Unsupervised learning reveals two phenotypes of left ventricle contraction in patients with heart failure

Cardiac resynchronization therapy (CRT) is an effective method of treating severe forms of congestive heart failure (CHF). Despite careful selection of patients for CRT, about 30% - 40% of patients have no sufficient effect of CRT. We present preliminary results on using a clustering algorithm based on echocardiography data in CHF patients before and after CRT to define distinct phenotypes in the population. We used the Fourier series for analysis of the dynamic changes in the left ventricular shape during cardiac cycle, a nonlinear dimensionality reduction method, and clusterization. Two phenotypes in the CHF patients, which exhibited differences in complexity of LV contraction pattern were found. A small population of patients did not allow us to distinguish sub-populations in terms of their response to CRT. The results suggest additional imaging data providing more information on the structural abnormalities in myocardium and detailed clinical data obtained in patients should be accounted for better patient stratification and prediction of their response to CRT.