Early prediction method, system, electronic device and medium for rheumatoid arthritis

Abstract

This disclosure provides a method, system, electronic device, and medium for early prediction of rheumatoid arthritis, including: feature screening of routine clinical laboratory indicators collected from patients with positive anti-CCP antibodies to obtain target routine clinical laboratory indicators; dividing the target routine clinical laboratory indicators after SHAP interpretation analysis into training and testing sets according to a preset ratio; training a traditional machine learning model and a deep learning model based on the training set; evaluating the performance of the trained model based on the testing set; inputting the target routine clinical laboratory indicators into the selected logistic regression model, and outputting the predicted incidence probability, risk level, and contribution of the indicators for rheumatoid arthritis. This disclosure integrates anti-CCP antibodies with routine clinical laboratory indicators and combines SHAP interpretation analysis for early prediction of rheumatoid arthritis in patients with positive anti-CCP antibodies, filling the gap in risk stratification of the anti-CCP antibody-positive population and improving the performance of the model.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent medical prediction technology, and in particular to an early prediction method, system, electronic device and medium for rheumatoid arthritis. Background Technology

[0002] Rheumatoid arthritis (RA) is a chronic, systemic autoimmune disease characterized by progressive inflammation of the synovial membrane, cartilage, and bone destruction. It has a high disability rate and severely impacts patients' quality of life and social productivity. Early diagnosis and timely intervention can significantly delay joint destruction and improve prognosis, but clinical practice still faces multiple challenges: early symptoms lack specificity and are easily confused with other musculoskeletal symptoms, making accurate identification difficult; traditional laboratory indicators for rheumatology have limited diagnostic value, inflammatory markers lack specificity, and rheumatoid factor has poor sensitivity and stability, failing to meet the needs of early diagnosis.

[0003] The development of artificial intelligence technology has provided new directions for the early prediction and personalized treatment of rheumatoid arthritis (RA). Many studies have applied machine learning to tasks such as RA treatment response, refractory prediction, and differential diagnosis. While existing research has achieved some predictive efficacy, it largely relies on high-tech, high-cost detection methods such as proteomics and specialized imaging, making it difficult to implement routinely in clinical practice.

[0004] Although interpretable artificial intelligence has been initially applied in some RA prediction studies, there are still significant gaps. Most studies focus on model performance but neglect clinical interpretability, making it difficult for the models to be understood and applied clinically.

[0005] In summary, the current field of early RA prediction still suffers from several key shortcomings: First, predictive indicators are costly and have poor accessibility, and high-value routine testing data have not been fully explored; second, research objectives are misaligned with clinical needs, focusing primarily on post-diagnosis prognosis and lacking early risk stratification for high-risk groups such as those with positive anti-CCP (cyclic citrullinated peptide) antibodies; third, the "black box" problem of models is prominent, lacking clinically acceptable interpretation mechanisms, which restricts practical application; and fourth, there is an information gap between anti-CCP antibodies and routine testing indicators, failing to form an efficient joint prediction scheme. These pain points collectively make it difficult for existing testing technologies to achieve low-cost, widely available, highly accurate, interpretable, and practical intelligent prediction of early RA risk. Summary of the Invention

[0006] The technical problem to be solved by this disclosure is to overcome the shortcomings of existing RA prediction methods, which are unable to predict the early risk of high-risk groups such as those with positive CCP antibodies, and to provide an early prediction method, system, electronic device and medium for rheumatoid arthritis.

[0007] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0008] The first aspect of this disclosure provides a method for early prediction of rheumatoid arthritis, the early prediction method comprising:

[0009] Clinical routine laboratory indicators were collected from patients who tested positive for anti-CCP antibodies. These clinical routine laboratory indicators included at least one of the following: complete blood count indicators, inflammatory indicators, biochemical indicators, autoantibody indicators, and demographic characteristics.

[0010] Feature screening is performed on the aforementioned routine clinical laboratory indicators to obtain target routine clinical laboratory indicators;

[0011] SHAP interpretation analysis was performed on the target routine clinical laboratory indicators;

[0012] The target routine clinical laboratory indicators after SHAP interpretation analysis were divided into training set and test set according to the preset ratio.

[0013] Traditional machine learning models and deep learning models are trained based on the training set;

[0014] The performance of trained traditional machine learning models and deep learning models is evaluated based on the test set to obtain evaluation metrics. The best-performing logistic regression model is then selected based on the evaluation metrics, which include at least AUC-ROC and MCC.

[0015] The clinical routine test indicators of patients who are positive for anti-CCP antibodies are input into the logistic regression model, and the model outputs the predicted incidence probability, risk level, and contribution of clinical routine test indicators to rheumatoid arthritis in these patients.

[0016] Optionally, the early prediction method further includes:

[0017] In response to the predicted incidence probability of rheumatoid arthritis being less than a first preset probability value, the risk level is determined to be low risk, and a prompt message for routine follow-up is generated.

[0018] or,

[0019] In response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to a first preset probability value and less than a second preset probability value, the risk level is determined to be medium risk, and a prompt message for enhanced follow-up and monitoring is generated.

[0020] or,

[0021] In response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to a second preset probability value, the risk level is determined to be high risk, and prompts for early treatment and close follow-up are generated.

[0022] Optionally, the step of performing feature screening on the routine clinical laboratory indicators to obtain the target routine clinical laboratory indicators includes:

[0023] Pearson correlation analysis was used to remove redundant features with correlation coefficients greater than a preset threshold for routine clinical laboratory indicators, and LASSO logistic regression was used for feature screening, retaining features with non-zero coefficients as target routine clinical laboratory indicators.

[0024] Optionally, the early prediction method further includes:

[0025] For the aforementioned routine clinical laboratory indicators, mean imputation was used to handle missing values, Z-score standardization was used to handle continuous variables, and one-hot coding was used to handle categorical variables.

[0026] Optionally, the early prediction method further includes:

[0027] The average absolute value of the SHAP value of the target routine clinical laboratory indicator is obtained as the contribution of the routine clinical laboratory indicator.

[0028] Based on the contribution analysis of the aforementioned routine clinical laboratory indicators, the degree of influence of the target routine clinical laboratory indicators on the incidence probability of rheumatoid arthritis was analyzed.

[0029] Optionally, the step of training the traditional machine learning model and the deep learning model based on the training set includes:

[0030] The training set was subjected to hyperparameter tuning using 5-fold cross-validation to obtain the target hyperparameters.

[0031] Traditional machine learning models and deep learning models are trained based on the target hyperparameters and the training set.

[0032] And / or,

[0033] The evaluation metrics also include at least one of accuracy, recall, precision, F1 score, specificity, and sensitivity.

[0034] A second aspect of this disclosure provides an early prediction system for rheumatoid arthritis, the early prediction system comprising:

[0035] The data acquisition module is used to collect routine clinical test indicators from patients who are positive for anti-CCP antibodies. These routine clinical test indicators include at least one of the following: blood routine indicators, inflammatory indicators, biochemical indicators, autoantibody indicators, and demographic characteristics.

[0036] The feature filtering module is used to perform feature filtering on the clinical routine test indicators to obtain target clinical routine test indicators.

[0037] The first processing module is used to perform SHAP interpretation analysis on the target routine clinical laboratory indicators;

[0038] The partitioning module is used to divide the target routine clinical laboratory indicators after SHAP interpretation and analysis into training and test sets according to a preset ratio.

[0039] The training module is used to train traditional machine learning models and deep learning models based on the training set;

[0040] The evaluation module is used to evaluate the performance of trained traditional machine learning models and deep learning models based on the test set, obtain evaluation metrics, and select the logistic regression model with the best performance based on the evaluation metrics. The evaluation metrics include at least AUC-ROC and MCC.

[0041] The prediction module is used to input the clinical routine test indicators of the patients who are positive for anti-CCP antibodies into the logistic regression model, and output the predicted incidence probability, risk level and contribution of the clinical routine test indicators of the patients who are positive for anti-CCP antibodies.

[0042] Optionally, the early prediction system further includes:

[0043] The first generation module is used to determine the risk level as low risk and generate a prompt message for routine follow-up in response to the predicted incidence probability of rheumatoid arthritis being less than a first preset probability value.

