This was a single-center retrospective cohort study conducted at the First Affiliated Hospital of Zhengzhou University. Consecutive patients undergoing their first eligible hospitalization in the Department of Cardiology between 1 January 2020 and 31 December 2024 were screened. Eligible patients were aged ≥ 75 years and had coexisting HF and T2DM. HF was defined by compatible symptoms or signs together with natriuretic peptide elevation and/or cardiac imaging evidence; T2DM was defined by a documented history of T2DM, glucose-lowering therapy, or laboratory criteria consistent with diabetes. Inclusion additionally required sufficient baseline information for core-variable assessment. Patients were excluded if expected survival was < 24 h; heart transplantation or ventricular-assist-device implantation occurred during the hospitalization; active malignant disease, severe hematological disease, severe acute infection, or systemic inflammatory disease was present; or clinical data were insufficient for core-variable assessment.

Index admission was defined as the first hospitalization during the study period that met all eligibility criteria; subsequent readmissions were not treated as separate index records. The cohort-construction process and the subsequent modelling and treatment-stratification workflow are shown in Figure 1, with baseline characteristics summarized in Table 1. Follow-up was obtained from the electronic medical record and structured telephone interviews. Follow-up time was calculated from admission to death or the last confirmed contact. Patients whose 12-month status could not be determined were not entered into the 12-month binary endpoint set, whereas patients who died within 365 days were included as events even when the full interval had not elapsed.

The study was approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University (approval No. 2025-KY-1217-001). The requirement for informed consent was waived because of the retrospective design. The study was registered with the Chinese Clinical Trial Registry (ChiCTR2500109213).

Candidate predictors comprised routinely collected baseline demographics, vital signs, comorbidities, laboratory measures, and echocardiographic variables available before discharge treatment assignment. Baseline values were defined as the first recorded measurements after admission during the index hospitalization. Endpoint labels, follow-up status, survival time, and discharge medications were excluded from model fitting so that prediction remained anchored to baseline prognosis.

Missing baseline covariates were handled using single imputation within each training fold^[15]. Numeric variables were completed using training-fold medians with missingness indicators, and categorical variables using the training-fold mode. Scaling and one-hot encoding were also fitted only on training data and applied unchanged to validation data, including within the inner validation loop. Endpoint and treatment-exposure fields were not imputed.

Original missingness was calculated before imputation. Variables with at least 30% missingness were excluded from the primary analysis feature set but retained in the complete imputed candidate feature set. This design prioritized more consistently observed predictors in the primary analysis while preserving a broader sensitivity benchmark. The three feature sets contained 7, 51, and 58 variables, respectively. Feature rosters used source-wide missingness, whereas imputation was fitted within training folds.

A multiple-imputation sensitivity analysis evaluated the fixed gradient-boosting model with the primary analysis feature set. Twenty stochastic chained-equations imputations of numeric predictors were generated within each of the 25 outer training samples using Bayesian ridge conditional models with posterior sampling; the fitted models were then applied to the corresponding validation data without using validation outcomes. Complete categorical variables were entered as one-hot auxiliary predictors. Skewed positive laboratory variables were log-transformed for imputation and restored to their recorded scales before model fitting; imputation support bounds were learned from training observations. Observed values were retained, and missingness indicators were included. Twenty held-out probabilities per repeat were averaged for evaluation. Ten- vs. twenty-imputation averages assessed Monte Carlo stability. This analysis tested the fixed model's missing-data sensitivity and did not repeat selection across the 30 candidates.

The primary outcome was all-cause mortality within 365 days of admission. Three feature sets were compared. The first was a guideline-informed core clinical feature set designed to capture major prognostic domains, including demographic characteristics, ischaemic heart disease history, cardiac function, haemodynamic stress and renal function. It comprised age, sex, albumin, LVEF, NT-proBNP, estimated glomerular filtration rate (eGFR) and previous myocardial infarction. Albumin was additionally included to reflect nutritional reserve and overall systemic disease burden. The second was the primary analysis feature set, which included eligible baseline variables with an original missingness rate of < 30% after excluding patient identifiers, outcome and follow-up fields, derived duplicate variables, and discharge treatment variables. The third was the complete imputed candidate feature set, which retained all eligible baseline predictors after missing-data imputation and was used for sensitivity benchmarking. The seven-variable core set was investigator-defined using guideline-informed prognostic domains.

Ten supervised learning algorithms were systematically compared: L2-penalized logistic regression, elastic net, radial-basis-function support vector machine, k-nearest neighbors, Gaussian naive Bayes, decision tree, random forest, extremely randomized trees, gradient boosting, and AdaBoost. Numeric variables were standardized and categorical variables were one-hot encoded within each training fold. The same predefined model settings were applied across all folds and repeats, without fold-specific hyperparameter searching. The full benchmark is presented in Figure 2. Key hyperparameters are provided in Supplementary Table 1. NT-proBNP entered the primary model in its recorded units.

