1105Machine-learning analysis of myocardial deformation patterns to predict incident heart failure or death in the general population

Abstract

Introduction: Different measurements derived from myocardial strain data have been identified as predictors of outcome in a broad spectrum of cardiac diseases, including heart failure (HF). We hypothesize that the comprehensive analysis of entire deformation patterns, rather than scalar indices (peak, time-to-peak values) extracted from them, can be more informative in identifying subjects at a higher risk of future events. Methods: In 1997 subjects enrolled in the Atherosclerosis Risk in Communities study (ARIC) we assessed strain patterns at 12 left ventricular locations (2 basal, 2 mid and 2 apical segments; from the 2ch and 4ch apical views) over a cardiac cycle using an unsupervised machine learning algorithm (multiple kernel learning) that positions subjects based on similarities in deformation. A K-means algorithm identified 4 clusters, for which we compared baseline characteristics and the primary outcome of death or HF event. Results: The unsupervised analysis of deformation patterns identified 4 clinicallydistinct clusters (Figure) with distinct clinical characteristics. One such cluster (Cluster 3) comprised the highest proportion of hypertensive patients (85.2%, p<0.0001) with prior myocardial infarction (3.5%, p=0.04) and atrial fibrillation (37.2%, p<0.0001); the lowest ejection fraction (61.1 (56.0-65.7) %, p<0.0001) and longitudinal strain (-12.9 (-14.4 to -11.7)%, p<0.0001); and the highest values of NT-proBNP (439 (145-1065) pg/mL, p<0.0001), left ventricular mass index (89 (73-111) g/m2, p<0.0001) and left atrial volume index (30.3 (24.3-38.4) ml/m2, p<0.0001). Cluster 3 was associated with a 4.1-fold increase in the risk of primary outcome (HR 4.11 (2.60-6.50), p<0.0001), which persisted after adjusting for age, sex, systolic blood pressure, prevalent coronary heart disease and prevalent atrial fibrillation (HR 2.62 (1.54-4.46), p<0.001). Conclusion: Our results serve as a proof-of-concept that unsupervised machine learning-based analysis of deformation patterns can agnostically identify subjects at a substantially higher risk of incident HF or death and confirm prior clinical knowledge. (Figure Presented) .