A machine learning-based frequent gout flare prediction method, device, and medium

Abstract

This application discloses a machine learning-based method, device, and medium for predicting frequent gout attacks, relating to the field of machine learning technology. The method includes: extracting radiomics features from regions of interest corresponding to DECT images; performing feature selection on the radiomics features to obtain a feature subset; training a pre-defined multi-model group based on the feature subset through multi-model cross-validation until each model converges, thereby determining the radiomics model and the DECT volume model; constructing a clinical prediction model for gout patients based on factor regression analysis according to factors influencing gout attacks; and determining the probability of gout attacks in gout patients through multi-model nomograph validation based on the radiomics model, the DECT volume model, and the clinical prediction model. This application solves the technical problem in existing technologies where gout prediction cannot effectively identify and predict frequent gout attacks using CT images.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a method, device and medium for predicting frequent gout attacks based on machine learning. Background Technology

[0002] Gout is the most common type of arthritis in adults. Clinically, gout attacks are generally defined as ≥2 attacks per year, while attacks <2 per year are considered infrequent. Controlling gout attacks mainly relies on lowering serum uric acid levels (ULT). Although ULT lowers serum uric acid concentration, it may trigger acute gout attacks, so stratified treatment is necessary based on the frequency of gout attacks.

[0003] Gouty tophi are associated with gout attacks. Since many gout patients experience side effects from anti-inflammatory drugs, there is a need to predict the frequency of gout attacks. Current techniques, such as dual-energy CT (DECT) analysis of nodule burden, can predict attack risk. However, DECT reconstructed images focus on quantifying deposition volume and do not fully utilize the original images, making it impossible to directly determine whether a patient experiences frequent attacks. Radiomics analysis also lacks an explanation of the biological mechanisms underlying the model's predictions, preventing current techniques from predicting the prognosis of gout patients after ulcerative colitis (ULT). Therefore, there is an urgent need for a multi-machine learning model based on DECT and clinical characteristics to predict the frequency of gout attacks after ULT in gout patients. Summary of the Invention

[0004] This application provides a machine learning-based method, device, and medium for predicting frequent gout attacks, which solves the technical problem in the prior art that gout prediction cannot effectively identify and predict frequent gout attacks through CT images.

[0005] In a first aspect, embodiments of this application provide a machine learning-based method for predicting frequent gout attacks. The method includes: acquiring DECT images of the corresponding index joints of gout patients and extracting radiomics features of the corresponding regions of interest in the DECT images; performing feature selection on the radiomics features using LASSO-regularized logistic regression to obtain a feature subset corresponding to the radiomics features; training a pre-defined multi-model group until each model converges through multi-model cross-validation based on the feature subset to determine the radiomics model and the DECT volume model; constructing a clinical prediction model for gout patients through factor regression analysis based on pre-defined influencing factors of gout attacks; and determining the probability of gout attacks in gout patients through multi-model nomograph validation based on the radiomics model, the DECT volume model, and the clinical prediction model.

[0006] In one implementation of this application, the radiomics features of the region of interest (ROI) corresponding to the DECT image are extracted, specifically including: acquiring the original cross-sectional image of the DECT image; extracting the MSU deposition region corresponding to the original cross-sectional image, and calibrating the soft tissue surrounding the MSU deposition region to obtain the ROI; wherein, the ROI includes: the MSU deposition region and the soft tissue surrounding the MSU deposition region; calculating the feature values ​​corresponding to each voxel in the ROI, and assigning the feature values ​​to the corresponding voxels to determine the radiomics features; wherein, the radiomics features include: first-order statistics of intensity distribution, texture features, and shape features of three-dimensional geometric attributes.

[0007] In one implementation of this application, feature selection is performed on radiomics features using LASSO-regularized logistic regression to obtain feature subsets corresponding to the radiomics features. Specifically, this includes: dividing the radiomics features into standardized subsets of radiomics features, and training a LASSO regression fitting model based on each subset of radiomics features until the model converges to obtain the model validation error; determining the optimal penalty parameter based on the model validation error, and inputting the optimal penalty parameter and the radiomics features into the converged LASSO regression fitting model to obtain the feature subsets.

[0008] In one implementation of this application, based on a feature subset, a pre-defined multi-model group is trained until each model converges through multi-model cross-validation to determine the radiomics model and the DECT volume model. Specifically, this includes: constructing each machine learning model based on the feature subset, optimizing the model parameters of each machine learning model through cross-validation, predicting the probability of frequent attacks in gout patients, and obtaining the radiomics model; wherein, each machine learning model includes: logistic regression model, support vector machine model, random forest model, multilayer perceptron model, k-nearest neighbor model, and extra-tree model; calculating the volume of tophi corresponding to the DECT image, and constructing the DECT volume model based on the tophi volume.

