on intra-articular injection history records, and calculate joint degeneration-related thrombosis susceptibility scores based on the hyaluronic acid injection exposure characteristics and the preset joint lesion severity mapping rules. The inflammation-coagulation coupling assessment module is used to calculate the coagulation activation degree value based on the D-dimer level, calculate the inflammation amplification coefficient based on the neutrophil to lymphocyte ratio, and perform interactive feature calculations on the coagulation activation degree value and the inflammation amplification coefficient to generate an inflammation-coagulation coupling feature vector. The metabolic disturbance correction module is used to quantify the contribution of metabolic disorders to endothelial function based on fasting blood glucose values, and calculate the metabolic-inflammatory co-risk offset by combining the inflammation amplification coefficient in the inflammation-coagulation coupling feature vector to generate a metabolically corrected risk feature set. The sparse predictor screening module is used to input the joint degeneration-related thrombosis susceptibility score, inflammation-coagulation coupling feature vector, metabolically corrected risk feature set and anesthesia duration record as candidate features into a penalized regression model with cross-validation. The module uses regularization path contraction to screen independent predictors whose coefficients do not approach zero, and generates a set of screened predictors. The nomogram construction and calibration module is used to construct a multivariate logistic regression prediction model based on the screened predictor set, convert the regression coefficients of each predictor into a scoring scale, generate an individualized thrombosis risk probability nomogram, and verify the consistency between the predicted probability of the nomogram and the actual incidence rate through resampling calibration.

[0005] The technical effects and advantages of the present invention, a system for predicting distal deep vein thrombosis after knee surgery: This invention overcomes the limitations of traditional imaging assessment, providing an efficient and feasible alternative for preoperative risk assessment. Optimized biomarker processing strategies effectively eliminate predictive bias caused by distributional skew, fully unleashing the predictive potential of laboratory indicators. Interactive system modeling successfully captures the synergistic effects among the inflammation-coagulation-metabolic systems, enabling the predictive model to reflect, for the first time, the complex pathological process observed in clinical practice where multiple systems jointly promote thrombosis. A feature selection method based on rigorous statistical principles provides a highly stable and reliable set of predictive factors, eliminating the uncertainty of traditional experience-based selection. The nomogram scoring tool achieves a perfect transformation from complex statistical models to clinically practical tools, allowing accurate prediction results to be seamlessly integrated into daily medical decision-making processes. Personalized risk stratification and prevention strategy recommendations provide tailored prevention plans for each patient. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the knee joint postoperative distal deep vein thrombosis prediction system of the present invention; Figure 2 This is a diagram illustrating the binomial bias and regularization path of the present invention. Figure 3 This is a schematic diagram of the characteristic coefficient shrinkage path of the present invention; Figure 4 This is the system calibration curve for the present invention; Figure 5 This is a schematic diagram of the decision curve analysis of the prediction model of the present invention; Figure 6 This is a schematic diagram of the subject operating curve of the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] This application provides a system for predicting distal deep vein thrombosis after knee surgery. The system's execution entities include, but are not limited to, a clinical decision support platform, a postoperative risk warning system, a patient management center, and a multi-parameter clinical analysis platform, which can be considered general computing nodes in this application. The prediction system includes, but is not limited to, at least one cloud-based risk analysis engine, a personalized prediction model, and a multi-factor integrated predictor.

[0009] Please see Figure 1 In this embodiment of the invention, a system for predicting distal deep vein thrombosis after knee surgery includes: The perioperative data acquisition module is used to collect patients' coagulation function indicators, metabolic function indicators, inflammation ratio indicators, anesthesia duration records, and intra-articular injection history records. Coagulation function indicators include D-dimer levels, recording the dynamic balance between thrombus formation and degradation; metabolic function indicators include fasting blood glucose levels, reflecting the patient's glucose metabolism status and potential risk of endothelial function impairment; inflammation ratio indicators include the neutrophil-to-lymphocyte ratio, serving as a sensitive marker of systemic inflammation; anesthesia duration records provide indirect information on surgical complexity and patient immobilization time; and intra-articular injection history records contain key data on the severity of joint lesions and previous treatments. This multi-dimensional data is acquired in real-time through the electronic medical record system, laboratory information system, and dedicated collection forms, providing a comprehensive clinical parameter foundation for subsequent analysis and ensuring the integrity and accuracy of the predictive model.

