Introduction

Heart failure (HF) is a prevalent clinical syndrome associated with high rates of hospitalization and mortality [1–3]. HF readmissions represent a substantial burden on healthcare resources and are associated with significant costs and worse patient outcomes [4–6]. Reducing readmissions has therefore been recognized as a national priority in the United States, exemplified by initiatives such as the Hospital Readmissions Reduction Program (HRRP) [7]. This has prompted research efforts to identify predictors of readmission that could inform pre-discharge checklists or post-discharge care strategies aimed at lowering readmission rates. Common comorbidities in HF patients, such as atrial fibrillation (AF), chronic kidney disease (CKD), and chronic obstructive pulmonary disease (COPD), have been consistently associated with an elevated risk of adverse outcomes, including hospital readmissions [8–10].

However, a recent study by Scholten et al. challenged the importance of these comorbidities as predictors of readmission in HF patients. Their study found that most commonly recognized comorbidities were not significantly associated with increased 100-day readmission risk, and even for comorbidities that did show significant associations, such as AF, CKD, and COPD, their predictive power was limited, as demonstrated by low area under the curve (AUC) values in receiver operating characteristic (ROC) analyses [11]. These findings raise important questions about the clinical value of traditionally recognized predictors of HF readmission. Furthermore, since most prior studies focused exclusively on the conventional outcome measure of 30-day readmissions, it remains unclear whether the limited predictive ability demonstrated by Scholten et al. at 100 days is similarly evident at this earlier standard time-point.

To address these questions, we elected to reassess the association and predictive power (evaluated by ROC-AUC analysis) of traditional comorbidities linked to HF readmission at 30 and 100 days, in a large cohort of patients admitted with acute decompensated HF (ADHF) and followed for an extended period.

Methods

Results

Patient characteristics

The study cohort included 7,652 patients admitted with ADHF. Of these, 1,670 patients (21.8%) were readmitted within 30 days of discharge, and 2,959 patients (38.7%) were readmitted within 100 days.

Baseline demographic characteristics and comorbidity distributions stratified by 30-day and 100-day readmission status are presented in Table 1. Age did not differ significantly between readmitted and non-readmitted patients at either timepoint, with an overall mean of 76.0 ± 12.0 years. Sex distribution was also similar at 30 days (47.4% vs. 49.8% female, p = 0.078) but differed at 100 days, with a lower proportion of females among those readmitted (47.8% vs. 50.2%, p = 0.035). Several comorbidities were consistently more prevalent among readmitted patients. Notably, CKD was present in 38.1% of patients readmitted within 30 days compared to 32.3% of those not readmitted (p < 0.001), and in 39.1% vs. 30.1% at 100 days (p < 0.001). Anemia showed a similarly strong pattern (71.0% vs. 65.0% at 30 days, p < 0.001; 71.2% vs. 63.3% at 100 days, p < 0.001). Other comorbidities, including IHD, AF, DM, and COPD, were also significantly more prevalent among readmitted patients at both timepoints. PVD was not associated with 30-day readmission but showed higher prevalence at 100 days (8.2% vs. 6.6%, p = 0.009).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Comorbidities, Age, and Sex Stratified by 30-Day and 100-Day Readmission Status.

https://doi.org/10.1371/journal.pone.0329829.t001

Comorbidity overlap

Beyond individual comorbidity characteristics, extensive overlap was observed between conditions, as illustrated in Fig 1. A prominent cluster was evident among vascular-related comorbidities, with DM co-occurring in 74.5% of patients with PVD, 64.0% of those with CKD, and 57.3% of those with IHD. Anemia demonstrated broad overlap across all comorbidities, present in 67.7% of patients with CKD, 66.6% with PVD, and 63.4% with DM. In contrast, COPD showed consistently low overlap with other conditions, ranging from 9.6% to 21.3%, indicating relative isolation from the vascular cluster. Notably, overlaps were often asymmetric—for example, while 64.0% of patients with CKD also had DM, only 42.2% of DM had CKD. Similarly, 59.6% of patients with IHD had anemia, compared to just 41.8% of those with anemia having IHD. The strongest asymmetry was observed between PVD and IHD: 63.0% of patients with PVD had IHD, while only 11.2% of IHD patients had PVD.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Overlap Between Comorbidities in the Study Cohort.

