Stress hyperglycemia ratio and glycemic variability for predicting mortality and new-onset rhythm related events in severe heart failure: a multicenter machine-learning cohort study

Stress hyperglycemia ratio (SHR) and glycemic variability (GV) reflect acute glucose dysregulation and are linked to adverse cardiovascular outcomes. Whether the combined assessment of SHR and GV provides additional prognostic information beyond either marker alone remains uncertain. This multicenter cohort study evaluated their individual and joint associations with adverse outcomes across different glycemic states, with machine-learning analyses performed as exploratory analyses. A total of 3202 patients aged > 60 years with severe HF were included from the MIMIC-IV database and two Chinese hospitals, Qingdao University Hospital and Qingdao Municipal Hospital. GV was calculated using glucose measurements obtained during the first 48 h after ICU admission, and follow-up began at the 48-hour landmark. Patients were categorized into six SHR-GV combination groups based on the median SHR and GV tertiles. The primary outcomes were 30-day and 365-day all-cause mortality. The secondary outcome was a prespecified composite of new-onset rhythm-related events, including atrial fibrillation, ventricular arrhythmias, cardiac arrest, and third-degree atrioventricular block. Given the clinical heterogeneity of its components, this composite outcome was considered exploratory. Associations were examined using Cox proportional hazards models, logistic regression, restricted cubic spline analyses, subgroup analyses, and exploratory mediation analyses. For the exploratory fixed-time prediction of 365-day mortality, machine-learning models were developed and internally validated in the MIMIC-IV cohort using five-fold cross-validation and externally validated in an independent Chinese hospital cohort. Model discrimination, calibration, and clinical utility were assessed, and SHAP analysis was used to interpret feature contributions. Median age was 75 years; 59.56% male. Overall, 320 and 448 patients died within 30 and 365 days. Patients with concomitantly high SHR and GV had the highest observed mortality. In the overall cohort and in the prediabetes and diabetes subgroups, the combined SHR/GV model did not significantly improve discrimination compared with either marker alone. In the normal glucose regulation subgroup, the combined model showed modestly higher discrimination for 365-day mortality (AUC = 0.681) than SHR alone (AUC = 0.659) or GV alone (AUC = 0.655). Combined assessment of SHR and GV was associated with mortality and the exploratory prespecified composite of rhythm-related events in elderly patients with severe HF. In exploratory subgroup analyses, the combined model showed modestly higher discrimination for 365-day mortality in patients with NGR, whereas no significant improvement was observed in the overall cohort or other glycemic subgroups.