[0009] In one implementation of this application, a clinical prediction model for gout patients is constructed based on pre-defined influencing factors of gout attacks through factor regression analysis. Specifically, this includes: performing univariate logistic regression analysis on the influencing factors of gout attacks to obtain a first correlation between the influencing factors and the frequency of gout attacks; screening influencing factors based on the first correlation to determine the first associated factors for frequent gout attacks; performing multivariate logistic regression analysis on the influencing factors of gout attacks to obtain a second correlation between the influencing factors and the frequency of gout attacks; and determining the second associated factors for frequent gout attacks based on the second correlation, and constructing a clinical prediction model corresponding to the first associated factors and the second correlation.

[0010] In one implementation of this application, the probability of gout attacks in gout patients is determined through multi-model nomograph validation based on a radiomics model, a DECT volume model, and a clinical prediction model. Specifically, this includes: obtaining the radiomics score output by the radiomics model, and integrating the value of the radiomics score with the variables corresponding to the DECT volume model and the clinical prediction model to form a clinical-radiomics nomograph; and determining the probability of gout attacks corresponding to the risk axis in the clinical-radiomics nomograph by cumulative mapping of variable scores based on the clinical-radiomics nomograph.

[0011] In one implementation of this application, after training a pre-defined multi-model group until each model converges based on a feature subset and through multi-model cross-validation to determine the radiomics model and the DECT volume model, the method further includes: acquiring proteomics validation data of tophi tissue corresponding to gout patients, and determining the protein abundance of the proteomics validation data through volcano plot and heatmap analysis to determine the protein abundance differences among gout patients; grouping gout patients into risk groups using radiomics Rad-score to obtain high-risk and low-risk gout patients; and determining the molecular profile differences between high-risk and low-risk gout patients based on the protein abundance differences.

[0012] In one implementation of this application, after determining the molecular profile differences between high-risk and low-risk gout patients based on protein abundance differences, the method further includes: performing gene set enrichment analysis on high-risk and low-risk gout patients to obtain enrichment data for high-risk gout patients; wherein, the enrichment data of high-risk gout patients in complement-related pathways and ECM regulatory pathways are significantly enriched; based on the enrichment data, the predictive performance of the radiomics model is determined through predictive effect analysis.

[0013] Secondly, embodiments of this application also provide a machine learning-based gout frequent attack prediction device, characterized in that the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement a machine learning-based gout frequent attack prediction method.

[0014] Thirdly, embodiments of this application also provide a non-volatile computer storage medium for predicting frequent gout attacks based on machine learning, storing computer-executable instructions, characterized in that the computer-executable instructions, when executed, can implement a method for predicting frequent gout attacks based on machine learning.

[0015] This application provides a machine learning-based method, device, and medium for predicting frequent gout attacks. By analyzing the radiomics of corresponding joint indicators in gout patients, a multi-model prediction method for frequent gout attacks is constructed based on radiomics, DECT volume models, and clinical prediction models. The biological relationship between gout patients with frequent attacks and the radiomics model is verified through proteomics. This solves the technical problem in the prior art that gout prediction cannot effectively identify and predict frequent gout attacks through CT images. It removes the black box effect from the image-based radiomics model, realizes the prediction of frequent gout attacks through radiomics analysis, and improves the predictive ability of gout prediction models. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of a machine learning-based method for predicting frequent gout attacks is provided in this application embodiment; Figure 2 A schematic diagram of segmentation and 3D reconstruction of MSU deposition on DECT provided for an embodiment of this application; Figure 3 This application provides an example of an image omics feature selection map based on LASSO regularized logistic regression. Figure 4 A feature coefficient bar chart for radiomics provided in this application embodiment; Figure 5 ROC curves of various machine learning algorithms under radiomics features are provided in this embodiment of the application. Figure 6 A schematic diagram illustrating the ability of radiomics scoring to differentiate the frequency of gout attacks and risk stratification in an embodiment of this application; Figure 7 A KM curve diagram of high-risk and low-risk gout patients provided in this application embodiment; Figure 8 ROC curve of the working characteristics of an image omics model provided in this application embodiment; Figure 9 A model calibration curve provided for an embodiment of this application; Figure 10 A clinical-radiomics nomograph provided for embodiments of this application; Figure 11 A schematic diagram of a representative HE staining of a tophi provided in an embodiment of this application; Figure 12A proteomic landscape of tophi during an acute attack, provided as an embodiment of this application; Figure 13 A schematic diagram of GSEA analysis provided in an embodiment of this application; Figure 14 This is a schematic diagram of the internal structure of a machine learning-based gout attack prediction device provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] This application provides a machine learning-based method, device, and medium for predicting frequent gout attacks. By analyzing the radiomics of corresponding joint indicators in gout patients, a multi-model prediction method for frequent gout attacks is constructed based on radiomics, DECT volume models, and clinical prediction models. The biological relationship between gout patients with frequent attacks and the radiomics model is verified through proteomics. This solves the technical problem in the prior art that gout prediction cannot effectively identify and predict frequent gout attacks through CT images. It removes the black box effect from the image-based radiomics model, realizes the prediction of frequent gout attacks through radiomics analysis, and improves the predictive ability of gout prediction models.

