The Jiangxi-ADHF II cohort represents a physician-initiated retrospective study. Its primary objectives are to maximize the utility of clinical data from hospital records of ADHF patients, explore novel methodologies for early risk stratification in this population, and generate valuable research evidence for improving adverse clinical outcomes. The Jiangxi-ADHF II cohort consecutively enrolled 3,484 hospitalized patients diagnosed with ADHF at Jiangxi Provincial People’s Hospital between January 2018 and January 2024. Diagnostic criteria followed the then-current European Society of Cardiology guidelines for acute and chronic heart failure management, with the latest version available at admission serving as reference. Regarding study design, the study protocol received formal ethical approval from the Ethics Committee of Jiangxi Provincial People’s Hospital (No. 2024-01). Regarding data utilization, written informed consent was obtained from all participants or their legally authorized representatives. The study complied with the ethical principles outlined in the Declaration of Helsinki and adhered to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines for reporting observational research findings.

In the current study, we aimed to investigate the association between the CALLY index and short-term outcomes in ADHF patients. The participant screening workflow and implementation process are summarized in Figure 1 and outlined as follows: (1) We excluded subjects with significant fluid and sodium retention secondary to non-cardiac conditions, including patients with liver cirrhosis, uremia, and those with chronic kidney disease undergoing hemodialysis treatment (n = 273). (2) Patients with malignancies were excluded, as their limited life expectancy could significantly impact the study outcomes (n = 160). (3) Given the potential confounding effect of reperfusion therapy on short-term prognosis, individuals who had undergone percutaneous coronary intervention within 3 months prior to enrollment were excluded (n = 102). (4) Patients with implanted cardiac pacemakers were excluded due to potential autonomic nervous system dysregulation (n = 121). (5) Pregnant women and minors (aged <18 years) were also excluded (n = 26). (6) Participants with missing CALLY index data were excluded (n = 1,577).

We evaluated baseline data of participants within 24 h of admission, encompassing the New York Heart Association (NYHA) functional classification assessed at admission, sociodemographic characteristics (age and gender), lifestyle habits (drinking and smoking status), personal medical history (hypertension, diabetes, stroke, and CHD), and echocardiographic parameters—specifically left ventricular ejection fraction (LVEF).

Venous blood samples were collected by nursing professionals after admission and transported to the Medical Laboratory Center of Jiangxi Provincial People’s Hospital, where they were analyzed using automated analyzers by trained laboratory technicians. The panel of routine blood tests and biochemical markers included: white blood cell count, neutrophil count, monocyte count, lymphocyte count, red blood cell count (RBC), platelet count (PLT), Alb, creatinine (Cr), blood urea nitrogen (BUN), alanine aminotransferase, aspartate aminotransferase (AST), fasting plasma glucose (FPG), CRP, and N-terminal pro-B-type natriuretic peptide (NT-proBNP).

This study primarily assessed the 30-day all-cause mortality rate among ADHF patients following hospitalization. Using the admission date as the reference time point (Day 0), participants’ survival status and the occurrence date of the endpoint event (death) were systematically recorded throughout the 30-day follow-up period.

As CRP testing is not routinely performed in ADHF patients, this observational study consequently excluded many participants due to missing CRP values. To enhance methodological transparency, we further compared the baseline characteristics between participants with and without CRP measurements in this study. The comparison results (Supplementary Table 1; most p-values >0.05) indicated that the majority of baseline characteristics showed similar distributions between the missing and non-missing CRP groups. These findings suggest that the missing CRP data occurred randomly and were independent of both observed and unobserved factors.

In the current study, partial missing values were observed for the covariates LVEF, Cr, BUN, and FPG, with a maximum missing rate of 3.75% (detailed in Supplementary Table 2). Given the relatively low proportion of missing data, the study retained the original dataset for analysis to preserve data authenticity and minimize potential bias.

R (version 4.2.1) and Empower^® (version 4.2) statistical software were utilized for data analysis in this study, and a two-tailed significance level of 5% was employed. Baseline characteristics of the study population are presented as frequency (percentage), mean ± standard deviation, or median and interquartile range.