[0010] This invention places great emphasis on the legality and compliance of data acquisition and processing. Perioperative data collection adheres to strict medical data protection standards, and all clinical data acquisition is conducted with the patient's full informed consent. The system design employs a multi-layered security architecture to ensure data security throughout its entire lifecycle. In the data processing stage, patient personal identification information is rigorously de-identified to ensure that only clinically necessary medical parameters are used in the analysis process. Furthermore, techniques such as hierarchical access control, access management, and data encryption are employed to prevent unauthorized access.

[0011] The Joint Degeneration Alternative Marker Quantification Module extracts hyaluronic acid injection exposure characteristics from intra-articular injection history records. Based on the mapping rules between hyaluronic acid injection exposure characteristics and pre-defined joint lesion severity, it calculates a joint degeneration-related thrombosis susceptibility score. This module first extracts the frequency, time, and dosage information of hyaluronic acid injections from patient medical records to construct injection exposure characteristics. Then, it queries a pre-defined mapping rule library to convert the injection exposure characteristics into a quantitative representation of joint lesion severity. Finally, it calculates a joint degeneration-related thrombosis susceptibility score, reflecting the potential contribution of chronic inflammation and microvascular disturbances to thrombosis. This innovative score avoids the complexity and subjectivity of traditional imaging assessments, providing an objective alternative marker based on clinical treatment history and offering a new dimension for thrombosis risk assessment.

[0012] The inflammation-coagulation coupling assessment module calculates coagulation activation values ​​based on D-dimer levels and an inflammation amplification coefficient based on the neutrophil-to-lymphocyte ratio. It then performs interactive feature calculations on the coagulation activation values ​​and the inflammation amplification coefficient to generate an inflammation-coagulation coupling feature vector. This module first standardizes the original indicators to eliminate differences in units and magnitudes; then it calculates the clinically significant derived values ​​for each indicator; finally, through interactive feature calculations, it generates a coupling feature vector reflecting the mutually reinforcing relationship between the inflammation and coagulation systems. This vector captures the key pathological loop of inflammation-mediated coagulation activation and inflammation amplification caused by the coagulation cascade, providing a system biology perspective for accurate prediction.

[0013] The metabolic disturbance correction module quantifies the contribution of metabolic disorders to endothelial function based on fasting blood glucose levels. It combines this with the inflammation amplification coefficient in the inflammation-coagulation coupling feature vector to calculate the metabolic-inflammatory synergistic risk offset, generating a metabolically corrected risk feature set. This module first assesses the degree to which blood glucose levels deviate from the normal range; then, it quantifies the vascular damage effect of metabolic disorders using a pre-defined endothelial injury dose-response curve; finally, it considers the synergistic effect of the inflammatory state to calculate the final risk offset, generating the corrected feature set. This correction process considers the nonlinear interaction between metabolic abnormalities and inflammatory states, more accurately reflecting the phenomenon observed in clinical practice where hyperglycemia and inflammation synergistically promote thrombosis.

[0014] The sparse predictor selection module uses joint degeneration-related thrombosis susceptibility scores, inflammation-coagulation coupling feature vectors, metabolically corrected risk feature sets, and anesthesia duration records as candidate features input to a penalized regression model with cross-validation. It then uses regularized path analysis to narrow down and select independent predictors whose coefficients do not approach zero, generating a set of selected predictors. This module employs penalized regression methods such as LASSO or Elastic Net to automatically filter a large number of candidate features, retaining the most predictive variable combinations. The selection process uses cross-validation to determine the optimal penalty strength, ensuring the model's generalization ability; simultaneously, regularized path analysis observes the trend of feature coefficients changing with penalty strength, identifying robust predictors. This data-driven feature selection method avoids the subjectivity of traditional empirical selection, improving the model's reliability and interpretability.

[0015] The nomogram construction and calibration module is used to build a multivariate logistic regression prediction model based on a screened set of predictors. It converts the regression coefficients of each predictor into scoring scales, generates a personalized nomogram of thrombosis risk probability, and verifies the consistency between the predicted probabilities and actual incidence rates through resampling calibration. This module first establishes the final prediction model; then, it transforms the statistical model into an intuitive scoring tool for easy clinical application; finally, it verifies the model's calibration performance using resampling techniques such as the expansion method, ensuring consistency between predicted probabilities and actual risks. The generated nomogram not only provides an overall risk assessment but also displays the relative contribution of each factor, supporting clinicians in developing targeted prevention strategies and achieving truly personalized medical decisions.

