LAUSR

Basic research in cardiovascular medicine has yielded dramatic insights into physiology, leading to therapeutic advances and a significant decrease in cardiovascular mortality for the past 50 years. Nonetheless, it is increasingly recognized that even highly efficacious therapies have heterogeneity of effect at the individual level. In addition, there is significant variation in the use of evidence-based therapies and outcomes in routine clinical practice. These factors limit the potential impact of scientific advances when implemented in care. The increasing availability of digital medical data, coupled with powerful analytic methods such as machine learning, hold promise to support more personalized medicine and effective population health management. If successful, translational medicine in the era of big data and machine learning could truly span bench to bedside to population and optimize the end results of healthcare, that is, the triple aim of delivering better patient care, improving population health, and reducing cost. In this article, we review the opportunities and challenges of using big data and machine learning to deliver personalized medicine that meets the triple aim. We put forward the position that research scientists across the translational spectrum will need to be engaged and lead research studies of multiple types for these opportunities to be met. The promulgation of EHRs, increasing availability of digital health data from sources such as apps and biosensors, and the explosion of genomic sequencing have contributed to the increasing availability of big data. These massive—and growing—data sets lend themselves to the application of analytic methods, such as machine learning to accomplish complex, iterative pattern recognition and development of predictive algorithms. By applying machine learning methods to big data, there is a hope that human intelligence can be mimicked, that is, artificial intelligence. For artificial intelligence to mimic human intelligence, it will require computers to not just recognize patterns in structured data but also be capable of natural language processing. Whether such powerful artificial intelligence can grow from deep learning and cognitive computing for next several years to decades remains uncertain. Ideally, big data could support personalized medicine, where machine learning algorithms predict individual patient risk, and more accurately and precisely identify which patients will benefit most from specific therapies. For example, could data from EHR, biosensors, and genomics improve the predictive power of the pooled cohort equations for incident cardiovascular events? Therapeutic recommendations based on computer-refined phenotypes (precision medicine), better population health management, accelerated identification of drug targets, and augmentation of the diagnostic ability of clinician are all potential applications. Currently, there is great hype but scant evidence for these potential applications. This suggests a critical role for cardiovascular researchers across the translational medicine spectrum to validate and prove effectiveness and safety of big data applications before broad deployment in practice. Without such evidence, the potential for big data and machine learning in cardiovascular medicine may not be realized. Convincing evidence will require studies that demonstrate improved patient outcomes which can be attributed to big data approaches, for example, observational studies or randomized trials. The rise of large biobanks with genomic data on millions of subjects linked to their EHR or curated phenotypes, and sometimes with prospective follow-up, constitutes one real opportunity. The UK Biobank has detailed clinical data with thousands of variables linked to genomic data on half a million subjects. Using these data, researchers made public genomewide association results for >2000 traits on ≈337 000 individuals (www.nealelab.is/uk-biobank). This democratization of data is meant to accelerate biological discovery by enabling focused hypotheses and quick validation of findings from one cohort using another. The United States is following with even larger biobanks down the pipeline, such as the Million Veteran Project and the All of Us Research Project, each of which is aiming to recruit one million subjects with extensive genomic and phenotypic data generation. Availability of population genomic data has improved the power to detect new associations and invalidated prior The opinions expressed in this article are not necessarily those of the editors or of the American Heart Association. From the MedStar Heart and Vascular Institute, Georgetown University, Washington, DC (W.S.W.); Division of Cardiology and Center for Genomic Medicine, Massachusetts General Hospital, Boston (A.C.F.); The Broad Institute of Harvard and MIT, Cambridge, MA (A.C.F.); Harvard Medical School, Boston, MA (A.C.F.); and University of Colorado School of Medicine, Aurora (J.S.R.). Correspondence to William S. Weintraub, MD, MedStar Heart and Vascular Institute, MedStar Washington Hospital Center, Suite 4B1, 110 Irving St NW, Washington, DC 20010. Email william.s.weintraub@ medstar.net Viewpoints