This study employed the Boruta algorithm for feature selection, an all-relevant feature selection method based on random forests. Unlike traditional minimal-optimal feature selection methods (e.g., LASSO regression) that aim to identify the smallest feature subset for optimal prediction, the Boruta algorithm offers significant advantages in nonlinear modeling, handling mixed-type variables, resilience to multicollinearity, and operational stability (35–37). These capabilities allow it to identify all relevant features. By comparing the importance of original features against a set of randomly generated “shadow” features, Boruta can robustly determine whether each original feature has a true association with the outcome and rank the key features. This makes it particularly suitable for highly complex, multi-variable interaction scenarios such as healthcare (35–37). Three multivariate-adjusted Cox regression models were constructed to evaluate the associations of the CALLY index and PNI with mortality in ADHF patients. Covariates included in the models were gender, age, hypertension, diabetes, stroke, CHD, NYHA classification, drinking status, smoking status, LVEF, neutrophil count, monocyte count, RBC, PLT, AST, Cr, BUN, UA, FPG, and NT-proBNP. To ensure comparability of the associations between PNI and CALLY index with mortality, we further evaluated the relationship between quartile groups of both indices and study outcomes, calculating corresponding hazard ratios (HRs) and 95% confidence intervals. Notably, all included covariates passed multicollinearity testing (Supplementary Tables 3, 4). Additionally, visual assessment of survival curves associated with PNI and the CALLY index (Figure 2) confirmed the applicability of the proportional hazards assumption.

Following confirmation of the roles of the PNI and CALLY index in 30-day mortality risk assessment, we further constructed receiver operating characteristic curves to systematically assess and compare the predictive performance of these indices and their constituent components for 30-day mortality events in ADHF patients. Corresponding area under the curve (AUC) values, specificity, sensitivity, and optimal thresholds were calculated. In addition, we further compared the predictive performance of the CALLY index, the PNI, and the established ADHERE (Acute Decompensated Heart Failure National Registry) risk score for short-term mortality (38). The AUC and net reclassification improvement were calculated to evaluate their incremental predictive performance.

We employed a 4-knot restricted cubic spline to model the dose–response relationship between the CALLY index and 30-day mortality risk in ADHF patients. A likelihood ratio test was used to examine potential nonlinear effects. Upon detecting a nonlinear association, we employed recursive algorithms to identify inflection points where risk significantly changed. Subsequently, piecewise Cox regression models were constructed to quantify the strength of associations between the CALLY index and short-term mortality risk in ADHF patients on either side of these inflection points.

Finally, we conducted stratified analyses to evaluate the association between the CALLY index and 30-day mortality risk in ADHF patients. Stratification factors included age (grouped by median value), gender, LVEF (categorized using a 50% cutoff), and comorbidities (hypertension, diabetes, stroke, and CHD). Given that multiple subgroup comparisons were conducted in this study, we applied the Bonferroni correction to control for the inflation of type I errors caused by multiple comparisons. A total of 14 independent tests were involved in this analysis; therefore, the statistical significance threshold for subgroup results was set at p < 0.0036. Likelihood ratio tests were applied to compare differences across strata and assess the presence of interaction effects.

To test the robustness of our findings, we further evaluated the association between the CALLY index and 30-day mortality risk in ADHF patients under various priori assumptions:

Table 1 displays the baseline characteristics of the study population stratified by CALLY index quartiles. The median age of participants was 71 years, with a male predominance (58.04% vs. 41.96%). The cohort predominantly comprised individuals from Jiangxi Province, China. Compared to participants with higher CALLY index scores, those with lower scores were generally older, had a higher prevalence of chronic comorbidities (CHD, diabetes, stroke, and hypertension), were more likely to be male and smokers, and exhibited higher levels of CRP, white blood cell count, neutrophil count, monocyte count, AST, BUN, Cr, UA, FPG, NT-proBNP, alongside lower levels of RBC and lymphocyte count (Table 1).

During a median observation period of 30 days, 109 (8.9%) participants experienced mortality events. Mortality rates stratified by CALLY index quartiles were 21.90, 9.15, 2.94, and 1.63% while those for PNI quartiles were 20.92, 6.86, 5.23, and 2.61%. Kaplan–Meier analysis further visualized the 30-day survival curves for CALLY index and PNI quartile groups, revealing that patients in the lower CALLY index groups (Q1 and Q2) exhibited relatively worse survival outcomes compared to the lower PNI groups (Q1 and Q2) (Figure 2).