[0016] The modules are connected via wired and / or wireless means to enable data transmission between them.

[0017] In this embodiment of the invention, the detailed implementation steps for calculating the joint degeneration-related thrombosis susceptibility score based on the characteristics of sodium hyaluronate injection exposure and a preset mapping rule for the severity of joint lesions include: The presence or absence of sodium hyaluronate injections within a pre-defined preoperative time window is extracted from intra-articular injection history records and encoded as a binary exposure identifier. The time window is typically set at 12 months preoperatively, a range sufficient to capture the progression of chronic lesions without being excessively prolonged to the point of information dilution. The extraction process first sorts the patient's treatment records chronologically to identify intra-articular injections within the window period; then, it identifies the type of injected medication, determining whether it was sodium hyaluronate; finally, the presence or absence is encoded as a binary identifier (present = 1, absent = 0). While this binary encoding simplifies complex treatment information, it provides a good indication of the presence and severity of joint lesions, while avoiding the problem of excessively sparse data.

[0018] The system uses a binary exposure marker to map the severity of joint lesions according to a rule that maps sodium hyaluronate injection exposure as a surrogate indicator for advanced osteoarthritis or recurrent synovitis. The mapping rule is constructed based on large-scale clinical research data, reflecting the statistical association between sodium hyaluronate treatment and the severity of joint lesions. The rule considers two key pathological states: advanced osteoarthritis (representing cartilage wear and bone changes) and recurrent synovitis (representing a chronic inflammatory state of the joint), both of which are associated with an increased risk of thrombosis. The mapping process is implemented through a lookup table; when the binary marker is 1, the system automatically assigns the corresponding lesion type weight, providing a basis for subsequent scoring calculations. This surrogate marker method avoids reliance on complex imaging examinations, improving the system's applicability and efficiency.

[0019] A joint degeneration-related thrombosis susceptibility score is obtained by weighting and summing the chronic inflammation weight and microvascular disturbance weight corresponding to alternative indicators. The two weights quantify the contribution of different pathological mechanisms to thrombosis risk: the chronic inflammation weight reflects how persistent low-level inflammation promotes coagulation system activation; the microvascular disturbance weight reflects the impact of periarticular blood flow changes and endothelial dysfunction on thrombus formation. The weight values ​​are determined based on medical literature and expert consensus, typically with a chronic inflammation weight of 0.6 and a microvascular disturbance weight of 0.4, reflecting the dominant role of inflammatory factors in the pathological process. The weighted summation formula is: ; in, For joint degeneration-related thrombosis susceptibility score, Weighting for chronic inflammation. To replace the inflammatory component value in the indicators, Weighting for microvascular disorder. The vascular component value is used as a substitute indicator.

[0020] The final score range is usually 0-10. The higher the value, the greater the risk of joint degeneration-related thrombosis, providing a quantitative representation of local joint factors for subsequent integrated prediction.

[0021] In this embodiment of the invention, the detailed implementation steps for generating an inflammation-coagulation coupling feature vector by performing interactive feature calculations on the coagulation activation degree value and the inflammation amplification coefficient include: Logarithmic transformation of D-dimer levels was performed to eliminate the influence of right-skewed distribution, mapping the transformed values ​​to coagulation activation levels. As a fibrin degradation product, D-dimer typically exhibits a right-skewed distribution in clinical data; high abnormal values ​​elongate the distribution tail, affecting statistical analysis. Logarithmic transformation is a standard method for processing such biomarkers, making the distribution closer to normal and facilitating subsequent analysis. The transformation formula is: ; in, The value after logarithmic transformation. The original D-dimer value is incremented by 1 to ensure effective transformation even if the original value is zero. The transformed value is then linearly mapped to a coagulation activation level value within the range of 0-1, making it consistent with the magnitude of other features and facilitating subsequent calculations. This process preserves the relative relationship of D-dimer level changes while eliminating the excessive influence of extreme values, thus improving the stability and representativeness of the feature.

