Machine-learning risk stratification for 12-month mortality and SGLT2 inhibitor-related absolute risk differences in elderly patients with heart failure and type 2 diabetes

Aim: Older adults with heart failure (HF) and type 2 diabetes mellitus (T2DM) face high 12-month mortality. We developed a routine-variable model for mortality prediction and explored SGLT2 inhibitor (SGLT2i)-related absolute risk differences across model-defined risk strata. Methods: This retrospective study included 500 hospitalized patients aged ≥ 75 years with HF and T2DM. Three feature sets (guideline-informed core, primary analysis, and complete imputed candidate) and ten algorithms were compared using repeated five-fold outer/four-fold inner cross-validation with training-fold preprocessing. Cox and multiple-imputation sensitivity analyses were performed. SGLT2i associations were assessed using propensity-score overlap weighting. Results: The binary modelling cohort included 434 patients with 167 deaths (38.5%). The nested selection workflow achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.729, area under the precision-recall curve of 0.612, and Brier score of 0.206. Cox sensitivity analysis yielded a censoring-adjusted 365-day AUROC of 0.755; multiple imputation gave an AUROC of 0.758 for the fixed gradient-boosting reference. Among 381 patients, weighted 12-month mortality was 16.1% in SGLT2i users vs. 31.1% in non-users, a difference of 15.0 percentage points (95%CI: 5.9-23.9). Stratum-specific differences were 6.5, 18.1, and 19.7 percentage points, with wide confidence intervals. A six-variable interactive interface supported risk communication. Conclusion: Routine clinical variables supported moderately discriminative mortality risk stratification under nested validation. SGLT2i use was associated with lower adjusted 12-month mortality overall; larger studies are needed to establish whether associations differ across risk strata. External and prospective validation remains necessary.