This prospective observational study was conducted at Thong Nhat Dong Nai General Hospital and Can Tho Central General Hospital between July 2025 and January 2026. The study population comprised patients hospitalized with acute decompensated heart failure (ADHF) and/or heart failure with reduced ejection fraction (HFrEF) who met the prespecified eligibility criteria. Patients were eligible if they were admitted with ADHF regardless of left ventricular ejection fraction (LVEF) phenotype and/or had HFrEF documented during the index hospitalization. Underlying HF was diagnosed by the treating cardiologists based on compatible symptoms and signs, echocardiographic assessment, and routine clinical investigations during the index hospitalization. ADHF was recorded when patients were admitted because of new or worsening symptoms and signs of HF requiring inpatient treatment. LVEF was assessed by transthoracic echocardiography during the index hospitalization and was used to classify HF phenotype as HFrEF, heart failure with mildly reduced ejection fraction (HFmrEF), or heart failure with preserved ejection fraction (HFpEF). For this study, HFrEF was defined as LVEF ≤40%, HFmrEF as LVEF 41% to 49%, and HFpEF as LVEF ≥50%. Patients with HFmrEF and HFpEF were included only if they were hospitalized with ADHF during the index hospitalization. Therefore, ADHF and HFrEF were not considered mutually exclusive categories. ADHF was considered a clinical presentation, whereas LVEF phenotype was considered an echocardiographic classification. Patients with end-stage chronic kidney disease (estimated glomerular filtration rate [eGFR]<15 mL/min/1.73 m²) or receiving dialysis, those with conditions known to affect sST2 levels, including systemic diseases such as lupus, rheumatoid arthritis, and chronic obstructive pulmonary disease, as well as those lost to follow-up, were excluded. Hospital admission and discharge decisions were made by the treating physicians according to routine clinical practice, and patients were discharged when they were considered clinically stable and suitable for outpatient follow-up. In addition, a non-HF comparison group was recruited for exploratory comparison of baseline characteristics and serum sST2 concentrations. These participants had no known history of HF based on medical record review and clinical history. They underwent clinical assessment to exclude HF, with routine electrocardiogram assessment as part of hospital evaluation. The absence of HF was further supported by N-terminal pro-B-type natriuretic peptide testing and transthoracic echocardiography. A total of 218 patients with ADHF and/or HFrEF and 59 non-HF controls were included in the final analysis.

The study protocol was reviewed and approved by the Biomedical Research Ethics Committee of Can Tho University of Medicine and Pharmacy, Vietnam, for both participating centers. Ethical approval was granted for the study conducted at Can Tho Central General Hospital under approval number 25.326.HV/PCT-HĐĐĐ, dated June 30, 2025, and for the study conducted at Thong Nhat Dong Nai General Hospital under approval number 25.274.HV/PCT-HĐĐĐ, dated June 30, 2025. The study was conducted in accordance with the ethical principles for biomedical research involving human participants. Written informed consent was obtained from all participants before enrollment. All patient data were coded and kept confidential, and the results were reported only in aggregate form without identifying individual participants. No financial compensation or payment was provided to participants. The manuscript and supplementary materials do not contain any identifiable individual participant images or personal information.

The primary outcome was the 30-day postdischarge MACE. For the main outcome analysis, MACE was defined as a patient-level composite end point of all-cause death or rehospitalization for HF during follow-up [13]. During the 30-day postdischarge period, outcome data were initially ascertained by an independent assessor through telephone contact with patients or their family members. This assessor was not involved in sST2 measurement, statistical analysis, or model development and was not provided with baseline sST2 results. Reported events were subsequently verified through a review of hospital and electronic medical records. Rehospitalization for HF was defined as hospital admission due to worsening symptoms or signs of HF requiring inpatient treatment. Deaths during follow-up were identified through telephone follow-up and verified using available medical records whenever possible. Patients were subsequently classified into the MACE and non-MACE groups according to the occurrence of the composite end point within 30 days after discharge. Each patient contributed only once to the primary MACE end point. If both HF rehospitalization and death occurred within 30 days, the patient was counted once in the composite end point and was not double-counted.

