Machine learning based prediction for the recurrence of uretero-pelvic junction obstruction in pediatric patients undergoing open Anderson-Hynes dismembered pyeloplasty

Authors

Abstract

Aim To create a multivariate prediction model based on machine learning to find predictors of recurrence after Anderson–Hynes dismembered pyeloplasty.
Methods Patients younger than 15 who underwent primary open Anderson-Hynes Dismembered Pyeloplasty between 2011 and 2020 were evaluated. Logistic regression, support vector machine, and random forest were used to train a classifier for predicting recurrence, and the feature importance analysis methods were performed to understand which predictors have more weight in the models.
Results Of the patients, 134 were boys, and 43 were girls, with a mean age of 30.4 (1-168) months. Recurrence developed in 15 / 177 (8.4%) of the patients. Postoperative anteroposterior renal pelvis diameter and intraoperative urine aspiration volume were the strongest predictors of recurrence. The Random Forest model achieved the best accuracy (AUC = 0.94) in predicting recurrence in patients under 15.
Conclusion To our knowledge, this is the first study to investigate whether the amount of urine aspirated from the intraoperative renal pelvis is predictive of recurrence. We found that the probability of recurrence of ureteropelvic junction obstruction increased as the amount of urine aspirated from the intraoperative renal pelvis increased.

Keywords

Introduction

The success rate of Anderson-Hynes dismembered pyeloplasty is between 90-100% ^1,2. However, recurrence may develop in 5-10% of patients ^3,4. In studies conducted to predict which patients may develop recurrence, factors such as the type of incision, not performing intraoperative retrograde pyelography , the presence of preoperative diversion, the presence of early postoperative complications , the history of endopyelotomy , and urine leakage from the anastomosis in the early postoperative period were found to be effective on recurrence ^3,4,5,6.
In recent years, studies based on machine learning or deep learning have gained momentum in different areas of urology ^7,8. In the pediatric urology field, there are studies based on machine learning or deep learning, although it is much more limited than in other fields of urology. AI applications in pediatric urology include vesicoureteral reflux quantification, hypospadias scoring, UPJ obstruction diagnosis enhancement, and UTI prediction ^{9,10,11,12,13,14}.
To our knowledge, only one study in the literature investigated the predictive factors for recurrence in patients who underwent pyeloplasty using machine learning methods. In this study by Drysdale et al., many features such as the anterior- posterior diameter of the renal pelvis in the first and second postoperative follow-up, the use of drains, the use of narcotic analgesics, the use of ureteral catheters/double-j stents, etc., were investigated, and a machine learning model was trained ¹⁵.
Accurate prediction of recurrence using machine learning models could enable early diagnosis and intervention. We aimed to develop ML models (SVM and random forests) to identify feature effects and predict recurrence with acceptable accuracy in pyeloplasty patients. We present this article in accordance with the TRIPOD reporting checklist.

Materials and Methods

Two hundred five patients younger than 15 who underwent primary Anderson-Hynes Dismembered pyeloplasty between September 2011 and March 2020 were evaluated. Our study is a single-center study. This center is a reference pediatric urology clinic of a university hospital. Only prospectively collected pediatric urology data, compiled by urologists, hospital automation systems, and patient files, were used in the evaluation.
Pyeloplasty was performed following EAU guidelines for patients with ureteropelvic junction stenosis based on progressive hydronephrosis, decreased renal function, or symptoms [available at: https://uroweb.org/guidelines/ paediatric-urology]. All patients underwent Anderson-Hynes dismembered pyeloplasty with dorsal lumbotomy by a single pediatric urologist. Trans-anastomotic diversion was performed using ureteral catheters or double-J stents. Complications were classified according to Clavien-Dindo criteria ¹⁶. Postoperative follow-up included ultrasonography at 4-6 weeks and radionuclide imaging at 3-6 months (after stent removal). Recurrence was defined as impaired drainage on scintigraphy or progressive hydronephrosis on ultrasonography. Key measured parameters included intraoperative pelvic urine volume (cc), excised segment length (cm), preoperative renal pelvis A-P diameter (mm) by ultrasonography, differential renal function (%) by diuretic scintigraphy, and postoperative A-P diameter (mm). Patient features and outcomes are detailed in Table 1. The predicted outputs are the patients’ state of cured vs recurrence, which gives us a binary classification setting. This can be posed as a detection problem in which the model tries to “detect” cases of recurrence. Posing the data modeling problem as detection enables us to visualize and validate results for both types of errors: false alarms in estimating recurrence for a cured patient versus missing a recurrence in a patient, and estimating a patient as cured. Receiver Operating Characteristics (ROC) curves give us a robust and transparent evaluation method, and the Area under the Curve (AUC) is a single success metric that we can use to compare different machine learning models.
We trained logistic regression (L2 penalty with C = 1), radial basis support vector machine (C = 1, gamma = 2), decision tree (gini split, max depth = 5, min samples internal = 2, min samples leaf = 1), random forest (max depth = 5, number of estimators = 10) on the dataset (70% train, 30% test split). All available features were used in training for the first set of experiments to understand the importance of features and the maximum performance we can achieve with the dataset. For finding the best performing model, a random grid search on random forest meta-parameters was performed using cross-validation (number of estimators, max depth, min samples internal, min samples leaf). The random grid search was done in a coarse- to-fine fashion by manually starting the optimization multiple times, starting from the best meta-parameters from the previous run. This “best” fitting model is used for investigating the importance of features on prediction. Permutation importance procedure is used to estimate the effect of each one of our features on prediction performance ¹⁷. The Scikit- learn library was used for he experiments ¹⁸.
Ethical Approval
This study was approved by the Ethics Committee of Çukurova University (Date: 2023-05-05, No: 133/38).