[0044] or,

[0045] The second generation module is used to determine the risk level as medium risk in response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to a first preset probability value and less than a second preset probability value, and to generate a prompt message for enhanced follow-up and monitoring.

[0046] or,

[0047] The third generation module is used to determine the risk level as high risk in response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to a second preset probability value, and to generate prompts for early treatment and close follow-up.

[0048] Optionally, the feature screening module is used to use Pearson correlation analysis to remove redundant features with correlation coefficients greater than a preset threshold for routine clinical laboratory indicators, and to use LASSO logistic regression for feature screening, retaining features with non-zero coefficients as target routine clinical laboratory indicators.

[0049] Optionally, the early prediction system further includes:

[0050] The second processing module is used to process the clinical routine test indicators by using mean imputation to handle missing values, Z-score standardization to process continuous variables, and one-hot coding to process categorical variables.

[0051] Optionally, the early prediction system further includes:

[0052] The acquisition module is used to acquire the average absolute value of the SHAP value of the target routine clinical test indicator as the contribution of the routine clinical test indicator.

[0053] The analysis module is used to analyze the impact of the target routine clinical laboratory indicators on the incidence probability of rheumatoid arthritis based on the contribution of the aforementioned routine clinical laboratory indicators.

[0054] Optionally, the training module includes:

[0055] The acquisition unit is used to perform hyperparameter tuning on the training set using 5-fold cross-validation to obtain the target hyperparameters.

[0056] A training unit is used to train traditional machine learning models and deep learning models based on the target hyperparameters and the training set.

[0057] And / or,

[0058] The evaluation metrics also include at least one of accuracy, recall, precision, F1 score, specificity, and sensitivity.

[0059] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and for running on the processor, wherein the processor executes the computer program to implement the early prediction method for rheumatoid arthritis described in the first aspect.

[0060] The fourth aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the early prediction method for rheumatoid arthritis as described in the first aspect.

[0061] The fifth aspect of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the early prediction method for rheumatoid arthritis as described in the first aspect.

[0062] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of this disclosure.

[0063] The positive and progressive effects of this disclosure are as follows:

[0064] This disclosure integrates anti-CCP antibodies with routine clinical laboratory indicators, and then uses SHAP interpretive analysis to train and test traditional machine learning models and deep learning models. The best-performing logistic regression model is selected to predict rheumatoid arthritis in patients who are positive for anti-CCP antibodies. This achieves refined risk stratification, fills the gap in risk stratification for the anti-CCP antibody positive population, and achieves global interpretability and individual traceability of the prediction process through SHAP interpretive analysis, thereby improving the performance of the model. Attached Figure Description

[0065] Figure 1 A flowchart of an early prediction method for rheumatoid arthritis provided in Embodiment 1 of this disclosure.

[0066] Figure 2 This is a schematic diagram of the logistic regression model provided in Embodiment 1 of this disclosure.

[0067] Figure 3 This is a schematic diagram of SHAP bee colony aggregation provided in Embodiment 1 of this disclosure.

[0068] Figure 4 This is a schematic diagram of the modules of the early prediction system for rheumatoid arthritis provided in Embodiment 2 of this disclosure.

[0069] Figure 5 This is a schematic diagram of the electronic device for implementing an early prediction method for rheumatoid arthritis according to Embodiment 3 of this disclosure. Detailed Implementation

[0070] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0071] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the claims or the context of the embodiments, and should not be construed as an unnecessary limitation. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0072] In this embodiment of the disclosure, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information comply with relevant laws and regulations and do not violate public order and good morals.

[0073] Example 1

[0074] Figure 1A flowchart of an early prediction method for rheumatoid arthritis provided in Embodiment 1 of this disclosure is shown below. Figure 1 As shown, this early prediction method includes:

[0075] S1. Collect routine clinical laboratory indicators from patients who are positive for anti-CCP antibodies. Routine clinical laboratory indicators include at least one of the following: complete blood count indicators, inflammatory indicators, biochemical indicators, autoantibody indicators, and demographic characteristics.

[0076] In this embodiment, early prediction of rheumatoid arthritis can be made based on routine clinical test indicators of 500 prospective cohort patients. First, the subjects were screened according to the following criteria: Inclusion criteria for subjects: (1) Age 18-70 years; (2) Positive anti-CCP antibody (≥17 U / mL) but not yet meeting the 1987 / 2010 RA classification criteria; (3) Signed informed consent form; Exclusion criteria for subjects: (1) Previously diagnosed with RA or other autoimmune diseases; (2) Comorbid infections, tumors or other diseases that may affect test indicators; (3) Lost to follow-up during the follow-up period.

[0077] Baseline laboratory tests were collected for 29 routine clinical laboratory indicators, including at least the following five categories: blood routine indicators (9 items): WBC, NEU, LYM, NLR, PLR, HGB, HCT, MCV, RDW; inflammatory indicators (5 items): CRP, ESR, FIB, Ferritin, PCT; biochemical indicators (8 items): ALB, GLO, AG, UA, TC, HDL-C, LDL-C, FBG; autoantibody indicators (3 items): RF, anti-CCP antibody, ANA; demographic characteristics (4 items): age, sex, BMI, smoking history.

[0078] Outcome variable definition: With 12 months of follow-up as the node, the following two binary outcomes were defined: Anti-CCP antibody positive outcome: Anti-CCP antibody titer ≥17U / mL was defined as positive (cohort positive rate 65.8%, 329 / 500); Early RA onset outcome: Meeting the 2010 ACR / EULAR classification criteria (total score ≥6 points) was defined as positive (cohort positive rate 81.4%, 407 / 500).

[0079] S2. Perform feature screening on routine clinical laboratory indicators to obtain target routine clinical laboratory indicators;

[0080] In an optional implementation, S2 includes:

[0081] Pearson correlation analysis was used to remove redundant features with correlation coefficients greater than a preset threshold for routine clinical laboratory indicators, and LASSO logistic regression was used for feature screening, retaining features with non-zero coefficients as target routine clinical laboratory indicators.

[0082] In this implementation, Pearson correlation analysis was used to remove redundant features with correlation coefficients >0.8 for routine clinical laboratory indicators (such as WBC and NEUT, with NEUT retained); LASSO logistic regression (alpha=0.01) was used for feature selection, retaining features with non-zero coefficients; and combined with the opinions of clinical experts, 15 core features (i.e., target routine clinical laboratory indicators) were ultimately retained: anti-CCP antibody, RF, CRP, ESR, RDW, PDW, PLT, age, gender, ALB, ALT, UA, TC, NEUT, and LYMPH.

[0083] S3. Perform SHAP interpretation analysis on the target routine clinical laboratory indicators;

[0084] In this embodiment, the SHAP interpretation analysis includes the following scenarios: Global interpretation: The KernelExplainer (general interpreter) is used to calculate the average |SHAP value of each target clinical routine test indicator. For example, the importance ranking is as follows: anti-CCP antibody (0.42) > RF (0.35) > CRP (0.28) > ESR (0.24) > RDW (0.21) > age (0.18) > PDW (0.16) > ALB (0.14). Local interpretation: One patient was randomly selected from the test set. The SHAP plot showed that the high risk (predicted incidence probability of 0.78) was mainly driven by positive anti-CCP antibody (SHAP=+0.32), elevated CRP (+0.21), and elevated RDW (+0.15), while normal ALB (-0.08) played a protective role. Dependency analysis: RDW and SHAP values ​​have a non-linear relationship. When RDW < 12.5%, the SHAP value is close to 0, and when RDW > 13.5%, the SHAP value increases sharply.

[0085] S4. Divide the target routine clinical laboratory indicators after SHAP interpretation and analysis into training set and test set according to the preset ratio.

[0086] In this implementation, stratified random sampling is used. For example, the target clinical routine test indicators after SHAP interpretation analysis are divided into a training set (350 cases) and a test set (150 cases) in a 7:3 ratio to ensure that the proportions of the two groups are consistent.

