Commentary on “Clinical management implications from 33 579 urinary stones: Novel patterns in composition, comorbidities, seasonal variation, and machine learning-based urosepsis prediction”

Dear Editor, We read with great interest the article by Chang et al[1]. The authors are to be congratulated for assembling a large stone-analysis dataset and for exploring comorbidity patterns, temporal trends, and a machine learning model for urosepsis risk. These efforts address important questions in contemporary stone management. We would, however, like to raise several methodological points that may help contextualize their conclusions. This study is compliant with the TITAN Guidelines 2025 – governing declaration and use of AI[2]. First, the stone-level analysis unit has important implications for interpreting comorbidity patterns. Recurrent episodes and multiple stones removed during one procedure were analyzed as separate observations, while demographic and clinical data were recorded at the time of stone analysis. Thus, individual patients may contribute multiple records with identical age, sex, and comorbidities, yet prevalences of diabetes, gout, cardiovascular disease, and urinary tract infection are presented as if each record represented a unique individual. This clustered structure violates the independence assumption of standard χ2 tests and logistic regression and can lead to underestimated uncertainty[3]. Explicitly framing results as stone-level associations and, where feasible, providing cluster-adjusted analyses would help readers judge the robustness of the reported gender-, seasonal-, and location-specific trends. Second, extensive hypothesis testing across many comorbidities, seasons, age strata, and stone locations raises concerns about multiplicity and clinical relevance. Although the authors state that they used the Benjamini–Hochberg procedure to control the false discovery rate, the main tables report numerous unadjusted P values, and it is unclear which associations remain significant after correction. In such a high-dimensional setting, even small absolute differences are likely to become statistically significant without being clinically meaningful, and multiple hypothesis testing requires particular caution[4]. Concentrating on prespecified key comparisons, reporting effect sizes with confidence intervals, and clearly indicating which findings remain significant after adjustment would better highlight the most actionable signals. Third, the machine learning-based urosepsis prediction model warrants cautious interpretation. Among 33 579 stone samples, only 108 urosepsis events were observed, yet many candidate predictors were screened using Lasso, XGBoost, and logistic regression, followed by stepwise selection for the final model. This low events-per-variable ratio, combined with repeated feature selection on the same data and a single 6:4 random split, increases the risk of overfitting and optimistic performance estimates, even though the reported validation AUC of 0.802 appears favourable. Moreover, only discrimination metrics are presented; calibration and decision-curve analyses are not reported, limiting assessment of probability accuracy and net clinical benefit, which are emphasized in contemporary guidance on prediction model development[5,6]. More comprehensive evaluation of calibration and clinical utility, alongside external or temporal validation, would be desirable before considering clinical implementation. Fourth, there appear to be inconsistencies in the reporting of effect estimates from the final logistic regression model. The authors state that stepwise regression identified six significant predictors, including “male gender (OR = −0.726)” and “absence of UTI (OR = −2.010).” As odds ratios cannot be negative, these values presumably represent regression coefficients on the log-odds scale rather than odds ratios. Transparent reporting guidelines recommend clearly distinguishing coefficients from derived odds ratios and accompanying them with 95% confidence intervals[6]. Clarifying whether the reported numbers are β coefficients or odds ratios [exp(β)], and providing confidence intervals, would improve interpretability and reproducibility, especially for a proposed risk stratification tool for a rare but serious complication. Finally, several broader claims regarding the impact of this work might be tempered in light of these limitations. The study is described as “Asia’s largest” of its kind and the “first machine learning-based urosepsis prediction model,” and is said to “challenge current guideline assumptions.” While the dataset is large and the integrative approach is innovative, the single-center, stone-level design, small number of urosepsis events, and lack of external validation may constrain generalizability and immediate practice-changing potential. Framing the work as an exploratory contribution that generates hypotheses and provides a framework for future multicenter, patient-level and prospectively validated studies would be more consistent with current expectations for prediction model research In summary, Chang et al provide valuable epidemiological insights into stone composition and propose a framework for infection risk prediction in urolithiasis. Addressing issues of analysis unit, multiplicity, model complexity, and reporting clarity through cluster-adjusted analyses, more stringent validation strategies and refined presentation of effect estimates could enhance the robustness and clinical applicability of their conclusions. We thank the authors for their work and look forward to future investigations that build on this foundation to optimize risk stratification and perioperative management in stone disease.