Results

After ten patients with additional pathology (ureterovesical junction stenosis, vesicoureteral reflux, double collecting system, ectopic kidney) and 18 patients operated on laparoscopically were excluded from the study, 177 patients under 15 years of age who underwent primary Anderson-Hynes Dismembered open pyeloplasty were included in the study ⁴. 134 (75.7%) of the patients were male, and 43 (24.3%) were female. The average operation time (time from incision to end of skin suturing) was 120.86 (60-300) minutes, and the average length of stay was 95.15 (48-420) hours. The number of patients who developed recurrence was 15 (8.4%), and complications were observed in 18 (10.1%) patients. Details are in Table 2.
Postoperative complications were observed in eighteen (10.1%) patients. Postoperative urinary tract infection (Clavien-Dindo grade 2) developed in ten (5.6%) patients, managed with empiric antibiotic therapy. One patient was re-admitted due to hematuria and managed conservatively with catheterization (Clavien-Dindo grade 2). Urine leakage from the drain occurred in five patients (2.82%), all managed with double J stent placement (Clavien-Dindo grade 3b). One patient developed postoperative ileus requiring nasogastric decompression (Clavien-Dindo grade 3a). One patient experienced colic pain after ureteral catheter removal, managed with double J stent placement (Clavien-Dindo grade 3b). No mortality occurred. Results regarding complications are listed in Table 2.
Among a large set of machine learning models tested (only 4 models shown in Figure 1), random forest performed best in detecting recurrence in patients (AUC 0.94). Logistic regression, as the simplest model, performed poorly, and radial basis SVM is the second best. Even for these non-optimized models, the overall trend in ROC curves suggests that it is possible to reach zero miss (i.e., true positive rate of 1.0) with a small false alarm ratio (i.e., false positive rate of around 0.15).
The best random forest model trained via a random grid search procedure (only on the training dataset) shows significant improvement on detection performance (Figure 2, AUC 0.97 on test data). This model is capable of perfect recall (no misses on recurrence in patients) with only 5% false alarms (incorrectly assigning recurrence for cured patients).
Feature importance analysis on the best random forest model suggests the 3 most important metrics in the predictions of recurrence: intraoperatively measured pelvic volume, complication score, and postoperative antero-posterior pelvic diameter measured with ultrasonography. Pelvic volume is the most important parameter in the permutation-based importance analysis, which indicates that this feature is the major dimension in the model.
To further investigate these three feature dimensions and create a transparent decision process, we fitted a single decision tree on the whole dataset with only the most important 3 dimensions above, and visualized the tree (Figure 3).
Again, a random grid search was used to maximize the performance of the model. This decision tree is expected to give worse predictive performance (AUC of 0.97, Figure 2) but it is capable of showcasing how each of these 3 features is used in the decision process of the model. Pelvic volume sits at the highest node in the tree, deciding the probability of recurrence based on a threshold (72.5). If pelvic volume is larger than this threshold, it is definitely a recurring patient (right branch of the tree). Below that pelvic volume threshold (left branch of the tree), first A-P diameter (threshold 24.5) then complication scores (threshold 1.5) are used to decide whether there is a chance of recurring patient. It is a simple yet striking decision process illustrating the importance of these 3 feature dimensions (Figure 3).