[0019] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0020] Figure 1 A flowchart illustrating a machine learning-based method for predicting frequent gout attacks, provided as an embodiment of this application. Figure 1 As shown in the figure, the present application provides a machine learning-based method for predicting frequent gout attacks, which specifically includes the following steps: Step 101: Obtain DECT images of the corresponding index joints of gout patients and extract radiomics features of the corresponding regions of interest in the DECT images.

[0021] For example, frequent gout attacks (FGF) are defined as ≥2 attacks per year, while infrequent gout attacks (INGF) are defined as ≤1 attack per year. Uric acid-lowering therapy (ULT) is the most effective means of controlling gout attacks. Initiating ULT reduces serum uric acid (SU) concentration and mobilizes deposited monosodium urate (MSU) load, thereby exposing crystals to the innate immune system. This may induce gout attacks, increase the risk of cardiovascular events, and many gout patients experience side effects from anti-inflammatory drugs. Dual-energy CT (DECT) can distinguish between urate and calcium in the soft tissues surrounding bone. DECT-based radiomics focuses on MSU deposition and emphasizes the influence of abundant inflammatory cells and surrounding fibrovascular tissue in the coronary band on gout attacks. These features are unclear in analyses after ULT. This application improves the extraction of radiomics features corresponding to regions of interest in DECT images, providing a data foundation for the non-black-box construction of radiomics models.

[0022] Specifically, the radiomics features of the region of interest (ROI) corresponding to the DECT image are extracted, including: obtaining the original cross-sectional image of the index joint corresponding to the DECT image; extracting the MSU deposition region corresponding to the original cross-sectional image and calibrating the soft tissue surrounding the MSU deposition region to obtain the ROI; wherein the ROI includes: the MSU deposition region and the soft tissue surrounding the MSU deposition region; calculating the feature values ​​corresponding to each voxel in the ROI and assigning the feature values ​​to the corresponding voxels to determine the radiomics features; wherein the radiomics features include: first-order statistics of intensity distribution, texture features, and shape features of three-dimensional geometric attributes.

[0023] Figure 2 This is a schematic diagram of segmentation and 3D reconstruction of MSU deposition on DECT provided for an embodiment of this application.

[0024] In one embodiment, the primary clinical outcome is defined as two or more gout attacks occurring in the MTP1 joint within 12 months. A gout attack is defined as meeting the validated criteria for a gout attack, which is at least three of the following four criteria: the patient defines an attack based on their self-assessment of previous attacks; a visual analog scale (VAS, range: 0 [no pain] to 10 [most severe pain]) score of >3 for resting pain; swelling in at least one joint; and elevated local skin temperature and redness in at least one joint.

[0025] It is important to note that the distinction between frequent and infrequent gout attacks is crucial for differentiating the frequency of gout flare-ups. The updated EULAR gout management guidelines recommend different clinical treatments for patients with infrequent (INGF) and frequent (FGF) attacks. In the prior art, the risk of attacks can be reduced by gradually increasing the dosage of uric acid-lowering medications and using anti-inflammatory prophylaxis during the first 3 to 6 months of ULT (Ultra-Low Temperature Threat). However, many gout patients experience side effects from anti-inflammatory drugs (colchicine, NSAIDs, oral steroids). Therefore, there is a need to improve the ability to risk-stratify gout patients initiating ULT. This application does not provide any related treatment options; it only uses this example to explain why it is necessary to distinguish between frequent and infrequent gout attacks and to predict attacks in patients with frequent attacks.

[0026] First, DECT images of patients with gout after ULT were acquired. The examination site was the MTP1 joint, and the scanning parameters were as follows: tube voltage 140 kV / 80 kV; tube current 56 mAs / 234 mAs; rotation time 0.5 seconds; helical factor 0.7; collimator 0.6 mm × 64; matrix 512 × 512; reconstruction slice thickness 0.6 mm (<1 mm).

[0027] On the pseudocolor image, bone structures were marked in blue, cancellous bone in pink, and monosodium urate (MSU) crystals in green. 80 kVp and 140 kVp images were fused together to generate a composite image, which was then manually reviewed after generation; images with obvious artifacts (nail bed / callosity) were excluded from the analysis.

