Clinical Research & Data
Potential Clinical Opportunities of Machine Learning and Artificial Intelligence
Manreet K. Kanwar, MD, introduces clinicians to artificial intelligence (AI) and machine learning and the potential clinical opportunities for applying these concepts to clinical research as well as daily practice to improve patient outcomes. Dr. Kanwar is the director of mechanical circulatory support and heart transplantation at Allegheny General Hospital in Pittsburgh, PA. She is the first author on a recently published paper in Journal of Heart and Lung Transplantation titled, “Machine Learning, Artificial Intelligence and Mechanical Circulatory Support: A Primer for Clinicians.”
Dr. Kanwar’s paper, which uses mechanical circulatory support to help clinicians see this information in context, introduces the current reach of machine learning and explains the vocabulary of machine learning, existing applications, and current barriers to the uptake of this technology.
During this interview, Dr. Kanwar defines 3 terms that she feels can help demystify this topic for the clinical community. In a nutshell:
- “Big data” is an extremely large and complex data set, which before the introduction of electronic health records, did not exist in our clinical lives; the scale of big data is such that it cannot be processed into clinically viable methodology by one person with an Excel spreadsheet
- “Supervised learning” is telling a machine what you want to learn from a particular data set; for example, looking at pre-LVAD parameters to predict a known outcome such as mortality
- “Unsupervised learning” is asking a machine to discover patterns that we cannot currently detect from a big data set; for example, deep neural networking or cluster analyses
Dr Kanwar explains how machine learning is different from traditional data analytics, noting, “I would say that the key difference between the two is the approach to the question being asked.” Traditional analysis, she states, is hypothesis driven: If you have A, B, and C, then you will have X outcome. Machine learning, she emphasizes, “is not going to put humans into neat little boxes like that.” Machine learning is probability driven: if you have A, B, and C parameters, the probability that such an event will happen is X. It mimics human decision-making by taking into account that variables are dynamic, impacting not just outcome, but each other. “Which is how, frankly, as clinicians, we think,” she states.
Dr. Kanwar concludes by discussing opportunities for clinicians to begin to integrate these concepts into meaningful patient outcomes. The biggest limitation, she explains, is access to these large data sets because electronic health records are documentation tools that are not geared toward research and outcomes are only as good as the data being analyzed. “If you have access to a fairly complete data set,” which she emphasizes, has to be large in scope, “then I think the possibilities are limitless.”
“It’s all about pattern recognition,” she explains, “So if you see a pattern happening, you can hopefully, rather than follow it after it has happened, you can implement decisions based on the recognition of that.” She provides the example of picking up patterns from a large amount of EKG data to predict patient outcomes before they happen. Dr. Kanwar states that she believes that pattern recognition and prediction of future events, based on data on hand at the moment, are at this time the two most exciting opportunities for integrating these concepts into improved patient outcomes.