A method for predicting postoperative periprosthetic infection of artificial joint prosthesis of rheumatoid arthritis

Abstract

This invention discloses a method for predicting periprosthetic joint infection (PJI) after rheumatoid arthritis surgery, belonging to the field of medical technology. This method can be applied to predict the risk of PJI after rheumatoid arthritis surgery. The method includes: feature selection using the Boruta algorithm; and the development of an ensemble machine learning algorithm based on a combination of genetic algorithms and the SMOTEENN algorithm for risk prediction, thereby improving the accuracy and practicality of PJI risk prediction in rheumatoid arthritis patients after joint replacement surgery. This method can provide doctors with more accurate predictions and personalized treatment plans, optimize preoperative management of joint replacement surgery in rheumatoid arthritis patients, and thus maximize the prevention of PJI, improving patients' joint function and quality of life.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and in particular to a method for predicting periprosthetic infection after rheumatoid arthritis surgery. Background Technology

[0002] Periprosthetic joint infection (PJI) is a serious complication following total joint replacement surgery, significantly impacting joint function and quality of life. With the increasing prevalence of joint replacement surgery, the incidence of PJI has risen, increasing the financial burden on patients and placing greater demands on medical resources. This is particularly true for patients with rheumatoid arthritis, who face a higher risk of PJI during total joint replacement surgery.

[0003] The risk of post-operative joint replacement (PJI) in rheumatoid arthritis (RA) patients is approximately 1.6 times higher than in osteoarthritis (OAI) patients. This is likely due to the immunosuppressive therapies often received by RA patients, including DMARDs and corticosteroids. Specifically, 46% of RA patients are receiving biologic DMARDs, 67% are receiving non-biologic DMARDs, and 25% are receiving glucocorticoids. RA patients are also more prone to postoperative anemia and are more likely to require blood transfusions due to bone marrow suppression caused by chronic disease or medication use, which can increase the risk of infection. Furthermore, the fragile soft tissues around the knee joint in RA patients may make them more susceptible to infection after joint replacement surgery. Therefore, assessing the risk of PJI in this group of RA patients is particularly crucial.

[0004] PJI (Persistent Jejunal Infection) results from an imbalance between bacterial contamination and host resistance. Although host resistance is difficult to quantify, studies have shown that multiple factors, including sex, body mass index (BMI), smoking behavior, diabetes, steroid use, and immunosuppression, are closely related to PJI risk. These factors are of significant reference value in clinical assessment, and integrating them into a comprehensive risk prediction system is crucial for making risk assessment more accurate and practical.

[0005] In recent years, risk prediction systems based on literature reviews and case analyses have emerged, but they are insufficient in terms of accuracy and practicality, and rely heavily on human experience. Summary of the Invention

[0006] To address the issues of insufficient accuracy and poor practicality in postoperative infection risk prediction in existing technologies, the present invention aims to provide a method for predicting periprosthetic infection after rheumatoid arthritis surgery, thereby at least partially solving the aforementioned problems.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A method for predicting periprosthetic infection after rheumatoid arthritis surgery, the method comprising the following steps:

[0009] Collect raw data, then perform feature filtering to obtain a dataset, including a training set and a validation set;

[0010] Multiple machine learning models are trained by combining them using the training set, and then the optimal combination of machine learning models is determined by the ensemble algorithm based on the genetic algorithm to obtain an infection risk prediction model.

[0011] Postoperative infection risk is predicted using the infection risk prediction model.

[0012] In some preferred embodiments, the original data is first preprocessed to obtain preprocessed data, and then the preprocessed data is subjected to feature filtering. The preprocessing of the original data includes one or more of the following: deletion, cleaning, feature encoding, error correction, unit unification, and interpolation.

[0013] In some preferred embodiments, feature selection is performed through the following steps:

[0014] The data is resampled multiple times using a self-service method. Each self-service resampling generates a different subset of data, and a random forest is constructed using these subsets.

[0015] In each random forest, a variable importance index is used to calculate the importance of each original feature;

[0016] A shadow feature is introduced, which is generated by randomizing the original feature. The shadow feature is used to simulate the random feature and is compared with the original feature.

[0017] Compare the importance of each original feature with its corresponding shadow feature. If the importance of an original feature is significantly higher than that of its shadow feature, then the original feature is considered important.