[0028] Then, the original cross-sectional image of the corresponding joint in the DECT image was acquired; the MSU deposition area corresponding to the original cross-sectional image was extracted, and the soft tissue surrounding the MSU deposition area was calibrated to obtain the region of interest (ROI). Since tophi represent aggregates of MSU crystals surrounded by chronic granulomatous inflammatory tissue, the ROI was delineated at all levels including the MSU deposits and surrounding soft tissue.

[0029] Finally, the images are imported into analysis software such as 3D-Slicer, and their radiomics plugins are used to calculate the feature values ​​corresponding to each voxel in the region of interest. The feature values ​​are then assigned to the corresponding voxels to extract all radiomics features. Among these features, voxel-based features include first-order statistics describing intensity distribution, texture features characterizing spatial heterogeneity, and shape features depicting three-dimensional geometric properties.

[0030] Step 102: Perform feature selection on radiomics features using LASSO-regularized logistic regression to obtain the feature subsets corresponding to the radiomics features.

[0031] For example, in order to select key features and construct a predictive model for frequent occurrences of identical diagnoses, this application uses LASSO-regularized logistic regression to perform feature selection on radiomics features, thereby extracting the feature subsets corresponding to the radiomics features.

[0032] Specifically, feature selection is performed on radiomics features through LASSO-regularized logistic regression to obtain feature subsets corresponding to the radiomics features. This includes: dividing the radiomics features into standardized radiomics feature subsets, and training a LASSO regression fitting model based on each radiomics feature subset until the model converges, obtaining the model validation error; determining the optimal penalty parameter based on the model validation error, and inputting the optimal penalty parameter and the radiomics features into the converged LASSO regression fitting model to obtain the feature subsets.

[0033] Figure 3 This application provides an example of an image omics feature selection map based on LASSO regularized logistic regression.

[0034] Figure 4 This application provides a feature coefficient bar chart for radiomics.

[0035] In one embodiment, radiomics features are divided into standardized subsets of radiomics features, and a LASSO (Least Absolute Shrinkage and Selection) regression fitting model is trained based on each subset of radiomics features until the model converges, thus obtaining the model validation error.

[0036] The optimal tuning parameter (λ) of the LASSO model was selected based on the minimum criterion using 10-fold cross-validation. The vertical line represents the optimal value of the LASSO tuning parameter (λ), which is 0.0256, and log(λ) = -3.683.

[0037] Furthermore, by plotting the 19 selected radiomics features corresponding to the vertical lines, in the bar chart related to the feature coefficients, the vertical axis represents the feature name and the horizontal axis represents the coefficient value; blue represents positive coefficients and red represents negative coefficients.

[0038] Finally, based on the model validation error, the optimal penalty parameter is determined, and the optimal penalty parameter and radiomics features are input into the converged LASSO regression fitting model to obtain the feature subset.

[0039] Step 103: Based on the feature subset, train the preset multi-model group until each model converges through multi-model cross-validation to determine the radiomics model and the DECT volume model.

[0040] For example, multiple interest-bearing period learning models are trained using feature subsets under radiomics, and the model with the best performance is selected as the radiomics model based on the training results. Based on the tophi volume of the MTP1 joint in DECT data, a DECT volume model related to the tophi volume is constructed.

[0041] Specifically, based on feature subsets, a pre-defined multi-model group is trained until each model converges through multi-model cross-validation to determine the radiomics model and the DECT volume model. This includes: constructing each machine learning model based on the feature subsets, optimizing the model parameters of each machine learning model through cross-validation, and predicting the probability of frequent gout attacks in patients to obtain the radiomics model; where each machine learning model includes: logistic regression model, support vector machine model, random forest model, multilayer perceptron model, k-nearest neighbor model, and extra-tree model; calculating the volume of tophi corresponding to the DECT image, and constructing the DECT volume model based on the tophi volume.

[0042] Figure 5 ROC curves of various machine learning algorithms under radiomics features provided in this application embodiment.

[0043] Figure 6 This is a schematic diagram illustrating the ability of radiomics scoring to differentiate the frequency of gout attacks and risk stratification, as provided in an embodiment of this application.

[0044] In one embodiment, a prediction model is constructed based on a subset of features using various machine learning algorithms, including logistic regression (LR), support vector machine (SVM), random forest, multilayer perceptron model, k-nearest neighbor model, and additional tree model, and the model parameters are optimized using five-fold cross-validation.