Serum sST2 concentration was measured using the Elabscience Human sST2 (Soluble ST2) enzyme-linked immunosorbent assay (ELISA) Kit (catalog number E-EL-H6082), a sandwich ELISA assay intended for research use, in accordance with the manufacturer’s instructions. According to the manufacturer’s documentation, the assay sensitivity was 0.19 ng/mL, and the detection range was 0.31‐20 ng/mL. Manufacturer-reported intra-assay coefficients of variation were 4.43%, 4.36%, and 6.31% across low, medium, and high concentration samples, respectively, and inter-assay coefficients of variation were 8.22%, 6.76%, and 6.58%, respectively. Blood samples were collected at hospital admission during the index hospitalization for patients with HF and at enrollment for non-HF comparison participants. Serum samples were processed according to routine laboratory procedures and the manufacturer’s instructions. Samples with optical density values exceeding the upper limit of the standard curve were not extrapolated. Instead, these samples were initially diluted 1:2 using the kit-provided Reference Standard & Sample Diluent and reassayed. If the measured concentration remained above the standard curve range, additional serial dilution was performed as needed until the measured concentration fell within the quantifiable range. Final concentrations were calculated by multiplying the measured concentration by the corresponding total dilution factor. Samples and standards were analyzed in duplicate, optical density was measured at 450±2 nm, and sST2 concentrations were calculated from a 4-parameter logistic standard curve. sST2 values were recorded in nanograms per milliliter and analyzed as a continuous variable to evaluate their association with 30-day postdischarge MACE.

Baseline variables were collected at admission and included demographic characteristics, clinical profile, medical history, and laboratory parameters relevant to HF prognosis. Age was recorded in years, and BMI was calculated as weight in kilograms divided by height in meters squared. Functional status was assessed according to the New York Heart Association classification [7]. Smoking was defined as a patient who had smoked at least 100 cigarettes during their lifetime and had smoked within 30 days before hospitalization [14]. Alcohol consumption was defined as any self-reported intake of alcoholic beverages, including beer, wine, or spirits, during the 12 months before hospitalization [15]. Comorbidities and clinical history, including hypertension, diabetes mellitus, dyslipidemia, prior myocardial infarction, stroke, atrial fibrillation, smoking, and alcohol use, were obtained from medical records and direct patient interview. LVEF was assessed by transthoracic echocardiography during the index hospitalization, preferentially using the biplane Simpson’s method when image quality was adequate [16]. Serum sodium, urea, and creatinine were measured using routine biochemical assays at admission. eGFR was calculated from serum creatinine using the Chronic Kidney Disease Epidemiology Collaboration equation [17].

All statistical analyses were performed in R (R Foundation for Statistical Computing). Data management and variable transformation were conducted primarily with dplyr, while tabular regression outputs were organized with broom and tibble. Continuous variables were summarized as mean (SD) or median (IQR), as appropriate, and categorical variables as frequencies and percentages. Comparisons between patients with HF and controls were performed using 2-tailed independent-samples t tests or Mann-Whitney U tests for continuous variables, and chi-square tests or Fisher exact tests for categorical variables, as appropriate. For sST2 comparisons between groups, adjusted analyses were performed using analysis of covariance controlling for confounders. In addition, baseline characteristics were compared between the MACE and non-MACE groups using the 2-tailed independent-samples t test for continuous variables and the chi-square test or Fisher exact test, or Fisher Freeman Halton exact test for categorical variables, as appropriate. Crude associations with MACE were explored using univariable logistic regression, and crude odds ratios (ORs) with 95% CIs were reported. Candidate predictors for multivariable modeling were chosen based on clinical plausibility together with univariable screening, with preference given to variables with P<.20 while avoiding the redundant representation of the same biological domain and limiting model complexity in view of the available number of events. A 2-sided P value <.05 was considered statistically significant for the final multivariable analyses [18,19]. Model development and reporting were guided by transparent prediction-model reporting principles, including TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) and TRIPOD+AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis plus Artificial Intelligence) recommendations [20,21]. With 47 events, parsimonious models were limited to 3 to 4 predictors, corresponding to approximately 15.7 to 11.8 events per variable, which is consistent with conventional events per variable guidance for logistic regression [20,21]. Sample size adequacy for prediction-model development was additionally assessed using the pmsampsize framework for binary outcomes proposed by Riley et al [22].

