Construction method for risk aggregation prediction model of urinary tract infection after pelvic plasty of children, and related system and equipment

Abstract

The invention relates to a construction method of a child pyeloplasty postoperative urinary tract infection risk aggregation prediction model and a related system and device, and belongs to the technical field of pyeloplasty postoperative urinary tract infection risk assessment. According to the method, the LightGBM model and the TabNet model are combined, and the efficient data processing capability of the LightGBM and the deep feature learning capability of the TabNet are utilized, so that the respective defects can be made up, the prediction performance and adaptability of the whole model are improved, and the prediction efficiency of the whole model is improved. The model fusion strategy adopted by the invention can effectively integrate the decision tree advantages of LightGBM in processing a small amount of key features and the neural network advantages of TabNet in processing complex feature interaction, and the model prediction performance is significantly improved. The aggregation model-based urinary tract infection risk diagnosis system for children after pyeloplasty provided by the invention can perform early warning on the urinary tract infection risk of children suffering from pyeloplasty, optimize treatment and improve prognosis, and has an important application value in urinary tract infection risk diagnosis for children after pyeloplasty.

Description

Technical Field

[0001] The present application relates to a method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children and related systems and equipment, belonging to the technical field of risk assessment of urinary tract infection after pyeloplasty. Background Art

[0002] Ureteropelvic junction obstruction (UPJO) refers to an obstruction at the junction between the kidney and ureter, resulting in reduced urine flow from the renal pelvis to the ureter. UPJO is the most common cause of pathological hydronephrosis in children, with an incidence of approximately 1:1500. Approximately 20% to 50% of children with UPJO eventually require surgical intervention. If left untreated, it may lead to hydronephrosis, chronic infection or urolithiasis, and often progressive renal insufficiency. Although the success rate of UPJO surgery is high, there are also many complications. The main complications of UPJO surgery are urinary tract infection (UTI), urine extravasation and leakage, pyelonephritis, bleeding and recurrent UPJO.

[0003] Several studies have shown that UTI increases the risk of anastomotic restenosis after pyeloplasty in children with UPJO and is considered to be one of the important factors for postoperative anastomotic restenosis, and the incidence of UTI remains high. UTI is mainly caused by Escherichia coli and is one of the most common types of infection in children. Previous studies have described risk factors for UTI in children, including male sex, weight, BUN level, recurrence of UTI within 3 months, long-term indwelling urinary catheter, double J tube placement, bilateral double J tube placement, etc. However, the development of predictive models for UTI after unilateral pyeloplasty is very limited, especially in large-sample cohort studies. There is currently no artificial intelligence model for UTI after pyeloplasty.

[0004] Over the past decade, the rate of advancement in artificial intelligence (AI) has increased exponentially. Increases in machine learning processor power and the amount of available data have laid the foundation for the development of deep learning and AI tools. The healthcare industry collects vast amounts of data, such as medical records and patient test results. With recent advances in AI, these technologies are now increasingly being used to inform clinical decision-making. With the development of highly efficient ensemble learning models, machine learning (ML) has achieved significant breakthroughs in the accurate and efficient processing of tabular data, making these methods particularly well-suited for clinical applications.

[0005] For example, ML algorithms can be used to analyze the prediction and diagnostic evaluation of bladder cancer, as well as the staging and recurrence prediction of kidney cancer and testicular cancer. In terms of kidney disease, ML has also shown good predictive application effects. Although the application of artificial intelligence technology in the field of urology has gradually increased, there are still few reports on its application in the field of functional urology. A review article on the application of machine learning in healthcare published in the Lancet magazine in 2024 showed that medical staff and patients believe that ML risk prediction models can add benefits in the field of healthcare. However, it is undeniable that the research and application of artificial intelligence technologies represented by ML may change the clinical practice of urology and improve the overall outcomes of patients.

[0006] TabNet is an end-to-end deep learning (DL) architecture with an encoder. It establishes a sequential multi-step architecture, which facilitates decision-making in selecting relevant features. In recent years, TabNet has been increasingly used in clinical prediction. For example, Zhang et al. used TabNet to predict tacrolimus dosage after renal transplantation; Yu et al. developed a predictive model for rapatinib in breast cancer patients; Cahan et al. established a 30-day all-cause mortality model for patients with pulmonary embolism; Asadi-Pooya et al. used TabNet to distinguish different types of epilepsy; and Chen et al. developed a mammographic diagnosis model for breast cancer. However, our search of medical applications of TabNet in the literature still found significantly fewer reports than on traditional machine learning models. This may be because medical data structures are often composed of small sample datasets from single centers, while deep learning models require large data sets for training and parameter tuning.

[0007] Although ML and DL each have their own advantages and disadvantages in the development and application of medical modeling, there is currently no artificial intelligence model for UTI after hydronephrosis surgery in children. This may result in the inability to provide early prediction and warning for patients after hydronephrosis surgery, resulting in the inability to receive timely and effective treatment for UTI, increasing the risk of long-term functional effects.

[0008] Ureteropelvic junction stenosis is a common condition in children. Laparoscopic or open unilateral pyeloplasty is the standard surgical approach for this condition. Based on our center's experience, over the past 10 years, among over 700 children undergoing unilateral pyeloplasty, the infection rate exceeded 30% within six months after surgery, significantly impacting postoperative recovery, quality of life, and family financial burden. Current methods for predicting postoperative infection rely heavily on empirical judgment and limited clinical parameters, lacking a multifactorial clinical prognostic model. Therefore, establishing an accurate prediction model for urinary tract infection after pyeloplasty in children is crucial. Summary of the Invention

[0009] In order to solve the above problems, a method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children and related systems and equipment are provided. The aggregated model can effectively integrate the decision tree advantages of LightGBM in processing a small number of key features and the neural network advantages of TabNet in processing complex feature interactions, significantly improving the model prediction performance. It has important application value in the diagnosis of the risk of urinary tract infection after pyeloplasty in children.

[0010] According to one aspect of the present application, a method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children is provided, the method comprising the following steps:

[0011] S1. Screen cases of patients who underwent pyeloplasty due to unilateral hydronephrosis from historical cases and divide the cases into training set and validation set;

[0012] S2. Input the training samples into the TabNet deep learning network and train the model in regression mode to obtain a trained TabNet model; input the training samples into the LightGBM machine learning network for model training to obtain a trained LightGBM model;

[0013] S3. Input the training samples into the Meta Learner deep learning network and train the model in regression mode to obtain a trained Meta Learner model. The Meta Learner deep learning network model uses the trained TabNet model and LightGBM model as base models, and uses Logistic Regression as the meta-model for meta-model training.

[0014] S4. Input the validation set into the trained Meta Learner model, and evaluate the identification and calibration through cross-validation, where the evaluation of identification includes accuracy, precision, and recall, so as to obtain the aggregation model for diagnosing the risk of urinary tract infection after pyeloplasty in children.