[0022] The difference between the neutrophil-to-lymphocyte ratio and a predefined healthy baseline ratio was standardized to obtain the inflammation amplification factor. The neutrophil-to-lymphocyte ratio (NLR) is a sensitive marker of systemic inflammation, reflecting the balance between innate and adaptive immune responses. The healthy baseline ratio is typically set at 1.65 (based on the median from large-scale population studies), and deviations from this baseline reflect the degree of abnormality in the inflammatory state. Standardization was performed using the Min-Max method, mapping the difference to the 0-1 interval, as shown in the formula: ; in, This represents the inflammation amplification factor. The ratio of neutrophils to lymphocytes in the patient. The ratio of healthy baseline, This is a clinically significant threshold (usually set to 4.0, representing a clear inflammatory state). This process converts continuous inflammatory indicators into clinically significant amplification factors; the larger the value, the more severe the inflammatory state and the more likely it is to amplify the risk of coagulation.

[0023] The product of the coagulation activation level and the inflammation amplification factor is calculated as the interaction feature component. The coagulation activation level, inflammation amplification factor, and interaction feature component are then combined to form an inflammation-coagulation coupling feature vector. The interaction feature component is a key calculation for capturing the mutually reinforcing relationship between the inflammation and coagulation systems, reflecting a synergistic effect that exceeds simple superposition. The product operation is a simple and effective method for simulating such nonlinear interaction effects, capable of amplifying the effects when high inflammation and high coagulation coexist. The final feature vector integrates information from three dimensions: coagulation state, inflammation state, and their interaction effect, comprehensively describing the patient's inflammation-coagulation coupling state. This ternary vector reflects the complex pathological process of thrombosis more comprehensively than a single indicator, providing input features from a systems biology perspective for the predictive model and improving the model's ability to identify high-risk states.

[0024] In this embodiment of the invention, the detailed implementation steps for calculating the metabolic-inflammatory co-risk offset based on the contribution of fasting blood glucose levels to the damage of metabolic disorders to endothelial function, combined with the inflammation amplification coefficient in the inflammation-coagulation coupling feature vector, include: Fasting blood glucose levels were compared to a preset normal glucose metabolism threshold, and the deviation exceeding the threshold was calculated as the degree of glucose metabolism abnormality. The normal glucose metabolism threshold was set at 5.6 mmol / L based on clinical guidelines; values ​​above this indicate at least a state of impaired fasting glucose. The deviation was calculated using a piecewise function, taking into account the different pathological significance of different blood glucose levels: linear increases were slow within the normal range, while increases accelerated beyond the threshold, reflecting the non-linear accelerating effect of elevated blood glucose on vascular damage. This approach aligns with clinical observations; even mild hyperglycemia significantly increases the risk of thrombosis, and this risk increases more pronounced with further increases in blood glucose. The formula for calculating the degree of glucose metabolism abnormality is: ; in, The degree of abnormal glucose metabolism. This refers to fasting blood glucose levels. This is a reference value (usually 4.0 mmol / L). This is within the normal threshold (5.6 mmol / L). and These are the weighting coefficients, and This reflects the accelerated effect of blood glucose elevation above the threshold.

[0025] The contribution of endothelial function impairment is obtained by querying a pre-defined dose-response curve based on the degree of glucose metabolism abnormality. The dose-response curve is a mapping function constructed based on clinical research data, describing how different degrees of glucose metabolism abnormality affect vascular endothelial function. The curve is typically S-shaped, reflecting the non-linear characteristics of the impairment effect: low-level abnormality produces a slight effect, moderate abnormality leads to a significant increase in impairment, and high-level abnormality causes impairment to reach a saturation level. This curve considers the adaptation mechanisms and impairment thresholds of vascular endothelial cells to metabolic stress, and is more consistent with the response patterns of biological systems. Curve lookup is achieved through pre-computed piecewise functions or lookup tables to ensure computational efficiency. The contribution of endothelial function impairment ranges from 0 to 1; a larger value indicates more severe endothelial injury and a higher tendency for thrombosis.

