Prognostic value of a thyroid–endothelial phenotype in HFpEF: a large-scale interpretable machine learning cohort study revealing biological dissociation
J Yang, J Zhang, Y H ZhangAbstract
Background
Heart failure with preserved ejection fraction (HFpEF) is increasingly recognised as a heterogeneous syndrome, yet identifying rare, high-risk clinical phenotypes within broad cohorts remains a challenge. Crucially, it remains unclear whether the biomarkers defining these data-driven subtypes are the same ones that drive their long-term prognostic risk in routine clinical cohorts.
Aims
We sought to operationalise a high-risk thyroid-endothelial phenotype using AutoML-optimised classifiers and to examine if the biological drivers of phenotype membership are concordant with those of adverse outcomes, with implications for risk stratification.
Methods
In 1,754 retrospective HFpEF patients, we performed hypothesis-driven Ward's clustering of standardised TSH, free triiodothyronine (FT3), and Big endothelin-1 (BigET) to evaluate the thyroid-endothelial axis independently of NT-proBNP. XGBoost and TabNet classifiers, refined via evolutionary AutoML and 5-fold cross-validation, were trained for phenotype reproduction and outcome discrimination. SHapley Additive exPlanations (SHAP) were used to quantify biomarker contributions. Prognostic independence was assessed using multivariable Cox models adjusted for NT-proBNP, eGFR, and clinical covariates. Given the low prevalence of P2, outcome models were intentionally parsimonious.
Results
We identified a high-risk subgroup (Phenotype 2 [P2], 7.9%) characterised by a low-FT3/high-BigET signature. P2 patients had significantly higher event rates than Phenotype 1 (46.0% vs 32.6%; 57/124 vs 479/1470; P<0.001). After adjusting for NT-proBNP and eGFR, P2 remained independently associated with adverse outcomes (HR 1.659; 95% CI 1.242-2.217; P<0.001; bootstrap 95% CI 1.268-2.212). While classifiers replicated P2 assignment with high accuracy (Out-of-fold AUC: 0.999), outcome discrimination was expectedly modest (AUC 0.680). SHAP analysis revealed a clear biological dissociation: BigET was the dominant contributor to phenotype definition (mean |SHAP| 4.09), whereas FT3 emerged as the primary contributor to prognostic risk (0.43).
Conclusions
Interpretable machine learning enables the identification of niche HFpEF subtypes while revealing a "defined-by-endothelium, risk-driven-by-thyroid" pattern. This statistical dissociation suggests that markers defining specific HFpEF clusters may not align with their prognostic drivers, requiring a more nuanced approach to risk stratification. The phenotype is deployable using routine blood tests, enabling potential early risk enrichment beyond natriuretic peptide–based assessment, though external validation remains essential.Fig1 Out-of-fold ROC phenotype vs outcomeFig2 SHAP dissociation definition vs risk