[0015] It should be noted that the training of the TabNet deep learning network and the LightGBM machine learning network in step S2 can be carried out simultaneously or successively, because there is no dependency between the training of the two. The training of the MetaLearner deep learning network in step S3 needs to be carried out after the training of both the TabNet deep learning network and the LightGBM machine learning network is completed, because the Meta Learner deep learning network model is based on the trained TabNet model and LightGBM model.

[0016] Optionally, the case includes the following clinical characteristic information: 1) operation time; 2) post-operative drainage method: nephrostomy or DJ tube; 3) age; 4) weight; 5) height; 6) BMI; 7) total protein Tp; 8) albumin Alb; 9) globulin Glb; 10) blood uric acid Ua; 11) blood creatinine Cr; 12) blood urea nitrogen Bun; 13) glomerular filtration rate eGFR; 14) neutrophil ratio N%; 15) lymphocyte ratio L%; 16) anteroposterior diameter of renal pelvis APD; 17) renal parenchymal thickness.

[0017] It should be noted that according to the SHAP results of the clinical agent importance analysis in the aggregate model in the present application, gender, surgical method, number of surgeries, history of urinary tract infection, history of nephrostomy, history of abdominal pain, cause of ureteropelvic junction stenosis, and side of hydronephrosis are basically not important in the aggregate prediction model of urinary tract infection risk after pyeloplasty in children in the present application. Therefore, the clinical characteristics can be simplified and the above 17 clinical characteristic indicators can be used. However, it should be noted that those skilled in the art can also add other clinical characteristic indicators.

[0018] In addition, it should be pointed out that the clinical characteristic information in the case is used to train the TabNet deep learning network, LightGBM machine learning network and Meta Learner deep learning network, and can also be used to input the clinical characteristic information of the case to be predicted into the trained Meta Learner model, thereby outputting the risk prediction results of clinical urinary tract infection diagnosed in children after pyeloplasty. It is understandable that the case can also include other clinical characteristic information in addition to the above-mentioned records.

[0019] Optionally, the case includes the following clinical characteristics: 1) gender; 2) surgical method: laparoscopic or open; 3) operation time; 4) postoperative drainage method: nephrostomy or DJ tube; 5) number of operations: first or non-first; 6) history of urinary tract infection: yes or no; 7) history of nephrostomy: yes or no; 8) history of abdominal pain: yes or no; 9) causes of ureteropelvic junction stenosis: congenital stenosis, aberrant vascular compression, ureteral polyps, ureteral valves, high ureter or ureteral rupture; 10) side of hydronephrosis: left or right; 11) age; 12) weight; 13) height; 14) BMI; 15) total protein (Tp); 16) albumin (Alb); 17) globulin (Glb); 18) blood uric acid (Ua); 19) blood creatinine (Cr); 20) blood urea nitrogen (Bun); 21) glomerular filtration rate (eGFR); 22) neutrophil ratio (N%); 23) lymphocyte ratio (L%); 24) anteroposterior diameter (APD) of the renal pelvis; and 25) renal parenchymal thickness.

[0020] Optionally, in the TabNet deep learning network in step S2: the data flows through the batch normalization BN layer and then enters the feature converter, the feature converter includes n gated linear unit GLU blocks, each of the GLU blocks consists of a fully connected FC layer, a BN layer and a GLU activation module.

[0021] Optionally, the feature converter includes four gated linear unit (GLU) blocks, two of which are shared and the other two are independently configured.

[0022] Optionally, the feature converter processes the batch normalized features and then outputs them to the attention converter in step i through the segmentation layer;

[0023] The attention transformer includes a fully connected FC layer, a BN layer, a priori scale layer, and a sparsemax layer.

[0024] Layer, the attention converter receives input from the segmentation layer, which is then processed by the fully connected FC layer and the BN layer, and then passes through the prior scale layer. The prior scale layer summarizes the feature magnitudes in the previous decision step, which is expressed as:

[0025] Where γ represents the relaxation parameter;

[0026] The attention converter calculates a learnable mask vector M[i] based on the feature a[i-1] processed in the previous step for sparse feature selection:

[0027] M[i]=sparsemax(P[i-1]·h i (a[i-1]))

[0028] Among them, P[i-1] is the prior scale term, h i () is a trainable function that encapsulates the fully connected FC layer and BN layer. The sparsemax layer normalizes the coefficients to enforce sparse feature selection;

[0029] In order to further regulate the feature sparsity, the sparsity regularization term L in the form of entropy is adopted sparse :

[0030]

[0031] The mask vector is fed into the feature transformer, which processes the filtered features and separates them into two outputs:

[0032] [d[i],a[i]]=f i (M[i]·f)

[0033] Where d[i] represents the output of the decision step and a[i] provides information for subsequent steps.

[0034] Optionally, the screening of historical cases in step S1 must simultaneously meet the following requirements: 1) unilateral hydronephrosis undergoing pyeloplasty; 2) meeting the diagnosis of ureteropelvic junction stenosis; 3) complete clinical feature data;

[0035] Exclusion criteria were any of the following: 1) bilateral hydronephrosis; 2) failure to use conventional anti-infection regimens; 3) complete ureteral dilatation; 4) duplicated kidney; 5) combined vesicoureteral reflux; 6) DRF < 10% of the affected kidney before surgery; 7) pediatric bladder and rectal dysfunction (BBD); and 8) incomplete clinical feature data.

[0036] Optionally, the risk prediction result of clinical urinary tract infection diagnosed in the child after pyeloplasty is clinical urinary tract infection occurring within 6 months after the operation.

[0037] Optionally, in step S4, discrimination and calibration are evaluated by 5-fold cross validation, wherein:

[0038]

[0039] TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively; P and N represent positive and negative, respectively.

[0040] According to one aspect of the present application, a system for diagnosing the risk of urinary tract infection after pyeloplasty in children based on an aggregate model is provided. The system is implemented based on any of the above-mentioned methods for constructing an aggregate prediction model for the risk of urinary tract infection after pyeloplasty in children. The system comprises:

[0041] A data input module, wherein the data input module is used to input clinical characteristic data of the case;

[0042] A data processing module, wherein the data processing module is used to execute the constructed aggregation model;

[0043] A data output module is used to output the diagnosis results of the risk of urinary tract infection after pyeloplasty in children.

[0044] According to one aspect of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the processor implements any one of the steps of the above-mentioned method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children.

[0045] According to one aspect of the present application, a readable storage medium is provided, which stores a computer program, and the computer program is suitable for loading by a processor to execute any step of the above-mentioned method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children.

