Dr. May Wang, professor in the Joint Biomedical Engineering Department of Georgia Tech and Emory University and AIMed founder Dr. Anthony Chang answer the questions they’re most commonly asked about the education of AI in healthcare
Q: How did you first become interested in artificial intelligence?
AC: I was always fascinated by math and science (computer, math, and chess club memberships in high school with fascination of Star Trek, especially the omniscient supercomputer of Captain Kirk). During my clinical training as a pediatric cardiologist, my mentor Dr. William Norwood guided me into the area of biomedical signals and complex mathematics (in particular chaos and complexity theory). After this research focus, I had the vision of a supercomputer that would incorporate all of the patients’ data and information as well as learn all of the experts’ knowledge and wisdom (I named this supercomputer “Leo” after Leonardo Da Vinci, the Renaissance genius who first described a congenital heart defect in 1513). All of these early forays into data science led to my current passion in the use of data science to elucidate diseases and improve outcomes.
MW: My interest was in data science. Growing up in Zhong-Guan-Cun (Chinese Silicon Valley) of Beijing China, I have always enjoyed playing number games and excelled in math throughout entire K-12 years. But I always wanted to become a medical doctor. My parents were both researchers in Chinese Academy of Sciences with degrees from a top engineering school Shanghai Jiaotong University. They strongly suggested me to pursue engineering so to best utilize my math talent. As a result, I attended Tsinghua University (the Chinese MIT) to study Pattern Recognition and AI in Digital Signal Processing as my undergrad major. During senior year, when I derived machine-learning algorithms to find image signal patterns, and implemented with limited computer hours available, I first got really excited about the discipline that was chosen for me.
Q: What educational programs did you pursue and what were the positive and negative aspects of this education?
MW: I attended Georgia Institute of Technology for my graduate study in Digital Signal Processing of ECE. Because my math performance was excellent, the math department chair was trying to persuade me to get a math degree as well. So I pursued ECE and Math dual programs to deepen my knowledge in theoretical pattern recognition, optimization, and statistics. The positive aspect is those few years of rigorous training has helped me establish solid theoretical foundation. One summer, AT&T Bell Labs DSP Center Director gave a lecture at GT and I was fortunate to get a Summer Intern opportunity in his center. My assignment was to assist video compression algorithm design. At that time, due to limited communication challenge bandwidth, video must be compressed before transmission. Within a few weeks, I designed a video compression-decompression algorithm that significantly improves quality after decompression. However, during field trial, this algorithm takes at least one minute to train. Because the human tolerance for communication delay is 200ms, this algorithm was useless.
That was the first time I realized the weakness of doing theoretic data science without real-world knowledge. This realization was transformational for me. Since then, I have devoted myself to do data analytics for solving real world problems.
When Georgia Tech started Biomedical Engineering Department with Emory University, I was recruited back to be a faculty. From then on, my childhood dream and my personal interest has converged. I am very excited to work in Biomedical Big Data and AI with a focus of Biomedical and Health Informatics for pHealth (personalized, predictive, preventative, and precision health).
AC: I was always interested in biostatistics and focused on this domain during my years in pursuing my Masters in Public Health at UCLA. When the IBM supercomputer Watson effortlessly defeated the human contestants on Jeopardy! in 2011, I sensed a renaissance of artificial intelligence and entered the Masters of Science in Biomedical Data Science program at Stanford School of Medicine. I am interested in pursuing additional education in the area of integrated cognitive architecture in the near future.
The biggest takeaway of all of this education in a very different domain than medicine is that my mode of thinking as a clinician is influenced for the better (improved Kahneman’s system 1 to system 2 balance). While a less formal and structured curriculum based on courses on the Internet can be better suited for some, I chose the more structured program for its opportunities of collaborating with others in class projects as well as its networking opportunities. The effort to pursue a degree demands sacrifices in the midst of busy clinical training or hectic clinical schedules, but the reward is also of immense value for the long term. It is ultimately a very personal decision based on individual assessment.
Q: How does one get started in this domain of AI in medicine and healthcare?