[0026] The contribution of endothelial dysfunction to inflammation is nonlinearly coupled with an inflammation amplification factor. When the inflammation amplification factor exceeds a preset inflammation activation threshold, an amplification factor is applied to the contribution of endothelial dysfunction, generating a metabolic-inflammatory co-risk offset. This step is innovative in the model, capturing the phenomenon observed in clinical practice where metabolic disorders and inflammatory states synergistically promote thrombosis. The nonlinear coupling operation considers the interaction effect of two pathological processes: when inflammation levels are low, metabolic factors act independently; however, when inflammation exceeds the activation threshold, the endothelial damage effect is significantly amplified, reflecting how inflammatory mediators enhance the damage of metabolic stress to blood vessels. The inflammation activation threshold is typically set to 0.5, representing a moderate level of inflammation. The formula for calculating the co-risk offset is: ; in, This represents the metabolic-inflammatory co-risk offset. Contribution to endothelial dysfunction This represents the inflammation amplification factor. The inflammation activation threshold, This is an indicator function; it returns 1 when the condition is met and 0 otherwise. The amplification factor is usually set to 1.5-2.5, determined based on clinical validation.

[0027] This collaborative risk offset provides a core component of the metabolically corrected risk feature set, reflecting the comprehensive impact of multi-system interactions on thrombosis and improving the model's ability to capture complex pathological processes.

[0028] In this embodiment of the invention, the detailed implementation steps for narrowing down and screening independent predictors whose coefficients do not approach zero through regularization path narrowing include: In penalized regression models, the penalty strength is gradually increased along the increasing path of the regularization parameter, and the coefficient shrinkage trajectory of each candidate feature is recorded. Penalized regression models employ LASSO (L1 regularization) or Elastic Net (a combination of L1 and L2 regularization) methods, controlling model complexity by adding a penalty term to the objective function. Regularization parameter... The severity of the punishment is determined. The values ​​gradually increase from near 0, corresponding to the transition from a complex model with no penalty to a simple model with high penalty. The coefficient shrinkage trajectory records how each feature coefficient changes with... The changes reflect the robustness and importance of features. The coefficients of important features remain non-zero even under high penalties, while the coefficients of redundant or weakly correlated features rapidly shrink to zero. This process not only filters features but also reveals the relative importance ranking among them, providing insights into understanding the prediction mechanism.

[0029] The prediction error corresponding to each regularization parameter value is calculated through cross-validation. The regularization parameter that satisfies the rule of minimum error plus one standard error is selected as the optimal penalty value. Regularization parameter ( The selection process is as follows: Figure 2 The diagram shows the binomial bias and regularization path. It illustrates the binomial bias (cross-validation error) of different log-regularization parameters during the training of a penalized regression model (such as LASSO). The dashed line on the left (e.g., LogLambda≈-4.5) corresponds to the "minimum error" point. The '0' represents the model's optimal fit performance on the training set; the dashed line on the right (e.g., LogLambda≈-3.8) corresponds to the point of "minimum error plus one standard error". Under the premise that the model performance loss is acceptable (within one standard error), it provides the strongest regularization penalty, thus obtaining the simplest and most robust model in terms of generalization ability.

[0030] Cross-validation is an effective method for evaluating a model's generalization ability. A common approach is 10-fold cross-validation, which divides the data into 10 parts, alternating between using 9 parts for training and 1 part for testing. For each... The value is used to calculate the average prediction error of cross-validation, resulting in an error curve. The minimum error point corresponds to... value( This approach offers optimal predictive performance but may lead to overly complex models; while the "minimum error plus one standard error" rule selects... value( This approach significantly simplifies the model while maintaining near-optimal performance. This rule is widely accepted in statistics and is considered the gold standard for feature selection, balancing predictive performance and interpretability while avoiding the risk of overfitting.

[0031] The shrinkage path of the eigenvalues ​​is as follows Figure 3 The diagram illustrates the feature coefficient contraction path. Each curve represents the trajectory of the regression coefficient of a candidate predictor (such as "joint degeneration-related thrombosis susceptibility score," "inflammation-coagulation coupling feature," etc.) as LogLambda increases (regularization strength increases). With increased penalty, the coefficients of unimportant or redundant features are gradually compressed to zero (the curve contracts to the horizontal axis), thus achieving automatic feature selection. The optimal regularization parameter is selected... (correspond Figure 2 At the point indicated by the dashed line on the right, only a subset of features have coefficients that are not zero. These retained features constitute the final "set of pre-selected predictors".

[0032] Under the optimal penalty value, candidate features with absolute coefficient values ​​greater than zero are retained as independent predictors, while redundant variables whose coefficients have shrunk to zero are excluded.