[0018] Iterative process: Delete all shadow features and repeat the above steps until all original features are marked as "important" or "unimportant" so that feature selection can be performed based on importance ranking.

[0019] In some preferred embodiments, the dataset is split into a training set and a validation set based on K-fold cross-validation, including:

[0020] Using K-fold cross-validation, the dataset is randomly divided into several folds of roughly equal size. Each fold is used as the validation set, and the remaining folds are used as the training set.

[0021] In some preferred embodiments, the method further includes the following steps before model training:

[0022] The training set is balanced using the SMOTEENN algorithm so that the ratio of infected to non-infected samples in the training set meets a preset range.

[0023] In some preferred embodiments, the step of balancing the training set using the SMOTEENN algorithm includes:

[0024] SMOTE starts

[0025] ① Select random data from a few categories;

[0026] ② Calculate the distance between the random data and its K nearest neighbors;

[0027] ③ Multiply the difference in the number of samples by a random number between 0 and 1, and then add the result to the minority class as a synthetic sample;

[0028] ④ Repeat steps ② and ③ until the required proportion for the minority group is met;

[0029] SMOTE ends, ENN begins.

[0030] ⑤ Determine K as the number of nearest neighbors. If not determined, then K=3;

[0031] ⑥ Find the K-nearest neighbor of the observation among other observations in the training set, and then return the majority class from the K-nearest neighbors;

[0032] ⑦ If the class of an observation is different from the majority class of its K-nearest neighbors, the observation and its K-nearest neighbors will be removed from the training set;

[0033] ⑧ Repeat steps ⑤ and ⑥ until the ratio between infected and non-infected samples meets the preset range.

[0034] In some preferred embodiments, the step of training multiple machine learning models in combination and then using a genetic algorithm to determine the optimal combination based on an ensemble algorithm includes:

[0035] ① Population initialization: A population containing N samples is created by random generation. The population is a combination of sub-models including at least one machine learning model.

[0036] ② Fitness calculation: The fitness score of the sub-model combination is obtained by using ensemble learning method to verify the combination of sub-models through the validation set.

[0037] ③ Selection operation: Based on the fitness of individuals in the population, individuals with high fitness scores are selected from the current population through a roulette wheel selection method, simulating the process of natural selection. By calculating the sum of the fitness of all individuals Σfi, the relative fitness of each individual fi / Σfi is calculated, similar to softmax. Then, a random number between 0 and 1 is generated, and the number of times each individual is selected is determined based on the position of the random number in the probability region.

[0038] ④Mating operation: Use a certain mating probability threshold to control whether to take single-point crossover to generate new individuals;

[0039] ⑤ Mutation operation: Randomly generates mutation points;

[0040] ⑥ The above steps are iterated M times, the iteration is terminated, the optimal individual is determined, and the combination of the optimal individuals is the infection risk prediction model.

[0041] In some preferred embodiments, in step ②, the fitness score is calculated using f3-score, where f3-score = 10 × precision × recall / (9 × precision + recall).

[0042] In some preferred embodiments, the best individuals are combined into a bottom-level learner, and a model file is retained for the ensemble model composed of meta-learners using soft voting. Then, the interface is encapsulated, and the postoperative infection risk is predicted through the encapsulated interface.

[0043] In some preferred embodiments, the method is applied to predict the risk of periprosthetic infection after rheumatoid arthritis surgery. The screened features include blood glucose, weight, hemoglobin, AST, bilirubin, ESR, medication use, CRP, gender, and joint type.

[0044] The beneficial effects of the present invention using the above technical solution are as follows: The present invention utilizes AI technology to develop an innovative PJI risk prediction tool, which can more accurately predict the risk of PJI after rheumatoid arthritis surgery, provide clinicians with stronger decision support, optimize the preoperative management of joint replacement for rheumatoid arthritis patients, thereby effectively preventing the occurrence of PJI and improving patients' joint function and quality of life. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0046] Figure 2 This is a schematic diagram illustrating the importance of the BORUTA feature in this invention.

[0047] Figure 3 This is a schematic diagram of the genetic algorithm in this invention.

