Distinction Between Heart Failure With Preserved Ejection Fraction, Heart Failure With Reduced Ejection Fraction, Athletes and Controls by a Supervised Machine Learning Model

Background: Physiological and pathological left ventricular (LV) remodeling exhibit distinct molecular mechanisms and lead to different clinical outcomes. However, there are no established imaging methods to distinguish these LV phenotypes.

Methods: In this prospective single-centered cohort patients (age ≥ 18 years), underwent clinical assessment and cardiac magnetic resonance (CMR) in a 3T MRI scanner. We employed random forests and decision trees, supervised machine learning techniques, to classify participants into one of four predefined groups (Athletes, Controls, HFpEF, HFrEF). The model was trained (70:30 train/test proportion) and evaluated using a confusion matrix. Variables used as predictors follow: extracellular volume fraction (ECV), native T1, LV cardiomyocyte mass [(1-ECV) × total LV mass] —a marker of total intracellular space—and cardiomyocyte size (tau)—a marker of hypertrophy, indexed LA volume, indexed end-diastolic RV volume, indexed end-diastolic LV volume (LVEDVi).

Results: A total of 171 participants were recruited: HFrEF (N = 43), HFpEF (N = 45), athletes (N = 43), and controls (N = 40). The median age varied among the groups [HFrEF (age, y): 53 (IQR 47.0-62.0); HFpEF: 63.0 (IQR 58.0-68.5); athletes: 30 (IQR 23.0-42.0); controls: 38.5 (IQR 30.25-49.25); p< 0.001] as well as sex [HFrEF (male, %): 20 (46.5%); HFpEF: 23 (51.1%); athletes: 37 (86.0%); controls: 15 (38.5%); p< 0.001]. Native T1 was highest in HFrEF patients (1279.5 ms, IQR 1236.5 - 1309.3) and lowest in athletes (1170.4 ms, IQR 1151.3 - 1202.3), with no significant difference between HFpEF patients (1250.0 ms, IQR 1224.5 - 1296.0) and controls (1249.0 ms, IQR 1183.3-1276.0). ECV was highest in HFrEF patients (31.0%, IQR 27.9 – 34.2), lowest in athletes (25.9%, IQR 24.9 – 29.7) and controls (27.2%, IQR 24.6 – 31.2), and intermediate in HFpEF participants (30.5%, IQR 26.1 - 34.7). HFrEF patients had the greatest LV cardiomyocyte mass (99.5 g/m2, IQR 73.6-131.7), but the lowest cardiomyocyte size (tau) (0.113 s, IQR 0.081-0.162). HFpEF patients [80.8 g/m2 (IQR 64.6-101.7); 0.146 s (IQR 0.126-0.173)] and athletes [82.0 g/m2 (IQR 71.8-92.7); 0.157 s (IQR 0.114-0.211)] had increased LV cardiomyocyte mass but cardiomyocyte size similar to controls. And controls had normal cardiomyocyte mass (52.3 g/m2, IQR 42.1-63.5), and cardiomyocyte size (0.137 s, IQR 0.097-0.185). The accuracy of the random forests model reached 75.6%. Native T1 values below 1206 ms, LVEDVi lower than 110 ml/m2 and LV cardiomyocyte mass greater than 37 g/m2 were found to be accurate discriminators of athletes from the remaining LV phenotypes. LV cardiomyocyte mass greater than 34 g/m2 discriminates HFpEF from controls and LVEDVi greater than 110 ml/m2 was the best discriminator of HFrEF from the remaining groups.

Conclusion: Multiparametric CMR tissue characterization allows for the differentiation of physiological and pathological LV remodeling phenotypes among HF spectrum patients and normal subjects.

Pannel A: Decision tree for classification of prelabeled groups: athletes, control, HFpEF and HFrEF. Pannel B: Decision tree for classification of prelabeled groups (removing HFrEF): athletes, control and HFpEF. Prediction variables used were: ECV, native T1, tau, cardiomyocyte mass, indexed LA volume, indexed end-diastolic RV volume, indexed end-diastolic LV volume. The variables kept on the tree were the ones with the most discriminative power according to the supervised learning method. Before a terminal node, cells present in the first line the group category defined by the algorithm, in the second line it displays the percentage of individuals of a given group (in the same order presented on the legend), and in the third line the percentage of total individuals within a box relative to the all participants. The algorithm stops when splitting no longer adds value to the predictions. The original data was split into 70:30 train/test proportion to build the tree. CM: cardiomyocyte; ECV: extra-cellular volume fraction; HFpEF: heart failure with preserved ejection fraction; HFrEF: heart failure with reduced ejection fraction; LVEDVi: indexed left ventricular end diastolic volume; RVEDVi: indexed right ventricular end diastolic volume.