To reduce the risk of overfitting, model development followed a prespecified parsimonious candidate-model strategy rather than unrestricted stepwise selection. All candidate models were fitted on the same complete-case dataset to ensure a fair comparison of model performance. There were no missing data for the variables included in the candidate prediction models. All 218 patients had available sST2 measurements, baseline clinical and laboratory variables, and 30-day outcome status. Based on the clinically prioritized core predictors age and sST2, 11 prespecified logistic regression models were examined. Final model selection was based on a joint assessment of optimism-corrected discrimination, calibration, overall fit, parsimony, and clinical interpretability. When candidate models showed comparable optimism-corrected discrimination, preference was given to the model with fewer predictors, a more favorable calibration slope, and greater feasibility for conversion into a simple point-based score. Apparent discrimination for each candidate model was quantified with the area under the receiver operating characteristic curve (AUC) using pROC, with 95% CIs estimated by the DeLong method. Overall fit was additionally assessed with the Akaike information criterion (AIC) and Brier score. Although hypertension showed a signal in univariable analysis, categorical variables with sparse subgroup counts or unstable estimates, particularly hypertension and New York Heart Association class, were not prioritized for the prespecified parsimonious multivariable models. This decision was made because some categories contained relatively small numbers of patients, resulting in wide CIs and reduced estimate stability, with the potential to increase model complexity and overfitting without clear incremental predictive value [18].

Shortlisted candidate models were then subjected to internal validation using bootstrap resampling with the rms package. Apparent and optimism-corrected performance were evaluated with 1000 bootstrap resamples, focusing on AUC, Brier score, calibration intercept, and calibration slope. The final model was selected according to the prespecified rule described above, with priority given to comparable optimism-corrected discrimination, calibration slope, parsimony, and clinical interpretability. The final coefficients and adjusted ORs were then estimated using standard multivariable logistic regression with the glm function under a binomial link. The assessment of calibration was emphasized because adequate calibration is essential for the clinical usefulness of risk prediction models [18,19].

A sensitivity analysis was performed using Firth penalized logistic regression with the logistf package to assess the robustness of coefficient estimates and inference under small-sample conditions and potential sparse-data bias [23,24]. Discrimination of the final logistic model was evaluated by AUC with a 95% CI using pROC. Calibration was examined both numerically and graphically using rms, including the calibration intercept and slope as summary indices and bootstrap-based calibration plots for visual assessment of agreement between predicted and observed risk. Figures were generated with ggplot2; when a smoothed receiver operating characteristic curve was displayed for visual presentation, the reported AUC and 95% CI were still calculated from the original empirical receiver operating characteristic curve.

