We used data from the Kochi YOSACOI study, which enrolled 1,061 consecutive patients with acute decompensated HF (ADHF) in Kochi, Japan, from May 2017–December 2019. In addition, we used data on clinical outcomes for all-cause mortality within 2 years with follow-ups through December 2021. Details of the Kochi YOSACOI study have been described previously (18). Briefly, the Kochi YOSACOI study was a collaborative effort of six hospitals that provide acute care for cardiovascular diseases in Kochi Prefecture, where the percentage of people over 65 years of age has reached 35%. All participating hospitals practiced acute HF treatment according to standard guidelines (11). Eligibility criteria for inclusion in the registry were age ≥20 years and hospitalization for ADHF at one of the participating hospitals. Based on the Framingham criteria, ADHF was diagnosed by the presence of at least two major criteria, including symptoms and physical examination, chest radiography, and echocardiographic findings, or the presence of one major and two minor criteria.

This study was approved by the Medical Research Ethics Committee of Kochi University of Medical Science (Approval No. 28–68) and the Medical Research Ethics Committee of Tokushima University Graduate School of Biomedical Sciences (Approval No. Z120). The study complied with the tenets of the Declaration of Helsinki, and informed consent was obtained from all patients or their families. Confidentiality and anonymity of patient data were maintained throughout.

The clinical characteristics of the patients have been described previously in detail (18). Data were collected by investigators at participating hospitals during the enrollment period. We obtained information on patient demographics, etiology of HF, medical history, long-term treatment, HF symptoms and vital signs at admission and discharge, discharge prescriptions, laboratory and echocardiographic data, and other clinical parameters. We used echocardiographic data from the time when HF status was stabilized during hospitalization. After HF stabilization, left ventricular ejection fraction (LVEF) was echocardiographically determined during hospitalization, and HF with an LVEF of ≤40% was categorized as HFrEF. Patients' nutritional status was assessed using the Geriatric Nutritional Risk Index (GNRI), a simple measure of nutritional status in older adults, calculated using the following equation: GNRI = 14.89 × serum albumin (g/dl) + 41.7 × body mass index/22 (19). RAS inhibitor use was defined by the prescription of angiotensin-converting enzyme inhibitors or angiotensin II receptor blockers. The RAS inhibitor and beta-blocker use was defined as patients receiving a prescription for RAS inhibitor (enalapril, lisinopril, and candesartan) and beta-blockers (carvedilol and bisoprolol) at the time of discharge. Because patients with HFrEF in this study were mainly older adults, MRAs were assumed to be used less frequently because of the risk of adverse events due to hyperkalemia when MRAs are added to RAS inhibitors. Therefore, we did not include MRAs in this study. In addition, ARNI and SGLT 2 inhibitors were not approved in Japan during the enrollment period (20).

Of the patient cohort (n = 1,061), we excluded 38 patients with missing all-cause mortality data within 2 years, 43 with missing drug data, and 97 with missing LVEF data at admission. Of the remaining 883 patients, all patients with HFrEF (n = 314) were included and divided into two groups according to all-cause mortality within 2 years: the all-cause mortality within 2 years group (n = 80) and the survivor group (n = 234; Figure 1).

We performed univariate and multivariate analyses of the association between RAS inhibitor and beta-blocker use and 2-year mortality. The interaction of RAS inhibitor and beta-blocker use with age was also analyzed to determine whether the efficacy of these medications depends on age. Additionally, all patients with HFrEF (n = 314) were divided into two groups according to cardiac death within 2 years: cardiac death within 2-year group (n = 41) and survival group (n = 273). Furthermore, Kaplan–Meier survival analysis was used to assess the effects of RAS inhibitors and beta-blockers on 2-year all-cause mortality and 2-year cardiac mortality, respectively, stratified by age.

All data are expressed as medians with interquartile ranges (IQRs) for non-normally distributed variables or frequencies (percentages) for categorical data. Differences in continuous variables were assessed using Student's t-test or the Mann–Whitney U test. Pearson's chi-squared test was used to compare categorical variables, and Fisher's exact test was used when the expected frequency was <5. Hazard ratios (HRs) and 95% confidence intervals (CIs) were also determined using logistic regression analysis. A P-value < 0.05 was indicated significance for all tests. Cumulative event rates were assessed using the Kaplan–Meier method. Multivariate Cox regression analysis was used to evaluate the adjusted relative risk of the variables. To investigate the nonlinear associations and interactions among risk factors associated with all-cause mortality in patients with HFrEF via an ML model, we used Light Gradient Boosting Machine (LightGBM), a variant of Gradient Boosting Decision Trees (GBDTs) (21). ML techniques offer the advantage of enabling the identification of important variables, including potentially obscured variable interactions and data patterns that contribute to patient outcomes. These may not be easily discovered by researchers using traditional statistical approaches (22). We developed a mortality prediction model for patients with HFrEF using previously described variables as explanatory variables and assessed the occurrence of all-cause mortality as the object variable. ML models are known for their high predictive accuracy, but they can also overfit. Overfitting occurs when a model performs well on the data used to develop it but fails to predict unseen data. To mitigate this risk, evaluating the model's performance on an independent dataset is important. Therefore, when building the mortality prediction model, we randomly allocated 80% of the dataset for model development and reserved the remaining 20% for validation. We evaluated the ML models using accuracy verification data and confirmed that the model in this study was unlikely to be overfitting. We used Shapley additive explanations (SHAP) values to quantify the influence of interactions between features (23). The primary focus of this analysis was the interaction between age and other variables. Using SHAP interaction plots, we explicitly showed how age and the variables of interest affect predictions and explored interaction effects that might modify the mortality risk associated with age. Statistical analyses were conducted using the R (version 3.3.1, R Foundation for Statistical Computing, Vienna, Austria. http://www.Rproject.org) and Python (version 3.8.10, Python Software Foundation, https://www.python.org) software.