The principal performance evaluation used five-fold stratified outer cross-validation repeated five times. A four-fold inner loop compared the 30 algorithm-feature-set combinations within each outer training sample; highest inner AUROC, followed by Brier score as a secondary ordering criterion, determined the candidate evaluated on the untouched outer fold. Imputation, encoding, and scaling were fitted within the corresponding training data. Repeat-level metrics were summarized as means and standard deviations. Following the initial exploratory comparison in the same cohort, gradient boosting with the primary analysis feature set was retained as a fixed reference model for interpretation and the digital display. Its repeated-validation and averaged out-of-fold estimates were reported separately. Confidence intervals for fixed-reference averaged predictions used 2,000 patient-level bootstrap samples of saved outcome-prediction pairs without repeating model fitting or selection; these intervals are conditional on the saved predictions.

AUROC assessed rank discrimination, whereas area under the precision-recall curve (AUPRC) was emphasized because mortality was not equally distributed across outcome classes. The Brier score, calibration intercept, and calibration slope assessed probability accuracy. Decision-curve analysis for the fixed gradient-boosting reference compared model-guided intensified review with review-all and review-none strategies across descriptive risk thresholds of 0.05-0.80 [Figure 3]. The thresholds represented hypothetical decision preferences for evaluating net benefit, with p/(1-p) specifying the relative penalty assigned to a false-positive decision^[16]. The range was a broad descriptive interval, not an established clinical cutoff. The hypothetical action was intensified clinical review or follow-up rather than initiation or withholding of SGLT2i. An operational threshold would require the benefits and burdens of the intended action to be specified.

Time-to-event sensitivity analysis used the seven-variable core clinical set in Cox regression among all patients with verified observation times, including those censored before 365 days. Admission was time zero; deaths through day 365 were events, and other observations were censored at last confirmed contact or day 365. Numeric gaps and missingness indicators were processed within training folds, categorical predictors used reference coding, and NT-proBNP was transformed as ln(1 + x). Repeated five-fold validation provided held-out 365-day probabilities; concordance and cumulative/dynamic AUROC with inverse censoring weights were calculated. Kaplan-Meier estimates used tertiles of averaged held-out Cox probabilities. Rank-based Schoenfeld tests assessed proportional hazards in the final cohort fit. Records without verifiable follow-up time were not treated as known survivors or valid censored observations.

For a separate six-variable interpretability display, variance inflation factors (VIFs) were calculated for age, BMI, NT-proBNP, NYHA grade, albumin, and LVEF. This display was distinct from the guideline-informed core clinical feature set used in model benchmarking. A VIF below 5 was considered acceptable. Permutation importance summarized the decrease in AUROC in five held-out folds of the first repeat, using ten shuffles per variable per fold. NT-proBNP was transformed as ln(1 + x) for the VIF calculation, and missing numeric values were median-imputed for this descriptive audit. VIF estimates are provided in Supplementary Table 2.

The six-variable interface used age, BMI, NT-proBNP, NYHA functional class, albumin, and LVEF, chosen for their clinical interpretability and routine availability. Together, they represented age-related vulnerability, nutritional reserve, cardiac load, symptom severity, and systolic function. The fixed full gradient-boosting reference model supported interpretation, and the six-variable model supported interactive risk communication. The predictor audit and interpretation outputs are presented in Figure 4. The three model-development feature rosters are listed in Supplementary Table 3.

The adjusted SGLT2i analysis included 381 patients with documented discharge exposure, ascertainable follow-up and positive post-discharge observation, without requiring survival for 30 days. The endpoint was all-cause mortality within 365 days of discharge; shorter observed follow-up was right-censored. Baseline risk was represented by archived out-of-sample predictions from the fixed gradient-boosting reference model. Tertiles defined three groups of 127 patients; the cohort, score source, grouping rule, and endpoint were fixed before the adjusted comparison. These strata differ from the fitted-score tertiles used in the original 358-patient comparison and were not optimized against treatment differences. Propensity models were fitted overall and within each stratum using admission year, age, sex, BMI, systolic blood pressure, pulse, NYHA class, LVEF, ln(1 + NT-proBNP), eGFR, albumin, glycated haemoglobin, chronic kidney disease, atrial fibrillation and previous myocardial infarction. Missing propensity covariates were median/mode completed. Overlap weights were 1 minus the propensity score for users and the propensity score for non-users. Weighted Kaplan-Meier estimates provided 365-day mortality and non-user minus user risk differences. Covariate balance, propensity overlap, and effective sample size were reported. Confidence intervals used 1,000 treatment-stratified patient bootstrap samples with propensity refitting, conditional on fixed risk scores and strata. Unrecorded prescribing reasons were not imputed.