A simplified point-based score was subsequently derived from the final parsimonious model. First, uniform shrinkage was applied to the regression coefficients using the bootstrap-corrected calibration slope from the final model. The shrunken coefficients were then converted into a pragmatic integer score using clinically interpretable reference values and increments. For each predictor, the shrunken coefficient was first rescaled according to its clinically defined increment: 5 years for age above 60 years, 5 ng/mL for sST2 above 20 ng/mL, and 10 mL/min/1.73 m² for eGFR below 60 mL/min/1.73 m². Let w_j denote the rescaled shrunken coefficient for predictor j. Integer points were assigned by dividing each w_j by the smallest absolute rescaled coefficient and rounding to the nearest integer: Points_j = round (w_j/min(|w|)). Using this approach, age, sST2, and eGFR were assigned 1, 2, and 1 points, respectively. The total score was obtained by summing the component points across predictors. Negative component points were not allowed. Scores above the maximum observed value in the derivation cohort were not used to generate separate risk estimates and should be interpreted as ≥12 until externally validated. A secondary logistic regression model with total score as the sole predictor was then fitted to translate score values into estimated probabilities of MACE. In this score-only logistic model, the intercept was −2.592, and the coefficient for each 1-point increase in the SAGE-HF score was 0.250, corresponding to an OR of 1.284 per 1-point increase. The predicted probability was calculated as logit(p)=–2.592 + 0.250 × SAGE-HF score. The discrimination of the simplified score was assessed by AUC with 95% CI using pROC, and its calibration was evaluated with Brier score, calibration intercept, calibration slope, and calibration plots using rms [18,19]. Decision curve analysis was performed to evaluate the clinical usefulness of the SAGE-HF score by estimating net benefit across threshold probabilities from 0.05 to 0.50 and comparing the score with treat-all and treat-none strategies [25].

In this two-center prospective study, serum sST2 concentrations were significantly higher in patients with HF compared with non-HF controls. In the two-center prospective HF cohort, 21.6% (47/218) of the patients experienced 30-day postdischarge MACE, comprising 19.3% (42/218) HF rehospitalizations and 2.3% (5/218) deaths. Although hypertension was less frequent in the MACE group, this crude inverse association should be interpreted cautiously and should not be considered evidence of a protective effect. Hypertension was recorded as a binary history-based comorbidity and may not reflect hemodynamic status during the index hospitalization. This finding may reflect residual confounding, selection or collider bias, the high overall prevalence of hypertension in the cohort, and the small number of patients without hypertension. In a sensitivity analysis, admission systolic blood pressure was not associated with 30-day MACE (OR 0.99 per mm Hg, 95% CI 0.985‐1.013; P=.89), supporting the interpretation that the inverse association for categorical hypertension was unlikely to represent a true protective effect. Patients with events were older and had higher admission sST2 and worse renal indices. Among prespecified candidate models, a parsimonious model including age, sST2, and eGFR achieved acceptable discrimination (AUC 0.741) and favorable internal validation with good calibration characteristics. A simplified integer score (SAGE-HF score) derived from the final model produced a clinically interpretable risk gradient, with predicted event probabilities rising from low single digits to approximately 60% across the observed score spectrum.

Age, renal function, and sST2 each contributed distinct prognostic information in this cohort. Patients with 30-day MACE were older than those without MACE (mean 74.9, SD 13.0 y vs 68.3, SD 12.1 y), and age remained independently associated with the end point, with an adjusted OR of 1.04 (95% CI 1.01‐1.07) per year. This finding is consistent with the recognized vulnerability of older adults with HF, who often have greater comorbidity burden, frailty, and reduced physiologic reserve [26]. Renal dysfunction was also more prominent in patients with events. The MACE group had lower eGFR than the non-MACE group (median 44.0, IQR 30.78-61.94 mL/min/1.73 m² vs 62.24, IQR 45.53-84.77 mL/min/1.73 m²), and eGFR remained independently associated with MACE, with an adjusted OR of 0.98 (95% CI 0.97‐1.00) per 1 mL/min/1.73 m². This supports the clinical relevance of cardiorenal interaction in early postdischarge risk after HF hospitalization [27]. Admission sST2 was higher in patients who developed MACE than in those who did not (median 27.61, IQR 23.36-30.62 ng/mL vs 22.68, IQR 15.94-30.27 ng/mL) and remained independently associated with 30-day MACE after adjustment for age and eGFR, with an adjusted OR of 1.06 (95% CI 1.01‐1.11) per ng/mL. This finding is consistent with prior literature linking higher sST2 concentrations to adverse outcomes in acute and chronic HF settings [9,28]. Mechanistically, sST2 reflects the activation of the IL-33/ST2 pathway and may capture myocardial stress, inflammation, and profibrotic remodeling [8,29]. In this cohort, the contribution of sST2 was observed despite the inclusion of renal function in the model, supporting its potential value as a complementary biomarker in short-term risk stratification.