[0045] A total of 1130 radiomics features were extracted from the Regions of Interest (ROIs) images. Features with intra-observer ICC (within observers) and inter-observer ICC (between observers) greater than 0.8 were included in the analysis. Nineteen features were finally selected using LASSO. Predictive models based on radiomics features were constructed using machine learning algorithms such as logistic regression (LR), support vector machine (SVM), multilayer perceptron (MLP), K-nearest neighbors, and RandomForest. Five-fold cross-validation was performed, and the best-performing RandomForest model was finally selected.

[0046] Figure 6The rad-score showed a statistically significant difference between patients with frequent seizures (FGF) and those with infrequent seizures (INGF) (training set: rad-score of FGF patients: 0.59±0.01; rad-score of INF patients: 0.41±0.01, p<0.001; validation set: rad-score of FGF patients: 0.54±0.02; rad-score of INF patients: 0.44±0.01, p<0.001).

[0047] Patients were divided into high-risk and low-risk groups based on the optimal rad-score cutoff determined by the Youden index. In the validation set, the proportion of patients with FGF in the high-risk group was significantly higher than that in the low-risk group (p<0.001).

[0048] Furthermore, the DECT volume model was mainly achieved by measuring the volume of tophi in the MTP1 joint from the DECT data. The MTP1 volume model was established based on the volume of tophi in the first metatarsophalangeal joint (MTP1) measured by DECT.

[0049] Step 104: Based on the pre-defined influencing factors of gout attacks, construct a clinical prediction model for gout patients through factor regression analysis.

[0050] For example, in order to associate the imaging features of gout patients with the frequency of attacks and the causes or influencing factors of frequent attacks, this application constructs a clinical prediction model for gout patients through factor regression analysis, in order to select relevant clinical indicators related to frequent attacks among the influencing factors of gout attacks and construct the corresponding clinical prediction model.

[0051] Specifically, based on the pre-defined influencing factors of gout attacks, a clinical prediction model for gout patients is constructed through factor regression analysis. This includes: performing univariate logistic regression analysis on the influencing factors of gout attacks to obtain the first correlation between these factors and the frequency of gout attacks; based on the first correlation, screening influencing factors to determine the first associated factors for frequent gout attacks; performing multivariate logistic regression analysis on the influencing factors of gout attacks to obtain the second correlation between these factors and the frequency of gout attacks; and based on the second correlation, determining the second associated factors for frequent gout attacks and constructing a clinical prediction model corresponding to the first associated factor and the second correlation.

[0052] In one embodiment, based on preliminary judgment from screening data, patients with frequent gout attacks have a longer disease course, higher VAS scores, and higher levels of erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), and blood glucose (GLU). Furthermore, patients with frequent attacks have a higher proportion of subcutaneous tophi, a higher proportion of family history of gout, and a higher proportion of smoking history.

[0053] Furthermore, patients with frequent gout attacks on imaging often have larger tophi under DECT.

[0054] First, a univariate logistic regression analysis was performed on the factors influencing gout attacks. The analysis showed that multiple factors, including disease duration (OR = 1.048, 95% CI: 1.000-1.098, p = 0.049), VAS score (OR = 1.242, 95% CI: 1.091-1.414, p = 0.001), subcutaneous tophi (OR = 0.469, 95% CI: 0.247-0.889, p = 0.020), and erythrocyte sedimentation rate (ESR) (OR = 1.045, 95% CI: 1.009-1.081, p = 0.014), were significantly associated with the risk of frequent gout attacks.

[0055] Then, multivariate logistic regression analysis showed that VAS score remained an independent risk factor for frequent gout attacks (OR = 1.183, 95% CI: 1.027-1.363, p = 0.02).

[0056] Finally, based on the first and second associated factors, this application constructs a clinical prediction model that includes six clinical indicators: disease duration, VAS, UA, ESR, eGFR, and BMI.

[0057] Step 105: Based on the radiomics model, DECT volume model, and clinical prediction model, the probability of gout attacks in gout patients is determined through multi-model nomograph validation.

[0058] For example, the integration of multimodal data may complement the heterogeneity of the microenvironment at different scales and enhance the predictive ability of existing models. Based on the radiomics model, DECT volume model and clinical prediction model obtained through training and analysis, this application constructs a nomograph of the model for joint evaluation through multivariate logistic regression analysis. By constructing the radiomics model, DECT volume model and clinical prediction model and the corresponding nomograph, this application realizes the prediction of gout attack probability in patients with frequent gout attacks based on radiomics DECT.

[0059] Specifically, based on the radiomics model, the DECT volume model, and the clinical prediction model, the probability of gout attacks in gout patients is determined through multi-model nomograph validation. This includes: obtaining the radiomics score output by the radiomics model and integrating the radiomics score with the variables corresponding to the DECT volume model and the clinical prediction model to form a clinical-radiomics nomograph; and determining the probability of gout attacks corresponding to the risk axis in the clinical-radiomics nomograph by cumulative mapping of variable scores based on the clinical-radiomics nomograph.