The original 358-patient, 30-day-survivor comparison summarized deaths over all available follow-up using fixed fitted-model tertiles, Wilson mortality intervals and Newcombe risk-difference intervals. Its absolute risk reduction and number needed to treat summaries are provided separately in Supplementary Table 4. The corresponding early-death sensitivity analysis retained the original score boundaries [Supplementary Figure 1]. The adjusted analysis used all eligible discharge patients, out-of-sample scores, and a fixed 365-day horizon; the populations, strata, and endpoint definitions are reported separately.

Adjusted treatment estimates are reported in Figure 4B-D and Supplementary Table 5. The Figure 5 interface uses a six-variable ridge surrogate of the fixed gradient-boosting reference model. Within each outer training sample, inner out-of-fold teacher predictions, clipped to 0.01-0.99 and logit-transformed, supplied the targets; the fitted surrogate was evaluated on the outer held-out fold. Ridge alpha was 2.0. Calibration, Brier score, and absolute prediction errors were evaluated together with correlation. Coefficients, intercept, units, ln(1 + NT-proBNP) transformation and standardization are provided in Supplementary Table 6. All analyses were performed in Python.

To support reproducibility, saved output files were used for manuscript tables, figure source data, out-of-fold prediction metrics, risk-stratum summaries and package-level audit reports. This structure allows the reported numbers in the manuscript, figures, supplementary tables, and web display to be traced back to their source files.

The clinical question and cohort construction are summarized in Figure 1A and B. The initial cohort comprised 500 eligible patients. Follow-up sufficient for endpoint ascertainment was available for 469 patients, whereas 31 patients were lost to follow-up. Overall, 198 deaths occurred during the available follow-up. The 12-month modelling cohort comprised 434 patients, of whom 167 died within 12 months (38.5%) [Figure 1B]. Among the 469 patients with ascertainable outcomes, 35 patients without a recorded death had less than 365 days of follow-up and therefore could not be classified for the 12-month binary endpoint; they were excluded from binary modelling. Patients with undetermined 12-month outcomes were not classified as survivors. Baseline characteristics stratified by 12-month all-cause mortality status are shown in Table 1.

Across 30 algorithm-feature-set combinations, descriptive mean AUROC ranged from 0.535 to 0.767 [Figure 2A]. The complete-set random forest had the highest mean AUROC (0.767). Within the primary analysis set, random forest and gradient boosting had mean AUROCs of 0.758 and 0.753, respectively [Figure 2B]. Gradient boosting had a lower mean Brier score (0.195 vs. 0.200) and a calibration slope closer to 1 (0.885 vs. 1.581). This comparison informed its retention as the fixed reference for probability estimation, interpretation, and the digital display. Candidate selection was evaluated through the principal nested validation workflow. All candidate estimates and outer-fold selection frequencies are in Supplementary Table 1.

The nested candidate-selection workflow achieved mean repeat AUROC 0.729 (SD 0.018), AUPRC 0.612 (SD 0.028), and Brier score 0.206 (SD 0.008). These are the principal performance estimates [Table 2], describing the selection workflow rather than a single fixed algorithm. The fixed gradient-boosting reference had mean repeat AUROC 0.753 and AUPRC 0.632; its averaged out-of-fold predictions gave AUROC 0.762 (95%CI: 0.716-0.805) and AUPRC 0.642 (95%CI: 0.566-0.720), as shown in Figure 2C and D.

For the fixed gradient-boosting reference model, averaged out-of-fold probabilities yielded a Brier score of 0.191, calibration intercept 0.017, and slope 0.975. The calibration curve and risk-decile plot displayed agreement and deviations across deciles [Figure 3A and B]. Decision-curve analysis compared net benefit across hypothetical thresholds for intensified review or follow-up, rather than treatment eligibility [Figure 3C]. Mortality and non-mortality groups had overlapping predicted-risk distributions, with higher values among those who died [Figure 3D]. These fixed-reference plots are distinct from the principal nested-workflow estimates in Table 2.

Held-out permutation importance ranked NYHA functional class, BMI, NT-proBNP, age, albumin and total bilirubin highest; LVEF ranked eighth [Figure 4A]. These variables represented complementary information on age, nutritional status, haemodynamic burden, symptom severity and cardiac or systemic function. The six variables used in the simplified risk display showed limited collinearity, with a maximum VIF of 1.322 [Supplementary Table 2]. The six-variable set was distinct from both the seven-variable core comparator and the full primary model.

