Author + information
- Jagmeet P. Singh, MD, ScM, DPhil, Deputy Editor, JACC: Clinical Electrophysiology∗ ()
- ↵∗Address for correspondence:
Dr. Jagmeet P. Singh, Cardiology Division, Massachusetts General Hospital, Harvard Medical School, 55 Fruit Street, Boston, Massachusetts 02114.
Progressive digitization coupled with the growing demand for value has created a need for artificial intelligence (AI)-aided clinical care. The recent rise of AI to the forefront is a direct consequence of an improvement in our computing abilities and the concurrent availability of big data. Simpler forms of AI (also known as applied AI) defined by the ability to perform a single intelligent task, that is confined to a code, have been in our lives for the past several years. In the world of electrophysiology, some examples include algorithm-driven identification and therapy for arrhythmias by implantable devices, sensor-guided modulation of pacing adjustments to AV and VV intervals, and risk-prediction algorithms for sudden death and heart failure readmissions.
More evolved forms of AI comprise machine learning, deep learning, and neural networks, which are conceptually not too distinct from each other. Machine learning (a term thrown around quite frivolously) involves engaging a machine (computer) to learn to perform a task without a pre-existing code (1,2). To enable this, a compilation of a large number of clinical case profiles are used to predict a specific outcome. The machine algorithm works its way through these multitude of case scenarios, while continuing to learn and adapt to achieve the goal at hand. This involves working off very large datasets of covariates and outcomes with repeated adjustment of weights of input data until a task is learned (1). A well-known example of this is “image recognition,” where algorithms after analyzing millions of pictures, are able to recognize forms, shapes, patterns, and more. This automated intelligence in imaging and signal processing has already made its inroads into the practice of medicine through adeptly categorizing dermatological, radiological, ophthalmological, and cardiovascular disease states (3–5).
Deep learning, a form of machine learning, refers to approaches that learn ways in which data can be represented (2–4). An example of this could be where retinal images, beyond just providing information about the ophthalmological disease state can also divulge information regarding age, sex, and risk for major cardiac adverse events (3). In practice, deep learning models are usually implemented as complex neural networks that have many tunable parameters. Neural networks refer to a type of machine learning that creates algorithms inspired by the structure of the human brain, with multiple layers and levels of abstraction. Neuro-evolution in its highest form is where AI is used to build AI with an evolutionary (genetic) algorithms to continually optimize neural networks. Essentially it is an algorithm that can continually mutate to solve problems.
There are several possibilities for AI to interface with electrophysiology. To name just a few, these could include: 1) redefining the categories of disease; 2) newer risk-stratifying approaches for specific outcomes and therapies; 3) empowering the natural evolution of care from the current transactional mode to one of continuous care; and 4) enabling a form of precision medicine (targeted ablation and pacing) that was previously not possible. Beyond machine learning, AI-inspired constructs for an intelligent electrophysiology lab will integrate voice commands, imaging modalities, and virtual reality (6). This would enable interactive holographic viewing within the cardiac chambers of scar, circuits, foci of arrhythmias, and real-time activation sequences to target our therapy.
Using heart failure as a “learning” example, we tend to trivialize its complexity by treating it as a singular disease. We broadly categorize it into heart failure with preserved ejection fraction and heart failure with reduced ejection fraction, especially for our decisions regarding the appropriateness of defibrillator therapy. A recent study using machine learning of data from over 40,000 patients in the Swedish Heart Registry categorized patients into 4 clusters quite different from the conventional phenotypes. Of note, this approach resulted in markedly better risk stratification and prognostication than ejection fraction did (7). Not surprisingly, within these clusters, there was significant overlap of left ventricular function. AI-assisted reclassification of disease will need us to revisit guidelines and question the validity of our clinical trials from a few decades ago. This novel clustering approach with dense baseline data will get further supercomplicated (but more precise) with the integration of continuous data from implantable and wearable devices within the electronic health care record. Similar large dataset-driven phenotyping of atrial fibrillation (AF) will call for revising the taxonomy from its current simplistic form of paroxysmal and persistent AF. This could encourage cluster-driven trials with drugs, devices, and catheter ablation. On the procedural front, there are many ways that AI will enhance the delivery of therapy. For this, there will need to be an integration of demographics, clinical characteristics, structural, procedural, and outcome data. Learning algorithms will recommend the type and location of the pacing modality (e.g., His bundle, right ventricle, coronary sinus, multipoint, multisite, or endocardial pacing) most likely to help a heart failure patient. Similarly, sizable databases of periablation atrial activation sequences or ventricular mapping can help create AI-assisted targeting of locations that may successfully terminate rhythm disturbances.
As health care becomes more patient-centric, the clinical care model for our (arrhythmia and heart failure) patients will advance from the conventional transactional to a continuous care model. We all vividly remember the prolonged uncomfortable transition phase of integrating remote monitoring into clinical practice. Notably, its adoption continues to remain fairly nonuniform. The primary impediment beyond early-on reimbursement issues has been operational. For those of us that have made remote monitoring a part of our lives, we are inundated on a daily basis with device alerts (mandating a response) that contribute to the “burnout” that many practicing electrophysiologists experience. This is where AI-facilitated stratification with self-learning algorithm-driven responses could make our lives considerably better. There is already some work in this area that has been validated. A simple algorithm integrated into the electronic health record to deal with device AF alerts, led to <2% of approximately 2,000 AF alerts needing active physician input (8). Imagine this scaled across all device alerts! That would make this data deluge so much more bearable and manageable. At the same time, many of these devices are becoming smarter. They have within them validated sensors that respond in real time and could potentially be autonomous in delivering care. Driven by the escalating complexity, AI-directed care will range from engaging the patient in self-management strategies to using the help of a health care navigator, nurse, nurse practitioner, or a physician. Continuous advances in wearable and implantable devices and sensors will enable the digitization of the human body. And it will be only through the use of AI that we will be able to handle and make sense of this deluge of big data.
The future is here. Beyond just recognizing it, we need to embrace it and become willing partners. To ensure that these efforts in machine learning can progress at a quick pace, we need to identify and prioritize “repetitive tedious tasks” as low-hanging fruits for early AI-based approaches. This could be as 1-dimensional as streamlining scheduling within office practices to cutting-edge endeavors for the delivery of continuous personalized care. There is an urgency for us to bring together multidisciplinary groups to generate and share “large clean datasets” to expedite the creation of meaningful algorithms. The primary goal being to construct a field where everything can be measured, so that AI-driven interventions can be proactive, preventative, and precise. However, it will be on us to keep technology tamed and to preserve our humane touch. As has been wisely said, “technology is a useful servant, but a dangerous master.”
Dr. Singh has served as a consultant for Biotronik, Boston Scientific, Medtronic, St. Jude Medical, LivaNova Group, Respicardia Inc., Impulse Dynamics, EBR Inc., and Toray Inc.
The author attests he is in compliance with human studies committees and animal welfare regulations of the author’s institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the JACC: Clinical Electrophysiology author instructions page.
- 2019 American College of Cardiology Foundation
- Johnson K.W.,
- Soto J.T.,
- Glicksberg B.S.,
- et al.
- Poplin R.,
- Varadarajan A.V.,
- Blumer K.,
- et al.
- Tao Q.,
- Yan W.,
- Wang Y.,
- et al.
- Silva J.N.A.,
- Southworth M.,
- Raptis C.,
- Silva J.
- Ahmad T.,
- Lund L.H.,
- Rao P.,
- et al.