Heatmap illustrating the percentage of patients with each comorbidity (rows) who also had a second comorbidity (columns). Values represent row-wise percentages; that is, each cell indicates the proportion of patients with the row condition who also had the column condition. A prominent vascular cluster was observed among diabetes mellitus, ischemic heart disease, chronic kidney disease, and peripheral vascular disease. Anemia demonstrated broad overlap across most conditions, while chronic obstructive pulmonary disease (COPD) showed relatively low overlap with others, suggesting relative isolation.

https://doi.org/10.1371/journal.pone.0329829.g001

Association between comorbidities and readmission risk

Subdistribution hazard ratios (sHRs) with 95% confidence intervals (CIs) for 30-day and 100-day readmissions are presented in Table 2. At 30 days, CKD (sHR 1.16, 95% CI 1.04–1.28, p = 0.007), AF (sHR 1.19, 95% CI 1.08–1.32, p < 0.001), COPD (sHR 1.20, 95% CI 1.06–1.35, p = 0.005), and anemia (sHR 1.20, 95% CI 1.07–1.36, p = 0.002) were significantly associated with higher readmission risk. IHD showed a borderline association (sHR 1.10, 95% CI 1.00–1.22, p = 0.055), while DM, PVD, female sex, and age were not significantly associated with 30-day readmission.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Fine-Gray Competing Risks Models for 30-Day and 100-Day Readmissions.

https://doi.org/10.1371/journal.pone.0329829.t002

At 100 days, the pattern was similar but with additional significant associations. CKD (sHR 1.23, 95% CI 1.13–1.32, p < 0.001), AF (sHR 1.25, 95% CI 1.16–1.35, p < 0.001), COPD (sHR 1.31, 95% CI 1.20–1.44, p < 0.001), anemia (sHR 1.24, 95% CI 1.13–1.35, p < 0.001), IHD (sHR 1.15, 95% CI 1.07–1.24, p < 0.001), and DM (sHR 1.09, 95% CI 1.01–1.17, p = 0.030) were significantly associated with increased readmission risk at this timepoint. PVD, female sex, and age were not significantly associated with 100-day readmission.

Predictive power of comorbidities for readmission

The ROC AUC was assessed for each comorbidity that remained independently significant in the Fine-Gray multivariable competing risks models. For 30-day readmission (Fig 2a), AUC values for individual comorbidities ranged from 0.516 (lowest, for COPD) to 0.529 (highest, for CKD). Incremental combinations of comorbidities were then tested, starting with the two strongest predictors—anemia and CKD—adding DM as a third, and finally IHD as the fourth. The maximal AUC achieved for 30-day readmission, with all four comorbidities combined, was 0.555.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Receiver Operating Characteristic (ROC) Curves for Readmission Prediction by Comorbidities.

AUC values are shown for individual comorbidities that remained independently significant in the multivariable Fine-Gray models, as well as for their incremental combinations. For 30-day readmission (Figure a), the maximal AUC achieved was 0.555, combining the four significant comorbidities: CKD, COPD, DM, anemia. For 100-day readmission (Figure b), the maximal AUC achieved was 0.593, combining the six significant comorbidities: CKD, DM, anemia, ischemic heart disease, peripheral vascular disease, and COPD.

https://doi.org/10.1371/journal.pone.0329829.g002

For 100-day readmission (Fig 2b), a similar pattern was observed. AUC values for the individual six comorbidities that were found to be significant in this timeframe, ranged from 0.528 (lowest) for DM, to 0.545 (highest) for CKD. The maximal AUC achieved for 100-day readmission, with all six comorbidities combined, was 0.593.

Discussion

In this study, we demonstrated that despite well-established associations between common comorbidities and HF readmission risk, their ability to predict which patients will be readmitted remained poor. At both 30 and 100 days after admission for ADHF, the discriminative power of these comorbidities, assessed by AUC analysis, did not exceed 0.6. These findings challenge the clinical utility of these comorbidities as effective tools for readmission risk stratification.