[0048] Figure 4 This is a schematic diagram comparing the scores of PJ patients and non-PJ patients in this invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0050] Example

[0051] A method for predicting periprosthetic infection after rheumatoid arthritis surgery, the method being applied to predict the risk of periprosthetic infection after rheumatoid arthritis surgery, the method comprising the following steps:

[0052] S1. Collect raw data.

[0053] Before collecting raw data, inclusion and exclusion criteria needed to be determined. After obtaining valid access to the mimic database, the inclusion criteria were: ① diagnosis of rheumatoid arthritis, ② history of joint replacement surgery, ③ readmission after joint replacement surgery. The exclusion criteria were: ① readmission within 4 weeks, ② incomplete clinical data, ③ history of related joint infections. A total of 133 patient records were collected, including 22 PJ patients.

[0054] Further, the specific patient information to be collected includes: patient number, primary disease diagnosis, surgery time, surgery type, readmission time, readmission diagnosis, age, gender, joint type, height, weight, BMI, medication history, white blood cell count, red blood cell count, platelet count, neutrophil count, lymphocyte count, hemoglobin, blood glucose, albumin, globulin, ALT, ALP, AST, total bilirubin, serum creatinine, blood urea, C-reactive protein (CRP), prothrombin time (PT), activated partial thromboplastin time (APTT), erythrocyte sedimentation rate (ESR), immunoglobulin M (IgM), immunoglobulin A (IgA), immunoglobulin G (IgG), rheumatoid factor (RF), C3, and C4.

[0055] Based on the above information, several raw data sets were constructed.

[0056] S2. Preprocess the raw data.

[0057] The preprocessing of the constructed raw data includes one or more of the following: deletion, cleaning, feature encoding, error correction, unit unification, and interpolation. After preprocessing the raw data, preprocessed data is obtained.

[0058] Data preprocessing is an important step that aims to clean, transform, and prepare data to better suit the requirements of machine learning algorithms.

[0059] "Deletion" refers to removing useless data. Useless data indicators include hospitalization time, hospital number, and diagnosis code, which are diagnostic data format indicators. These can indicate the source of the sample data, but they have no effect on machine learning prediction, so the corresponding data is deleted. At the same time, data with more than 50% missing features are also deleted.

[0060] Cleaning refers to correcting erroneous data, including outliers, and unifying data from multiple sources.

[0061] Feature encoding refers to converting non-numerical features into numerical representations, such as gender and medication history, which are represented by 0 or 1.

[0062] Error correction refers to the detection and handling of outliers in data. Outliers can be identified and handled through statistical methods or outlier detection algorithms.

[0063] Unit unification: refers to unifying data into commonly used units.

[0064] Imputation refers to using the median to impute missing values ​​in continuous or categorical variables.

[0065] S3. Perform feature filtering on the preprocessed data.

[0066] The original data in step S1 has many dimensions of feature information, but not all dimensions of feature information have analytical value, so feature filtering is required.

[0067] In this embodiment, the Boruta algorithm is used to screen out features that are highly correlated with the target variable. By comparing the original features with "false features" generated in a random order, the specific features that are truly important are determined. This includes the following steps:

[0068] ① Perform multiple self-sampling operations on the data, generating different data subsets each time, and then use these data subsets to construct a random forest.

[0069] ② In each random forest, the importance of each original feature is calculated using the variable importance index.

[0070] ③ Introduce shadow features. Shadow features are generated by randomizing the original features. Shadow features are used to simulate random features and are compared with the original features.

[0071] ④ Compare the importance of each original feature with its corresponding shadow feature. If the importance of an original feature is significantly higher than that of its shadow feature, then the original feature is considered to be important.

[0072] ⑤ Iterative process: Delete all shadow features and repeat the above steps until all original features are marked as "important" or "unimportant" so that feature selection can be performed based on importance ranking.

[0073] In this embodiment, the importance ranking of various indicators based on the Boruta algorithm is obtained from feature selection, such as... Figure 2 As shown, blood glucose, weight, hemoglobin, AST, bilirubin, ESR, medication use, and CRP are more important than shadow variables. In addition, basic features such as gender and joint type are also included in the final prediction model.

[0074] After feature filtering of the raw data, a dataset is formed.

[0075] S4. Split the dataset to obtain the training set and the validation set.