Discussion

In our study, we found that the intraoperatively aspirated urine volume from the renal pelvis > 72.5ml, the postoperative A-P diameter > 24.5mm, and the presence of postoperative complications were strong predictors of recurrence. These discoveries were made by training a state-of-the-art machine learning model (random forest) on our relatively large dataset. Train/test split of the dataset and cross-validation procedure ensure the generalization performance of our model and the validity of our feature analyses. Overall, the trained model gave very satisfactory prediction capability for recurrence (zero misses with only 5% false alarms).
We found that intraoperative urine volume aspirated from the renal pelvis > 72.5cc was most associated with recurrence. To our knowledge, this is the first study investigating this relationship. This discovery is validated by both random forest feature importance analysis and decision tree analysis. When we trained a random forest model without this feature, predictive performance significantly dropped, with false alarm rates increasing from 5% to 30% while maintaining zero missed recurrences. This finding suggests that surgeons should consider these patients at higher risk for recurrence when they aspirate more than 72.5 ml of urine during surgery and adjust their follow-up protocols accordingly (e.g., more frequent ultrasonography checks, early scintigraphy follow-up).
Although no studies have performed intraoperative renal volume/ pelvic volume measurement, imaging studies demonstrate its clinical significance. Danuser et al. found that pyelocalyceal volume > 50 ml reduced endopyelolithotomy success ¹⁹. Elbaset et al. showed that renal pelvis volume ≥ 50 mm3 correlated with supranormal differential renal function ²⁰. Bai et al. demonstrated that renal volume measurements could predict surgical recommendations in pediatric hydronephrosis ²¹. We recommend routine intraoperative pelvic urine volume measurement and documentation. Future prospective studies should evaluate ultrasonographic volume measurements as predictors.
In the study conducted by Drysdale et al., which used the ML model to predict recurrence, postoperative A-P diameter was an essential predictor for recurrence. Our study qualitatively confirms this finding with a different dataset and different models, although it is not possible to directly compare quantitative results. On the individualized application web page published as open access, the parameters that need to be entered for recurrence prediction include the A-P diameters at the first postoperative and second controls ¹⁵. We found that postoperative renal pelvis A-P diameter > 24.5 mm was a strong predictor for recurrence, in accordance with the study by Drysdale et al.
Recent studies have demonstrated the effectiveness of machine learning algorithms in predicting surgical intervention requirements for pediatric UPJ obstruction cases. Alici et al. achieved remarkable results with XGBClassifier, obtaining ROC- AUC values of 0.977 and accuracy rates of 95.4% in predicting surgical necessity ²². Their findings similarly identified scintigraphic obstruction presence and ultrasonographic parameters as key predictive factors. The integration of non-invasive imaging with machine learning represents a significant advancement in UPJ obstruction diagnosis. Mohan et al. emphasized the importance of real-time ultrasound imaging combined with artificial intelligence algorithms for early detection, highlighting the potential for reducing patient discomfort associated with traditional invasive diagnostic methods ²³.
We investigated whether complications are reliable predictive features for recurrence due to pyeloplasty’s reconstructive nature. Our machine learning experiments confirm that complications are important predictive features, with both random forest feature importance analysis and decision tree training demonstrating their critical role in predicting recurrence. Clinically, we performed redo pyeloplasty in two (40%) of five patients who developed urine leakage, encountering highly fibrotic structures at the ureteropelvic junction. Additionally, recurrence developed in five (50%) of ten patients with postoperative urinary tract infection, suggesting that complications may impair anastomotic healing. Based on these statistical results and surgical observations, careful follow-up is warranted for patients developing postoperative complications, particularly urine leakage and urinary tract infections. We suggest this relationship should be investigated in future prospective studies.
The threshold values determined in our decision tree (such as pelvic volume ≤ 72.5 ml) are optimal cutoff points mathematically determined by the model, but it should be noted that such sharp boundaries may not be found in clinical practice. While these threshold values are important guidelines for physicians in determining risk, they should be considered as part of a gradual risk spectrum, not as definitive ‘right’ or ‘wrong’ limits.