[0046] The beneficial effects of this application include but are not limited to:

[0047] 1. According to the method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children and the related systems and equipment of this application, the clinical prediction model for UTI after pyeloplasty provided in this application helps in early prediction and timely treatment, and ultimately reduces the risk of UTI affecting long-term function.

[0048] 2. According to the construction method of the aggregated prediction model of urinary tract infection risk after pyeloplasty in children and the related systems and equipment of this application, the respective advantages of ML and DL can be combined to improve the effect of UTI prediction after hydronephrosis surgery. It improves the prediction performance for specific tasks and data sets by dynamically adjusting the contribution of different base models, and dynamically realizes knowledge collaboration.

[0049] 3. According to the construction method of the urinary tract infection risk aggregation prediction model after pyeloplasty in children and the related systems and equipment of this application, the model fusion strategy can effectively integrate the decision tree advantages of LightGBM in processing a small number of key features and the neural network advantages of TabNet in processing complex feature interactions, significantly improving the model prediction performance.

[0050] 4. According to the method for constructing an aggregated prediction model for the risk of urinary tract infection after pyeloplasty in children and the related systems and equipment of this application, the system provided by the scheme of this application can predict the risk of urinary tract infection after pyeloplasty in children, provide prospective clinical intervention, reduce the incidence of postoperative infection and re-obstruction rate, improve the prognosis of pyeloplasty, and reduce health and economic burdens. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0052] Figure 1 This is a performance result diagram of the TabNet prediction model involved in Example 4 of this application;

[0053] Figure 2 This is a graph showing the area under the ROC curve for different models involved in Example 4 of the present application;

[0054] Figure 3This is a graph showing the area under the ROC curve for different models involved in Example 4 of the present application;

[0055] Figure 4 This is a performance result diagram of the LightGBM model (25 features) involved in Example 4 of this application;

[0056] Figure 5 This is a performance result diagram of the LightGBM model (18 features) involved in Example 4 of this application;

[0057] Figure 6 This is a performance result diagram of the TabNet model (25 features) involved in Example 4 of the present application;

[0058] Figure 7 This is a performance result diagram of Meta Learner (25 features) involved in Example 4 of this application;

[0059] Figure 8 This is a graph showing the LightGBM feature importance analysis results involved in Example 5 of this application;

[0060] Figure 9 This is a diagram showing the TabNet feature importance analysis results involved in Example 5 of this application;

[0061] Figure 10 This is a SHAP result diagram of the first 20 features of Meta Learner involved in Example 6 of this application;

[0062] Figure 11 This is a feature dependency analysis result diagram of the Meta Learner involved in Example 6 of the present application;

[0063] Figure 12 This is a schematic diagram of the TabNet encoder structure involved in Example 2 of this application;

[0064] Figure 13 This is a schematic diagram of the TabNet decoder structure involved in Example 2 of this application;

[0065] Figure 14 This is a schematic diagram of the TabNet feature converter structure involved in Example 2 of this application;

[0066] Figure 15 This is a schematic diagram of the TabNet attention converter structure involved in Example 2 of this application;

[0067] Figure 16 This is a structural diagram of the Meta Learner involved in Example 3 of this application. DETAILED DESCRIPTION

[0068] The present application is described in detail below with reference to examples, but the present application is not limited to these examples. Unless otherwise specified, the raw materials and catalysts in the examples of the present application are purchased through commercial channels.

[0069] In this application, data from more than 1,000 pediatric patients who underwent open or classic three-port laparoscopic unilateral pyeloplasty in the Department of Urology of the Children's Hospital Affiliated to Capital Institute of Pediatrics from January 2012 to January 2023 were collected. After inclusion and exclusion, 764 patients were enrolled. A total of 25 clinical features were collected, including the general condition, medical history, surgical conditions, hydronephrosis measurement values, etiology, postoperative drainage methods, immunological indicators, renal function indicators and blood cell indicators. After feature engineering was completed using LightGBM, five traditional machine learning algorithms, including Logistic regression, SVM, random forest, XGBoosting and LightGBM, were used for modeling. At the same time, the deep learning TableNet algorithm was also used for feature importance analysis and modeling. In this application, the performance and feature analysis of different machine learning and deep learning models in the risk prediction of acute upper urinary tract infection (acute pyelonephritis) were compared, and a new integrated learning method was proposed, which can combine the respective advantages of traditional machine learning methods and deep learning methods, significantly improve the performance of the prediction model, and combine SHAP to complete the visualization of the integrated black box model.

[0070] Of the 764 enrolled patients, 265 developed urinary tract infections after surgery, 217 of which occurred within 3 months after surgery. Of the 265 UTI cases, 73 were diagnosed as acute upper urinary tract infections (acute pyelonephritis) and 192 as acute lower urinary tract infections. Among the entire cohort, 41 patients (5.4%) experienced re-obstruction during mid- to long-term follow-up after surgery, of whom 22 had UTIs early (within 3 months) after surgery. In the non-infected group, only 19 patients (3.8%) experienced re-obstruction in the long term, a statistically significant difference (χ2 = 0.05). 2 =155.11, P<0.01). This indicates that postoperative UTI has a significant impact on the occurrence of re-obstruction, indicating that diagnosis of the risk of urinary tract infection after pyeloplasty in children is very necessary and has important practical application significance.

[0071] The accuracy, precision and AUC of TabNet after modeling all features (25Features) are better than those of traditional machine learning models such as LightGBM. In this application, LightGBM and TabNet are used as base models, logistic regression (LR) is used as the meta-model, and the model aggregation method is adopted to integrate the advantages of the two classifiers. The performance of the aggregated classifier is significantly better than that of the base model: the accuracy is 0.7974, the precision is 0.8200, the recall rate is 0.6508, and the area under the ROC curve is 0.8018. The prediction performance of the model is significantly improved and enhanced.

[0072] In this application, not only the first machine learning prediction model for urinary tract infection after pyeloplasty in children is proposed, but also the important role of eGFR and ALB in the prediction of UTI after pyeloplasty in children is proposed. The role of feature engineering in model building is explored, the modeling performance of traditional machine learning (LightGBM) and deep learning (TableNet) algorithms is compared, and an integrated learning model Meta Learner is proposed. The application of model aggregation method can integrate the advantages of the two classifiers and significantly improve the model prediction performance.

[0073] This study was reviewed and approved by the Institutional Review Board of the Capital Institute of Pediatrics. This study was reviewed by the Ethics Committee of this institution, with the ethics approval number: SHERLLM2023041. During the data acquisition period (June 1, 2023 to December 31, 2023), a total of 1033 cases of pediatric hydronephrosis were obtained. After applying the inclusion and exclusion criteria, a total of 764 cases were included in the cohort. Case inclusion criteria: (1) unilateral hydronephrosis undergoing pyeloplasty (2) diagnosis of ureteropelvic junction stenosis in accordance with the EAU and the "Expert Consensus on the Diagnosis and Treatment of Congenital Hydronephrosis" (3) complete clinical data and cooperation with postoperative follow-up. Case exclusion criteria: (1) bilateral hydronephrosis (2) no conventional anti-infection regimen (3) full ureteral dilatation (4) duplicate kidney (5) combined vesicoureteral reflux (6) preoperative dilated renal function (DRF) <10% in the affected kidney (7) pediatric bladder and rectal dysfunction (BBD) (8) incomplete clinical data. Patients were divided into two cohorts in an 8:2 ratio: a training set (n=612) and a validation set (n=152).