We analyzed 314 patients with HFrEF. The median age was 76 (IQR, 68.0–83.0) years, 41.1% (129/314) were aged ≥80, 57.6% (181/314) were prescribed RAS inhibitors, and 79.0% (248/314) were prescribed beta-blockers (Table 1). Moreover, 25.5% (80/314) experienced all-cause mortality within 2 years. Patients in the all-cause mortality group had a higher age, lower GNRI, higher rate of New York Heart Association (NYHA) class III/IV at discharge, higher rate of a prior HF-related admission, and higher rate of chronic obstructive pulmonary disease (COPD) than did patients in the survivor group. Patients in the all-cause mortality group had lower estimated glomerular filtration rate (eGFR) and sodium levels but higher B-type natriuretic peptide (BNP) levels than did patients in the survivor group. Patients in the all-cause mortality group had lower proportions of RAS inhibitor (35% vs. 65.4%; P < 0.001) and beta-blocker (60.0% vs. 85.1%; P < 0.001) use than did patients in the survivor group.

We investigated factors related to all-cause mortality in patients with HFrEF (Table 2). The univariate analysis showed that advanced age, GNRI, NYHA class III/IV at discharge, prior HF-related admission, COPD, high BNP levels, low eGFRs, low sodium levels, RAS inhibitor use, and beta-blocker use were associated with all-cause mortality. The multivariate analysis showed that advanced age (adjusted HR, 1.049; 95% CI, 1.022–1.078; P < 0.001), prior HF-related admission (adjusted HR, 1.873; 95% CI, 1.160–3.025; P = 0.01), and RAS inhibitor use (adjusted HR, 0.436; 95% CI, 0.26–0.729; P = 0.002), but not beta-blocker use (adjusted HR, 0.807; 95% CI, 0.464–1.406; P = 0.45; Table 2), were independently associated with all-cause mortality.

The variables RAS inhibitor and beta-blocker use were tested for their interactions with the variable age. The interaction between RAS inhibitor use and age was not significant (P = 0.814) (Table 3), but that between beta-blocker use and age was significant (P = 0.031). The results of the subgroup analyses for the interactions of age showed a differential effect of age on RAS inhibitor and beta-blocker use among patients with HF.

GBDT models were constructed for each patient to predict the probability of all-cause mortality within 2 years. The GBDT model's performance, evaluated through the Area Under the Curve (AUC) for the validation data, was 0.72 (95% CI, 0.568–0.871). The performance shows that our GBDT models were not vulnerable to overfitting. SHAP values were calculated for each patient-derived model to quantify feature importance and interaction effects. Calculation and ranking of SHAP values for each variable of patient data revealed that older age, low sodium levels, high BNP levels, low GNRI, prior HF-related admission, RAS inhibitor use, and beta-blocker use at discharge were the most important features for a predicted increased all-cause mortality within 2 years across patients with HFrEF (Figure 2).

When examining the SHAP interaction values, an interaction was noted between age and beta-blocker use (Figure 3). In patients aged 80 years and older, the risk of mortality within 2 years differed depending on whether beta-blockers were used or not. Specifically, in patients over 80 years of age, not using beta-blockers slightly reduced the age-related risk of mortality. Thus, compared with no use, beta-blocker use increased the risk of age-related mortality. In other words, the results of this analysis showed that the use of β-blockers altered the risk of mortality, either by increasing or decreasing it, with a threshold of approximately 80 years of age.

We performed an age-dependent subgroup analysis of the effects of RAS inhibitor and beta-blocker use on all-cause mortality. RAS inhibitor use was associated with a reduced risk in both age groups. By contrast, beta-blocker use was associated with a reduced risk in the age <80 years group (adjusted HR, 0.486; 95% CI, 0.247–0.956; P = 0.037), but the risk tended to increase in the age ≥80 years group (adjusted HR, 1.966; 95% CI, 0.97–3.985; P = 0.061; Table 4).

Kaplan–Meier survival analysis showed that RAS inhibitor and beta-blocker use reduced all-cause mortality in all age groups (P < 0.001; Figures 4A-1,B-1). Furthermore, RAS inhibitors significantly reduced all-cause mortality in patients aged <80 years and in those aged ≥80 years (P < 0.001; Figures 4A-2,A-3). Contrastingly, beta-blocker use significantly reduced all-cause mortality in patients aged <80 years (P < 0.001, Figure 4B-2) but not in those aged ≥80 years (P = 0.319, Figure 4B-3).

On the other hand, RAS inhibitor use significantly reduced the incidence of cardiac death (P < 0.001; Figure 5A-1), while beta-blocker use showed no significant differences in all age groups (P = 0.28; Figure 5B-1). Furthermore, RAS inhibitors significantly reduced cardiac death in patients aged <80 years and in those aged ≥80 years (P < 0.001; Figures 5A-2,A-3). Contrastingly, beta-blocker use significantly reduced cardiac death in patients aged <80 years (P < 0.001, Figure 5B-2) but not in those aged ≥80 years (P = 0.246, Figure 5B-3).