[0087] S5. Train traditional machine learning models and deep learning models based on the training set;

[0088] In this embodiment, the traditional machine learning model includes at least one of logistic regression (LR), random forest (RF), gradient boosting (GB), support vector machine (SVM), and k-nearest neighbors (KNN); the deep learning model includes at least one of multilayer perceptron (MLP), residual network (ResNet), Transformer, autoencoder classifier (AE-Classifier), and temporal convolutional network (TCN).

[0089] S6. Based on the test set, evaluate the performance of the trained traditional machine learning model and deep learning model to obtain evaluation metrics, and select the logistic regression model with the best performance based on the evaluation metrics. The evaluation metrics include at least AUC-ROC and MCC.

[0090] In an optional implementation, the evaluation metrics also include at least one of accuracy, recall, precision, F1 score, specificity, and sensitivity.

[0091] In this implementation, the performance of five traditional machine learning models and five deep learning models is compared on the same test set, and the optimal model is selected through a comprehensive evaluation of multiple indicators.

[0092] We used AUC-ROC and MCC as the main evaluation metrics, and accuracy, recall, precision, F1 score, specificity, and sensitivity as auxiliary evaluation metrics to evaluate the performance of five traditional machine learning models and five deep learning models.

[0093] Among them, AUC-ROC: used to measure the model's ability to distinguish between "high-risk" and "low-risk" individuals, and is the primary evaluation indicator. MCC: the most critical indicator when dealing with class imbalance (e.g., the incidence of RA is not 100%), effectively avoiding the "class collapse" problem (e.g., TCN and ResNet classify all samples as positive to obtain high recall). Sensitivity / Recall: important for screening tools to not miss truly high-risk patients. Specificity: avoids causing unnecessary anxiety and medical intervention for low-risk patients. F1 Score: the harmonic mean of sensitivity and specificity, used for comprehensive evaluation. The evaluation process includes: Internal validation: using 5-fold cross-validation to evaluate model stability on the training set. Independent test set: using 30% of the reserved data as the test set to evaluate the final performance of the model. Statistical test: using the DeLong test to compare whether the difference in AUC of different models is statistically significant (p<0.05), to prove that LR (logistic regression) AUC=0.8524 is significantly better than other models.

[0094] In this embodiment, the expression for accuracy is shown in formula (1):

[0095] (1)

[0096] in, The values ​​represent accuracy, TP represents true positives, FP represents false positives, FN represents false negatives, and TN represents true negatives.

[0097] The expression for precision is shown in formula (2):

[0098] (2)

[0099] in, Indicates accuracy.

[0100] The expression for recall rate is shown in formula (3):

[0101] (3)

[0102] in, This indicates the recall rate.

[0103] The expression for the F1 score is shown in formula (4):

[0104] (4)

[0105] in, This represents the F1 score.

[0106] The expression for MCC is shown in formula (5):

[0107] (5)

[0108] It should be noted that MCC∈[-1,1], where 0 represents random prediction, is the most robust comprehensive evaluation index under class imbalance.

[0109] In addition, a calibration curve was plotted using the predicted incidence rate and actual outcome from the test set to ensure that the model's predicted incidence rate matches the actual incidence rate.

[0110] An assessment revealed that, for example Figure 2 As shown, the logistic regression model (AUC=0.8524) performs best among all models, with its curve at the top; the curves of traditional machine learning models ( Figure 2 The section marked by the dashed line (in the middle) is generally superior to the curve family of deep learning models ( Figure 2 (The solid line part); Transformer (AUC=0.812) is the only deep learning model whose curve overlaps with the main region of traditional machine learning models; while the curves of TCN (AUC = 0.520) and autoencoder classifier (AUC=0.583) are close to random diagonals and have almost no discriminative ability.

[0111] Furthermore, regarding the identification of class collapse risk, confusion matrix analysis revealed that while TCN and ResNet achieved extremely high recall rates (0.995-1.000), this was due to a class collapse strategy that predicted almost all samples as positive, rather than actual discriminative ability. The logistic regression model achieved the best balance between true negative recognition rate (43.0%) and true positive recognition rate (94.3%); this finding reveals the indispensability of MCC as a robust assessment indicator in imbalanced data.

[0112] In this embodiment, the trained model is encapsulated as a RESTful API and integrated into the laboratory information system. It can be prospectively validated on 120 newly enrolled anti-CCP antibody-positive patients from January to June 2024. The AUC reached 0.836, and the calibration curve showed that the predicted incidence rate was in good agreement with the actual incidence rate (Hosmer-Lemeshow test p=0.352).

[0113] In a specific example, the evaluation metric for the model obtained based on the test set is:

[0114] Logistic Regression: AUC=0.8524, Accuracy=78.5%, Sensitivity=81.2%, Specificity=76.3%, F1=0.792;

[0115] XGBoost: AUC=0.8437, Accuracy=77.8%, Sensitivity=79.8%, Specificity=76.1%, F1=0.781;

[0116] MLP: AUC=0.8215, Accuracy=75.3%, Sensitivity=77.5%, Specificity=73.6%, F1=0.753;

[0117] Optimal Model Selection: Considering the above performance and interpretability requirements, logistic regression was selected as the core prediction model.

[0118] S7. Input the clinical routine test indicators of the patients who are positive for anti-CCP antibodies into the logistic regression model, and output the predicted incidence probability, risk level and contribution of clinical routine test indicators of rheumatoid arthritis of the patients who are positive for anti-CCP antibodies.

[0119] In this embodiment, in the process of screening the logistic regression model with the best performance, not only based on 29 routine clinical test indicators, but also combined with anti-CCP antibody and the feature interaction terms between anti-CCP antibody and routine clinical test indicators, the logistic regression model most suitable for the "fusion feature set" is selected through ensemble / comparative learning.

[0120] For characteristic interactions between anti-CCP antibodies and routine clinical laboratory indicators, such as anti-CCP positivity and CRP: capturing the synergistic effect of "autoimmune activation" and "acute inflammation." Anti-CCP titer and ESR: capturing the combined effect of "specific autoantibody levels" and "non-specific inflammatory burden." Anti-CCP positivity and ALB: capturing the risk amplification effect of autoantibody positivity in a low-protein background (weakened protective factors).

[0121] Table 1 shows a performance comparison of logistic regression models built based on different features within a specific quantum context. For example, the single-feature model uses only the feature "anti-CCP antibody" to construct a simple univariate logistic regression model. The standard model uses only 29 routine clinical laboratory indicators (excluding anti-CCP antibody) to construct a multivariate logistic regression model. The fusion model is a complete logistic regression model that includes "routine clinical laboratory indicators + anti-CCP antibody + feature interaction terms".

[0122] Table 1

[0123] Single feature model 0.73（0.66-0.78） 64 70 Conventional model 0.79（0.75-0.83） 75 74 Fusion Model 0.85（0.82-0.88） 82 77

[0124] As shown in Table 1, the logistic regression model selected based on routine clinical test indicators, anti-CCP antibody, and feature interaction terms has the best performance.

[0125] In this embodiment, based on 29 routine clinical test indicators of 500 prospective cohort patients, with ≥17 U / mL anti-CCP antibody at 12 months as the outcome label, after preprocessing, 5 traditional machine learning models and 5 deep learning models were compared. The logistic regression model with the best performance was selected through multi-indicator evaluation. Combined with SHAP interpretability analysis, key biomarkers were identified to provide support for early rheumatoid arthritis risk stratification and clinical decision-making.

[0126] This implementation method integrates anti-CCP antibodies with routine clinical test indicators, and then uses SHAP interpretive analysis to train and test traditional machine learning models and deep learning models. The best-performing logistic regression model is selected to make early predictions of rheumatoid arthritis in patients who are positive for anti-CCP antibodies. This achieves refined risk stratification, fills the gap in risk stratification for the anti-CCP antibody positive population, and achieves global interpretability and individual traceability of the prediction process through SHAP interpretive analysis, thereby improving the performance of the model.

[0127] In an optional implementation, the early prediction method further includes:

[0128] In response to the predicted incidence probability of rheumatoid arthritis being less than the first preset probability value, the risk level is determined to be low risk, and a prompt message for routine follow-up is generated.