[0060] Figure 7 This application provides a KM curve diagram for high-risk and low-risk gout patients.

[0061] Figure 8 ROC curve of the working characteristics of an image omics model provided in this application embodiment.

[0062] Figure 9 This is a model calibration curve provided for an embodiment of this application.

[0063] Figure 10 This application provides a clinical-radiomics nomograph.

[0064] In one embodiment, patients were divided into low-risk and high-risk groups based on the Youden index (rad-score) in the training set as a cutoff value, and the same cutoff value was applied to the validation set. Kaplan-Meier analysis showed that patients in the low-risk group took significantly longer to reach the primary endpoint (≥2 gout attacks) than patients in the high-risk group in both the training and validation sets.

[0065] ROC curves of the radiomics model operating characteristics show that the area under the curve (AUC) values ​​of the radiomics model, clinical model, and MTP1 volumetric model in the training cohort are 0.89 (0.84-0.93), 0.67 (0.60-0.73), and 0.61 (0.54-0.67), respectively; and the area under the curve (AUC) values ​​of the validation cohort are 0.74 (0.65-0.84), 0.75 (0.66-0.85), and 0.63 (0.52-0.73), respectively.

[0066] The results above indicate that the radiomics model outperforms the other two models in the training set. Although the predictive efficiency of the radiomics model decreased in the validation set, it still has a predictive advantage over the simple MTP1 volume model (Dronwyn 0.74 (0.65-0.84) vs 0.63 (0.52-0.73), p < 0.05).

[0067] In the validation set, NRI and IDI results showed that the radiomics model combined with the clinical model improved both net reclassification and overall discriminative ability compared to the clinical model alone, i.e., NRI: 0.162, 95% CI: (0.006–0.316), P<0.05; IDI: 0.098, 95% CI: (0.047–0.157), P<0.05.

[0068] Furthermore, the radiomics model combined with the MTP1 volume model also showed significant improvement in net reclassification and overall discriminative ability compared to the MTP1 volume model alone, i.e., NRI: 0.365, 95% CI: (0.211–0.518), P<0.05; IDI: 0.204, 95% CI: (0.129–0.284), P<0.05.

[0069] The data highlighted above demonstrate the additional predictive value of radiomics, showing that imaging features can capture supplementary information beyond traditional clinical and volumetric parameters, thereby improving risk stratification and providing insights into the underlying biological mechanisms of frequent gout attacks.

[0070] The calibration curves showed good consistency, and the Hosmer-Lemeshow test showed no statistical significance on the validation set (p = 0.47).

[0071] VAS score, disease duration, body mass index (BMI), erythrocyte sedimentation rate (ESR), uric acid (UA), estimated tubular filtration rate (eGFR), MTP1 volume, and radiomics score (rad-score) were included in the clinical-radiomic nomograph.

[0072] When using nomographs, a vertical line is drawn from the corresponding axis to the point axis based on the rad-score value. The points for other variables (VAS, disease duration, BMI, ESR, UA, eGFR, and rad-score) are summed, and a vertical line is drawn from this sum to the total point axis to determine the probability of a gout attack at the lowest point. The nomograph model demonstrated superior predictive ability compared to radiomics models in the validation cohort (AUC = 0.82; 95% CI, 0.74–0.9).

[0073] Furthermore, after training a pre-defined multi-model group until each model converges based on feature subsets and multi-model cross-validation to determine the radiomics model and DECT volume model, the method further includes: acquiring proteomics validation data of tophi tissue corresponding to gout patients, and using volcano plots and heatmaps to determine protein abundance differences in the proteomics validation data to identify differences in protein abundance among gout patients; grouping the gout patients into risk groups using radiomics Rad-score scores to obtain high-risk and low-risk gout patients; and determining the molecular profile differences between high-risk and low-risk gout patients based on differences in protein abundance.

[0074] Furthermore, after determining the molecular profile differences between high-risk and low-risk gout patients based on protein abundance differences, the method also includes: performing gene set enrichment analysis on high-risk and low-risk gout patients to obtain enrichment data for high-risk gout patients; among them, patients with frequent gout attacks showed significant enrichment in complement-related pathways and ECM regulatory pathways; based on the enrichment data, the predictive performance of the radiomics model was determined through predictive effect analysis.

[0075] Figure 11 This is a schematic diagram of a representative HE staining of a gouty tophi provided in an embodiment of this application.

[0076] Figure 12 This is a proteomic landscape diagram of a tophi during an acute attack, provided as an embodiment of this application.

[0077] Figure 13 This is a schematic diagram of GSEA analysis provided in an embodiment of this application.

