Machine learning and traditional analysis of the interaction between cardiovascular diseases and hematological malignancies

Abstract Background Cardiovascular diseases and clonal hematopoiesis of indetermine potential (CHIP), the premalignant state of hematological cancers, exhibit reciprocal interactions. While the presence of CHIP confer a risk for heart failure, it is unclear if cardiovascular diseases are associated with elevated risk hematological cancers. Purpose To determine the association between cardiovascular diseases (CVD) and cardiovascular risk, and hematological cancer. Methods We evaluated 27,231 adults, free of CVD and cancer, who participated in preventive healthcare program between 2000-2018. Each participant underwent a complete physical evaluation, blood test and exercise stress test at baseline. CVD was defines as ischemic heart disease, stroke, atrial fibrillation or hypertension. Then, we used machine learning and traditional survival analysis to evaluate the association between CVD and hematological cancer. For machine learning, we used the random forest method using Curat package in r, and incorporated all continuous variables gathered into the model (70% training, 30% test). For survival analysis, we first matches each CVD patient to a control patient by propensity scores to balance for age, sex, smoking, diabeles mellitus, renal function, body mass index and length of follow-up. Then, we compared the CVD and non-CVD patients using cox proportional hazard regression. Results Using machine learning, we found that the strongest predictors for hematological cancer in our cohort are low eGFR, age, high atherosclerosis cardiovascular disease score (ASCVD) and BMI. Of note, systolic and diastolic blood pressure were also strong predictors of hematological cancer (Figure A). Then, we matched 5500 CVD patients to controls with median follow-up of 9.2 years (Interquartile range 5-15 years). During follow-up, 1380 developed cancer. Of them, 186 developed hematological cancer. We used cox proportional hazard regression and Kaplan Mayer method to determine the risk and hazard ratio of developing hematological cancer (Figure B-C). Compared with subjected without CVD, patients with CVD were 91% more likely to develop cancer during follow-up, and 62% more likely to develop hematological cancer (Figure C, HR=1.61. 95%CI 1.21-2.16). Moreover, the 10-year risk to develop hematological cancer was 1.9% for patients with CVD and 1.0% for individuals without CVD (Figure D). Notably, the increased incidence of hematological cancers included 38% increase in cases of leukemia but only 11% increase in cases of lymphoma. Conclusions Using machine learning and traditional analysis we show that CVD risk and CVD were associated with increased incidence of hematological cancer. Understanding this relationship could lead to advances in cardiovascular and cancer treatment and could ultimately help to improve patient outcomes.