[0129] or,

[0130] In response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to the first preset probability value and less than the second preset probability value, the risk level is determined to be medium risk, and a prompt message for enhanced follow-up and monitoring is generated.

[0131] or,

[0132] In response to a predicted incidence probability of rheumatoid arthritis being greater than or equal to a second preset probability value, the risk level is determined to be high risk, and prompts for early treatment and close follow-up are generated.

[0133] In this embodiment, decision curve analysis (DCA) is used to determine the "net benefit" of intervention under different preset probability values. The interval that maximizes the net benefit is selected as the preset probability value. For example, the first preset probability value is set to 0.3, and the second preset probability value is set to 0.7. When the predicted incidence probability of rheumatoid arthritis is <0.3, the net benefit is negative or insignificant, which is considered low risk; when 0.3 ≤ the predicted incidence probability of rheumatoid arthritis is <0.7, the net benefit is positive and stable, which is considered medium risk; when the predicted incidence probability of rheumatoid arthritis is ≥0.7, the net benefit increases significantly, which is considered high risk.

[0134] Furthermore, when the predicted incidence probability of rheumatoid arthritis (RA) is < 0.3, the risk level is determined to be low, and the logistic regression model indicates an extremely low probability of developing RA within 12 months. Recommendations include routine follow-up, such as retesting anti-CCP antibodies and routine inflammatory markers every 6-12 months, and providing health education. When the predicted incidence probability of RA is 0.3 ≤ 0.7, the risk level is determined to be medium, and the logistic regression model indicates a certain risk of developing RA, but the confidence level is not high. Recommendations include enhanced follow-up and monitoring, such as retesting every 3-6 months, and conducting more detailed joint symptom consultations. Doctors may consider exploring the use of relatively safe drugs such as hydroxychloroquine for "early intervention" based on clinical judgment, but this is not a mandatory standard. When the predicted incidence probability of RA is ≥ 0.7, the risk level is determined to be high, and the logistic regression model is highly confident that RA will develop within 12 months. Early intervention is recommended; referral to the rheumatology department is strongly advised. The doctor may consider initiating treatment with disease-modifying antirheumatic drugs (DMARDs) such as methotrexate to prevent or delay arthritis flare-ups, while conducting close follow-up (every 1-3 months).

[0135] In an optional implementation, the early prediction method further includes:

[0136] For routine clinical laboratory indicators, mean imputation was used to handle missing values, Z-score standardization was used to handle continuous variables, and one-hot coding was used to handle categorical variables.

[0137] In an optional implementation, mean imputation is used to handle missing values ​​with a missing rate of <5%; Z-score standardization is used to handle continuous variables, the expression of which is shown in formula (6): (6), where z represents the standardized clinical routine test indicators; x represents the original clinical routine test indicators; μ represents the average of all clinical routine test indicators; and σ represents the standard deviation of the clinical routine test indicators. One-hot coding was used to process categorical variables (e.g., gender, smoking history).

[0138] In an optional implementation, the early prediction method further includes:

[0139] The average absolute value of the target routine clinical laboratory indicator's SHAP value is used as the contribution of the routine clinical laboratory indicator.

[0140] This analysis focuses on the contribution of routine clinical laboratory indicators to the incidence probability of rheumatoid arthritis.

[0141] In this embodiment, the KernelExplainer is used to calculate the SHAP value of each core feature (i.e., the target clinical routine test indicators), which is based on the Shapley value definition in game theory as shown in formula (7):

[0142] (7)

[0143] in, Indicates the first The SHAP value of each feature, Represents the complete feature set. Indicates using only a subset of features The expected value predicted by the logistic regression model at time can be decomposed into the baseline value plus the sum of the SHAP values ​​of each feature: S represents all subsets of features other than the current feature i; i represents the specific feature whose SHAP value is calculated.

[0144] Furthermore, the logistic regression model's expression for predicting the incidence rate of a single anti-CCP antibody-positive patient is shown in Equation (8):

[0145] (8)

[0146] in, This represents the predicted incidence rate of a single anti-CCP antibody-positive patient using a logistic regression model. This represents the average predicted probability of disease incidence from the logistic regression model, i.e., the baseline value, without considering any feature information. For example, the baseline value... (Average incidence rate in the cohort).

[0147] In this embodiment, such as Figure 3 As shown, the average absolute values ​​of the target clinical routine laboratory indicators are sorted from top to bottom, with red indicating high eigenvalues ​​and blue indicating low eigenvalues; n=100 test samples; the analysis results are as follows:

[0148] ESR ranks first in importance (SHAP values ​​range from approximately -0.4 to +0.35): high ESR values ​​(red dot clusters) are concentrated in the positive SHAP region, while low ESR values ​​(blue dot clusters) are distributed in the negative SHAP region, showing a clear positive driving pattern.

[0149] CRP ranks second in importance: it exhibits a positive pattern highly similar to ESR, and a high CRP value corresponds to a positive SHAP contribution.

[0150] ALB ranks third in importance: it exhibits a significant reverse pattern, with high ALB values ​​corresponding to negative SHAP contributions, making it the only protective factor among the top 6 features.

[0151] PCT, AG, and HGB also exhibit a negative driving pattern, reflecting the degree of influence of infection-inflammatory state differentiation, globulin elevation-related immune activation, and chronic disease anemia on the risk of RA, respectively.

[0152] The contribution of each routine clinical laboratory indicator was further quantified using a SHAP feature importance bar chart. For example, ESR: average |SHAP|=0.0974; CRP: average |SHAP|=0.0904; ALB: average |SHAP|=0.0623.

[0153] ESR, CRP, and ALB constitute the core biomarker triad, contributing approximately 38% of all feature prediction information. Notably, anti-CCP antibody ranked only 11th (average |SHAP|=0.0231), lower than many routine clinical test indicators, indicating that under the multi-feature combined prediction framework, the comprehensive information content of inflammation and metabolic indicators is no less than that of traditional single autoantibody biomarkers.

[0154] Furthermore, the nonlinear patterns and interaction effects of key features are further revealed through the SHAP dependency graph:

[0155] ESR: It exhibits a non-linear positive correlation. When the standardized value increases from -1 to +1.5, the SHAP value jumps from -0.3 to +0.35, and then tends to plateau, showing a saturation threshold effect.

[0156] CRP: A clear activation inflection point appears at approximately +0.5 of the standardized value; below this value, SHAP approaches 0.

[0157] ALB: It shows a monotonically decreasing negative correlation and crosses zero at approximately +0.5 of the standardized value, indicating that the risk of RA is significantly increased when ALB is below the normal median.

[0158] ESR and CRP interaction: High CRP values ​​cluster in high ESR regions, indicating a positive synergistic interaction between the two.

[0159] Furthermore, for example, the SHAP waterfall plot was used to perform individualized decomposition of four representative cases (baseline probability 0.799):

[0160] True positive case #95 (predicted incidence rate of 1.000): driven by a combination of low ALB (SHAP=+0.060) and high CRP (+0.047), combined with low HGB (+0.028) and mildly elevated ESR (+0.014).

[0161] True positive case #55 (predicted incidence probability of 1.000): High CRP (+0.038) combined with very low HCT (+0.037) constituted the main driving force, suggesting that acute inflammation combined with significant anemia is an important signal combination for high-confidence RA prediction;

[0162] True negative case #4 (predicted probability of onset is 0.002): High ALB (-0.167) contributed the strongest protective effect, while low ESR (-0.135) and low CRP (-0.117) synergistically reduced the probability to 0.002;

[0163] True negative case #52 (predicted incidence probability of 0.002): low ESR (-0.191) contributed the most to the single protective effect; it is worth noting that the anti-CCP antibody value was elevated (+2.11) in this case, but the contribution of SHAP was still negative (-0.065), indicating that in the context of low levels of multiple inflammatory markers, the elevation of a single anti-CCP antibody is not enough to dominate the prediction direction, demonstrating the advantage of multi-feature combined prediction over single marker threshold judgment;

[0164] The SHAP decomposition results of all four cases were highly consistent with the clinical pathology logic, confirming that the logistic regression model prediction mechanism has good clinical interpretability.