Our results are consistent with the findings of Scholten et al., who similarly reported low predictive power of common comorbidities for cardiovascular readmissions within 100 days among HF patients [11]. Their analysis, like ours, identified statistically significant associations for several comorbid conditions, yet these translated into weak discrimination when assessed by ROC-AUC. In their study, the maximal AUC achieved was 0.63 (95% CI 0.61–0.64) using logistic regression and 0.62 (95% CI 0.60–0.64) using Rasch analysis, both reflecting poor predictive performance. Other studies evaluating predictive models for HF readmission using comorbidities as input variables have reported similarly limited discriminative power, with AUC values generally ranging from 0.60 to 0.65 [9,10]. These findings align with our analysis, in which maximal AUC values were 0.555 for 30-day readmission and 0.593 for 100-day readmission, slightly lower but broadly comparable. The modest differences between studies may reflect variations in outcome definitions—such as cardiovascular-specific versus all-cause readmissions—as well as differences in modeling approaches, including our use of Fine-Gray competing risks regression to account for death as a competing event. Of note, the consistently poor discriminative performance across these varying methodologies and outcome definitions reinforces the conclusion that comorbidity profiles alone, regardless of analytic strategy, offer limited utility for predicting HF readmissions. The present study further supports this conclusion by demonstrating that the limited discriminative ability is not confined to the 100-day time frame but is already evident at the conventional 30-day readmission endpoint. Moreover, by assessing both individual comorbidities and their incremental combinations, our findings specifically tested whether additive modeling of these variables could improve risk stratification—an approach not the primary focus of previous studies—and found no substantial improvement.

While direct evaluations of predictive power remain limited, numerous observational studies and meta-analyses have consistently reported strong associations between comorbidities such as CKD, DM, IHD, and AF and adverse HF outcomes, including hospital readmissions [8,21,22]. However, as we showed, these associations, although statistically robust, do not necessarily imply strong predictive power. Several factors may contribute to this phenomenon. One is that these comorbidities are so prevalent in HF populations that they fail to meaningfully separate high- from low-risk individuals. When nearly half of all patients have a condition like DM or CKD, it raises average risk, but leaves little discriminatory signal between those who will and won’t be readmitted. Another factor is that these comorbidities reflect baseline vulnerability, but not the proximate causes of readmission, which often result from acute, dynamic events such as infections, arrhythmias, or lapses in adherence [23,24]. As a result, comorbidities may explain who is at risk for worse overall outcomes, but fail to identify who will cross the threshold into a specific readmission event within a defined timeframe.

Aside from these clinical explanations, our findings highlight a fundamental distinction between regression modeling and prediction analysis. Regression models test whether a variable shifts average risk across a population, but predictive discrimination depends on how well it separates individual outcomes. Even when mean risk differs significantly between groups, wide within-group variance can cause heavy overlap, limiting classification accuracy. Importantly, this within-group variance is not reflected by the 95% confidence interval (CI) typically reported in regression models, which describes the precision of the group mean estimate, not the spread of individual risks around that mean. Thus, a narrow 95% CI may coexist with wide individual-level variance that overlaps with risks in the comparator group, limiting discrimination and lowering predicting power [25,26]. Our study underscores the importance of reporting predictive metrics, not just regression outputs, when evaluating potential risk markers.

The selection of the 30-day and 100-day timeframes in the present analysis warrants consideration. While the 30-day period remains the standard benchmark in HF readmission research and policy, its exclusive use has been questioned. Magaña et al. introduced the concept of phased post-discharge vulnerability, identifying and describing a triphasic structure of risk following HF hospitalization: early (0–30 days), intermediate (31–90 days), and late (91–180 days) phases comorbidities in HF. In the current study, for a second timeframe beyond the early 30-day timeframe, we chose the 100-day timeframe as a pragmatic balance, aligning with the timeframe Scholten et. al chose, and aiming to capture both the intermediate and late phases within a single analysis window. Although repeating the analysis separately for these three phases might offer further granularity, we considered this beyond the scope of our study, which focused on evaluating the overall predictive power of comorbidities rather than the temporal evolution of readmission risk. The similarly poor predictive performance observed at both 30 and 100 days suggests that the limitations of comorbidity-based stratification are not restricted to a specific follow-up duration.