AC: I think watching some video clips of common topics such as machine learning and natural language processing as well as learning from an introductory course on artificial intelligence is a good start. I had the privilege of authoring a book with both the clinicians as well as the data scientists in mind (Intelligence-Based Medicine), so it is a good resource to attain a comprehensive and yet deeper knowledge without excessive esoteric material. The AIMed meetings are ideal gatherings for clinicians to jump into the domain and learn not only the data science but also related issues (ethics, bias, etc) and relevant technologies (cloud computing, edge AI, etc).
We have also started two-day comprehensive review courses (based on the book Intelligence-Based Medicine) aimed towards certification under the American Board of Artificial Intelligence in Medicine (ABAIM) with a multidisciplinary faculty; we keep the course enjoyable and interactive by having conversations rather than presentations of the material. Basically, clinician education in AI in healthcare is less about learning to code, but much more about understanding the panoply of AI tools available, the framework of designing AI solutions to clinical problems, and acquiring knowledge about data, databases, and biostatistics and digital infrastructure.
MW: From my experience, I suggest to start AI in medicine and healthcare using problem-based learning strategy first, followed by obtaining basic terminologies and conceptual knowledge. In terms of topic, the following four introductory topics would be beneficial: introduction to biostatistics, introduction to pattern recognition (or machine learning) for decision making, one programming language, and one on data visualization.
Q:What would you predict for AI in medicine and healthcare in the coming decade or two and how should clinicians prepare for this new paradigm?
MW: The 1980s was the second wave of AI with the development of many expert systems and CAD. However, looking back, due to lack of infrastructure, the successful stories of AI adoption were limited. There are at least five grand challenges in biomedical AI. The first is data harmonization and quality control, the second is data integration, the third is real time decision making, the fourth is explainable AI, and the last is causal inference modeling.
During the current wave of AI, there are more data, more computing power, and more models developed. Thus, there will be more real world AI successes. AI is likely to penetrate deeper and wider in medicine and healthcare in the coming decade or two.
For clinicians, learning the basic AI concepts and terminologies should be good for understanding the news and literature. Then using the problems in their medical specialty as the application to find AI tools would be the most effective way to help become clinician prepared.
AC: I think AI will be more and more prevalent in most, if not all, fields of medicine with the image-focused areas like radiology, cardiology, pathology, dermatology, and ophthalmology leading the way as AI utilization in image interpretation is relatively more mature. Exciting nascent areas of AI in medicine include: natural language processing in some communication tools, unsupervised learning for discovery of new diseases and subtypes, deep learning for drug discovery and repurposing, self-supervised learning for training of biomedical data, and embedded AI in wearable devices. Future medical education and clinical training can also benefit from the convergence of AI with extended reality. A clinician of the future should, therefore, learn the basics of artificial intelligence and its deployments and limitations. Towards the end of this decade, many more clinicians will be more actively involved in AI projects to improve patient outcomes. A few of this cohort will even take on the role of a medical AI architect (chief intelligence officer).
Q: How would you recommend that professional schools (such as medical, dental, and nursing schools) be ready for the upcoming era of AI in medicine and healthcare?
AC: I think health professional schools as well as residency and fellowship programs could institute a course or even program that is dedicated to the basics of biomedical informatics and data science with clinical examples and applications. I can also foresee several schools to have dual degree programs in place to encourage students to be dually educated in clinical as well as data science. I also think that eventually data science and artificial intelligence will be a focused area of discipline with a dedicated one to two year fellowship in a few subspecialties within this decade.
MW: Making “introduction to basic AI” a board certification topic for medical residents to study (e.g. like statistics), and designing problem based learning AI case studies are good training for the next generation practitioners.
Q: How do I keep up with the knowledge and utility of fast changing AI technologies/tools?
AC: I work very closely with our MI3 data scientist team and have regular meetings with them to discuss what is new in the field. In addition, I try to attend not only machine learning meetings, but also artificial intelligence meetings that cover a much wider spectrum of topics. The courses and meetings as well as the weekly newsletters and article of the week features that I am deeply involved in with AIMed are also very productive for me to learn about what others are doing as well as what areas of investigation are coming in the near future.
We also have the monthly and annual Medical Intelligence Society (MIS) meetings for clinicians and data scientists who have a special affinity to AI. In addition, as the privileged editor-in-chief of the Intelligence-Based Journal, I often learn about new methodologies and/or applications of artificial intelligence in medicine from centers around the world. Lastly, I meet with mentors (I call them my “AI muses”) several times a week to simply have open discussions about any AI topic.