Çoban F. et al., 2026: Risk factors and a predictive model for steinstrasse formation after shock wave lithotripsy in pediatric urolithiasis.

Çoban F, Kutlu H, Akbaba KT, Deniz ME, Kalyenci B.
Urolithiasis. 2026 Apr 10;54(1):86. doi: 10.1007/s00240-026-01975-6

Abstract

Steinstrasse is an important complication following extracorporeal shock wave lithotripsy; however, predictive models for risk factors in pediatric populations remain limited. This study aimed to develop and externally validate an explainable machine learning model for predicting steinstrasse formation after ESWL in children. This two-center retrospective study included 807 pediatric patients (< 18 years) who underwent ESWL for renal stones (2015–2025). The internal cohort comprised 685 patients (Adıyaman), while the external validation cohort consisted of 122 patients (Adana). Twenty-six preoperative variables were evaluated. Optimal predictors were identified through consensus-based feature selection using eight algorithms. Thirteen machine learning algorithms were compared using nested cross-validation. Model interpretability was assessed through SHAP, LIME, and partial dependence analyses. The overall steinstrasse incidence was 10.5% (85/807). Nine predictors were selected: S.T.O.N.E. Score, Triple-D Score, age, stone perimeter, stone depth, Hounsfield unit, stone volume, distal ureteral diameter, and skin-to-stone distance. XGBoost demonstrated optimal external validation performance (AUC = 0.949, sensitivity = 90%, specificity = 99.1%, MCC = 0.891) and exhibited excellent generalizability (ΔAUC = 0.001). Explainable artificial intelligence analyses identified age (mean|SHAP|=1.965) and distal ureteral diameter (mean|SHAP|=1.575) as dominant predictors with significant interaction (SHAP = 0.458). Decision curve analysis confirmed clinical utility across a threshold range of 2%–98%. This externally validated model accurately predicts steinstrasse formation using routine preoperative parameters. Integration of patient age, ureteral anatomy, and stone characteristics may guide decisions between ESWL, prophylactic stenting, or alternative endoscopic approaches in pediatric patients.

Comment Peter Alken

This paper is all mathematics in form of statistics and machine learning. The results present a convincing probability. For a non-machine-learning native like me this is a knock-out.

The learning model was developed on the data of 685 patients and subsequently evaluated externally with the data of 122 patients of a different hospital. There is no information on the machines used, details of the procedures or the persons performing the treatment. That would be classical reasons to turn down such a manuscript because the “essential” data to judge on the results and their comparability are missing. 

The counterstrike states “diversity was intentionally selected to evaluate the generalizability of the model across different populations and clinical practices.”

The knock-out  comes with the results and follows in the discussion : “Although procedure-specific variables such as energy level, shock wave number, and session number applied during ESWL were available in the dataset, these parameters were deliberately not included in the model development process due to their potential to cause data leakage by being influenced by clinical decisions and early treatment outcomes; this choice aimed to ensure the model can be reliably used in actual preoperative decision-making scenarios.

Weather reports are also probability-reports. We rely to a large extent on the sunshine they  predict or seem to promise. If it nevertheless rains and ruins the lady’s dress, we may blame the provider but – we, the users are responsible.

Peter Alken