[0078] In one embodiment, to explore the underlying biological mechanisms of this predictive model, proteomic analysis was performed on tophi samples from gout patients. According to the model prediction, 7 cases belonged to the high-risk group and 5 cases belonged to the low-risk group.

[0079] Intergroup comparisons identified 203 differentially expressed proteins with FC > 1.5 or < 0.67 and p < 0.05 (107 were upregulated and 96 were downregulated in the high-risk group).

[0080] In gene set enrichment analysis (GSEA), proteins significantly associated with frequent gout attacks were primarily enriched in complement-related pathways, including complement cascades, complement initiation, and the generation of C4 and C2 activators. Furthermore, enrichment was also identified in extracellular matrix (ECM) regulatory pathways.

[0081] Furthermore, the volcano plot shows the differentially expressed proteins between the high-risk and low-risk groups. Red and blue dots represent proteins that are significantly upregulated and downregulated in the high-risk group (FC>1.5 or FC<0.67, P<0.05), respectively, while gray dots represent proteins with no significant difference.

[0082] A heatmap of differentially expressed proteins, with rows representing proteins and columns representing patient samples. Red and blue indicate relatively high and low expression levels, respectively. Both proteins and samples are clustered hierarchically.

[0083] Gene set enrichment analysis (GSEA) results show significantly enriched pathways in the high-risk group in bubble plots. Bubble size represents the number of core enriched genes, and color represents the normalized enrichment score (NES). GSEA enrichment plots for the four most significantly enriched pathways show the distribution of enrichment scores and the core enriched genes.

[0084] It is worth noting that, based on radiomics scores, the degree of inflammatory cell infiltration in the joint tissues of high-risk patients was significantly higher than that of low-risk patients, suggesting that the high-risk group had a persistent subclinical local inflammatory state. Proteomics analysis identified differentially expressed proteins in the high-risk and low-risk groups that were significantly enriched in gout-related pathways such as complement cascade, complement activation, and C3 and C4 protein activation.

[0085] Proteomics analysis revealed molecular signatures between high-risk and low-risk gout patients, and GSEA analysis showed significant enrichment in complement-related pathways (e.g., complement cascade, initial triggering, and C4 / C2 activator formation) and extracellular matrix (ECM) regulatory pathways in the frequent attack group. These findings provide a biological context for the predictive performance of radiomics models.

[0086] Furthermore, this application establishes a DECT-based radiomics model that can effectively predict the frequent occurrence of ULT in gout patients. The radiomics model based on multi-dimensional features is superior to the single clinical model or the single MTP1 volume model.

[0087] This application is the first to construct and validate a machine learning-based radiomics model to predict frequent gout attacks, and provides proteomics analysis to determine that frequent gout attacks may be related to classic inflammatory pathways of gout, such as complement cascade, complement activation, C2 and C4 activation, and ECM regulation. This provides a theoretical basis for subsequent model updates and prediction accuracy.

[0088] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a machine learning-based gout frequent attack prediction device, the structure of which is as follows: Figure 14 As shown.

[0089] Figure 14 This is a schematic diagram of the internal structure of a machine learning-based gout frequent attack prediction device provided in an embodiment of this application. Figure 14 As shown, the device includes: At least one processor 1401; And a memory 1402 that is communicatively connected to at least one processor; The memory 1402 stores instructions that can be executed by at least one processor, which are executed by at least one processor 1401 to enable at least one processor 1401 to implement a method for predicting frequent gout attacks based on machine learning.

[0090] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium for predicting frequent gout attacks based on machine learning stores computer-executable instructions that, when executed, can implement a method for predicting frequent gout attacks based on machine learning.

[0091] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0092] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0093] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0098] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0099] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0100] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0101] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A machine learning-based method for predicting frequent gout attacks, characterized in that, The method includes: Obtain DECT images of the corresponding index joints of gout patients, and extract radiomics features of the corresponding regions of interest in the DECT images; By using LASSO-regularized logistic regression, feature selection is performed on the radiomics features to obtain the feature subsets corresponding to the radiomics features; Based on the feature subset, a preset multi-model group is trained until each model converges through multi-model cross-validation to determine the radiomics model and the DECT volume model. Based on the pre-defined factors influencing gout attacks, a clinical prediction model for the gout patients is constructed through factor regression analysis. Based on the radiomics model, the DECT volume model, and the clinical prediction model, the probability of gout attacks in the gout patients was determined through multi-model nomograph validation.

