Risk factor clusters driving hospitalization in heart failure
M Popescu, M T Tzikas, G C Ciorabai, S B BalanescuAbstract
Background
Heart failure (HF) is a major driver of recurrent hospitalizations (RH), yet predicting which patients will experience repeated admissions remains challenging. Many approaches focus on single comorbidities or isolated laboratory markers, which may miss higher-risk multifactorial phenotypes.
Purpose
To identify clusters of clinical and biochemical risk factors associated with RH and to derive a pragmatic, admission-time risk score reflecting association to the highest-risk phenotype.Methods. In this single-center retrospective study, 5130 records of patients admitted for HF, between 2018–2022 were screened for ≥3 hospitalizations (n=188). Clinical and laboratory/echocardiographic variables (total 24 variables) were analyzed:(1) prevalence description; (2) relative risk (RR) estimation for individual pathologies for RH above the cohort median (>3 vs 3 hospitalizations); (3) correlation testing and regression modelling; and (4) unsupervised K-means clustering on standardized variables after dimensionality reduction (Principal Component Analysis – PCA. A proximity-to-high-risk score was calculated using Euclidean distance in standardized PCA space from the centroid of the highest-risk cluster, scaled to 0–10 using: score = 10 × (1 − distᵢ/dist_max) (lower score = closer to high-risk centroid).Results.The mean age was 73 years with 52% males.Individual pathology-based RR estimates showed only modest effect sizes (RR range 0.88–1.36) and none reached p < 0.05, suggesting limited predictive value for individual factors.Clustering supported a 3-cluster solution (silhouette ≈ 0.077), separating patients into: (i) a "balanced" profile, (ii) a cardiometabolic (comorbidity-heavy) profile, and (iii) a high-risk systemic dysfunction profile characterized by renal dysfunction, high natriuretic peptides, inflammation markers, reduced ejection fraction, anemia, and multi-organ dysfunction (highest mean RH rate (6.14) (Figure 1).The derived proximity score performance demonstrated a Spearman correlation with readmission count of ρ = −0.14 (p = 0.06) (sign reflecting inverse distance scaling), and moderate discrimination (AUC = 0.64). When grouped into risk strata, admission frequency showed a graded pattern (Table 1).
Conclusion
In this recurrently hospitalized HF cohort, single clinical or laboratory parameters did not accurately identify patients with RH. Clustering revealed a clinically coherent high-risk phenotype dominated by inflammation, reduced ejection fraction, anemia, and multi-organ dysfunction and associated with the highest readmission burden. A proximity-based clinical score derived from this phenotype showed moderate discrimination and may support admission-time risk stratification. Larger, prospective, multi-center validation, is needed, as well as standardization of score directionality and calibration, assessment of incremental value over existing HF readmission models (including treatment variables and longitudinal trajectories).Figure 1.Clustering K means.For image description, please refer to the figure legend and surrounding text.Table 1.Readmission burdenFor image description, please refer to the figure legend and surrounding text.