Why the study?
Patients at risk for or in pre-clinical stages of heart failure are heterogeneous, and standard clinical tools often fail to stratify early systemic vulnerability. Whether unsupervised machine learning could identify hidden clinico-metabolic phenotypes to predict long-term overall survival in this population was uncertain.
Does unsupervised machine learning clinico-metabolic phenotyping predict long-term all-cause mortality in an at-risk, pre-clinical heart failure population?
Does unsupervised machine learning clinico-metabolic phenotyping predict long-term all-cause mortality in an at-risk, pre-clinical heart failure population?
Unsupervised machine learning integrating clinical, metabolic, renal, inflammatory, and gut-microbiome biomarkers identifies three distinct phenotypes that strongly stratify long-term mortality risk (C-Index 0.823) in patients at-risk for or with pre-clinical heart failure.
Hypothesis-generating for ML phenotyping in pre-clinical HF; requires prospective validation before clinical adoption.
Background/Introduction Patients at-risk for or in pre-clinical stages of heart failure represent a highly heterogeneous primary prevention population. Standard clinical tools often fail to stratify early systemic vulnerability, leaving a critical gap in preventative cardiology. Purpose We aimed to determine if unsupervised machine learning could identify hidden clinico-metabolic phenotypes to predict long-term overall survival in a at-risk, pre-clinical heart failure population. Methods We defined a primary prevention cohort from NHANES (2005–2010) linked to NCHS mortality records through 2019. The cohort included patients with hypertension, diabetes or obesity, excluding overt cardiovascular disease (prior myocardial infarction, stroke, coronary heart disease or clinical heart failure). Following an 80/20 train/test split, missing variables were iteratively imputed. Unsupervised K-Means clustering (k=3) integrated age, HbA1c and log-transformed biomarkers of inflammation (hs-CRP), renal stress (UACR), and gut-microbiome dietary exposure (enterolactone). A survey-weighted Random Survival Forest (1,000 trees) modeled right-censored all-cause mortality. Results The final cohort (n=8,011; mean age 52.4 years) had 17.2% mortality over a mean follow-up of 10.9 years. Clustering identified three distinct phenotypes. Biotype 1 (n=461) represented a young, low-risk phenotype (mean age 35.0, HbA1c 5.2%). Biotype 0 (n=547) exhibited high metabolic and renal stress (age 54.6, highest HbA1c 6.5%, highest hs-CRP 0.71 mg/L, UACR 19.5 mg/g). Biotype 2 (n=595) represented an older, metabolically stable phenotype (age 64.5, highest enterolactone 451.8 nmol/L, lowest hs-CRP 0.24 mg/L). Kaplan-Meier analysis of the unseen test cohort (n=1,603) demonstrated robust, non-overlapping survival trajectories (p<0.0001). The older/stable (Biotype 2) and high-metabolic stress (Biotype 0) phenotypes exhibited significantly higher mortality than the low-risk phenotype. The unified machine learning model achieved high predictive discrimination (Test C-Index: 0.823). Conclusion(s) Unsupervised machine learning successfully discovered distinct, biologically plausible clinico-metabolic phenotypes within the at-risk and pre-clinical heart failure population. This integrated prognostic framework (C-Index 0.823) offers robust early risk stratification for primary cardiovascular prevention.
No takes yet. Share an insight, caveat, or question.
Mavromanoli et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:
More on Heart failureMore from European Heart Journal