[0076] In this embodiment, the dataset is split into a training set and a validation set based on K-fold cross-validation, specifically including:

[0077] Using K-fold cross-validation, the dataset is randomly divided into several (e.g., 5) folds of roughly equal size. Each fold is used as the validation set, and the remaining folds are used as the training set.

[0078] S5. Balance the training set.

[0079] The SMOTEENN algorithm is used to balance the training set so that the ratio of infected to non-infected samples in the training set is within a preset range, such as approximately 1:1. This improves the model's performance on the minority class and avoids overfitting.

[0080] In this embodiment, the steps for balancing the training set using the SMOTEENN algorithm include:

[0081] SMOTE starts

[0082] ① Select random data from a few categories;

[0083] ② Calculate the distance between the random data and its K nearest neighbors;

[0084] ③ Multiply the sample size difference (which refers to the difference in the number of minority samples and the majority samples) by a random number between 0 and 1, and then add the result to the minority class as a synthetic sample;

[0085] ④ Repeat steps ② and ③ until the required proportion for the minority group is met;

[0086] SMOTE ends, ENN begins.

[0087] ⑤ Determine K as the number of nearest neighbors. If not determined, then K=3;

[0088] ⑥ Find the K-nearest neighbor of the observation among other observations in the training set, and then return the majority class from the K-nearest neighbors;

[0089] ⑦ If the class of an observation is different from the majority class of its K-nearest neighbors, the observation and its K-nearest neighbors will be removed from the training set;

[0090] ⑧ Repeat steps ⑤ and ⑥ until the ratio between infected and non-infected samples meets the preset range, for example, approximately 1:1.

[0091] S6. Multiple machine learning models are trained by combining them using a training set. Then, the optimal combination of machine learning models is determined by the ensemble algorithm based on a genetic algorithm to obtain an infection risk prediction model.

[0092] Machine learning models include the following 55 types:

[0093] Logistic regression: Ridge regression, Lasso regression, Elastic network regression, Unregularized;

[0094] Random Forest: Hyperparameter selection: ntree=10, 15, 20, 30, depth=20, 50, 100, 200;

[0095] Additional tree: Split point = entropy and Gini index, depth = 2, 3, 5, 10;

[0096] Decision tree: Split point = entropy and Gini index, depth = 2, 3, 5, 10;

[0097] Support Vector Machine: Kernel functions = polynomial kernel, RBF kernel, sigmoid kernel, linear kernel;

[0098] Multilayer perceptron: Hidden layers = 20, 50, 100, 200, (50, 50), (100, 100);

[0099] Gradient Boosting Tree (GBDT): ntree=20, 50, 100, 200, depth=3, 5.

[0100] For the 55 models mentioned above, each model is trained on each of the 5 training and validation sets of similar size in step S4, and the training is performed iteratively. Then, it is validated on each of the 5 validation sets to end the training.

[0101] like Figure 3 As shown, the specific steps to determine the optimal combination of machine learning models and obtain an infection risk prediction model include:

[0102] ① Population initialization: A population containing N (e.g., 5000) samples is created by random generation. The population is a combination of sub-models including at least one of the above machine learning models. Each machine learning model in the population is called an individual, and each individual is encoded as an array of 0s or 1s.

[0103] ② Fitness calculation: The fitness score of the sub-model combination is obtained by combining the sub-models in step ① and using the ensemble learning method to validate the combination through the validation set.

[0104] ③ Selection operation: Based on the fitness of individuals in the population, individuals with high fitness scores are selected from the current population through a roulette wheel selection method, simulating the process of natural selection. By calculating the sum of the fitness of all individuals Σfi, the relative fitness of each individual fi / Σfi is calculated, similar to softmax. Then, a random number between 0 and 1 is generated, and the number of times each individual is selected is determined based on the position of the random number in the probability region.

[0105] ④Mating operation: Use a certain mating probability threshold to control whether to take single-point crossover to generate new individuals.

[0106] ⑤ Mutation operation: Randomly generate mutation points.

[0107] ⑥ The above steps are iterated M times (e.g., 500 times), the iteration is terminated, the best individuals are determined, and these best individuals are decoded into specific combinations, which is the infection risk prediction model.