[0074] The data collected for the enrolled cases included the general condition of the children (sex, age, height, weight, body mass index (BMI), conditions before pyeloplasty surgery (history of abdominal pain, history of nephrostomy, history of urinary tract infection), surgical conditions (number of surgeries (first / non-first), operation time, surgical method (laparoscopic / open), side of hydronephrosis), anteroposterior diameter of the renal pelvis and renal parenchyma thickness measured by Doppler color ultrasound before surgery, causes of ureteropelvic junction stenosis (congenital stenosis, vagal vascular compression, ureteral polyps, ureteral valves, high-positioned ureters, ureteral rupture), postoperative drainage method (nephrostomy / DJ tube), immunological indicators (total protein, albumin, globulin), renal function indicators (serum creatinine, blood urea nitrogen, glomerular filtration rate, blood uric acid), and blood cell indicators (neutrophil ratio, lymphocyte ratio).

[0075] The outcome label for this study was clinical urinary tract infection (UTI) diagnosed after pyeloplasty, with the onset defined as UTI occurring within 6 months after surgery. UTI was diagnosed based on: ① The 2015 European Association of Urology (EAU) guidelines for urinary tract infection; and ② The 2016 Chinese Medical Association Pediatric Nephrology Group's Evidence-Based Guidelines for the Diagnosis and Treatment of Urinary Tract Infection.

[0076] Cefuroxime sodium (20 mg / kg) is administered intravenously within 0.5 hours before surgery to prevent infection. After surgery, cefuroxime sodium is administered intravenously for 5-7 days at a daily dose of 30-100 mg / kg, divided into 3-4 doses. Further treatment adjustments are made based on clinical symptoms and urine routine monitoring. Prophylactic antimicrobial therapy during DJ tube placement: Since there are currently no guidelines or consensus regarding prophylactic antimicrobial therapy for children with DJ tubes, our center adopts the EAU and AUA guidelines for VUR and the expert consensus of the Chinese Society of Pediatric Urology: ① Infants under 2 months of age: Cefaclor suspension, taken orally once daily at bedtime, at a dose of 1 / 3 the therapeutic dose. ② Infants 2 months to 7 years of age: Co-trimoxazole (cotrimoxazole), at a dose of 1 / 3 the therapeutic dose. ③ Infants older than 7 years of age: Amoxicillin-clavulanate potassium, taken orally once daily at bedtime, at a dose of 1 / 3 the therapeutic dose. During the period of DJ tube placement, urine routine and urine smear tests should be reviewed every 2 weeks, anti-infection regimens should be adjusted in a timely manner, and drug resistance monitoring should be performed.

[0077] This study used six machine learning and deep learning algorithms to develop a predictive model for urinary tract infection after pyeloplasty. The performance of the different algorithms was compared. A novel ensemble learning approach was proposed, combining the strengths of traditional machine learning and deep learning methods to improve the performance of the predictive model. Finally, the ensemble black-box model was visualized using SHAP. Five common machine learning algorithms were used: logistic regression, support vector machine (SVM), random forest, extreme gradient boosting tree model (XGBoosting), and lightweight gradient boosting tree model (LightGBM). Accuracy, precision, recall, and area under the receiver operating characteristic (ROC) curve (AUC) were used to evaluate the predictive accuracy and consistency of the model. Discrimination and calibration were assessed using 5-fold cross-validation.

[0078] The calculation formulas for accuracy, precision, and recall are as follows:

[0079]

[0080] TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively; P and N represent positive and negative, respectively.

[0081] The present application is described below through specific embodiments.

[0082] Example 1 Dataset Structure

[0083] The research data in this application were derived from pediatric patients who underwent surgery for hydronephrosis in a single center over a 10-year period. Based on the exclusion criteria, data from 764 patients were finally included for a retrospective cohort study. Among the 764 enrolled cases, 265 developed urinary tract infections after surgery, of which 217 occurred within 3 months after surgery and 48 occurred between 3 and 6 months after surgery. Of the 265 cases of urinary tract infection, 73 were diagnosed with acute upper urinary tract infection (acute pyelonephritis) and 192 were diagnosed with acute lower urinary tract infection. Among the upper urinary tract infection cases, 32 were diagnosed with acute focal bacterial nephritis and 7 were diagnosed with renal abscess. New renal scarring was found by DMSA examination.

[0084] Each case in this cohort was followed up for a long period of time (1 to 8.5 years, median 5.5 years). The incidence of urinary tract infection and re-obstruction after surgery was collected and recorded. The results showed that 41 cases (5.4%) in the entire cohort had re-obstruction in the mid- to long-term follow-up after surgery, of which 22 had urinary tract infection in the early postoperative period (within 3 months). In the group without postoperative infection, nearly 19 cases (3.8%) had re-obstruction in the long term. The difference was statistically significant (χ 2 =155.11, P<0.01). This indicates that early postoperative urinary tract infection has a significant impact on the occurrence of ureteropelvic junction re-obstruction in the medium and long term.

[0085] This study included patients and collected 25 clinical characteristics, including demographics, biochemical renal function tests, immunological parameters, and specific hydronephrosis markers. The baseline clinical and statistical descriptions of the cohort are shown in Table 1. The training and test cohorts were randomly assigned in a 4:1 ratio, and no statistical difference was found in the distribution of the 25 characteristics, as shown in Table 2. Enumeration data are presented as "cases," and intergroup comparisons were performed using the χ2 test or Fisher's exact test. Quantitative data are presented as M[Q1,Q3], and intergroup comparisons were performed using the independent sample t-test or rank-sum test. A P value < 0.05 was considered statistically significant.

[0086] The clinical characteristics corresponding to each row in Tables 1 and 2 are as follows: 1) gender; 2) surgical method: laparoscopic or open; 3) operation time; 4) postoperative drainage method: nephrostomy or DJ tube; 5) number of operations: first time or not first time; 6) history of urinary tract infection: yes or no; 7) history of nephrostomy: yes or no; 8) history of abdominal pain: yes or no; 9) cause of ureteropelvic junction stenosis: congenital ureteral stricture, vagal vein compression, ureteral polyps, ureteral valves, high ureter, or ureteral rupture. rupture); 10) side of hydronephrosis: left L or right R; 11) age; 12) weight; 13) height; 14) BMI; 15) total protein Tp; 16) albumin Alb; 17) globulin Glb; 18) blood uric acid Ua; 19) blood creatinine Cr; 20) blood urea nitrogen Bun; 21) glomerular filtration rate eGFR; 22) neutrophil ratio N%; 23) lymphocyte ratio L%; 24) anteroposterior diameter of the renal pelvis APD; 25) renal parenchymal thickness.

