Can AI and predictive models accurately predict stone-free status? a systematic review and meta-analysis.

OBJECTIVES

The emergence of artificial intelligence (AI) and predictive modeling offers prospects for clinical, anatomical, and imaging factor combination, like radiomics, to help with stone-free status (SFS) estimation and peroperative decision-making. The goal of this study was, therefore, to define the present performance range, determine sources of heterogeneity, and determine methodological practices permitting reliable implementation by varied circumstances.

METHODS

We searched six bibliographic databases through 19 September 2025. Studies deriving or validating AI/predictive models for SFS after ureteroscopy were eligible. Independent dual screening, duplicate data extraction, and risk-of-bias consideration using QUADAS-AI were conducted.

RESULTS

Five retrospective cohorts were included. Modeling approaches encompassed multivariable logistic regression, regularized/radiomics pipelines, gradient boosting, and ensembles. SFS definitions ranged from <2 mm residual (day-1 to 3 months) to ≤5 mm residual (1 month), determined by plain radiography, ultrasound, and/or CT. The pooled ratio-scale effect for stone size per 1 mm increase was 1.26 (95% CI 0.91-1.76; τ² ≈ 0.055; Q = 18.52; I² = 94.6%; prediction interval 0.03-49.45). Hydronephrosis (moderate-severe vs. mild/none) showed a pooled RR 2.72 (95% CI 0.96-7.72; τ² ≈ 0.821; Q = 65.40; I² = 96.9%; prediction interval 0.03-249.87). As continuous contrasts, stone size was larger in the non-stone-free group (SMD 1.36, 95% CI 0.85-1.86; τ² ≈ 0.096; I² = 72.9%; prediction interval -3.77 to 6.48), and HU was higher (SMD 0.64, 95% CI 0.39-0.90; τ² ≈ 0; Q = 0.73; I² = 0%; prediction interval -0.99 to 2.27).

CONCLUSIONS

Across studies evaluating AI and predictive models for ureteroscopy, discrimination was generally acceptable to excellent, and performance appeared highest in models integrating radiomics with anatomic/clinical descriptors. However, the degree of between-study heterogeneity (population mix, outcome definitions, imaging protocols, thresholds, and follow-up windows) was sufficiently large that pooled quantitative estimates should be considered clinically uninterpretable.