Letter to the Editor regarding: “Clinical management implications from 33, 579 urinary stones” Authors: Zuomin Wang, Qinwei Liu, Wangdong Deng Published: 2026 Dear Editor, We read with great interest the comprehensive study by Chang et al on urolithiasis patterns and urosepsis prediction in southern China[1]. While we commend the authors for analyzing 33 579 cases – the largest Asian cohort to date – we wish to highlight critical methodological concerns that may affect the clinical applicability of their findings, Transparency In the reporting of Artificial Intelligence – The TITAN guideline[2]. The authors acknowledge reduced case volumes during 2020–2022 but do not adequately address systematic selection bias introduced by the pandemic. COVID-19 likely delayed presentations of uncomplicated cases while prioritizing emergent conditions, fundamentally altering case-mix severity. The post-pandemic surge in calcium oxalate stones may reflect accumulated mild cases rather than true epidemiological shifts. Without stratified analyses by clinical presentation (emergent versus elective), surgical modality, and stone burden, the temporal comparability of these cohorts remains questionable, potentially undermining trend interpretations. Beyond temporal bias, the inclusion of “30+ comorbidities” raises concerns about ascertainment methodology. The authors do not specify whether diagnoses were standardized using validated criteria or extracted from heterogeneous medical records. Detection bias is particularly problematic – thyroid dysfunction rates likely reflect testing patterns in symptomatic patients rather than true prevalence. More critically, the temporal relationship between comorbidities and stone formation remains unclear. If conditions were diagnosed post-operatively during workup or hospitalization, their inclusion in preoperative risk models is methodologically inappropriate and clinically misleading. The claim of developing the “first machine learning-based urosepsis prediction model” requires more rigorous substantiation. No systematic literature search is presented to verify this novelty – similar models may exist in non-English publications or institutional databases. More concerning is the absence of comparative validation against established clinical tools (shown in Figure 1). How does their XGBoost model perform against qSOFA or SIRS criteria specifically in urolithiasis populations[3]? Without benchmarking against traditional logistic regression, the incremental value of algorithmic complexity remains unclear. Figure 1.Missing Benchmark Comparison: Clinical Scoring Systems Not Evaluated. The study reported discrimination metrics (AUC) for machine learning approaches (final optimized model 0.802, XGBoost 0.787) and logistic regression (0.786) but entirely omitted comparison with established clinical scoring systems routinely used for sepsis risk assessment (qSOFA, SIRS). Without head-to-head benchmarking, the claim of developing the “first” model and the clinical value of increased algorithmic complexity cannot be validated. Error bars represent 95% confidence intervals. Data extracted from study Figure 5 and Results text (n = 33,579 total cases; test set n = 13,432). Missing Benchmark Comparison: Clinical Scoring Systems Not Evaluated. The study reported discrimination metrics (AUC) for machine learning approaches (final optimized model 0.802, XGBoost 0.787) and logistic regression (0.786) but entirely omitted comparison with established clinical scoring systems routinely used for sepsis risk assessment (qSOFA, SIRS). Without head-to-head benchmarking, the claim of developing the “first” model and the clinical value of increased algorithmic complexity cannot be validated. Error bars represent 95% confidence intervals. Data extracted from study Figure 5 and Results text (n = 33,579 total cases; test set n = 13,432). Furthermore, while the reported AUC of 0.802 demonstrates acceptable discrimination, the authors provide no calibration assessment – a fundamental requirement for clinical prediction models. Discrimination and calibration are distinct properties: a model may rank patients correctly (high AUC) while systematically over- or under-predicting actual risk. A poorly calibrated model predicting 30% urosepsis probability when true incidence is 10% would trigger unnecessary prophylactic interventions, increasing antibiotic resistance and healthcare costs. Hosmer–Lemeshow tests, calibration plots, and decision curve analyses are essential to demonstrate clinical utility across different risk thresholds. Finally, the study performs hundreds of statistical comparisons across 30 + comorbidities, multiple stone types, gender, seasons, and age groups without mentioning multiple testing corrections. At the conventional α = 0.05 significance level, approximately 5% of reported associations likely represent false positives. The absence of Bonferroni[4], false discovery rate, or Benjamini-Hochberg adjustments makes it difficult to distinguish robust findings from statistical noise. These concerns do not diminish the study’s valuable contribution to Asian urolithiasis epidemiology but highlight methodological refinements necessary before clinical implementation. We recommend the authors provide: (1) sensitivity analyses excluding pandemic-period data or adjusting for clinical severity indicators; (2) explicit definitions and temporal verification of pre-operative comorbidities; (3) systematic head-to-head comparison with existing risk stratification tools and simpler statistical approaches; (4) comprehensive calibration metrics with proposed clinical decision thresholds; and (5) adjusted p-values clearly identifying associations robust to multiple testing correction.