[0087] Table 1 Baseline characteristics of the infection group and the non-infection group

[0088]

[0089] Table 1 continued

[0090]

[0091] Table 2 Univariate comparison of training cohort and validation cohort

[0092]

[0093] Table 2 continued

[0094]

[0095] Example 2 TabNet TabNet is a deep neural network specially designed for tabular data processing. Its encoder architecture (e.g. Figure 12The TabNet encoder is composed of a feature transformer, an attentive transformer, and a mask. The split block splits the processed output representation for the attention transformer in subsequent steps and for overall use. For each step, the feature selection mask provides interpretable information about the model's functionality, and the masks can be aggregated to obtain important global feature properties. In each step i, the processed information from the previous step i-1 guides feature selection and generates an aggregated feature representation for overall decision making. The model inputs a data batch with a specified batch size B and D-dimensional features without the need for global feature selection or normalization, where B is the batch size and D is the feature dimension. The BN operation, i.e. BatchNormal normalization, is performed to obtain BN_Features of dimension B×D. Then, based on serialized multi-step processing, there are N_steps rounds of decision making, and BN_Features is input in each round. The features selected in the i-th round are determined by the information passed from the i-1th round. The selected features are then feature processed and the results of all rounds of decision feature processing are fused, and the corresponding activation function outputs are connected according to the task type.

[0096] The data flows through the batch normalization layer (BN layer) and enters the feature transformer. The transformer contains n gated linear unit (GLU) blocks, each of which consists of a fully connected (FC) layer, a BN layer, and a GLU activation. As a specific implementation, Figure 14 As shown in the figure, when four GLU blocks are used, two are shared and two are independent, resulting in a total of 4 layers of network, 2 of which are shared by all decision steps and 2 are related to the decision steps, thereby optimizing parameter efficiency and robustness, and adding skip connections between consecutive blocks to maintain information flow stability.

[0097] The feature converter processes the batch normalized features and outputs them to the attention converter through the split layer (Split) in step i. Figure 15 As shown in Figure 1, the Attentive Transformer integrates a series of layers: FC, BN, Prior scales, and sparsemax. First, it receives input from the split layer, processes it through the FC layer and BN layer, and then passes it through the Prior scale layer. The Prior scale layer summarizes the feature magnitudes in the previous decision step and is expressed as:

[0098]

[0099] where γ represents the relaxation parameter.

[0100] like Figure 11 As shown, the attention transformer calculates a learnable mask vector M[i] based on the feature a[i-1] processed in the previous step for sparse feature selection:

[0101] M[i]=sparsemax(P[i-1]·h i (a[i-1]))

[0102] Here, P[i-1] is the prior scale term, and hi() is a trainable function that encapsulates the FC and BN layers. The sparsemax layer normalizes the coefficients, thereby enforcing sparse feature selection. In the Attentive Transformer, a single layer map is modulated using prior scale information, which summarizes the usage of each feature before the current decision step. Sparsemax is used to normalize the coefficients, enabling sparse selection of salient features.

[0103] In order to further regulate the feature sparsity, a sparsity regularization term (L sparse ):

[0104]

[0105] Where ∈ is a small constant that ensures the sparsity of the regularization term, and is not specifically required in the model. In the model of the present application, ∈ is set to e^-20.

[0106] The mask vector is fed into the feature transformer, which processes the filtered features and separates them into two outputs:

[0107] [d[di],a[i]]=f i (M[i]·f)

[0108] Where d[i] represents the output of the decision step and a[i] provides information for subsequent steps.

[0109] TabNet provides interpretability through local and global feature explanations. The model aggregates the step outputs to produce a scalar that represents the importance of the step to the final result:

[0110]

[0111] The local feature importance comes from summing the contributions of the steps of a single sample b. The global aggregate feature importance mask Magg (b,j)is the normalized feature contribution, calculated as follows:

[0112]

[0113] like Figure 13 The following figure shows a schematic diagram of the TabNet decoder. The input is the encoded representation of the data, and the output is the reconstructed features. Each step consists of a feature transformer. TabNet eliminates the need for feature engineering and is an improvement over traditional methods such as random forests and LightGBM. Unlike previous methods, TabNet can directly process tabular data without manual feature selection or conversion. This significantly reduces the preprocessing workload and can effectively process raw, unprocessed data.

[0114] Example 3 Meta Learner

[0115] However, TabNet's reliance on neural network architecture complicates model interpretation and requires careful tuning of hyperparameters to achieve optimal performance. Furthermore, TabNet is prone to overfitting, especially in small-sample medical scenarios. The network optimization process can easily fall into local minima during training, which can lead to decreased generalization and poor performance on unseen data.

[0116] To this end, this application proposes a new Meta Learner method, which introduces an advanced ensemble learning method based on multiple meta-models. It aims to enhance the predictive performance of specific tasks and datasets by dynamically adjusting the contributions of different meta-models. Unlike traditional single models or simple ensemble techniques, Meta Learner integrates cross-model knowledge and dynamically realizes knowledge collaboration.

[0117] Specifically, Meta Learner coordinates learning strategies in multiple branches, where each branch contains independent base models, including TabNet, SVM, logistic regression, random forest, LightGBM, and XGBoost. Each base model M i The input data X is processed independently and predictions are generated as follows:

[0118] y i =M i (X)

[0119] These predictions are then combined using a meta-network that aggregates the outputs of all base meta-models:

[0120]

[0121] The meta-model employs a debate ensemble technique, where the predictions of each base model are weightedly combined to produce a final prediction. This hierarchical learning architecture enables us to capture and utilize information from different levels of abstraction, resulting in a comprehensive understanding of the data structure and underlying patterns.

[0122] As a specific implementation method, the Meta Learner in the present application solution selects a combination of the deep learning TabNet model and the LightGBM model with the best performance in traditional machine learning. The meta network aggregates the TabNet model and the LightGBM model, such as Figure 16 shown.

[0123] LightGBM and TabNet are trained separately, and then used as base models. In this specific implementation, Logistic regression is used as the meta-model for meta-model training.

[0124] Meta-learning, or learning to learn, is the science of systematically observing how different machine learning methods perform across a wide range of learning tasks, and then learning from this experience, or metadata, to learn new tasks much faster than would otherwise be possible. This not only greatly speeds up and improves the design of machine learning pipelines or neural architectures, it also allows hand-designed algorithms to be replaced with new methods that learn in a data-driven manner.