The adjusted discharge-treatment cohort included 381 patients, comprising 198 SGLT2i users and 183 non-users; 84 deaths occurred within 365 days of discharge. Overlap-weighted mortality was 16.1% in users and 31.1% in non-users, with a risk difference of 15.0 percentage points (pp) (95%CI: 5.9-23.9). Overall effective sample sizes were 156.6 and 147.1, respectively. Maximum absolute standardized covariate differences were below 0.002 after weighting [Supplementary Table 5].

Each risk stratum contained 127 patients. Adjusted non-user minus user risk differences were 6.5 (-6.5 to 17.4), 18.1 (-4.5 to 39.6) and 19.7 (-1.7 to 40.6) pp in the low-, intermediate- and high-risk strata, respectively [Figure 4B-D and Table 3]. Point estimates were larger in the intermediate- and high-risk strata, but each stratum-specific interval included zero. The high-minus-low contrast was 13.2 pp (95%CI: -10.0 to 36.2).

The original full-follow-up comparison and its early-death sensitivity analysis are presented separately in Supplementary Table 4 and Supplementary Figure 1.

Cox sensitivity analysis included all 469 patients with verified observation times, including 35 patients censored before day 365, and recorded 167 deaths through day 365. Mean repeat concordance was 0.724, and censoring-adjusted 365-day AUROC was 0.755. Kaplan-Meier 12-month mortality was 12.9%, 36.8% and 57.7% across tertiles of held-out Cox predictions (log-rank P < 0.001; Supplementary Figure 2A and B). No individual rank-based proportional-hazards test had P < 0.05 (minimum P = 0.067). These estimates assess sensitivity to the observed follow-up structure using the seven-variable core set [Supplementary Table 7].

For the fixed gradient-boosting reference, twenty-imputation mean repeat AUROC was 0.758, AUPRC 0.634, and Brier score 0.193, compared with 0.753, 0.632, and 0.195 using training-fold median/mode completion [Supplementary Figure 2C and D]. Ten- vs. twenty-imputation averaged probabilities differed by a mean of 0.37 pp. This sensitivity analysis evaluated the fixed model without repeating selection among the 30 candidates [Supplementary Table 8].

To enhance the translational accessibility of the risk-stratification framework, we developed an interactive digital interface based on six routinely available clinical variables—age, BMI, NT-proBNP, NYHA functional class, albumin and LVEF—to display admission-origin 12-month mortality estimates [Figure 5]. The ridge surrogate approximated the fixed gradient-boosting reference model’s out-of-fold predictions and preserved their overall ordering (held-out Pearson correlation, 0.821). In held-out assessment, the surrogate achieved an AUROC of 0.758, a Brier score of 0.193, and a calibration slope of 1.211; mean absolute and root mean squared differences from fixed-reference risk were 9.8 and 12.5 pp, respectively [Supplementary Table 6]. The interface reduced input burden to six variables and displayed all six standardized log-odds contributions. The treatment panel presented the adjusted cohort-level 12-month mortality differences separately from the individual mortality estimate; treatment-cohort strata were not assigned by the six-variable score.

Several limitations should be considered. This retrospective, single-center cohort was modest relative to the number of candidate predictors and algorithms, and external validation was not performed. The fully nested estimates therefore provide the principal internal assessment, while the fixed gradient-boosting reference and its digital surrogate require evaluation in independent populations. In addition, 31 patients lacked verifiable follow-up and were excluded from endpoint analyses. The Cox sensitivity analysis retained 35 patients with shorter event-free observation as censored records, but selection related to unavailable follow-up cannot be excluded. Multiple imputation evaluated numeric missing-data uncertainty for the fixed reference model rather than the full candidate-selection workflow or treatment analysis. SGLT2i exposure was observational. Overlap weighting balanced recorded covariates, but unmeasured prescribing factors, including contraindications, clinical judgment, adherence, and treatment costs^[26], may influence the associations. The original comparison also required 30-day post-discharge survival; its separate early-death sensitivity analysis and the revised discharge-origin comparison address the importance of explicit observation windows^[27]. Stratum-specific confidence intervals were wide and did not establish heterogeneous treatment associations. Bootstrap intervals for the fixed-reference predictions were conditional on saved outcome-prediction pairs and did not repeat model development; nested assessment included candidate selection. Treatment bootstrap intervals refitted propensity models but remained conditional on fixed risk strata, and overlap weighting targets patients with similar measured treatment-assignment profiles^[28]. Finally, the six-variable surrogate had a mean absolute difference of 9.8 pp from the reference predictions. Independent calibration, prospective usability testing and larger multicentre cohorts are needed before routine clinical implementation.