Using Model‐Based Analysis and Physiology‐Informed Machine Learning to Refine the Stratification of Phenotypes in Heart Failure

Abstract

A model‐based approach was developed to elucidate etiological differences between and within patient groups representing the two dominant heart failure diagnoses: heart failure with reduced (HFrEF) and preserved (HFpEF) ejection fraction. A closed‐loop model of the cardiovascular system informed by patient‐specific transthoracic echocardiography (TTE) and right heart catheterization (RHC) data was used to identify key parameters representing cardiovascular mechanics and hemodynamics. Thirty‐one patient records (10 HFrEF, 21 HFpEF) were obtained from the Cardiovascular Health Improvement Project database at the University of Michigan. Model simulations were tuned to match RHC and TTE pressure, volume, and cardiac output measurements in each patient. The underlying physiological model parameters were compared to model‐based norms and between HFrEF and HFpEF diagnoses. Our results confirm the main mechanistic parameter driving HFrEF is reduced left ventricular (LV) contractility, whereas HFpEF exhibits a more heterogeneous phenotype. To determine subgroups within the HFpEF diagnosis, we conducted a principal component analysis on the optimized parameters, which combined with machine learning techniques including k‐means and hierarchical clustering methods reveal (i) a group of HFrEF‐like HFpEF patients (HFpEF1) that share characteristics with HFrEF, (ii) a HFpEF group that exhibit classical characteristics of patients with diastolic dysfunction (HFpEF2), and (iii) a group of HFpEF patients that do not consistently cluster (NCC) among the machine learning techniques. These subgroups cannot be distinguished from the clinical data alone. However, reanalyzing the clinical data of each of these newly determined subgroups reveals that elevated systolic and diastolic LV volumes seen in both HFrEF and HFpEF1 may be used to identify HFrEF‐like HFpEF patients. Similar HFpEF groupings have been identified by other studies using specialized clinical measures (e.g., from non‐routine echocardiography and heart biopsies) whereas our study analyzes data from standard clinical procedures, such as TTE and RHC. Hence, our methodology has great translational potential to be broadly accepted in the clinic. These results suggest that physiology‐informed model‐based analysis of standard clinical data in conjunction with unsupervised machine learning can distinguish subgroups of HFpEF as separate phenotypes. Moreover, this methodology has the potential to retrospectively assess patient selection in past HFpEF clinical studies, aid in the optimization of prospective HFpEF subgroup selection for future clinical trials, and elucidate patient‐specific treatment strategies.