[0125] The challenge of meta-learning is to learn from previous experience in a systematic, data-segmented manner. First, metadata describing previously learned tasks and previously learned models must be collected. This includes the precise algorithm configuration used to train the model, including hyperparameter settings, pipeline composition, and / or network architecture; resulting model evaluations such as accuracy and training time; learned model parameters such as the trained neural network weights; and measurable properties of the task itself, also known as meta-features. Learning from this prior metadata is then necessary to extract and transfer knowledge to guide the search for optimal models for new tasks.

[0126] Specifically, in the specific implementation of this application, LightGBM and TabNet are used together as the base model. LightGBM is also a lightweight gradient boosting machine based on GBDT, with the following characteristics:

[0127] LightGBM uses a histogram-based algorithm that bins continuous feature (attribute) values ​​into discrete bins. This speeds up training and reduces memory usage. The advantages of a histogram-based algorithm include: 1) Lower computational cost for each split: Pre-sorting-based algorithms have a time complexity, but this only involves a fast summation operation. Once the histogram is constructed, the histogram-based algorithm has a time complexity, and the number of bins is much smaller than the data; 2) Further acceleration using histogram subtraction: To obtain the histogram of a leaf in a binary tree, the histograms of its parents and neighbors are subtracted, so it only needs to construct a histogram for a single leaf. This can then be done with a small cost through histogram subtraction; 3) Reduced memory usage: Continuous values ​​are replaced with discrete bins. If the number of bins is small, small data types such as uint8_t can be used to store the training data, eliminating the need to store additional information for pre-sorted feature values; 4) Reduced communication costs for distributed learning.

[0128] Reference link for training the LightGBM model in this example https: / / lightgbm.readthedocs.io / en / stable / Features.html The content of

[15] is used for the training of LightGBM. For LightGBM, please refer to the article LightGBM: A Highly Efficient Gradient Boosting Decision Tree (Authors: Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, Tie-Yan Liu). It should be noted that in this embodiment, feature engineering is performed before LightGBM model training, but feature engineering can also be omitted.

[0129] Example 4 Feature Engineering and Multi-method Modeling Analysis Results

[0130] 1) Feature Engineering

[0131] Traditional machine learning models require feature engineering before they are built. This process begins by ranking the importance of all features based on the underlying model, performing feature screening, and setting up cross-validation. Using network parameter search methods, typically using accuracy as a criterion, features with a significant impact on model results are incorporated into the final model. However, the deep learning model TabNet eliminates the need for feature engineering screening. It inputs all features into the final model, performing only a ranking and analysis of their importance.

[0132] The results show that feature engineering can significantly improve the predictive performance of traditional machine learning models. For example, before feature engineering, the LightGBM model achieved accuracy of 0.70, precision of 0.69, recall of 0.70, and an area under the receiver operating characteristic (ROC) curve of 0.74. After feature engineering, the LightGBM model achieved accuracy of 0.71, precision of 0.71, recall of 0.71, and an area under the receiver operating characteristic (ROC) curve of 0.78.

[0133] 2) Prediction results of different models

[0134] The performance comparison results of the prediction models constructed using different algorithms are shown in Table 3 below, where the results of TabNet are as follows: Figure 1 As shown, the ROC results of different models are as follows Figure 2 shown.

[0135] The performance ranking of six machine learning and deep learning algorithms shows that, from low to high, the order of model performance is: Logistic Regression, SVM, Random Forest, XGBoost, LightGBM without feature engineering, LightGBM with feature engineering, and finally TableNet.

[0136] Table 3 Performance comparison of prediction models constructed using different algorithms

[0137]

[0138] 3) Results of LightGBM feature engineering and TabNet modeling before and after

[0139] The accuracy, precision and AUC of TabNet after modeling all features (25Features) are better than those of the LightGBM model, while the recall rate is lower. LightGBM and TabNet are used as base models, logistic regression (LR) is used as a meta-model, and the model aggregation method is adopted to integrate the advantages of the two classifiers. The performance of the combined classifier is significantly better than that of the base model (Accuracy: 0.7974, Precision: 0.8200, Recall: 0.6508, ROC AUC: 0.8018). This shows that the integration method can significantly improve the prediction performance, especially when combining different classification algorithms. The results of the area under the ROC curve are shown in Figure 2. Figure 3 shown.

[0140] The performance of the LightGBM model (25 features) before feature engineering on the test set showed Accuracy: 0.6959, Precision: 0.6933, Recall: 0.6959, and ROC AUC: 0.7444. Figure 4 The LightGBM model (18 features) after feature engineering has a test set performance of Accuracy: 0.7059, Precision: 0.7043, Recall: 0.7059, and ROC AUC: 0.7794. Figure 5 As shown. The Tablenet model (25 features) test set performance shows Accuracy: 0.7516, Precision: 0.8571, Recall: 0.4762, ROC AUC: 0.7918. Figure 6 As shown. LightGBM and Tabnet are used as base models, Logistic Regression is used as meta-model, and the performance of the aggregate model Meta Learner (25 features) test set shows Accuracy: 0.7974, Precision: 0.8200, Recall: 0.6508, ROC AUC: 0.8018. The results are as follows Figure 7 shown.

[0141] In summary, Meta Learner has the best model performance.

[0142] Example 5 Feature Importance Analysis and Visualization Results of LightGBM and TabNet

[0143] 1) LightGBM feature importance analysis

[0144] The results are as follows Figure 8 As shown in the figure, the feature importance results based on the LightGBM model show that estimated glomerular filtration rate (eGFR) is the most influential feature, followed by blood urea nitrogen value (Bun), serum albumin (ALB), neutrophil ratio (N%), operation time (time), serum creatinine value (cr), renal parenchymal thickness (Renal), serum globulin (GLB), blood uric acid value (ua) and lymphocyte ratio (L%); in addition, the anteroposterior diameter of the renal pelvis (apd), body mass index (bmi), serum total protein (tp), age (age), weight (weight), drainage method (drainge), height (height), and surgical method (surgery) also have a certain influence in the prediction model.

[0145] 2) TabNet feature importance analysis

[0146] The results are as follows Figure 9The results of the feature importance analysis based on the Tabnet model showed that serum albumin (ALB) was the most influential feature, followed by estimated glomerular filtration rate (eGFR), serum globulin (GLB), serum total protein (tp), serum creatinine (cr), lymphocyte ratio (L%), anteroposterior diameter of the renal pelvis (apd), drainage method (drainge), surgical side (side), and blood urea nitrogen (Bun). In addition, age (age), body mass index (bmi), height (height), history of abdominal pain (abdominalgia), serum uric acid (ua), cause (cause), number of surgeries (number of oerations), and neutrophil ratio (n) also had a certain influence in the prediction model.

