Background and aims: Efficient analysis strategies for complex network with cardiovascular disease (CVD) risk stratification remain lacking. We sought to identify an optimized model to study CVD prognosis using survival conditional inference tree (SCTREE), a machine-learning method.
Methods and results: We identified 5379 new onset CVD from 2006 (baseline) to May, 2017 in the Kailuan I study including 101,510 participants (the training dataset). The second cohort composing 1,287 CVD survivors was used to validate the algorithm (the Kailuan II study, n = 57,511). All variables (e.g., age, sex, family history of CVD, metabolic risk factors, renal function indexes, heart rate, atrial fibrillation, and high sensitivity C-reactive protein) were measured at baseline and biennially during the follow-up period. Up to December 2017, we documented 1,104 deaths after CVD in the Kailuan I study and 170 deaths in the Kailuan II study. Older age, hyperglycemia and proteinuria were identified by the SCTREE as main predictors of post-CVD mortality. CVD survivors in the high risk group (presence of 2-3 of these top risk factors), had higher mortality risk in the training dataset (hazard ratio (HR): 5.41; 95% confidence Interval (CI): 4.49-6.52) and in the validation dataset (HR: 6.04; 95%CI: 3.59-10.2), than those in the lowest risk group (presence of 0-1 of these factors).
Conclusion: Older age, hyperglycemia and proteinuria were the main predictors of post-CVD mortality.
Trial registration: ChiCTR-TNRC-11001489.
Keywords: Artificial intelligence; Cardiovascular disease; Diabetes mellitus; Machine learning; Metabolic abnormalities.
Copyright © 2020. Published by Elsevier B.V.