This study employed the Boruta algorithm for feature selection, utilizing shadow features as reference benchmarks to systematically compare the Z-scores of actual features against those of shadow features. Variables with significantly higher Z-scores than shadow features were labeled green (critical features), while those without significant differences were labeled red (non-critical features). Variables falling in the yellow zone represented tentative factors. Using the Boruta algorithm, we identified 18 variables most significantly associated with 30-day mortality risk in ADHF patients (Figure 3, green zone). The analysis revealed that the CALLY index emerged as one of the most influential factors affecting 30-day mortality in ADHF patients, exhibiting higher predictive importance than the PNI and ranking second only to FPG.

Table 2 summarizes the associations between the CALLY index, PNI, and 30-day mortality in ADHF patients. The study demonstrated inverse associations between both the CALLY index and PNI with 30-day mortality risk in ADHF patients across unadjusted and adjusted models: Specifically, each unit increase in the CALLY index was associated with an 8% reduction in 30-day mortality risk (HR 0.92, 0.86–0.97); similarly, each unit increase in PNI corresponded to a 6% risk reduction (HR 0.94, 0.91–0.97). To better quantify the impact of the CALLY index and PNI on 30-day mortality risk in ADHF patients, we calculated the mortality risks stratified by quartiles of these parameters. The results demonstrated that compared to the highest CALLY index quartile (Q4), patients in the lowest quartile (Q1) exhibited a 321% increased risk of 30-day mortality (HR 4.21, 1.59–11.13). Similarly, compared to the highest PNI quartile (Q4), those in the lowest PNI quartile (Q1) had a 232% increased risk (HR 3.32, 1.42–7.77). In summary, lower CALLY index and lower PNI were independent risk factors for short-term mortality risk in ADHF patients, with lower CALLY index conferring a posing mortality risk than lower PNI.

We evaluated the predictive accuracy of PNI, the CALLY index, and their constituent components for 30-day mortality in ADHF patients by calculating the AUC, sensitivity, specificity, and optimal thresholds. As shown in Table 3, the CALLY index demonstrated the highest predictive accuracy for short-term mortality events compared to PNI and individual components of the CALLY index (Figure 4; AUC = 0.80), with an optimal threshold determined to be 2.79.

As shown in Supplementary Tables 5, 6, we compared the predictive performance of the CALLY index model against the PNI and ADHERE models. The analysis revealed that the CALLY index model achieved an AUC of 0.80, significantly outperforming the PNI (AUC = 0.74) and ADHERE (AUC = 0.61) models (DeLong test, p < 0.05). Continuous net reclassification analysis demonstrated that, compared with the PNI model and ADHERE model, the CALLY index model achieved a significant net improvement, with a net reclassification improvement greater than 0.2 (p < 0.05).

We further constructed a restricted cubic spline curve to visualize the association between the CALLY index and 30-day mortality risk in ADHF patients (Figure 5). The results revealed a strong correlation between low CALLY index levels and elevated 30-day mortality risk. Notably, nonlinearity testing demonstrated a significant nonlinear association between the CALLY index and 30-day mortality in ADHF patients (p for nonlinearity <0.001). This nonlinear pattern was further characterized by a saturating effect (L-shaped pattern), where 30-day mortality risk plateaued after the CALLY index surpassed a certain threshold. Through recursive algorithm, we identified an inflection point at 3.14 for the association between the CALLY index and 30-day mortality risk, with segmented Cox regression analyses (Table 4) revealing distinct risk patterns across this threshold: pre-inflection (CALLY index <3.14), each unit increase was associated with a 39% reduction in mortality risk (HR 0.61, 0.49–0.75), whereas post-inflection (CALLY index >3.14), the risk reduction became non-significant at 1% (HR 0.99, 0.95–1.03).

