A non-contrast CT radiomics-clinical machine learning model for predicting spontaneous ureteral stone passage: a retrospective cohort study

Decisions about observation or early intervention for ureteral stones are still driven mainly by stone size, location and clinical judgement. We developed and internally validated a model combining non-contrast computed tomography (NCCT) radiomics with routine clinical variables to estimate the probability of spontaneous stone passage within 4 weeks under conservative management. We also explicitly evaluated the availability of medical expulsive therapy (MET) and medication data, performed sensitivity analyses accounting for crossovers to intervention, and explored incremental value in the intermediate stone-size subgroup. We retrospectively analysed 274 patients with solitary ureteral stones measuring 10 mm or smaller who underwent conservative treatment between January 2021 and August 2025. Time zero was defined as the date of the baseline NCCT at which conservative management was decided. Patients were grouped according to 4-week passage status. From manually segmented NCCT stone regions, 851 radiomic features were extracted. Reproducible and non-redundant features were screened by observer-agreement testing, Spearman correlation analysis and least absolute shrinkage and selection operator regression. Clinical, radiomics and combined models were built with elastic net regression, support vector machine, random forest and LightGBM. Hyperparameters were tuned by 5-fold cross-validation, and the final models were assessed by bootstrap internal validation, calibration analysis, decision curve analysis and the DeLong test. Eight radiomic features were retained for modelling. In bootstrap validation, the combined random forest model achieved an AUC of 0.928 (95% CI 0.896–0.959), which was higher than that of the clinical LightGBM model (AUC 0.837, 95% CI 0.789–0.884; P < 0.001) and the radiomics random forest model (AUC 0.856, 95% CI 0.809–0.903; P = 0.029). The combined model showed acceptable calibration (Hosmer–Lemeshow P = 0.591) and provided net benefit across threshold probabilities of 5%-99%. SHAP analysis identified CT texture features and ureteral wall thickness as important contributors to prediction. In the sensitivity analysis counting the 18 crossovers as non-passage, the combined model yielded an AUC of 0.914 (95% CI 0.878–0.949). In the exploratory 5–8 mm proximal/middle ureter subgroup ( n = 55), the combined model achieved an AUC of 0.881 (95% CI 0.782–0.979). The ablation analysis showed that adding texture features to clinical variables plus size/HU/UWT changed the AUC from 0.901 (95% CI 0.864–0.937) to 0.928 (95% CI 0.896–0.959); the simple size-location benchmark achieved an AUC of 0.812 (95% CI 0.758–0.866). MET prescription data were available in 186 (67.9%) patients and missing/unknown in 88 (32.1%); among patients with available data, MET use did not differ significantly between passage and non-passage groups (68.2% vs. 61.0%, P = 0.317). In this single-centre cohort, adding NCCT radiomic features to clinical variables improved prediction of spontaneous ureteral stone passage in patients initially managed conservatively. However, the model remains hypothesis-generating, is limited by unmeasured conservative-management factors including MET, and requires external validation before clinical use. The model may provide preliminary decision support but is not yet ready for clinical adoption.