A key finding of this study is that a 3-predictor model incorporating age, admission sST2, and eGFR provided clinically interpretable short-term risk stratification in a cohort with 47 MACE among 218 hospitalized patients with HF. The final model achieved an apparent AUC of 0.741 (95% CI 0.664‐0.819), and its performance remained stable after internal validation with 1000 bootstrap resamples, with an optimism-corrected AUC of 0.725, Brier score of 0.155, calibration intercept of 0.009, and calibration slope of 0.930. These values indicate acceptable discrimination and calibration for an early postdischarge outcome, while also showing the expected degree of optimism in a relatively small derivation cohort. Compared with the 4-predictor model including serum sodium, the 3-predictor model had comparable or better optimism-corrected discrimination and a more favorable calibration slope, supporting the decision to prioritize parsimony.

The simplified SAGE-HF score further translated this model into a bedside format. In the derivation cohort, the predicted 30-day MACE probability increased from 7.0% at a score of 0 to 60.1% at a score of 12. When provisional risk strata were applied, the observed MACE risk was 2.2% in the low-risk group, 22.0% in the intermediate-risk group, and 37.0% in the high-risk group, with exact binomial 95% CIs of 0.1% to 11.5%, 14.9% to 30.6%, and 24.3% to 51.3%, respectively. This gradient is clinically relevant because the overall 30-day MACE rate was 21.6%, comprising 42 HF rehospitalizations and 5 deaths. However, because the low-risk stratum included only 1 event and its upper confidence limit exceeded 10%, the proposed <10% threshold should be interpreted as provisional rather than definitive.

These findings should be interpreted in the context of previous HF prediction literature. Prior studies have emphasized that 30-day readmission after HF hospitalization is difficult to predict because early events are driven by heterogeneous clinical and health-system factors [4]. Machine learning and traditional models for 30-day HF readmission have also shown variable performance and limited external validation across settings [30]. Against this background, the SAGE-HF model does not eliminate the need for clinical judgment, but it offers a low-burden and biologically interpretable approach using 3 readily available predictors. The inclusion of sST2 is consistent with literature showing that sST2 reflects myocardial stress, inflammation, fibrosis, and remodeling and has prognostic value in HF [28]. In this cohort, sST2 remained associated with MACE after adjustment for age and eGFR, supporting its potential role as a complementary biomarker rather than a replacement for standard clinical assessment.

From a postdischarge care perspective, the practical value of the SAGE-HF score lies in its ability to identify patients who may need more intensive early follow-up. This potential clinical usefulness was also supported by decision curve analysis, which showed a favorable net benefit of the SAGE-HF score across a clinically relevant range of threshold probabilities. Timely postdischarge care and structured transition-of-care strategies have been emphasized in HF management and have been associated with lower readmission rates in HF populations [31,32]. In this cohort, patients with scores ≥7 had an observed 30-day MACE risk of 37.0%, suggesting that this group may be considered for closer follow-up, medication optimization, early reassessment, or structured telephone monitoring after discharge. However, these risk categories remain exploratory and should not be used as definitive clinical thresholds until validated in larger and geographically diverse cohorts.