In addition to evaluating predictive performance, our study also identified comorbidity clustering patterns that may help illustrate the biological heterogeneity of HF. Notably, the relative isolation of COPD from vascular comorbidities such as DM, IHD, and CKD suggests that COPD may contribute to HF through distinct mechanisms. While vascular comorbidities are predominantly driven by atherosclerosis, endothelial dysfunction, and metabolic dysregulation, [27,28]. COPD is more closely associated with chronic hypoxia, pulmonary vascular remodeling, and right ventricular strain—features that may give rise to a different form of cardiac stress within the HF spectrum [29,30]. These contrasting pathways may help explain the looser clustering of COPD with vascular conditions in our cohort, despite shared systemic risk factors such as smoking and inflammation. Conversely, the broad overlap of anemia across comorbidity groups likely reflects its role as a nonspecific marker of overall disease burden, potentially driven by chronic inflammation, nutritional deficiencies, or renal impairment [31,32]. These structural patterns strengthen the notion that considering biological context and phenotypic clustering—rather than simple comorbidity counts—may be necessary to advance prognostic modeling and care stratification in HF. It is also worth noting that advances in heart failure pharmacotherapy may influence the downstream effects of comorbidities. For instance, sacubitril/valsartan has been associated with improvements in physical frailty among patients listed for heart transplantation, highlighting the potential of contemporary treatments to mitigate the clinical impact of baseline vulnerability [33]. Such developments further complicate the interpretation of static comorbidity profiles, emphasizing the need for risk models that reflect the evolving therapeutic landscape.

Our findings, together with those of Scholten et al., bring into sharp focus a critical question: should we abandon comorbidity-based risk models and look for other, perhaps more dynamic and individualized alternatives? Several recent modeling approaches suggest alternative or complementary paths. Frailty-based models, particularly those using multidimensional constructs encompassing cognitive, social, and physical domains, have shown stronger predictive value than traditional comorbidity counts [34–36]. Machine-learning tools using administrative and clinical data—such as eXtreme Gradient Boosting (XGBoost)—have reported improved accuracy in predicting HF readmissions, with AUCs reaching 0.763 [37]. Additional approaches have explored other predictors such as natriuretic peptides, [38] functional status metrics, [39] or medication adherence [40] as dynamic indicators of short-term vulnerability. While these tools each have limitations—ranging from complexity to generalizability—they reflect a clear shift toward multidimensional, patient-centered models. In this context, comorbidities may still serve a useful role, but likely as components of integrated frameworks rather than standalone predictors.

This study has several limitations. First, its retrospective design and reliance on administrative coding may have led to misclassification of comorbidities or outcomes. Second, the analysis was conducted at a single center, which may limit generalizability to other populations or healthcare settings. Third, we focused on comorbidity profiles and did not include additional variables such as laboratory data, functional measures, medication adherence, or social determinants of health—all of which may contribute meaningfully to readmission risk and are increasingly incorporated into contemporary risk prediction models to capture the dynamic and multidimensional nature of patient vulnerability. Fourth, ejection fraction (EF) phenotyping (e.g., HF with reduced vs. preserved EF) was not included in the analysis because EF data were available for only 31% of patients—insufficient for inclusion in the Fine-Gray competing risk model used in this study and too limited to support even univariable exploration of comorbidity patterns by phenotype. Finally, although multiple imputations were applied to address missing data for anemia status, residual confounding from unmeasured or imperfectly measured variables cannot be excluded.

In conclusion, while traditional comorbidities in patients with ADHF remain independently associated with increased readmission risk, their overall predictive value is limited, even when combined. These findings underscore the limitations of comorbidity-based risk models when used in isolation and highlight the need for more comprehensive predictive strategies that incorporate dynamic clinical, behavioral, and physiologic dimensions.