[0108] In step ②, the fitness score is calculated using the f3-score, which is calculated as f3-score = 10 × precision × recall / (9 × precision + recall).

[0109] The optimal combination of sub-models determined in this step consists of the following 5 sub-models: logistic regression (lasso regression), extra tree (split point = entropy, depth = 10), extra tree (split point = entropy, depth = 3), support vector machine (kernel function = linear kernel), and multilayer perceptron (hidden layer = 20).

[0110] S7. After the model training is completed, the best individuals from step S6 are combined into the bottom layer learners, and the ensemble model composed of soft-voting meta learners is saved as the Model file.

[0111] S8. Encapsulate the Model file from step S7 into an interface, and encapsulate the website as "http: / / 8.222.247.162:3838 / RA / ". Use the encapsulated interface to predict the risk of periprosthetic infection after rheumatoid arthritis surgery.

[0112] When performing risk prediction, the following information is input through the encapsulated interface: gender, joint type, blood glucose, weight, hemoglobin, AST, bilirubin, ESR, medication use, and CRP, as shown in Table 1; the output of the interface is a risk score.

[0113] Input methods for Table 1-10 indicators

[0114] Indicator Name Input method gender Qualitative input (1=male, 0=female) Joint type Qualitative input (1 = hip joint, 0 = knee joint) weight Quantitative input (kg) Use of medications related to rheumatoid arthritis Enter the following options: (None, Glucocorticoids, Methotrexate, Other) C-reactive protein level Quantitative infusion (mg / L) fasting blood glucose Quantitative input (mmol / L) AST Quantitative input (U / L) hemoglobin Quantitative input (g / L) bilirubin Quantitative input (umol / L) ESR Quantitative input (mm / h)

[0115] The method for predicting the risk of periprosthetic infection in rheumatoid arthritis artificial joints disclosed in this embodiment of the invention (referred to as the final model) was evaluated and compared with some of the 55 machine learning models in step S6, as shown in Table 2. Table 2 shows that the final model in this invention performs best in terms of AUC, recall, f2-score, and f3-score, and also performs well in terms of accuracy, precision, specificity, and f1-score.

[0116] Table 2 - Performance Evaluation Comparison of Different Models

[0117] logistic regression Random Forest extra tree Decision Tree Support Vector Machine Multilayer perceptron Gradient boosting tree Final Model auc 0.706 0.763 0.749 0.665 0.745 0.631 0.734 0.798 f1-score 0.415 0.455 0.561 0.383 0.500 0.327 0.468 0.548 accuracy 0.639 0.639 0.812 0.564 0.774 0.474 0.692 0.752 Accuracy 0.283 0.303 0.457 0.250 0.395 0.207 0.327 0.392 Recall rate 0.773 0.909 0.727 0.818 0.682 0.773 0.818 0.909 Specificity 0.613 0.586 0.829 0.514 0.793 0.414 0.667 0.721 f2-score 0.574 0.649 0.650 0.562 0.595 0.500 0.629 0.719 f3-score 0.659 0.758 0.687 0.667 0.636 0.607 0.711 0.803

[0118] Where, f1-score=1×precision×recall / (1×precision+recall), f2-score=5×precision×recall / (4×precision+recall), f3-score=10×precision×recall / (9×precision+recall).

[0119] This invention records data from the clinical electronic record system of the First Affiliated Hospital of Sun Yat-sen University on patients who underwent joint replacement surgery for rheumatoid arthritis between 2007 and 2021. Telephone follow-ups were conducted with 106 patients, including 4 PJ patients and 102 non-PJ patients. Further, relevant data such as age, blood glucose, weight, AST, PT, ESR, medication use, and CRP were collected and input into the encapsulated interface, such as... Figure 4 As shown, the scores of PJ patients were significantly higher than those of non-PJ patients (0.795 vs 0.417, p<0.001).