We conducted exploratory subgroup analyses to evaluate the association between the CALLY index and 30-day mortality in ADHF patients. Subgroups were stratified by age, gender, LVEF, and comorbidities (hypertension, diabetes, stroke, and CHD). The findings demonstrated that the association between the CALLY index and 30-day mortality in ADHF patients remained robust in the majority of subgroups (Table 5). However, after further Bonferroni correction (for 14 comparisons), a significant association between the CALLY index and 30-day risk was observed only in ADHF patients with diabetes (p < 0.0001). In addition, further interaction tests revealed a significant difference in the CALLY index-associated 30-day mortality risk among the diabetic subgroup (p-interaction <0.01): Compared to ADHF patients without diabetes, those with diabetes exhibited a significantly stronger inverse association between the CALLY index and short-term mortality outcomes (HR: diabetes 0.73 vs. nondiabetic 0.95). This suggests that diabetes significantly amplifies the mortality risk associated with low CALLY index scores in ADHF patients.

We further conducted age-adjusted nonlinear analyses, special population analyses, temporal sensitivity analyses, and data integrity assessments to evaluate the robustness of the association. The robustness of the association between the CALLY index and 30-day mortality in ADHF patients was confirmed, remaining largely unchanged even after incorporating an age-squared term (Table 6: sensitivity-1). After controlling for potential reverse causation, the association pattern between the CALLY index and ADHF patients remained unchanged (Table 6: sensitivity-2). Following additional adjustment for frailty as a potential confounder, the primary findings remained robust with no substantive alterations (Table 6: sensitivity-3). Finally, replication of the primary analysis in a multiple-imputed complete dataset confirmed the results’ robustness (Table 6: sensitivity-4).

This study possesses several notable strengths: First, the research topic demonstrates innovation, as the CALLY index incorporates indicators that are readily accessible in clinical practice and derived from large-sample cohort data, endowing it with both clinical utility and translational potential for short-term prognosis assessment in ADHF patients. Second, a rigorous study design was implemented, encompassing a multidimensional validation strategy that includes subgroup analyses, temporal sensitivity testing, and data quality validation. This comprehensive approach substantially strengthens the robustness of our findings.

This study has several limitations: (1) While this study elucidated the prognostic utility of the baseline CALLY index for short-term mortality risk in ADHF patients, it did not explore the dynamic evolution of this index during hospitalization. Future research is recommended to focus on the temporal evolution characteristics of the CALLY index and evaluate its dynamic association with clinical prognosis. The specific approaches are as follows: (i) Increase the frequency of monitoring Alb, lymphocyte count, and CRP during hospitalization, and assess the trajectory of the CALLY index based on these repeated measurement data; (ii) Conduct regular follow-ups and perform trajectory analysis to assess the dynamic relationship between the CALLY index and clinical outcomes. (2) Although we adjusted for numerous potential confounders—including smoking/drinking status, comorbidities, and cardiac function—residual confounding may persist and potentially influence the results (87). (3) Since CRP is not routinely measured in clinical practice for HF patients, a relatively high proportion of participants were excluded due to missing CRP data, which may introduce a certain degree of selection bias. Although a comparison of baseline characteristics between those with and without missing data confirmed that the data were missing at random, the possibility of Missing Not at Random cannot be entirely ruled out. Large-scale multicenter cohort studies are therefore warranted to validate our results. (4) As a retrospective cohort analysis (non-interventional observational study design), this research inherently carries methodological limitations: First, the study framework cannot evaluate comparative clinical efficacy across various therapeutic regimens administered post-admission in ADHF patients. Second, observational data analysis only permits correlational inference between the CALLY index and outcomes, precluding the establishment of causal relationships between therapeutic interventions and clinical endpoints (88). These findings essentially reflect biomarker fluctuation patterns during the natural disease progression. (5) This study has geographical limitations: as the study population was predominantly recruited from Jiangxi Province in southern China, the generalizability of conclusions to geographically diverse regions in northern China and ethnically diverse populations requires confirmation through multicenter, cross-regional validation studies. Furthermore, as our study was conducted in a hospital-based setting, the generalizability of our findings may be limited to Chinese populations with similar healthcare-seeking behaviors. Differences in culture, diet, and healthcare systems between China and other regions may affect the applicability of our results to other populations. (6) A large number of subgroup analyses were conducted in the current study, which increases the risk of type I errors (i.e., false-positive findings). Although these analyses provide valuable insights for exploring potential effect modifiers, their results are based on a relatively small sample size and significant outcomes were only observed in the diabetes subgroup, and thus should be interpreted with caution and require further validation in independent cohorts.