2. The method for predicting frequent gout attacks based on machine learning according to claim 1, characterized in that, Extracting radiomics features of the region of interest corresponding to the DECT image, specifically including: Obtain the original cross-sectional image of the index joint corresponding to the DECT image; The MSU deposition region corresponding to the original cross-sectional image group is extracted, and the soft tissue surrounding the MSU deposition region is calibrated to obtain the region of interest; wherein, the region of interest includes: the MSU deposition region and the soft tissue surrounding the MSU deposition region; The feature values ​​corresponding to each voxel in the region of interest are calculated, and the feature values ​​are assigned to the corresponding voxels to determine the radiomics features; wherein, the radiomics features include: first-order statistics of intensity distribution, texture features, and shape features of three-dimensional geometric attributes.

3. The method for predicting frequent gout attacks based on machine learning according to claim 1, characterized in that, Feature selection is performed on the radiomics features using LASSO-regularized logistic regression to obtain a feature subset corresponding to the radiomics features, specifically including: The radiomics features are divided into standardized subsets, and a LASSO regression fitting model is trained based on each subset until the model converges, thus obtaining the model validation error. Based on the model validation error, the optimal penalty parameter is determined, and the optimal penalty parameter and the radiomics features are input into the convergent LASSO regression fitting model to obtain the feature subset.

4. The method for predicting frequent gout attacks based on machine learning according to claim 1, characterized in that, Based on the aforementioned feature subset, a pre-defined multi-model group is trained until each model converges through multi-model cross-validation to determine the radiomics model and the DECT volume model, specifically including: Based on the feature subset, each machine learning model is constructed, and the model parameters of each machine learning model are optimized through cross-validation to predict the probability of frequent attacks in the gout patient, so as to obtain the radiomics model; wherein, each machine learning model includes: logistic regression model, support vector machine model, random forest model, multilayer perceptron model, k-nearest neighbor model, and extra tree model. Calculate the volume of the tophi corresponding to the DECT image, and construct the DECT volume model based on the volume of the tophi.

5. The method for predicting frequent gout attacks based on machine learning according to claim 1, characterized in that, Based on the pre-defined influencing factors of gout attacks, a clinical prediction model for gout patients is constructed through factor regression analysis, specifically including: Univariate logistic regression analysis was performed on the factors influencing gout attacks to obtain the first correlation between the factors influencing gout attacks and the frequency of gout attacks. Based on the first correlation, the factors influencing the onset of gout in patients are screened to determine the first associated factor for frequent gout attacks in patients. Multivariate logistic regression analysis was performed on the factors influencing gout attacks to obtain the second correlation between the factors influencing gout attacks and the frequency of gout attacks. Based on the second correlation, a second associated factor for frequent gout attacks is determined, and a clinical prediction model corresponding to the first associated factor and the second correlation is constructed.

6. The method for predicting frequent gout attacks based on machine learning according to claim 1, characterized in that, Based on the radiomics model, the DECT volume model, and the clinical prediction model, the probability of gout attacks in the gout patients was determined through multi-model nomotype validation, specifically including: Obtain the radiomics score output by the radiomics model, and integrate the value of the radiomics score with the variables corresponding to the DECT volume model and the clinical prediction model to form a clinical-radiomics nomograph; Based on the clinical-radiomics nomograph, the probability of gout attacks corresponding to the risk axis in the clinical-radiomics nomograph is determined by cumulative mapping of variable scores.

7. The method for predicting frequent gout attacks based on machine learning according to claim 1, characterized in that, After training a pre-defined multi-model group until each model converges based on the aforementioned feature subset through multi-model cross-validation to determine the radiomics model and the DECT volume model, the method further includes: Proteomics validation data of tophi tissues corresponding to the gout patients were obtained, and protein abundance was determined by analyzing volcano plots and heatmaps to identify differences in protein abundance among the gout patients. The gout patients were grouped into high-risk and low-risk gout patients using the Rad-score radiomics score. Based on the differences in protein abundance, the molecular profile differences between the high-risk gout patients and the low-risk gout patients were determined.

8. The method for predicting frequent gout attacks based on machine learning according to claim 7, characterized in that, After determining the molecular profile differences between high-risk and low-risk gout patients based on the protein abundance differences, the method further includes: Gene set enrichment analysis was performed on the high-risk and low-risk gout patients to obtain enrichment data for the high-risk gout patients; among them, the enrichment data of the high-risk gout patients in complement-related pathways and ECM regulatory pathways were significantly enriched. Based on the enriched data, the predictive performance of the radiomics model is determined through prediction effect analysis.

9. A machine learning-based device for predicting frequent gout attacks, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to implement a machine learning-based method for predicting frequent gout attacks as described in any one of claims 1-8.

10. A non-volatile computer storage medium for predicting frequent gout attacks based on machine learning, storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they can implement a machine learning-based method for predicting frequent gout attacks as described in any one of claims 1-8.

Images