[0120] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for predicting periprosthetic infection after rheumatoid arthritis surgery, characterized in that, The method includes the following steps: Collect raw data, then perform feature filtering to obtain a dataset, including a training set and a validation set; Multiple machine learning models are trained in combination using the training set, and then the optimal combination of machine learning models is determined by the ensemble algorithm based on the genetic algorithm to obtain the infection risk prediction model. The determined optimal sub-model combination consists of the following 5 sub-models: logistic regression based on the lasso method; extra tree, split point = entropy, depth = 10; extra tree, split point = entropy, depth = 3; support vector machine, kernel function = linear kernel; multilayer perceptron, hidden layer = 20. Postoperative infection risk is predicted using the infection risk prediction model described above. Feature filtering is performed using the following steps: The data is resampled multiple times using a self-service method. Each self-service resampling generates a different subset of data, and a random forest is constructed using these subsets. In each random forest, a variable importance index is used to calculate the importance of each original feature; A shadow feature is introduced, which is generated by randomizing the original feature. The shadow feature is used to simulate the random feature and is compared with the original feature. Compare the importance of each original feature with its corresponding shadow feature. If the importance of an original feature is significantly higher than that of its shadow feature, then the original feature is considered important. Iterative process: Delete all shadow features and repeat the above steps until all original features are marked as important or unimportant, so that feature selection can be performed based on importance ranking; The characteristics selected after screening include blood glucose, weight, hemoglobin, AST, bilirubin, ESR, medication use, CRP, gender, and joint type. Before training the model, the method further includes the following steps: The training set is balanced using the SMOTEENN algorithm so that the ratio of infected to non-infected samples in the training set meets a preset range.

2. The method according to claim 1, characterized in that: The original data is first preprocessed to obtain preprocessed data, and then the preprocessed data is subjected to feature filtering. The preprocessing of the original data includes one or more of the following: deletion, cleaning, feature encoding, error correction, unit unification, and interpolation.

3. The method according to claim 1, characterized in that: The dataset is split into a training set and a validation set based on K-fold cross-validation, including: Using K-fold cross-validation, the dataset is randomly divided into several folds of roughly equal size. Each fold is used as the validation set, and the remaining folds are used as the training set.

4. The method according to claim 1, characterized in that, The step of balancing the training set using the SMOTEENN algorithm includes: SMOTE starts ① Select random data from a few categories; ② Calculate the distance between the random data and its K nearest neighbors; ③ Multiply the difference in the number of samples by a random number between 0 and 1, and then add the result to the minority class as a synthetic sample; ④ Repeat steps ② and ③ until the required proportion for the minority group is met; SMOTE ends, ENN begins. ⑤ Determine K as the number of nearest neighbors. If not determined, then K=3; ⑥ Find the K-nearest neighbor of the observation among other observations in the training set, and then return the majority class from the K-nearest neighbors; ⑦ If the class of an observation is different from the majority class of its K-nearest neighbors, the observation and its K-nearest neighbors will be removed from the training set; ⑧ Repeat steps ⑤ and ⑥ until the ratio between infected and non-infected samples meets the preset range.

5. The method according to claim 1, characterized in that: The step of training multiple machine learning models using the training set in combination, and then determining the optimal combination of machine learning models using an ensemble algorithm based on a genetic algorithm, includes: ① Population initialization: A population containing N samples is created by random generation. The population is a combination of sub-models including at least one machine learning model. ② Fitness calculation: The fitness score of the sub-model combination is obtained by using ensemble learning method to verify the combination of sub-models through the validation set. ③ Selection operation: Based on the fitness of individuals in the population, individuals with high fitness scores are selected from the current population through a roulette wheel selection method, simulating the process of natural selection. By calculating the total fitness of all individuals Σfi, the relative fitness of each individual fi / Σfi is calculated, and a random number between 0 and 1 is generated. The number of times each individual is selected is determined based on the position of the random number in the probability region. ④Mating operation: Use a certain mating probability threshold to control whether to take single-point crossover to generate new individuals; ⑤ Mutation operation: Randomly generates mutation points; ⑥ The above steps are iterated M times, the iteration is terminated, the optimal individual is determined, and the combination of the optimal individuals is the infection risk prediction model.

6. The method according to claim 5, characterized in that: In step ②, the fitness score is calculated using f3-score, which is calculated as f3-score = 10 × precision × recall / (9 × precision + recall).

7. The method according to claim 5, characterized in that: The best individuals are combined into the underlying learner, and the ensemble model composed of soft voting as meta learners is stored in the Model file. Then, the interface is encapsulated, and the postoperative infection risk is predicted through the encapsulated interface.

Images