Limitations

This study has several important limitations that should be acknowledged when interpreting our findings. First, the retrospective single-center design limits generalizability and introduces potential selection bias. Our patient population may not be representative of other institutions with different referral patterns, surgical practices, or follow-up protocols. Second, we could not perform external validation with independent datasets, which is crucial for assessing the true predictive performance of machine learning models before clinical implementation. Without external validation, there is a risk of overfitting to our specific dataset. Third, our study was restricted to open dismembered pyeloplasty cases, excluding minimally invasive approaches such as laparoscopic or robotic techniques that are increasingly utilized in contemporary pediatric urology practice. The predictive factors and recurrence patterns may differ substantially between surgical approaches. Fourth, our model relied primarily on intraoperative measurements without systematic preoperative volumetric assessments. The absence of standardized preoperative imaging protocols limits our ability to establish correlations between preoperative findings and intraoperative observations, representing a missed opportunity for earlier risk stratification. Fifth, the relatively small number of recurrence cases (n = 15, 8.4%) creates an imbalanced dataset that limits statistical power and makes it difficult to detect patterns associated with rare complications. This class imbalance may affect model sensitivity and positive predictive value. Finally, our median follow-up period of 36 months may not capture delayed failures that can manifest years after surgery, potentially underestimating true recurrence rates. These limitations highlight the need for prospective multicenter studies with larger sample sizes, external validation, inclusion of minimally invasive approaches, and systematic preoperative imaging analysis before clinical implementation.

Conclusion

Machine learning-based predictive models are gaining momentum across medicine. Our models accurately predicted recurrence in pediatric patients undergoing open Anderson- Hynes pyeloplasty. Future multicenter, prospective studies including open, laparoscopic, and robot-assisted methods should validate our findings. An open question is whether intraoperative urine aspiration volume correlates with preoperative ultrasonographic renal pelvic volume, which could be investigated in prospective studies to establish the relationship between ultrasonographic measurements and recurrence.