[0147] Example 6 Feature Importance Analysis and Feature Cross-Correlation after Model Aggregation

[0148] 1) Feature Importance Analysis

[0149] To further analyze the positive or negative impact of the identified important features on the prediction and to investigate the relationship between them, SHAP (shapley additive explanations) was performed using Python 3.7.0.

[0150] SHAP is a post hoc interpretation method that draws on principles from game theory. SHAP evaluates the impact of individual features by calculating their marginal contribution to the model. The fundamental goal of SHAP is to quantify the impact of each feature on the prediction of an instance. A higher SHAP value indicates a higher probability of AKI. For each feature, a point is created based on its positive (red) or negative (blue) contribution to the model output for each patient, resulting in a dotted line for each feature across all patients.

[0151] The top 20 features based on SHAP values ​​had a significant impact on the model output. The results showed that eGFR was the most influential feature, followed by postoperative drainage method, serum urea nitrogen, serum albumin, renal parenchymal thickness, neutrophil ratio, serum uric acid level, serum total protein, surgical time, serum globulin, serum creatinine level, anteroposterior diameter of the renal pelvis, lymphocyte ratio, age, height, BMI index, weight, surgical method, surgical side, and patient gender.

[0152] To reveal the overall positive or negative impact on the model output, a SHAP summary plot of the first 20 features of the aggregated model was created, such as Figure 10As shown in the figure, eGFR has a negative correlation with the model prediction results. When the eGFR value increases, the SHAP value tends to decrease, indicating that higher eGFR values ​​have a negative impact on the model prediction results. When the postoperative drainage draine is 1 (i.e., DJ tube is used), SHAP tends to increase. When the bun value increases, the SHAP value tends to decrease.

[0153] 2) Feature dependency analysis

[0154] Based on the feature importance score of the aggregate model, the most important features among the top 20 features were selected to create a feature dependency graph. The four features selected were estimated glomerular filtration rate (egfr), postoperative drainage method (drainge), serum urea nitrogen (bun), and serum albumin level (alb). Figure 11 As shown in the figure, all three figures show a trend of decreasing SHAP values ​​as eGFR values ​​increase, indicating that higher eGFR values ​​have a negative impact on the model prediction results. In the low eGFR value range, SHAP values ​​are mostly positive, and in the high eGFR value range, SHAP values ​​are mostly negative.

[0155] Figure a shows the draining feature (draininge), which demonstrates that higher drainage values ​​are more common in the low eGFR range, while lower drainage values ​​are more common in the high eGFR range. Figure b shows the bun feature, and Figure c shows the alb feature. Both show that bun and alb values ​​are relatively evenly distributed across the eGFR range. Figure d shows the glb feature, which demonstrates that glb values ​​are relatively evenly distributed across the different draining method (draininge) values.

[0156] In summary, LightGBM, or lightweight gradient boosting machine, is an efficient tree-based learning algorithm that optimizes computational speed and memory usage by constructing histograms and adopting a leaf-first strategy, significantly improving the efficiency of processing large-scale data. Furthermore, LightGBM supports efficient parallel computing, enabling it to perform well in big data environments. However, this model may not be as effective as neural network models on datasets with high feature dimensions or a large number of categorical features because it performs less automatic feature crosstalk and complex pattern recognition. TabNet, through its unique encoder and sparse learning mask mechanism, is able to effectively perform feature selection and representation learning on tabular data. This mechanism enables TabNet to capture complex data relationships without the need for intensive feature engineering, demonstrating its unique advantages in feature interactions and nonlinear pattern recognition. Furthermore, TabNet's online learning capability enables the model to be continuously updated, adapting to dynamically changing data environments.

[0157] The combination of these two models in this application, that is, the use of LightGBM's efficient data processing capabilities and TabNet's deep feature learning capabilities, can make up for their respective shortcomings to a certain extent, thereby improving the predictive performance and adaptability of the overall model. This model fusion strategy can effectively integrate the decision tree advantages of LightGBM when processing a small number of key features and the neural network advantages of TabNet when processing complex feature interactions. Specifically, through this fusion, when the feature dimension is high and there are more category features, the deep relationship between features can be learned through TabNet, and when the feature dimension is low or there are fewer key features, data analysis and prediction can be quickly and effectively achieved through LightGBM. In addition, the combined model can also show higher robustness and accuracy on different data sets, especially in the applications of medical data analysis and complex biomarker identification, this method has important practical significance.

[0158] The feature importance analysis results of the LightGBM model and the Tabnet model showed that eGFR and alb were common important features of the two models. The LightGBM model screened and ranked the 25 features and used the SHAP value for visualization analysis. The results showed that eGFR, blood urea nitrogen (Bun), and serum albumin concentration (Alb) were the top three important features. The feature importance analysis results of the Tabnet model showed that serum albumin concentration (Alb), eGFR, and Glb were the top three important features. The two indicators, eGFR and ALB, showed high importance in both the LightGBM and TabNet models.

[0159] Estimated glomerular filtration rate (eGFR) can be assessed by measuring substance clearance, circulating markers, and formulas that estimate glomerular filtration rate. In this study, eGFR was calculated using the updated Schwartz bedside eGFR formula (CKiDcr) for all cases. Both the LightGBM and TabNet models demonstrated that eGFR was significant and contributed significantly to the model's predictive power, which has important clinical implications for predicting infection after pyeloplasty. Previous studies have also demonstrated that eGFR can be a predictor of urinary tract infection in patients with diabetes. Although none of the patients in this study had a history of systemic disease, given that all underwent urinary tract devascularization and reconstruction surgery, an invasive procedure that damages the urethral mucosa and facilitates bacterial entry, thus increasing the risk of urinary tract infection, a critical eGFR value may not necessarily be clinically significant. Our results suggest that lower eGFR is associated with a higher risk of urinary tract infection, and lower eGFR levels may be associated with the severity of obstructive hydronephrosis. Previous studies have shown that a renal pelvis with a maximum anteroposterior diameter greater than 15 mm, accompanied by dilatation of the surrounding calyces and changes in the renal parenchyma, is a high-risk factor for urinary tract infection. Urinary tract obstruction may also impair the urinary tract's self-cleansing function in children, leading to excessive urine retention in the renal system, which promotes bacterial growth and reproduction. Studies have shown that urine retention caused by UPJO provides time and opportunity for bacteria to adhere to urothelial cells, proliferate, and infect the host.

