We read with interest the Publish-Ahead-of-Print study by Wang et al, examining the combined prognostic contribution of the stress hyperglycemia ratio (SHR) and glycemic variability (GV) for mortality among patients with sepsis across normal glucose regulation (NGR), prediabetes, and diabetes (International Journal of Surgery; doi: 10.1097/JS9.0000000000003525)1. Leveraging MIMIC-IV v3.1 (n = 4838), the authors derived SHR from admission glucose and HbA1c, quantified GV as the coefficient of variation during ICU stay, compared conventional and tree-based models with SHAP explanations, and explored time-heterogeneous effects using landmark analysis. The pairing of phenotype-stratified design with two complementary dysglycemia domains is clinically coherent and deserves recognition. Because phenotype-aware thresholds translate directly into alert burden, staffing and hypoglycemia risk, implementation-ready reporting has immediate service value for surgical and critical-care teams. The study is well-positioned within the Sepsis-3 emphasis on heterogeneous host responses and organ dysfunction2. By normalizing acute glucose to chronic glycemia, SHR captures relative metabolic stress more faithfully than an absolute threshold, particularly in nondiabetic states3. If language tools assisted drafting or analysis, a brief AI use/nonuse note in line with contemporary transparency guidance (e.g., TITAN) would align the report with current editorial expectations without altering its scientific message4. Moreover, prior evidence indicates that the relation between dysglycemia domains and mortality differs by diabetic status, supporting the authors’ phenotype-aware interpretation and their reported 28-day risk patterns (e.g., NGR with high SHR/high GV; prediabetes with low SHR/high GV; different patterning in diabetes)5. Measurement governance. GV is sensitive to sampling intensity and timing; reporting per-patient glucose test counts per ICU-day, the monitoring route (capillary vs. venous), and hypoglycemia handling would quantify surveillance bias and enhance cross-ICU comparability. The manuscript already details GV definitions and tertiles; adding sampling distributions would complete the picture. Numerical calibration, reproducibility, and local deployment. Discrimination is moderate across strata (AUC roughly 0.75–0.78), with landmark analyses indicating clearer separation beyond ~day 101. For adoption, calibration-in-the-large (intercept) and calibration slope with 95% CIs should be reported by phenotype and time horizon, alongside Brier scores, enabling judgment of absolute-risk accuracy rather than rank ordering alone. A minimal local recalibration path (intercept → slope) and explicit citation of the exact MIMIC-IV version, accompanied by a compact code/configuration note (even pseudocode), would materially facilitate replication and safe transport across institutions6. Decision-readiness and early-phase dynamics – plus missing data practice. Because diabetic status modifies the mortality association of hyperglycemia and variability, prespecifying phenotype-specific action thresholds (e.g., 10% for intensified monitoring; 20% for escalation) and presenting threshold tables – sensitivity, specificity, PPV/NPV, and expected actions per 100 patients – would connect predictions to workload and stewardship; pairing with decision-curve analysis would render net-benefit trade-offs explicit across thresholds and phenotypes7. As separation strengthens after ~day 10, incorporating time-updated features (rolling means/slopes, time-varying covariates) may improve early-window discrimination while preserving model simplicity. When missingness is nontrivial, multiple imputation by chained equations should be executed within each resampling split and pooled using Rubin’s rules, with all preprocessing (imputation, scaling, thresholding) nested to avoid leakage; complement with bootstrap optimism correction and mild uniform shrinkage via the calibration slope so gains translate realistically in new settings8. In summary, Wang et al deliver a careful, phenotype-aware analysis that jointly operationalizes SHR and GV for risk stratification in sepsis. By tightening measurement reporting, expanding numerical calibration with a minimal recalibration path, presenting phenotype-specific thresholds with threshold tables plus decision curves, and embedding robust imputation/optimism practices, the approach becomes more portable, auditable and ready for pragmatic adoption across ICUs.
Yang et al. (Wed,) studied this question.