INTRODUCTION
Hypertrophic Cardiomyopathy (HCM) is characterized by myocyte hypertrophy, myocyte disarray, interstitial/replacement fibrosis, and is associated with a high risk for atrial and ventricular arrhythmias. A large proportion (~25-30%) of HCM patients develop atrial fibrillation (AF) during their lifetime.[1] Notably, HCM patients with AF are at increased risk for stroke even in the setting of low CHA2DS2-VASc scores.[1-3] Furthermore, stroke can be the first manifestation of AF in HCM. Hence there is a need to understand AF pathophysiology and assess AF risk in HCM patients.[4]
Several small studies have reported an association between age,[5] NYHA class,[6] left atrial (LA) size/function,[5-7] EKG-P-wave dispersion,[7] N-terminal proB-type natriuretic peptide (NT-proBNP) levels,[7] fibrosis in the left ventricle (LV)[6] and AF in HCM patients. Clinical risk scores for AF prediction have been developed and validated using general populations from the Framingham Heart Study (FHS),[8] Cardiovascular Heart Study (CHS),[9] Atherosclerosis Risk in Communities Study (ARIC),[10] Multi-Ethnic Study of Atherosclerosis Study (MESA)[11], Reykjavik Study (AGES)[12] and Rotterdam Study (RS).[13] But it is unknown whether these models are effective in assessing AF risk in HCM, given the differences in cardiac physiology between HCM patients and the general population.
We hypothesized that a machine learning-based model developed using HCM patients would perform better than existing models (that were developed using clinical data from the general population) to assess risk factors for AF in HCM. In order to achieve this goal, as a first step, we retrospectively identify AF cases and clinical features associated with higher/lower risk of AF, using electronic health record (EHR) data of HCM patients who underwent deep clinical phenotyping by multi-modality imaging. Similar to other machine learning methods and our work in ventricular arrhythmias, [14] our method has the advantage of allowing for thorough, unbiased scanning of the data along several dimensions, which permits derivation of a classification algorithm that stratifies HCM cases based on their respective likelihood to present with AF. Importantly, this method is flexible, can be easily updated as additional clinical data becomes available, and provides insights into the pathophysiology of AF in HCM patients. The next step is prospective testing of our model to predict AF development in HCM.