Artificial Intelligence For FITs

March 28, 2018 | Ahmad Masri, MD

The introduction of electronic medical record systems and digital archiving of medical images has revolutionized health care delivery and led to the daily accumulation of terabytes of data. However, it took years before physicians and scientists explored the utility of this treasure trove of data. Lagging behind other industries for years, they are finally seeing many original investigations and review articles published in medical journals that tackle the use of artificial intelligence in medicine. Artificial intelligence was thrust into the limelight due to controversial projections that were made by some in hopes that it would eventually replace imagers (radiologists and cardiologists) in the future.

Artificial intelligence refers to the use of algorithms that iteratively learn from the data itself, allowing the machine to find connections, build models and make inferences without being programmed to do so. Broadly, artificial intelligence encompasses three main disciplines: machine learning, cognitive computing and deep learning. Machine learning algorithms are broadly classified into supervised learning, unsupervised learning and reinforcement learning. For an FIT, these complex terms sound discouraging. However, artificial intelligence is here to stay and the use of its methods will likely increase over time. Even more encouraging, the focus is now shifting towards using artificial intelligence to complement what physicians do rather than the earlier hype of “replacing physicians.” I asked four cardiologists who are leaders in the field to comment on the future role of artificial intelligence and give advice to FITs.

Partho P. Sengupta, MD, FACC, is the chief of cardiology, director of cardiac imaging and professor of medicine at the West Virginia University. He authored numerous articles on the use of artificial intelligence in medicine, focusing on the use of artificial intelligence in echocardiography.

PS: “In the year 2018, several projects in artificial intelligence are already focused towards improving throughput and efficiency in clinical labs. Besides standardized and automated image measurements, the focus is increasingly shifting towards deep learning techniques that can perform automated image recognition, segmentation and interpretation. Artificial intelligence will also enable integrating multitude of measurements, extracting unseen phenotypic characteristics for precise disease and risk estimation in individual cases. Future drug, device and therapy trials are expected to be focusing on the use of artificial intelligence for patient selection. With the advent of artificial intelligence, the role of imagers and echocardiographers will likely elevate as clinical specialists, who will have the skill-set as 'therapy modelers' and can advise the clinical teams on the best therapeutic strategy for a given individual.

In the past, I would advise fellows with interest in academic careers to spend time learning research methodologies. Now, I advise every fellow to befriend a data-scientist or learn at least an introductory course in machine learning.”

Oscar C. Marroquin, MD, FACC, is the chief clinical analytics officer for the University of Pittsburgh Medical Center (UPMC) Health Services Division and an associate professor of medicine, epidemiology and clinical and translational science at the University of Pittsburgh. He leads the big data and analytics efforts for the provide arm of UPMC and has extensive practical experience in the potential application of artificial intelligence in routine clinical care.

OM: “Artificial intelligence is already playing a role at our institution. We have built machine learning algorithms that identify patients who are at high risk of 7-day and 30-day re-hospitalizations. These models were generated by looking retrospectively at about 600,000 discharges from hospitals within our health care system. The model has been prospectively validated and we are actively using it in four of our hospitals. Every day at midnight, the data for each inpatient admission at UPMC hospitals is put through the algorithm and a risk of re-hospitalization is assigned (high, medium and low), and the information is made available to the providers caring for the patients (physicians and care managers). This system has resulted in better resource and time allocation for those at high risk to ensure good transitions of care, follow up, etc. Similarly, we have developed machine learning algorithms to do more “precise phenotyping” in different populations. Our clinicians are piloting these models in congestive heart failure, asthma and low back pain.

My advice to FITs is that they need to see artificial intelligence as a technology that will enable and augment what they do, not as a replacement. As such, it is imperative to learn more about the basics of artificial intelligence to know when to use it (if one is a researcher), or how to interpret studies that use these techniques, as they will keep coming and improving.”

Rahul C. Deo, MD, PhD, is an associate professor of medicine at the University of California, San Francisco. His work focuses on emerging computational approaches to analyze large-scale genomic data.  More recently, Deo and his team have been working on a computer vision pipeline for automated determination of cardiac structure, function and detection of disease using echocardiography.

RD: “Recent experiences in other industries have made it clear that we should anticipate automation in almost every situation where there is abundant digital training data and a clear objective. However, instead of trying to absorb the latest machine learning algorithms, physicians and trainees should be thinking carefully about how such automation would fit into clinical practice. Health care is inherently a more challenging application for artificial intelligence than most other industries and it will take the insight of current practitioners to pinpoint what we are lacking. When thinking about the future of artificial intelligence in medicine, it is helpful to consider two (somewhat mundane) examples where artificial intelligence is already at play in cardiology: automated echocardiography interpretation and clinical risk scores. Given the growing abundance of digital data and increased academic and corporate interest in this space, we should expect more variants of these two, specifically automated image interpretation systems (for echocardiography, chest radiographs and cardiac magnetic resonance imaging) and potentially continuously learning tools that predict diverse outcomes such as hospital readmissions or adverse clinical events. In some cases, these advances are likely to be disruptive, but it is not always obvious how they will integrate with the existing workflow given current complex incentives. This presents an exciting opportunity to shape the deployment of artificial intelligence to ensure that it improves care and addresses unmet needs.”

Sanjiv Jayendra Shah, MD, FACC, is a professor of medicine, director of the Northwestern Heart Failure With Preserved Ejection Fraction Program and director of the T1 Center for Cardiovascular Therapeutics at Northwestern University Feinberg School of Medicine. Shah has published studies on the use of novel machine learning techniques for improved classification and therapeutic targeting of heart failure syndromes.

SS: “We are most likely still far from a time when artificial intelligence (specifically machine learning) will be able to accurately predict cardiovascular outcomes, especially for heterogeneous syndromes such as heart failure. One of the problems is the fact that current data sources such as electronic health records capture only part of the patient's journey during interactions with the health care system, but future data on medication compliance, physiologic variables (via wearables) and diet tracking technologies may assist with understanding the patient more holistically. We need more ‘orthogonal’ data types to understand the patient from several unique and additive perspectives to more accurately predict outcomes, especially in the individual patient, as opposed to current prediction models that work well on average but are not very accurate on the individual patient level. FITs should learn more about artificial intelligence because of the increasing availability of large datasets, low-cost and high-speed computational systems, and continually evolving and improving techniques for machine learning analyses. These methods allow us to explore research questions from a unique multi-dimensional perspective that embraces the complexity of cardiovascular disease.”

The question for most FITs is, “Where to start?” Below are numerous resources that FITs can use to learn more about artificial intelligence.

Online Resources to Study and Explore Artificial Intelligence:

  • offers a range of courses for a small fee for each course
  • offers some free courses, while others are paid
  • offers three free courses online
  • offers multiple courses for a fee (some are free)

Selected Published Articles on Artificial Intelligence (original publications and review articles):

This article was authored by Ahmad Masri, MD, Fellow in Training (FIT) at the University of Pittsburgh Medical Center in PA.