[0033] By leveraging the automatic variable selection capabilities of penalized regression, the model in The coefficients of irrelevant or redundant features are precisely compressed to zero, while retaining only those features with true predictive value. This mechanism not only improves the model's generalization ability but also simplifies the model structure and enhances interpretability. The resulting set of independent predictors typically contains 3-7 key variables, far fewer than the initial candidate feature set, but retains the main predictive information. These predictors are statistically independent of each other, forming a robust predictive foundation, avoiding multicollinearity, and ensuring the model's stability and reliability. The generated predictor set is directly used for subsequent nomogram construction, transforming it into a clinically applicable scoring tool.

[0034] In this embodiment of the invention, the detailed implementation steps for verifying the consistency between the predicted probability and the actual occurrence rate of the nomogram through resampling calibration include: A predetermined number of resamplings with replacement are performed on the original dataset. A multivariate logistic regression prediction model is fitted to each resampled sample, and the predicted probabilities are calculated on the original sample. Resampling typically employs a bootstrap method, generating multiple dummy samples of the same size as the original dataset. The number of resampling iterations is usually set to 1000 to ensure statistical stability. For each dummy sample, a logistic regression model is constructed using selected predictors. This model is then applied to the original sample to calculate the predicted probability for each patient. This process simulates the model's performance differences across different samples, assessing its stability and uncertainty. The set of predicted probabilities forms a probability distribution, reflecting the confidence interval of the predictions and providing foundational data for subsequent calibration analyses. The advantage of the bootstrap method lies in its ability to fully utilize limited samples and obtain robust statistical inferences, making it particularly suitable for situations with limited sample sizes in medical research.

[0035] The predicted probabilities are grouped by interval, and the actual thrombosis incidence rate within each interval is calculated. A calibration curve is then plotted, and the calibration intercept and slope are calculated. The predicted probabilities are typically divided into 10 equally wide intervals (each grouped in 0.1 increments), or quantile grouping is used to ensure balanced sample sizes across groups. For each interval, the actual observed thrombosis incidence rate is calculated and compared to the interval's average predicted probability. Ideally, the predicted probability and the actual incidence rate should be consistent, forming a diagonal calibration curve. The calibration curve is fitted using logistic regression to obtain the calibration intercept and slope: the intercept reflects systematic bias (over- or under-prediction), with an ideal value of 0; the slope reflects the dispersion of the prediction (over- or under-dispersion), with an ideal value of 1. These two parameters are key indicators for evaluating the model's calibration performance and directly affect the reliability and applicability of the predictions in clinical decision-making.

[0036] In one specific embodiment, the calibration of the prediction model (i.e., the consistency between the predicted probability and the actual occurrence rate) is verified using a calibration curve, such as... Figure 4 The figure shows the system calibration curves. The "ideal curve" represents perfectly accurate predictions, the "apparent performance curve" represents the model's performance on the original data, and the "Bootstrap calibration curve" represents the performance after resampling calibration. As shown, the calibration curves are very close to the ideal curves. The calibration slope is close to 1 and the intercept is close to 0, while the Brier Score is 0.052 (the closer this value is to 0, the better). Overall, this indicates that the predicted probabilities of this invention have high accuracy, with no significant systematic overestimation or underestimation of risk, and high reliability for clinical decision-making.

[0037] The presence of systematic overestimation or underestimation bias in the nomogram is determined by whether the calibration intercept is close to zero and the calibration slope is close to unit. A calibration intercept significantly greater than zero indicates that the model systematically underestimates the risk, while a significantly less than zero indicates overestimation. A calibration slope significantly less than 1 indicates over-dispersion of predictions (overestimation of high-risk groups and underestimation of low-risk groups), while a significantly greater than 1 indicates under-dispersion of predictions. The judgment criterion is usually based on whether the 95% confidence interval contains the ideal value: if the 95% CI of the calibration intercept contains 0 and the 95% CI of the slope contains 1, the model is considered well calibrated. For models with calibration bias, the predicted probabilities can be adjusted using the calibration plot to ensure accuracy in clinical applications. This calibration process is a crucial step in model validation, ensuring that the predicted probabilities accurately reflect the actual risk and improving the reliability of clinical decision-making.

[0038] The discrimination index and net benefit curve of the nomogram were calculated, and the clinical decision-making advantage of the nomogram compared with the all-intervention strategy and the all-no-intervention strategy was verified within a preset threshold probability range.

