Unsupervised machine learning applied to CPETecho for subclassifying heart failure with preserved ejection fraction: a pathophysiological approach

Background Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous disease, and not all patients respond well to current treatment. Identifying subgroups of HFpEF patients may support more personalised and effective therapy. Using combined cardiopulmonary exercise testing and stress echocardiography (CPETecho), the entire oxygen pathway during exercise can be analysed and is an attractive way to characterise HFpEF pathophysiologically. Purpose To define clinically relevant HFpEF phenotypes with distinct exercise and outcome profiles. Methods We retrospectively included 913 HFpEF patients, as defined by published criteria, across 4 third-line hospitals and divided them into a derivation cohort (n= 623) and a validation cohort (n= 290). We applied unsupervised machine learning, including graph-based clustering following dimensionality reduction (66 continuous variables), and compared pathophysiological characteristics, exercise capacity and outcomes between phenotypes. Results In the derivation cohort, cluster analysis identified 5 distinct HFpEF phenotypes: (1) pre-HFpEF, (2) HFpEF with peripheral limitation, (3) garden-variety HFpEF, (4) obesity-HFpEF, and (5) renocardial-HFpEF. Pre-HFpEF patients exhibited only mild diastolic dysfunction with preserved biventricular systolic reserve. The HFpEF with peripheral limitation group showed a female predominance and impaired right ventricular function. Garden-variety HFpEF was characterised by a high prevalence of atrial fibrillation, reduced right ventricular reserve, and left atrial and ventricular remodelling. The obesity-HFpEF group exhibited impaired stroke volume reserve, whereas renocardial-HFpEF showed a male predominance, the highest NT-proBNP levels, more severe cardiac remodelling, and the lowest cardiac power output. Phenotypes 3 and 5 demonstrated the poorest exercise capacity. The composite outcome of all-cause death and unplanned cardiovascular hospitalisation differed significantly across phenotypes: at a mean follow-up of 3.0 years, 18% (pre-HFpEF) to 47% (renocardial-HFpEF; hazard ratio 2.88; 95% CI 1.62-5.11; p-value <0.001) experienced an outcome event. All clusters could be replicated in the validation cohort despite several differences in baseline patient characteristics. Conclusion We identified 5 distinct exercise phenotypes among patients with HFpEF using CPETecho and unsupervised machine learning. These phenotypes exhibit marked differences in pathophysiological features, exercise capacity, and clinical outcomes. This subclassification may serve as a foundation for more targeted and effective treatment strategies in HFpEF.