[0165] In an optional implementation, S5 includes:

[0166] Five-fold cross-validation was used to tune the hyperparameters of the training set to obtain the target hyperparameters.

[0167] Traditional machine learning models and deep learning models are trained based on target hyperparameters and training sets.

[0168] In this implementation, for example, we compare five machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), XGBoost (XGB), and Multilayer Perceptron (MLP).

[0169] The hyperparameter settings are as follows:

[0170] LR: Regularization parameter C=1.0, penalty term L2;

[0171] SVM: Kernel function RBF, C=1.0, gamma='scale';

[0172] RF: n_estimators=100, max_depth=5, min_samples_split=5;

[0173] XGB: n_estimators=100, learning_rate=0.1, max_depth=3;

[0174] MLP: Hidden layer (64, 32), activation function ReLU, optimizer Adam, learning rate 0.001, dropout=0.2.

[0175] In this implementation method, Stratified 5-fold Cross-Validation (CV) is used for hyperparameter tuning during model training. An early stopping mechanism (patience=10) is used to prevent overfitting and ensure that the ratio of positive to negative samples in each fold is consistent with the original dataset. That is, it ensures that the ratio of anti-CCP positive / negative and RA morbidity / non-morbidity in each fold of training is consistent with the overall data, preventing the model from learning spurious correlations.

[0176] This implementation method predicts early-stage rheumatoid arthritis based on routine clinical test indicators and interpretable machine learning models. By systematically comparing five traditional machine learning models and five deep learning models, the best-performing logistic regression model is selected. Four-dimensional SHAP interpretability analysis is introduced to address key technical pain points in existing technologies, such as high indicator costs, low accessibility, lack of risk stratification, and uninterpretable models.

[0177] Furthermore, specifically targeting high-risk RA individuals who are positive for anti-CCP antibodies, this method relies solely on routine clinical laboratory indicators, requiring no special equipment or additional blood collection, reducing testing costs by over 90%. It can be widely deployed in medical institutions at all levels, enabling early screening and risk prediction of RA with anti-CCP antibodies, significantly lowering the technological application threshold. For the first time, it integrates anti-CCP antibodies with routine clinical laboratory indicators to screen for the optimal logistic regression model. This logistic regression model improves AUC by 15-20% and sensitivity by 25-30% compared to single indicators, synergistically enhancing predictive accuracy. It also explores the information complementarity between anti-CCP antibodies and routine clinical laboratory indicators, achieving… The refined stratification of low, medium, and high risk guides differentiated follow-up and early intervention, filling the gap in risk stratification for individuals with positive anti-CCP antibodies. Through SHAP interpretability analysis, the prediction process achieves global interpretability and individual traceability, quantifying the contribution of each indicator and significantly improving clinicians' trust and adoption of the AI ​​system. This overcomes the "black box" problem of machine learning, transforming the "black box" model into a clinically readable decision-making basis. Furthermore, this implementation method possesses continuous learning and cross-institutional transfer capabilities, supporting incremental updates based on real-world data and cross-population adaptation, ensuring long-term model stability and multi-center promotion capabilities, resulting in significant socio-economic benefits.

[0178] Example 2

[0179] Corresponding to the aforementioned embodiment of an early prediction method for rheumatoid arthritis, this disclosure also provides an embodiment of an early prediction system for rheumatoid arthritis.

[0180] Figure 4 This is a schematic diagram of a module of an early prediction system for rheumatoid arthritis provided in Embodiment 3 of this disclosure, as shown below. Figure 4 As shown, this early prediction system includes:

[0181] The data acquisition module 21 is used to collect routine clinical test indicators from patients who are positive for anti-CCP antibodies. These routine clinical test indicators include at least one of the following: blood routine indicators, inflammatory indicators, biochemical indicators, autoantibody indicators, and demographic characteristics.

[0182] In this embodiment, early prediction of rheumatoid arthritis can be made based on routine clinical test indicators of 500 prospective cohort patients. First, the subjects were screened according to the following criteria: Inclusion criteria for subjects: (1) Age 18-70 years; (2) Positive anti-CCP antibody (≥17 U / mL) but not yet meeting the 1987 / 2010 RA classification criteria; (3) Signed informed consent form; Exclusion criteria for subjects: (1) Previously diagnosed with RA or other autoimmune diseases; (2) Comorbid infections, tumors or other diseases that may affect test indicators; (3) Lost to follow-up during the follow-up period.

[0183] Baseline laboratory tests were collected for 29 routine clinical laboratory indicators, including at least the following five categories: blood routine indicators (9 items): WBC, NEU, LYM, NLR, PLR, HGB, HCT, MCV, RDW; inflammatory indicators (5 items): CRP, ESR, FIB, Ferritin, PCT; biochemical indicators (8 items): ALB, GLO, AG, UA, TC, HDL-C, LDL-C, FBG; autoantibody indicators (3 items): RF, anti-CCP antibody, ANA; demographic characteristics (4 items): age, sex, BMI, smoking history.

[0184] Outcome variable definition: With 12 months of follow-up as the node, the following two binary outcomes were defined: Anti-CCP antibody positive outcome: Anti-CCP antibody titer ≥17U / mL was defined as positive (cohort positive rate 65.8%, 329 / 500); Early RA onset outcome: Meeting the 2010 ACR / EULAR classification criteria (total score ≥6 points) was defined as positive (cohort positive rate 81.4%, 407 / 500).

[0185] Feature filtering module 22 is used to perform feature filtering on routine clinical laboratory indicators to obtain target routine clinical laboratory indicators.

[0186] In an optional implementation, the feature screening module 22 is used to use Pearson correlation analysis to remove redundant features with correlation coefficients greater than a preset threshold for routine clinical laboratory indicators, and to use LASSO logistic regression to screen features and retain non-zero coefficient features as target routine clinical laboratory indicators.

[0187] In this implementation, Pearson correlation analysis was used to remove redundant features with correlation coefficients >0.8 for routine clinical laboratory indicators (such as WBC and NEUT, with NEUT retained); LASSO logistic regression (alpha=0.01) was used for feature selection, retaining features with non-zero coefficients; and combined with the opinions of clinical experts, 15 core features (i.e., target routine clinical laboratory indicators) were ultimately retained: anti-CCP antibody, RF, CRP, ESR, RDW, PDW, PLT, age, gender, ALB, ALT, UA, TC, NEUT, and LYMPH.

[0188] The first processing module 23 is used to perform SHAP interpretation analysis on the target routine clinical laboratory indicators;

[0189] In this embodiment, the SHAP interpretation analysis includes the following scenarios: Global interpretation: The KernelExplainer (general interpreter) is used to calculate the average |SHAP value of each target clinical routine test indicator. For example, the importance ranking is as follows: anti-CCP antibody (0.42) > RF (0.35) > CRP (0.28) > ESR (0.24) > RDW (0.21) > age (0.18) > PDW (0.16) > ALB (0.14). Local interpretation: One patient was randomly selected from the test set. The SHAP plot showed that the high risk (predicted incidence probability of 0.78) was mainly driven by positive anti-CCP antibody (SHAP=+0.32), elevated CRP (+0.21), and elevated RDW (+0.15), while normal ALB (-0.08) played a protective role. Dependency analysis: RDW and SHAP values ​​have a non-linear relationship. When RDW < 12.5%, the SHAP value is close to 0, and when RDW > 13.5%, the SHAP value increases sharply.

[0190] The partitioning module 24 is used to divide the target clinical routine laboratory indicators after SHAP interpretation analysis into training set and test set according to a preset ratio;

[0191] In this implementation, stratified random sampling is used. For example, the target clinical routine test indicators after SHAP interpretation analysis are divided into a training set (350 cases) and a test set (150 cases) in a 7:3 ratio to ensure that the proportions of the two groups are consistent.