This study has several strengths. The prospective design and inclusion of patients from 2 different centers enhance the reliability of the findings. The model was developed based on a prespecified and transparent strategy, with internal validation using bootstrap methods and comprehensive assessment of calibration. In addition, the inclusion of a non-HF comparison group allowed for an exploratory comparison of sST2 levels between patients with HF and individuals without HF. A Firth penalized logistic regression analysis was also performed to assess the robustness of the model and minimize potential bias related to the limited number of events. Furthermore, a simplified risk score was derived to facilitate clinical application, particularly in discharge planning and early postdischarge risk stratification. However, several limitations should be considered. First, the number of events was relatively small, which limited the number of predictors that could be included and may affect the stability of the model. To contextualize sample size adequacy for prediction-model development, an additional assessment using the pmsampsize framework was performed (Table S8 in Multimedia Appendix 1). With an anticipated C-statistic of 0.74, an outcome prevalence of 0.216, and 15 candidate predictor parameters before model reduction, the estimated minimum sample size was 1031 patients, corresponding to 223 events. In a sensitivity calculation using 19 candidate predictor parameters, the estimated requirement increased to 1306 patients, corresponding to 283 events. Because the present derivation cohort included 218 patients and forty-seven 30-day MACE, the study was underpowered for fitting a full multivariable prediction model containing all candidate predictors before reduction. Therefore, this study should be interpreted as an internally validated derivation study of a deliberately parsimonious score rather than a fully powered development study for a large multivariable prediction model. Second, the follow-up period was limited to 30 days, which may not fully reflect longer-term outcomes after HF hospitalization. In addition, death before rehospitalization may act as a competing event when rehospitalization is analyzed separately. Because the dataset recorded the occurrence of 30-day MACE rather than complete component-specific time-to-event data, a Fine-Gray competing-risk analysis could not be performed appropriately. Third, short-term outcomes after discharge may also be influenced by health care system factors, such as access to follow-up, medication availability, and care coordination, which were not fully captured in this study. In addition, the counterintuitive inverse association between recorded hypertension and 30-day MACE may reflect residual confounding, selection or collider bias, and the limitation of using a history-based binary comorbidity variable rather than detailed longitudinal blood pressure, treatment, or hemodynamic data. Fourth, serum sST2 was measured using a research-use Elabscience ELISA kit rather than the US Food and Drug Administration–cleared Presage ST2 Assay. Therefore, absolute sST2 concentrations in this study should not be directly compared with Presage-based cohorts, and the Presage 35 ng/mL cutoff is not directly transferable. The observed associations and the SAGE-HF score should be interpreted as assay-specific findings that require external validation, ideally including assay harmonization or comparison with clinically validated sST2 platforms. Fifth, the non-HF comparison group was not formally matched to the HF cohort by age, sex, or BMI. Therefore, comparisons between patients with HF and non-HF comparison participants should be interpreted as exploratory and were not used in the development, validation, or performance assessment of the SAGE-HF score. Sixth, although this study included patients from 2 hospitals, the number of centers was small and the findings may not be generalizable to other regions, hospital levels, or HF populations with different etiology patterns. Center-level influence was assessed using a fixed-effect sensitivity analysis, but external validation in larger and geographically diverse multicenter cohorts remains necessary. In addition, formal comparison with some established HF risk scores was limited because several required predictors for these tools, particularly the OPTIMIZE-HF postdischarge mortality score, were not systematically collected in this cohort. Finally, the proposed risk thresholds are exploratory and were derived from the current cohort. They should not be used as definitive clinical decision thresholds before external validation. In addition, the low-risk stratum contained only 1 MACE. Although the observed event rate was low, the exact 95% CI extended above 10%, indicating that the proposed low-risk threshold requires confirmation in larger external validation cohorts. Further studies are needed to validate these findings in larger and more diverse populations, with longer follow-up and adequate event numbers. Such work may help clarify the role of this model in routine clinical practice.

In this two-center prospective cohort of patients hospitalized for HF, 21.6% (47/218) experienced 30-day postdischarge MACE. Admission age, sST2, and eGFR were independently associated with risk and were integrated into the SAGE-HF score, which demonstrated acceptable predictive performance. The score provided clear risk stratification with a graded increase in predicted MACE probability. External validation is warranted before routine clinical application.