Arti ﬁ cial Intelligence in Cardiology

Arti ﬁ cial intelligence and machine learning are poised to in ﬂ uence nearly every aspect of the human condition, and cardiology is not an exception to this trend. This paper provides a guide for clinicians on relevant aspects of arti ﬁ cial intelligence and machine learning, reviews selected applications of these methods in cardiology to date, and identi ﬁ es how cardiovascular medicine could incorporate arti ﬁ cial intelligence in the future. In particular, the paper ﬁ rst reviews predictive modeling concepts relevant to cardiology such as feature selection and frequent pitfalls such as improper dichotomization. Second, it discusses common algorithms used in supervised learning and reviews selected applications in cardiology and related disciplines. Third, it describes the advent of deep learning and related methods collectively called unsupervised learning, provides contextual examples both in general medicine and in cardiovascular medicine, and then explains how these methods could be applied to enable precision cardiology and improve patient outcomes. (J Am Coll Cardiol 2018;71:2668 – 79) © 2018 The Authors. Published by Elsevier on behalf of the American College of Cardiology Foundation.

T he promise of artificial intelligence (AI) and machine learning in cardiology is to provide a set of tools to augment and extend the effectiveness of the cardiologist. This is required for several reasons. The clinical introduction of datarich technologies such as whole-genome-sequencing and streaming mobile device biometrics will soon require cardiologists to interpret and operationalize information from many disparate fields of biomedicine (1)(2)(3)(4). Simultaneously, mounting external pressures in medicine are requiring greater operational efficiency from physicians and health care systems (5). Finally, patients are beginning to demand faster and more personalized care (6,7). In short, physicians are being inundated with data requiring more sophisticated interpretation while being expected to perform more efficiently. The solution is machine learning, which can enhance every stage of patient care-from research and discovery to diagnosis to selection of therapy. As a result, clinical practice will become more efficient, more convenient, more personalized, and more effective. Furthermore, the future's data will not be collected solely within the health care setting. The proliferation of mobile sensors will allow physicians of the future to monitor, interpret, and respond to additional streams of biomedical data collected remotely and automatically. In this technology corner, we introduce common methods for machine learning, review several selected applications in cardiology, and forecast how cardiovascular medicine will incorporate AI in the future (Central Illustration).

RELATE TO STATISTICS?
Physicians have long needed to identify, quantify, and interpret relationships among variables to improve patient care. AI and machine learning comprise a variety of methods that allow computers to do just this, by algorithmically learning efficient representations of data. Here, we use the terms "artificial intelligence" and "machine learning" more or less synonymously, although more precisely machine learning can be understood as a set of techniques to enable AI. The difference between classical machine learning and classical statistics is less one of methodology than one of intent and culture. The

FEATURE SELECTION
Feature selection is essential for predictive modeling, and machine learning is particularly useful for it.
Consider the example of a physician who wishes to predict whether a patient with congestive heart failure will be readmitted to the hospital within 30 days of the index admission. This is a difficult problem where machine learning techniques have been shown to improve on traditional statistical methods (8,9).
Our hypothetical clinician possesses a large but "messy" electronic health record (EHR) dataset ( Figure 1)

DICHOTOMANIA
Clinicians generally work with dichotomized outcomes (e.g., "Should we give this patient a statin or not?") (14). However, framing clinical and scientific questions like this in some cases is imprecise and is called "improper dichotomization" (15,16). Two cases  The central promise of machine learning is to incorporate data from a variety of sources (clinical measurements and observations, biological -omics, experimental results, environmental information, wearable devices) into sensible models for describing and predicting human disease. The typical machine learning workflow begins with data acquisition, proceeds to feature engineering and then to algorithm selection and model development, and finally results in model evaluation and application.  64% of the value attainable by using raw, nondichotomized numbers instead (17). This issue is so widespread in biomedical publications that it is sometimes facetiously referred to as "dichotomania"  Table 1). In machine learning, one must choose a strategy (by selecting a particular  TREE-BASED METHODS. Tree-based methods are a widely applied set of powerful but deceptively simple algorithms. Clinicians will find these useful because they are often referred to as the best "off-the-shelf" machine learning algorithm, and they are often among the first algorithms that should be used. In contrast to regularized regression, tree methods are especially useful when the data are "tall," that is,   (38), and deep learning (39).
One of the most promising uses of unsupervised learning methods for cardiology is subtyping or "precision phenotyping" of cardiovascular disease (40,41). We use precision medicine as a term describing the synthesis of multiple sources of evidence to refine monolithic disease categories into more stratified and ultimately more personal disease concepts. Precision medicine in cardiology exists in contrast to precision medicine as understood in other fields such as cancer, where a series of somatic genetic mutations can clearly define a before and after state (40,41). In cardiology, most diseases are slow, heterogeneous, multimorbid, chronic processes where pathogenesis may begin decades before any ultimate disease manifestation. This is compounded by the issue that many disease concepts in cardiology such as heart failure or coronary artery disease are somewhat broadly defined and may be arrived at by different pathophysiological mechanisms. Unsupervised learning allows to us enable precision cardiology by learning subtypes of monolithic disease concepts, and we envision ultimately it will help to treat these subtypes differently and thus lead to improved outcomes.
In this context, cardiology is ripe for the application of unsupervised learning. For example, Li et al. (38) combined EHR with genetic data from a health system biobank to study type 2 diabetes mellitus. An Cardiologists make decisions for patient care from data, and they tend to have access to richer quantitative data on patients compared with many other specialties. Despite some potential pitfalls, it is becoming evident that the best way to make decisions on the basis of data is through the application of techniques drawn from AI. Cardiologists will thus need to incorporate AI and machine learning into the clinic. Indeed, as the amount of available patientlevel data continues to increase and we continue to incorporate new streams of complex biomedical data into the clinic, it is likely that AI will become essential to the practice of clinical medicine. This will probably happen sooner rather than later, as exemplified by the rapid adoption of automated algorithms for computer vision in radiology and pathology (52).
However, the incorporation of AI into cardiology is not something that clinicians should fear, but is instead a change that should be embraced. AI will drive improved patient care because physicians will be able to interpret more data in greater depth than ever before. Reinforcement learning algorithms will become companion physician aids, unobtrusively assisting physicians and streamlining clinical care.
Advances in unsupervised learning will enable far greater characterization of patients' disorders and ultimately lead to better treatment selection and improved outcomes. Indeed, AI may obviate much of the tedium of modern-day clinical practice, such as interacting with EHRs and billing, which will likely soon be intelligently automated to a much greater extent. Although currently machine learning is often performed by personnel with specialized training, in the future deploying these methods will become increasingly easy and commoditized. The expert knowledge of pathophysiology and clinical presentation that physicians acquire over their training and career will remain vital. Physicians should therefore take a lead role in deciding where to apply and how to interpret these models.