[0192] Training module 25 is used to train traditional machine learning models and deep learning models based on the training set;

[0193] In this embodiment, the traditional machine learning model includes at least one of logistic regression (LR), random forest (RF), gradient boosting (GB), support vector machine (SVM), and k-nearest neighbors (KNN); the deep learning model includes at least one of multilayer perceptron (MLP), residual network (ResNet), Transformer, autoencoder classifier (AE-Classifier), and temporal convolutional network (TCN).

[0194] Evaluation module 26 is used to evaluate the performance of trained traditional machine learning models and deep learning models based on the test set, obtain evaluation metrics, and select the best-performing logistic regression model based on the evaluation metrics. The evaluation metrics include at least AUC-ROC and MCC.

[0195] In an optional implementation, the evaluation metrics also include at least one of accuracy, recall, precision, F1 score, specificity, and sensitivity.

[0196] In this implementation, the performance of five traditional machine learning models and five deep learning models is compared on the same test set, and the optimal model is selected through a comprehensive evaluation of multiple indicators.

[0197] We used AUC-ROC and MCC as the main evaluation metrics, and accuracy, recall, precision, F1 score, specificity, and sensitivity as auxiliary evaluation metrics to evaluate the performance of five traditional machine learning models and five deep learning models.

[0198] Among them, AUC-ROC: used to measure the model's ability to distinguish between "high-risk" and "low-risk" individuals, and is the primary evaluation indicator. MCC: the most critical indicator when dealing with class imbalance (e.g., the incidence of RA is not 100%), effectively avoiding the "class collapse" problem (e.g., TCN and ResNet classify all samples as positive to obtain high recall). Sensitivity / Recall: important for screening tools to not miss truly high-risk patients. Specificity: avoids causing unnecessary anxiety and medical intervention for low-risk patients. F1 Score: the harmonic mean of sensitivity and specificity, used for comprehensive evaluation. The evaluation process includes: Internal validation: using 5-fold cross-validation to evaluate model stability on the training set. Independent test set: using 30% of the reserved data as the test set to evaluate the final performance of the model. Statistical test: using the DeLong test to compare whether the difference in AUC of different models is statistically significant (p<0.05), to prove that LR (logistic regression) AUC=0.8524 is significantly better than other models.

[0199] In this embodiment, the expression for accuracy is as shown in formula (1) in Example 1; the expression for precision is as shown in formula (2) in Example 1; the expression for recall is as shown in formula (3) in Example 1; the expression for F1 score is as shown in formula (4) in Example 1; and the expression for MCC is as shown in formula (5) in Example 1.

[0200] It should be noted that MCC∈[-1,1], where 0 represents random prediction, is the most robust comprehensive evaluation index under class imbalance.

[0201] In addition, a calibration curve was plotted using the predicted incidence rate and actual outcome from the test set to ensure that the model's predicted incidence rate matches the actual incidence rate.

[0202] An assessment revealed that, for example Figure 2 As shown, the logistic regression model (AUC=0.8524) performs best among all models, with its curve at the top; the curves of traditional machine learning models ( Figure 2 The section marked by the dashed line (in the middle) is generally superior to the curve family of deep learning models ( Figure 2(The solid line part); Transformer (AUC=0.812) is the only deep learning model whose curve overlaps with the main region of traditional machine learning models; while the curves of TCN (AUC = 0.520) and autoencoder classifier (AUC=0.583) are close to random diagonals and have almost no discriminative ability.

[0203] Furthermore, regarding the identification of class collapse risk, confusion matrix analysis revealed that while TCN and ResNet achieved extremely high recall rates (0.995-1.000), this was due to a class collapse strategy that predicted almost all samples as positive, rather than actual discriminative ability. The logistic regression model achieved the best balance between true negative recognition rate (43.0%) and true positive recognition rate (94.3%); this finding reveals the indispensability of MCC as a robust assessment indicator in imbalanced data.

[0204] In this embodiment, the trained model is encapsulated as a RESTful API and integrated into the laboratory information system. It can be prospectively validated on 120 newly enrolled anti-CCP antibody-positive patients from January to June 2024. The AUC reached 0.836, and the calibration curve showed that the predicted incidence rate was in good agreement with the actual incidence rate (Hosmer-Lemeshow test p=0.352).

[0205] In a specific example, the evaluation metric for the model obtained based on the test set is:

[0206] Logistic Regression: AUC=0.8524, Accuracy=78.5%, Sensitivity=81.2%, Specificity=76.3%, F1=0.792;

[0207] XGBoost: AUC=0.8437, Accuracy=77.8%, Sensitivity=79.8%, Specificity=76.1%, F1=0.781;

[0208] MLP: AUC=0.8215, Accuracy=75.3%, Sensitivity=77.5%, Specificity=73.6%, F1=0.753;

[0209] Optimal Model Selection: Considering the above performance and interpretability requirements, logistic regression was selected as the core prediction model.

[0210] The prediction module 27 is used to input the clinical routine test indicators of the patients who are positive for anti-CCP antibodies into the logistic regression model, and output the predicted incidence probability, risk level and contribution of clinical routine test indicators of rheumatoid arthritis of the patients who are positive for anti-CCP antibodies.

[0211] In this embodiment, in the process of screening the logistic regression model with the best performance, not only based on 29 routine clinical test indicators, but also combined with anti-CCP antibody and the feature interaction terms between anti-CCP antibody and routine clinical test indicators, the logistic regression model most suitable for the "fusion feature set" is selected through ensemble / comparative learning.

[0212] For characteristic interactions between anti-CCP antibodies and routine clinical laboratory indicators, such as anti-CCP positivity and CRP: capturing the synergistic effect of "autoimmune activation" and "acute inflammation." Anti-CCP titer and ESR: capturing the combined effect of "specific autoantibody levels" and "non-specific inflammatory burden." Anti-CCP positivity and ALB: capturing the risk amplification effect of autoantibody positivity in a low-protein background (weakened protective factors).

[0213] In a specific quantum context, the performance comparison of logistic regression models built based on different features is shown in Table 1 of Example 1: For example, Single-feature model: A simple univariate logistic regression model is built using only the feature "anti-CCP antibody". Conventional model: A multivariate logistic regression model is built using only 29 routine clinical test indicators (excluding anti-CCP antibody). Fusion model: A complete logistic regression model including "routine clinical test indicators + anti-CCP antibody + feature interaction terms".

[0214] As shown in Table 1, the logistic regression model selected based on routine clinical test indicators, anti-CCP antibody, and feature interaction terms has the best performance.

[0215] In this embodiment, based on 29 routine clinical test indicators of 500 prospective cohort patients, with ≥17 U / mL anti-CCP antibody at 12 months as the outcome label, after preprocessing, 5 traditional machine learning models and 5 deep learning models were compared. The logistic regression model with the best performance was selected through multi-indicator evaluation. Combined with SHAP interpretability analysis, key biomarkers were identified to provide support for early rheumatoid arthritis risk stratification and clinical decision-making.

[0216] This implementation method integrates anti-CCP antibodies with routine clinical test indicators, and then uses SHAP interpretive analysis to train and test traditional machine learning models and deep learning models. The best-performing logistic regression model is selected to make early predictions of rheumatoid arthritis in patients who are positive for anti-CCP antibodies. This achieves refined risk stratification, fills the gap in risk stratification for the anti-CCP antibody positive population, and achieves global interpretability and individual traceability of the prediction process through SHAP interpretive analysis, thereby improving the performance of the model.

[0217] In an optional implementation, the early prediction system further includes:

[0218] The first generation module is used to respond to the fact that the predicted incidence probability of rheumatoid arthritis is less than the first preset probability value, determine the risk level as low risk, and generate a prompt message for routine follow-up.

[0219] or,

[0220] The second generation module is used to determine the risk level as medium risk in response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to the first preset probability value and less than the second preset probability value, and to generate a prompt message for enhanced follow-up and monitoring.

[0221] or,

[0222] The third generation module is used to respond to the prediction of the incidence of rheumatoid arthritis being greater than or equal to the second preset probability value, determine the risk level as high risk, and generate prompts for early treatment and close follow-up.