[0160] In the past, serum albumin levels were generally considered an indicator of nutritional status. However, recent literature has demonstrated a close relationship between serum albumin concentration, disease severity, and mortality. Hypoalbuminemia can be considered an independent risk factor for morbidity and mortality, particularly in infectious diseases, where low serum albumin levels are associated with an increased risk of death in patients with severe sepsis. Low albumin concentrations reflect low serum protein levels, resulting in low serum bactericidal activity and susceptibility to bacterial infections. Falguera et al. have suggested that low serum albumin is a predictor of infection (such as community-acquired pneumonia). In this study, ALB results were obtained within 3 days before surgery, minimizing the impact of surgical intervention. Recent studies have also suggested that albumin levels are also an indicator of inflammatory status. A study by Lee et al. suggests that hypoalbuminemia may reflect baseline nutritional status or be associated with inflammatory changes during sepsis. Although the exact mechanisms remain to be fully elucidated, serum albumin has protective effects, such as antioxidant activity, maintenance of physiological homeostasis, and anti-inflammatory effects. A study by Li et al. showed that children with acute glomerulonephritis (AGN) who have low serum protein levels are more likely to develop urinary tract infections. In recent years, some researchers have proposed combining different biological test results to predict the risk of UTI. For example, the serum procalcitonin / albumin ratio can be used for early diagnosis and prediction of urosepsis and febrile UTI. Among patients with urosepsis, those with a higher procalcitonin / albumin ratio are more likely to develop urosepsis shock. The urea nitrogen / albumin ratio can be used to predict febrile UTI in infants, identify children at high risk for bacteremia, and facilitate management decisions for infants with febrile UTI. The results of this study suggest that ALB levels should be monitored in children undergoing pyeloplasty to early identify those at high risk for postoperative UTI and enable timely clinical management decisions.

[0161] The Meta Learner prediction model provided in this application can predict the risk of urinary tract infection after pyeloplasty in children, provide prospective clinical intervention, reduce the incidence of postoperative infection and re-obstruction rate, improve the prognosis of pyeloplasty, and reduce health and economic burdens. It has important application value in the diagnosis of urinary tract infection risk after pyeloplasty in children.

[0162] The foregoing is merely an embodiment of the present application, and the scope of protection of the present application is not limited by these specific embodiments, but is determined by the claims of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc. made within the technical ideas and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children, characterized in that: The method comprises the following steps: S1. Screen cases of patients who underwent pyeloplasty due to unilateral hydronephrosis from historical cases and divide the cases into training set and validation set; S2. Input the training samples into the TabNet deep learning network and train the model in regression mode to obtain a trained TabNet model; input the training samples into the LightGBM machine learning network for model training to obtain a trained LightGBM model; S3. Input the training samples into the Meta Learner deep learning network, and train the model in regression mode to obtain a trained Meta Learner model. The Meta Learner deep learning network model uses the trained TabNet model and LightGBM model as base models, and uses Logistic regression as a meta-model for meta-model training. S4. Input the validation set into the trained Meta Learner model, and evaluate the identification and calibration through cross-validation, wherein the evaluation of identification includes accuracy, precision and recall, so as to obtain the aggregation model for diagnosing the risk of urinary tract infection after pyeloplasty in children.

2. The method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to claim 1, characterized in that: The cases include the following clinical characteristics: 1) operation time; 2) postoperative drainage method: pyelostomy or DJ tube; 3) age; 4) weight; 5) height; 6) BMI; 7) total protein Tp; 8) albumin Alb; 9) globulin Glb; 10) blood uric acid Ua; 11) blood creatinine Cr; 12) Blood urea nitrogen Bun; 13) Glomerular filtration rate eGFR; 14) Neutrophil ratio N%; 15) Lymphocyte ratio L%; 16) Renal pelvic anteroposterior diameter APD; 17) Renal parenchymal thickness.

3. The method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to claim 1, characterized in that: In the TabNet deep learning network in step S2: the data flows through the batch normalization BN layer and then enters the feature converter, the feature converter includes n gated linear unit GLU blocks, each of the GLU blocks consists of a fully connected FC layer, a BN layer and a GLU activation module; Optionally, the feature converter includes four gated linear unit (GLU) blocks, two of which are shared and the other two are independently set.

4. The method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to claim 3, characterized in that: After the feature converter processes the batch normalized features, it is output to the attention converter through the segmentation layer in step i; The attention converter includes a fully connected FC layer, a BN layer, a prior scale layer, and a sparsemax layer. The attention converter receives input from the segmentation layer, which is then processed by the fully connected FC layer and the BN layer, and then passed through the prior scale layer. The prior scale layer summarizes the feature magnitudes in the previous decision steps, expressed as: Where γ represents the relaxation parameter; The attention converter calculates a learnable mask vector M[i] based on the feature a[i-1] processed in the previous step for sparse feature selection: M[i]=sparsemax(P[i-1]·h i (a[i-1])) Among them, P[i-1] is the prior scale term, h i () is a trainable function that encapsulates the fully connected FC layer and BN layer. The sparsemax layer normalizes the coefficients to enforce sparse feature selection; In order to further regulate the feature sparsity, the sparsity regularization term L in the form of entropy is adopted sparse : The mask vector is fed into the feature transformer, which processes the filtered features and separates them into two outputs: [d[i],a[i]]=f i (M[i]·f) Where d[i] represents the output of the decision step and a[i] provides information for subsequent steps.

5. The method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to claim 1, characterized in that: The screening of historical cases in step S1 must simultaneously meet the following requirements: 1) unilateral hydronephrosis undergoing pyeloplasty; 2) meeting the diagnosis of ureteropelvic junction stenosis; 3) complete clinical feature data; The exclusion criteria for cases were any of the following: 1) bilateral hydronephrosis; 2) no conventional anti-infection regimen; 3) full ureteral dilatation; 4) duplicated kidney; 5) combined vesicoureteral reflux; 6) DRF of the affected kidney before surgery < 10%; 7) pediatric bladder and rectal dysfunction (BBD); 8) incomplete clinical feature data.

6. The method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to claim 1, characterized in that: The urinary tract infection risk diagnosis aggregation model after pyeloplasty in children can output the prediction result of whether the case to be predicted will develop clinical urinary tract infection within 6 months after the operation.

7. The method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to claim 1, characterized in that: In step S4, identification and calibration are evaluated by 5-fold cross validation, where: Accuracy Accuracy Recall TP, TN, FP, and FN stand for true positive, true negative, false positive, and false negative, respectively; P and N stand for positive and negative, respectively.

8. A urinary tract infection risk diagnosis system for children after pyeloplasty based on an aggregate model, characterized in that: The system is obtained based on the method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children according to any one of claims 1 to 7, and the system comprises: A data input module, wherein the data input module is used to input clinical characteristic data of the case; A data processing module, wherein the data processing module is used to execute the constructed aggregation model; A data output module is used to output the diagnosis results of urinary tract infection risk after pyeloplasty in children.

9. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children are implemented as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the steps of the method for constructing a risk aggregation prediction model for urinary tract infection after pyeloplasty in children as described in any one of claims 1 to 7.

Images