References

1. Klingler HC, Remzi M, Janetschek G, Kratzik C, Marberger MJ. Comparison of open versus laparoscopic pyeloplasty techniques in treatment of uretero-pelvic junction obstruction. Eur Urol. 2003;44(3):340-5. doi:10.1016/s0302-2838(03)00297-5.
   PubMed Google Scholar
2. Salem YH, Majd M, Rushton HG, Belman AB. Outcome analysis of pediatric pyeloplasty as a function of patient age, presentation and differential renal function. J Urol. 1995;154(5):1889-93.
   PubMed Google Scholar
3. Braga LHP, Lorenzo AJ, Bägli DJ, et al. Risk factors for recurrent ureteropelvic junction obstruction after open pyeloplasty in a large pediatric cohort. J Urol. 2008;180(4 Suppl):1684-7. doi:10.1016/j.juro.2008.03.086.
   PubMed Google Scholar
4. Ceyhan E, Ileri F, Ceylan T, Aydin AM, Dogan HS, Tekgul S. Predictors of recurrence and complications in pediatric pyeloplasty. Urology. 2019;126:187-91. doi:10.1016/j.urology.2019.01.014.
   PubMed Google Scholar
5. Lucas SM, Sundaram CP, Wolf JS, et al. Factors that impact the outcome of minimally invasive pyeloplasty: Results of the multi-institutional laparoscopic and robotic pyeloplasty collaborative group. J Urol. 2012;187(2):522-7. doi:10.1016/j. juro.2011.09.158.
   PubMed Google Scholar
6. Seo IY, Oh TH, Lee JW. Long-term follow-up results of laparoscopic pyeloplasty. Korean J Urol. 2014;55(10):656-9. doi:10.4111/kju.2014.55.10.656.
   PubMed Google Scholar
7. Chen J, Remulla D, Nguyen JH, et al. Current status of artificial intelligence applications in urology and their potential to influence clinical practice. BJU Int. 2019;124(4):567-77. doi:10.1111/bju.14852.
   PubMed Google Scholar
8. Suarez-Ibarrola R, Hein S, Reis G, Gratzke C, Miernik A. Current and future applications of machine and deep learning in urology: A review of the literature on urolithiasis, renal cell carcinoma, and bladder and prostate cancer. World J Urol. 2020;38(10):2329-47. doi:10.1007/s00345-019-03000-5.
   PubMed Google Scholar
9. Kabir S, Pippi Salle JL, Chowdhury MEH, Abbas TO. Quantification of vesicoureteral reflux using machine learning. J Pediatr Urol. 2024;20(2):257-64. doi:10.1016/j.jpurol.2023.10.030.
   PubMed Google Scholar
10. Abbas TO, AbdelMoniem M, Khalil IA, Abrar Hossain MS, Chowdhury MEH. Deep learning based automated quantification of urethral plate characteristics using the plate objective scoring tool (POST). J Pediatr Urol. 2023;19(4):373.e1- 373.e9. doi:10.1016/j.jpurol.2023.03.033.
    PubMed Google Scholar
11. Blum ES, Porras AR, Biggs E, et al. Early detection of ureteropelvic junction obstruction using signal analysis and machine learning: A dynamic solution to a dynamic problem. J Urol. 2018;199(3):847-52. doi:10.1016/j.juro.2017.09.147.
    PubMed Google Scholar
12. Arlen AM, Alexander SE, Wald M, Cooper CS. Computer model predicting breakthrough febrile urinary tract infection in children with primary vesicoureteral reflux. J Pediatr Urol. 2016;12(5):288.e1-288.e5. doi:10.1016/j.jpurol.2016.03.005.
    PubMed Google Scholar
13. Zheng Q, Furth SL, Tasian GE, Fan Y. Computer-aided diagnosis of congenital abnormalities of the kidney and urinary tract in children based on ultrasound imaging data by integrating texture image features and deep transfer learning image features. J Pediatr Urol. 2019;15(1):75.e1-75.e7. doi:10.1016/j. jpurol.2018.10.020.
    PubMed Google Scholar
14. Abdovic S, Cuk M, Cekada N, et al. Predicting posterior urethral obstruction in boys with lower urinary tract symptoms using deep artificial neural network. World J Urol. 2019;37(9):1973-9. doi:10.1007/s00345-018-2588-9.
    PubMed Google Scholar
15. Drysdale E, Khondker A, Kim JK, et al. Personalized application of machine learning algorithms to identify pediatric patients at risk for recurrent ureteropelvic junction obstruction after dismembered pyeloplasty. World J Urol. 2022;40(2):593- 9. doi:10.1007/s00345-021-03879-z.
    PubMed Google Scholar
16. Dindo D, Demartines N, Clavien PA. Classification of surgical complications: a new proposal with evaluation in a cohort of 6336 patients and results of a survey. Ann Surg. 2004;240(2):205-13. doi:10.1097/01.sla.0000133083.54934.ae.
    PubMed Google Scholar
17. Breiman L. Random Forests. Machine Learning. 2001;45(1):5-32. doi:10.1023/A:1010950718922.
    PubMed Google Scholar
18. Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine learning in python. JMLR. 2011;12(85):2825-30.
    PubMed Google Scholar
19. Danuser H, Ackermann DK, Böhlen D, Studer UE. Endopyelotomy for primary ureteropelvic junction obstruction: Risk factors determine the success rate. J Urol. 1998;159(1):56-61. doi:10.1016/s0022-5347(01)64011-4.
    PubMed Google Scholar
20. Elbaset MA, Ezzat O, Elgamal M, et al. Supranormal differential renal function in adults with ureteropelvic junction obstruction: Does it really exist? Indian J Urol. 2020;36(3):205-11. doi:10.4103/iju.IJU_109_20.
    PubMed Google Scholar
21. Bai K, Hou Y, Zhang Z, et al. Ability of volume measures of hydronephrosis to predict need for surgery and evaluate renal function in children with ureteropelvic junction obstruction. Int J Urol. 2022;29(3):235-41. doi:10.1111/iju.14755.
    PubMed Google Scholar
22. Alici CA, Tokar B. Prediction of surgical necessity in children with ureteropelvic junction obstruction using machine learning. Ir J Med Sci. 2025;194(2):583-90. doi:10.1007/s11845-025-03895-7.
    PubMed Google Scholar
23. Mohan D, Jain V, Dayalani AN. Pelvic Uretero Junction Obstruction Detection Using Machine Learning Algorithm. 2025 Fifth International Conference on Advances in Electrical, Computing, Communication and Sustainable Technologies (ICAECT), Bhilai, India; 2025.p.1-7.
    PubMed Google Scholar

Declarations

Additional Information

Publisher’s Note
Bayrakol MP remains neutral with regard to jurisdictional and institutional claims.

Rights and Permissions

Creative Commons License

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). To view a copy of the license, visit https://creativecommons.org/licenses/by-nc/4.0/

View full license details →

About This Article