[0223] In this embodiment, decision curve analysis (DCA) is used to determine the "net benefit" of intervention under different preset probability values. The interval that maximizes the net benefit is selected as the preset probability value. For example, the first preset probability value is set to 0.3, and the second preset probability value is set to 0.7. When the predicted incidence probability of rheumatoid arthritis is <0.3, the net benefit is negative or insignificant, which is considered low risk; when 0.3 ≤ the predicted incidence probability of rheumatoid arthritis is <0.7, the net benefit is positive and stable, which is considered medium risk; when the predicted incidence probability of rheumatoid arthritis is ≥0.7, the net benefit increases significantly, which is considered high risk.

[0224] Furthermore, when the predicted incidence probability of rheumatoid arthritis (RA) is < 0.3, the risk level is determined to be low, and the logistic regression model indicates an extremely low probability of developing RA within 12 months. Recommendations include routine follow-up, such as retesting anti-CCP antibodies and routine inflammatory markers every 6-12 months, and providing health education. When the predicted incidence probability of RA is 0.3 ≤ 0.7, the risk level is determined to be medium, and the logistic regression model indicates a certain risk of developing RA, but the confidence level is not high. Recommendations include enhanced follow-up and monitoring, such as retesting every 3-6 months, and conducting more detailed joint symptom consultations. Doctors may consider exploring the use of relatively safe drugs such as hydroxychloroquine for "early intervention" based on clinical judgment, but this is not a mandatory standard. When the predicted incidence probability of RA is ≥ 0.7, the risk level is determined to be high, and the logistic regression model is highly confident that RA will develop within 12 months. Early intervention is recommended; referral to the rheumatology department is strongly advised. The doctor may consider initiating treatment with disease-modifying antirheumatic drugs (DMARDs) such as methotrexate to prevent or delay arthritis flare-ups, while conducting close follow-up (every 1-3 months).

[0225] In an optional implementation, the early prediction system further includes:

[0226] The second processing module is used to handle missing values ​​for routine clinical laboratory indicators using mean imputation, continuous variables using Z-score standardization, and categorical variables using one-hot coding.

[0227] In an optional implementation, mean imputation is used to handle missing values ​​with a missing rate of <5%; Z-score standardization is used to handle continuous variables, and its expression is shown in formula (6) in Example 1.

[0228] In an optional implementation, the early prediction system further includes:

[0229] The acquisition module is used to obtain the average absolute value of the SHAP value of the target routine clinical test indicator as the contribution of the routine clinical test indicator.

[0230] The analysis module is used to analyze the impact of target routine clinical laboratory indicators on the incidence probability of rheumatoid arthritis based on the contribution of routine clinical laboratory indicators.

[0231] In this embodiment, KernelExplainer is used to calculate the SHAP value of each core feature (i.e., the target clinical routine test indicators), which is based on the Shapley value definition in game theory as shown in formula (7) in Example 1.

[0232] Furthermore, the expression for the predicted incidence probability of a single anti-CCP antibody-positive patient by the logistic regression model is shown in formula (8) in Example 1;

[0233] In this embodiment, such as Figure 3 As shown, the average absolute values ​​of the target clinical routine laboratory indicators are sorted from top to bottom, with red indicating high eigenvalues ​​and blue indicating low eigenvalues; n=100 test samples; the analysis results are as follows:

[0234] ESR ranks first in importance (SHAP values ​​range from approximately -0.4 to +0.35): high ESR values ​​(red dot clusters) are concentrated in the positive SHAP region, while low ESR values ​​(blue dot clusters) are distributed in the negative SHAP region, showing a clear positive driving pattern.

[0235] CRP ranks second in importance: it exhibits a positive pattern highly similar to ESR, and a high CRP value corresponds to a positive SHAP contribution.

[0236] ALB ranks third in importance: it exhibits a significant reverse pattern, with high ALB values ​​corresponding to negative SHAP contributions, making it the only protective factor among the top 6 features.

[0237] PCT, AG, and HGB also exhibit a negative driving pattern, reflecting the degree of influence of infection-inflammatory state differentiation, globulin elevation-related immune activation, and chronic disease anemia on the risk of RA, respectively.

[0238] The contribution of each routine clinical laboratory indicator was further quantified using a SHAP feature importance bar chart. For example, ESR: average |SHAP|=0.0974; CRP: average |SHAP|=0.0904; ALB: average |SHAP|=0.0623.

[0239] ESR, CRP, and ALB constitute the core biomarker triad, contributing approximately 38% of all feature prediction information. Notably, anti-CCP antibody ranked only 11th (average |SHAP|=0.0231), lower than many routine clinical test indicators, indicating that under the multi-feature combined prediction framework, the comprehensive information content of inflammation and metabolic indicators is no less than that of traditional single autoantibody biomarkers.

[0240] Furthermore, the nonlinear patterns and interaction effects of key features are further revealed through the SHAP dependency graph:

[0241] ESR: It exhibits a non-linear positive correlation. When the standardized value increases from -1 to +1.5, the SHAP value jumps from -0.3 to +0.35, and then tends to plateau, showing a saturation threshold effect.

[0242] CRP: A clear activation inflection point appears at approximately +0.5 of the standardized value; below this value, SHAP approaches 0.

[0243] ALB: It shows a monotonically decreasing negative correlation and crosses zero at approximately +0.5 of the standardized value, indicating that the risk of RA is significantly increased when ALB is below the normal median.

[0244] ESR and CRP interaction: High CRP values ​​cluster in high ESR regions, indicating a positive synergistic interaction between the two.

[0245] Furthermore, for example, the SHAP waterfall plot was used to perform individualized decomposition of four representative cases (baseline probability 0.799):

[0246] True positive case #95 (predicted incidence rate of 1.000): driven by a combination of low ALB (SHAP=+0.060) and high CRP (+0.047), combined with low HGB (+0.028) and mildly elevated ESR (+0.014).

[0247] True positive case #55 (predicted incidence probability of 1.000): High CRP (+0.038) combined with very low HCT (+0.037) constituted the main driving force, suggesting that acute inflammation combined with significant anemia is an important signal combination for high-confidence RA prediction;

[0248] True negative case #4 (predicted probability of onset is 0.002): High ALB (-0.167) contributed the strongest protective effect, while low ESR (-0.135) and low CRP (-0.117) synergistically reduced the probability to 0.002;

[0249] True negative case #52 (predicted incidence probability of 0.002): low ESR (-0.191) contributed the most to the single protective effect; it is worth noting that the anti-CCP antibody value was elevated (+2.11) in this case, but the contribution of SHAP was still negative (-0.065), indicating that in the context of low levels of multiple inflammatory markers, the elevation of a single anti-CCP antibody is not enough to dominate the prediction direction, demonstrating the advantage of multi-feature combined prediction over single marker threshold judgment;

[0250] The SHAP decomposition results of all four cases were highly consistent with the clinical pathology logic, confirming that the logistic regression model prediction mechanism has good clinical interpretability.

[0251] In an optional implementation, the training module includes:

[0252] The acquisition unit is used to perform hyperparameter tuning on the training set using 5-fold cross-validation to obtain the target hyperparameters.

[0253] Training units are used to train traditional machine learning models and deep learning models based on target hyperparameters and training sets.

[0254] In this implementation, for example, we compare five machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), XGBoost (XGB), and Multilayer Perceptron (MLP).

[0255] The hyperparameter settings are as follows:

[0256] LR: Regularization parameter C=1.0, penalty term L2;

[0257] SVM: Kernel function RBF, C=1.0, gamma='scale';

[0258] RF: n_estimators=100, max_depth=5, min_samples_split=5;

[0259] XGB: n_estimators=100, learning_rate=0.1, max_depth=3;

[0260] MLP: Hidden layer (64, 32), activation function ReLU, optimizer Adam, learning rate 0.001, dropout=0.2.

[0261] In this implementation method, Stratified 5-fold Cross-Validation (CV) is used for hyperparameter tuning during model training. An early stopping mechanism (patience=10) is used to prevent overfitting and ensure that the ratio of positive to negative samples in each fold is consistent with the original dataset. That is, it ensures that the ratio of anti-CCP positive / negative and RA morbidity / non-morbidity in each fold of training is consistent with the overall data, preventing the model from learning spurious correlations.