[0039] The clinical net benefit of this predictive nomogram was quantified using decision curve analysis (DCA), and the results are as follows: Figure 5 The diagram shows a schematic of the decision curve analysis for the predictive model. It compares three strategies: "using a nomogram (i.e., this predictive nomogram)," "intervention for all patients" (total treatment), and "no intervention for any patient" (no treatment). Within a broad threshold probability range (approximately 5%-75%), the red nomogram curve consistently remains above the reference lines for the "total treatment" and "no treatment" strategies, indicating that decision-making based on this model yields a clear net clinical benefit. For example, at the threshold point where physicians consider the benefits and risks of intervention and non-intervention to be equivalent (typically 10%-20%), using this model to guide decision-making can avoid unnecessary preventative treatment for a large number of low-risk patients while ensuring timely intervention for high-risk patients, thereby maximizing clinical benefit.

[0040] Discriminant metrics primarily include AUC (Area Under the Curve), C-index (Conformity Index), and Brier score, which assess the model's ability to differentiate between high- and low-risk patients. The net benefit curve is quantified using decision curve analysis (DCA) at different risk thresholds to quantify the model's clinical value. DCA considers the different costs of erroneous decisions (the relative harm of false negatives and false positives) and calculates the net benefit predicted by the model at each risk threshold. The curve is typically compared to two reference strategies: full intervention (assuming all patients receive preventative measures) and no intervention (assuming no patients receive preventative measures). Ideally, the model should show a significant net benefit advantage within clinically relevant thresholds (typically 5%-30%). This analysis directly connects statistical performance and clinical value, helping physicians understand the model's effectiveness in practical decision-making and promoting evidence-based medicine practice.

[0041] The discriminative power of the predictive model in this invention (i.e., the ability to distinguish between patients with and without distal deep vein thrombosis) is evaluated using receiver operating characteristic (ROC) curves, such as... Figure 6 The figure shows a schematic diagram of the receiver operating characteristic (ROC) curve. The area under the blue ROC curve (AUC) is 0.889 (95% CI: 0.851 - 0.927), significantly higher than the random guessing reference line (gray area) shown diagonally. An AUC value greater than 0.8 indicates that the prediction system has excellent discriminative ability and can effectively identify patients at high risk of postoperative thrombosis.

[0042] This invention achieves accurate prediction of the risk of distal deep vein thrombosis after knee surgery through perioperative data acquisition, quantification of joint degeneration replacement markers, assessment of inflammation-coagulation coupling, correction of metabolic disturbances, screening of sparse predictive factors, and construction and calibration of nomograms. The multi-dimensional integrated prediction method of this invention can comprehensively assess the patient's inflammatory status, coagulation function, metabolic status, and degree of joint lesions, providing a systematic solution for precise stratification and individualized prevention of knee surgery patients.

[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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.

[0044] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0045] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A system for predicting distal deep vein thrombosis after knee surgery, characterized in that, include: The perioperative data acquisition module is used to collect patients' coagulation function indicators, metabolic function indicators, inflammation ratio indicators, anesthesia duration records, and intra-articular injection history records. The coagulation function indicators include D-dimer levels, the metabolic function indicators include fasting blood glucose levels, and the inflammation ratio indicators include the neutrophil to lymphocyte ratio. The joint degeneration replacement marker quantification module is used to extract hyaluronic acid injection exposure characteristics based on the intra-articular injection history record, and calculate the joint degeneration-related thrombosis susceptibility score based on the hyaluronic acid injection exposure characteristics and the preset joint lesion severity mapping rule. The inflammation-coagulation coupling assessment module is used to calculate the coagulation activation degree value based on the D-dimer level, calculate the inflammation amplification coefficient based on the neutrophil to lymphocyte ratio, and perform interactive feature operation on the coagulation activation degree value and the inflammation amplification coefficient to generate an inflammation-coagulation coupling feature vector. The metabolic disturbance correction module is used to quantify the contribution of metabolic disorders to endothelial function based on the fasting blood glucose value, and calculate the metabolic-inflammatory co-risk offset by combining the inflammation amplification coefficient in the inflammation-coagulation coupling feature vector to generate a metabolically corrected risk feature set. The sparse predictor screening module is used to input the joint degeneration-related thrombosis susceptibility score, the inflammation-coagulation coupling feature vector, the metabolically corrected risk feature set, and the anesthesia duration record as candidate features into a penalized regression model with cross-validation. The module uses regularization path contraction to screen independent predictors whose coefficients do not approach zero, and generates a set of screened predictors. The nomogram construction and calibration module is used to construct a multivariate logistic regression prediction model based on the set of screened predictors, convert the regression coefficients of each predictor into a scoring scale, generate an individualized thrombosis risk probability nomogram, and verify the consistency between the predicted probability and the actual incidence of the nomogram through resampling calibration.

2. The system according to claim 1, characterized in that, The calculation of a joint degeneration-related thrombosis susceptibility score based on the sodium hyaluronate injection exposure characteristics and a preset joint lesion severity mapping rule includes: Extract the presence or absence of sodium hyaluronate injection within the pre-preoperative preset time window from the intra-articular injection history record, and encode the presence or absence status as a binary exposure identifier. The severity mapping rule for joint lesions is queried based on the binary exposure identifier. The mapping rule maps sodium hyaluronate injection exposure as an alternative indicator for advanced osteoarthritis or recurrent synovitis. The joint degeneration-related thrombosis susceptibility score is obtained by weighting and summing the chronic inflammation weight and microvascular disorder weight corresponding to the alternative indicators.

3. The system according to claim 1, characterized in that, The step of performing interactive feature calculations on the coagulation activation degree value and the inflammation amplification coefficient to generate an inflammation-coagulation coupled feature vector includes: Logarithmic transformation is performed on the D-dimer level to eliminate the influence of right-skewed distribution, and the transformed value is mapped to the coagulation activation degree value. The difference between the neutrophil-to-lymphocyte ratio and the preset healthy baseline ratio is standardized to obtain the inflammation amplification factor. The product of the coagulation activation level value and the inflammation amplification coefficient is calculated as an interactive feature component, and the coagulation activation level value, the inflammation amplification coefficient, and the interactive feature component are combined into the inflammation-coagulation coupling feature vector.

4. The system according to claim 1, characterized in that, The contribution of metabolic disorders to endothelial function is quantified based on the fasting blood glucose value. Combined with the inflammation amplification coefficient in the inflammation-coagulation coupling feature vector, the metabolic-inflammatory co-risk shift is calculated, including: The fasting blood glucose value is compared with a preset normal glucose metabolism threshold, and the deviation exceeding the normal threshold is calculated as the degree of glucose metabolism abnormality. Based on the abnormality of glucose metabolism, the preset dose-response curve of endothelial injury is queried to obtain the corresponding contribution of endothelial function damage. The contribution of endothelial function impairment is nonlinearly coupled with the inflammation amplification coefficient. When the inflammation amplification coefficient exceeds a preset inflammation activation threshold, an amplification factor is applied to the contribution of endothelial function impairment to generate the metabolic-inflammatory synergistic risk offset.

5. The system according to claim 1, characterized in that, The method of narrowing down and screening independent predictors whose coefficients do not approach zero through regularization includes: In the penalized regression model, the penalty intensity is gradually increased along the increasing path of the regularization parameter, and the coefficient shrinkage trajectory of each candidate feature is recorded. The prediction error corresponding to each regularization parameter value is calculated by cross-validation, and the regularization parameter that satisfies the rule of minimum error plus one standard error is selected as the optimal penalty value. Under the optimal penalty value, candidate features with coefficients having an absolute value greater than zero are retained as independent predictors, while redundant variables whose coefficients have shrunk to zero are excluded.

6. The system according to claim 1, characterized in that, The step of verifying the consistency between the predicted probability and the actual occurrence rate of the nomogram through resampling calibration includes: Perform a preset number of resamplings with replacement on the original dataset, fit the multivariate logistic regression prediction model on each resampled sample, and calculate the prediction probability on the original sample. The predicted probabilities are grouped by intervals, the actual thrombosis incidence rate in each interval is calculated, a calibration curve is plotted, and the calibration intercept and calibration slope are calculated. Based on whether the calibration intercept is close to zero and whether the calibration slope is close to a unit value, determine whether the nomogram has a systematic overestimation or underestimation bias. The discrimination index and net benefit curve of the nomogram are calculated, and the clinical decision advantage of the nomogram relative to the all-intervention strategy and the all-no-intervention strategy is verified within a preset threshold probability range.

Images