[0262] This implementation method predicts early-stage rheumatoid arthritis based on routine clinical test indicators and interpretable machine learning models. By systematically comparing five traditional machine learning models and five deep learning models, the best-performing logistic regression model is selected. Four-dimensional SHAP interpretability analysis is introduced to address key technical pain points in existing technologies, such as high indicator costs, low accessibility, lack of risk stratification, and uninterpretable models.

[0263] Furthermore, specifically targeting high-risk RA individuals who are positive for anti-CCP antibodies, this method relies solely on routine clinical laboratory indicators, requiring no special equipment or additional blood collection, reducing testing costs by over 90%. It can be widely deployed in medical institutions at all levels, enabling early screening and risk prediction of RA with anti-CCP antibodies, significantly lowering the technological application threshold. For the first time, it integrates anti-CCP antibodies with routine clinical laboratory indicators to screen for the optimal logistic regression model. This logistic regression model improves AUC by 15-20% and sensitivity by 25-30% compared to single indicators, synergistically enhancing predictive accuracy. It also explores the information complementarity between anti-CCP antibodies and routine clinical laboratory indicators, achieving… The refined stratification of low, medium, and high risk guides differentiated follow-up and early intervention, filling the gap in risk stratification for individuals with positive anti-CCP antibodies. Through SHAP interpretability analysis, the prediction process achieves global interpretability and individual traceability, quantifying the contribution of each indicator and significantly improving clinicians' trust and adoption of the AI ​​system. This overcomes the "black box" problem of machine learning, transforming the "black box" model into a clinically readable decision-making basis. Furthermore, this implementation method possesses continuous learning and cross-institutional transfer capabilities, supporting incremental updates based on real-world data and cross-population adaptation, ensuring long-term model stability and multi-center promotion capabilities, resulting in significant socio-economic benefits.

[0264] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0265] Example 3

[0266] Figure 5 This is a schematic diagram of the structure of an electronic device according to Embodiment 3 of this disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the early prediction method for rheumatoid arthritis described in any of the above embodiments. Figure 5 The electronic device 90 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0267] like Figure 5 As shown, the electronic device 90 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 90 may include, but are not limited to: at least one processor 91, at least one memory 92, and a bus 93 connecting different system components (including memory 92 and processor 91).

[0268] Bus 93 includes a data bus, an address bus, and a control bus.

[0269] The memory 92 may include volatile memory, such as random access memory (RAM) 921 and / or cache memory 922, and may further include read-only memory (ROM) 923.

[0270] The memory 92 may also include a program tool 925 (or utility) having a set (at least one) program module 924, such program module 924 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0271] The processor 91 executes various functional applications and data processing by running computer programs stored in the memory 92, such as the early prediction method for rheumatoid arthritis provided in any of the above embodiments.

[0272] Electronic device 90 can also communicate with one or more external devices 94 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 95. Furthermore, electronic device 90 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 96. Figure 5 As shown, network adapter 96 communicates with other modules of electronic device 90 via bus 93. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with electronic device 90, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0273] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0274] Example 4

[0275] Embodiment 4 of this disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the early prediction method for rheumatoid arthritis provided in any of the above embodiments.

[0276] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0277] Example 5

[0278] Embodiment 5 of this disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the early prediction method for rheumatoid arthritis described in any of the preceding claims.

[0279] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0280] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0281] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A method for early prediction of rheumatoid arthritis, characterized by, The early prediction methods include: Clinical routine laboratory indicators were collected from patients who tested positive for anti-CCP antibodies. These clinical routine laboratory indicators included at least one of the following: complete blood count indicators, inflammatory indicators, biochemical indicators, autoantibody indicators, and demographic characteristics. Feature screening is performed on the aforementioned routine clinical laboratory indicators to obtain target routine clinical laboratory indicators; SHAP interpretation analysis was performed on the target routine clinical laboratory indicators; The target routine clinical laboratory indicators after SHAP interpretation analysis were divided into training set and test set according to the preset ratio. Traditional machine learning models and deep learning models are trained based on the training set; The performance of trained traditional machine learning models and deep learning models is evaluated based on the test set to obtain evaluation metrics. The best-performing logistic regression model is then selected based on the evaluation metrics, which include at least AUC-ROC and MCC. The clinical routine test indicators of patients who are positive for anti-CCP antibodies are input into the logistic regression model, and the predicted incidence probability, risk level, and contribution of clinical routine test indicators of rheumatoid arthritis of patients who are positive for anti-CCP antibodies are output.

2. The method for early prediction of rheumatoid arthritis according to claim 1, wherein The early prediction method also includes: In response to the predicted incidence probability of rheumatoid arthritis being less than a first preset probability value, the risk level is determined to be low risk, and a prompt message for routine follow-up is generated. or, In response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to a first preset probability value and less than a second preset probability value, the risk level is determined to be medium risk, and a prompt message for enhanced follow-up and monitoring is generated. or, In response to the predicted incidence probability of rheumatoid arthritis being greater than or equal to a second preset probability value, the risk level is determined to be high risk, and prompts for early treatment and close follow-up are generated.

3. The method for early prediction of rheumatoid arthritis according to claim 1, wherein The step of performing feature screening on the routine clinical laboratory indicators to obtain the target routine clinical laboratory indicators includes: Pearson correlation analysis was used to remove redundant features with correlation coefficients greater than a preset threshold for routine clinical laboratory indicators, and LASSO logistic regression was used for feature screening, retaining features with non-zero coefficients as target routine clinical laboratory indicators.

4. The method for early prediction of rheumatoid arthritis according to claim 1, wherein The early prediction method also includes: For the aforementioned routine clinical laboratory indicators, mean imputation was used to handle missing values, Z-score standardization was used to handle continuous variables, and one-hot coding was used to handle categorical variables.

5. The method for early prediction of rheumatoid arthritis according to claim 1, wherein The early prediction method also includes: The average absolute value of the SHAP value of the target routine clinical laboratory indicator is obtained as the contribution of the routine clinical laboratory indicator. Based on the contribution analysis of the aforementioned routine clinical laboratory indicators, the degree of influence of the target routine clinical laboratory indicators on the incidence probability of rheumatoid arthritis was analyzed.

6. The method for early prediction of rheumatoid arthritis according to claim 1, wherein The steps of training traditional machine learning models and deep learning models based on the training set include: The training set was subjected to hyperparameter tuning using 5-fold cross-validation to obtain the target hyperparameters. Traditional machine learning models and deep learning models are trained based on the target hyperparameters and the training set. And / or, The evaluation metrics also include at least one of accuracy, recall, precision, F1 score, specificity, and sensitivity.

7. An early prediction system for rheumatoid arthritis, characterized by, The early prediction system includes: The data acquisition module is used to collect routine clinical test indicators from patients who are positive for anti-CCP antibodies. These routine clinical test indicators include at least one of the following: blood routine indicators, inflammatory indicators, biochemical indicators, autoantibody indicators, and demographic characteristics. The feature filtering module is used to perform feature filtering on the clinical routine test indicators to obtain target clinical routine test indicators. The first processing module is used to perform SHAP interpretation analysis on the target routine clinical laboratory indicators; The partitioning module is used to divide the target routine clinical laboratory indicators after SHAP interpretation and analysis into training and test sets according to a preset ratio. The training module is used to train traditional machine learning models and deep learning models based on the training set; The evaluation module is used to evaluate the performance of trained traditional machine learning models and deep learning models based on the test set, obtain evaluation metrics, and select the logistic regression model with the best performance based on the evaluation metrics. The evaluation metrics include at least AUC-ROC and MCC. The prediction module is used to input the clinical routine test indicators of the patients who are positive for anti-CCP antibodies into the logistic regression model, and output the predicted incidence probability, risk level and contribution of the clinical routine test indicators of the patients who are positive for anti-CCP antibodies.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory for running on the processor, characterized in that, When the processor executes the computer program, it implements the method for early prediction of rheumatoid arthritis as described in any one of claims 1 to 6.

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

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for early prediction of rheumatoid arthritis as described in any one of claims 1 to 6.

Images