AI for The Rest of Us – An Interview with Sunandini Chopra, PhD – Part 1 of 2

We've been hearing a lot about artificial intelligence lately.  Let's face it - it can be confusing and intimidating, bringing up questions ranging from, “What is it?” “How does it work?” “What are applications in healthcare going to look like?” and “Is it going to take away my job?” Our guest is Sunandini Chopra, a PhD- level scientist from MIT, actively engaged in AI at IBM Watson and founder of the group, AI For The Rest of Us.

Joseph Anderson, MD: Sunadini Chopra, welcome. Could you tell us about AI for The Rest of Us?  It sounds very exciting and useful -  particularly for those of us in healthcare, where we keep hearing about how AI is going to have such a large impact.  How did this start?  What in your background and current experience led you to be interested in this?  

Sunandni Chopra, PhD: AI for The Rest of Us - for those who are not familiar - is a meetup group.  It's a community based in the Boston-Cambridge area.  We are right now a virtual community of over seven hundred people.  The way this group came together was when some of my friends and I started seriously reading and following the news in health care.  As you just mentioned, there's a lot of information and there's a lot of potential for machine learning applications to make a positive impact in healthcare. The space has just exploded.  I joined a company, professionally, that works in the AI sphere.  A group of us got very intrigued and we wanted to understand more about what was happening. As you know, artificial intelligence is something that has been around for decades.  We were curious to know and understand and learn why this is happening now.  How much of it is actually true on the ground?

So, with that objective in mind, we brought together a group of like-minded individuals in the Boston- Cambridge community.  We reached out to individuals in academia, industry, research and start-up entrepreneurs.  They shared with us some of the actual work that they were doing and the impact that they were making so that we could separate what we were reading about from what was actually happening on the ground.  

As you would understand from the name, it’s called, “AI for The Rest of Us.”  The goal was very education-focused and awareness-focused.  We wanted to make it open to people from different backgrounds - software engineers, data analysts, healthcare professionals, and researchers.  So really, the goal was to bring together like-minded people to learn more about how machine learning algorithms are being used in patient care or research and in the healthcare space from the people who are actually working on it.

JA: Like you said, many people understand roughly what AI is.  And they understand, of course, what healthcare is.  Specifically, let's talk about what the applications of AI in healthcare are going to look like. I think it a good place to start would be the doctor-patient encounter.  Clearly, the physician is not going to be replaced overnight by some sort of robot.  So, what are some of the early applications going to look like? 

SC: In terms of the early applications that are being created and what is out there, a lot of value is being generated for frontline clinicians to help them be more productive.  And, also, some of the back-end processes, too, are being made more efficient.

I would categorize applications in the healthcare space in the following three categories: 1) clinical, 2) operational and 3) administrative. 

In the clinical category, you have things that are potentially highly impactful, but also may be challenging to implement, such as clinical decision support at the point of care. Image analysis solutions are being developed to assist in decision making and diagnosis in radiology and pathology. There are also applications to help better understand medication dosage errors.  

For the operational applications, I would focus on some of the work that is being done in robotic assisted surgery, virtual nursing assistants, and clinical trial management.

 And lastly, in the administrative section, where actually, some of the most progress has been made. This involves understanding variation in care across organizations and helping with reporting and benchmarking, so that physicians can perform at the top of their fields.  I look at the impact that machine learning can have in healthcare and it is really across these three categories - clinical, operational and administrative.

JA: AI is roughly defined as “a system that is capable of learning or evolving and incorporating new information into future decisions.” One thing in health care that's quite prominent is the regulatory aspect.  Could you talk about what that's going to look like?  Many people may be concerned about what is perceived to be a “black box” aspect.  How is it possible to regulate something which is continually changing or evolving?

SC: That's a great question.  Regulation is complex and something that is still being continued to be understood.  There many aspects of the regulatory process that continue to remain unclear.  For example, with respect to the size of a training data set, what’s the right number of patients? How can the quality of those data sets  be maintained?  How do these solutions continue to have FDA approval as new data sets are being incorporated into the training models?  The list goes on.  Unfortunately, we do not have answers to all of those questions.  What I can say is that progress is being made and this is the subject of ongoing work. The FDA is currently evaluating solutions that are providing a diagnostic read, image analysis, and basically anything to do with a patient diagnosis or making a direct clinical judgment.  

JA: Early applications in pharma may be more focused on drug discovery and development of new compounds, where machine learning can play an active role. But what about once products are launched?  

SC: Once the product is launched, there is scope for improvement there as well.  As you said, machine learning is definitely making an impact in the drug discovery and development space, but one area where I do see the potential for some impact - even post launch - is in the observational study space.  What happens is that newly launched drugs - after successful phase three trials - sometimes have observational trial obligations, where they have to monitor adverse events.  This is done among patients in the real world in a non-clinical trial environment.  This space is very disorganized and can definitely benefit from disruption by machine learning and automation.  Machine learning can play a role in identifying patient cohorts and subgroups in which a particular molecule or drug is really effective.  It can be used to quantify certain adverse events in certain subgroups so we can we have a more agile process as we are learning about patients in the real world.

There are ways of incorporating these algorithms even post launch.  Unfortunately, a lot of that hasn't been done yet, but there is scope for improvement.  I think applications that are more prominent right now are certainly in research and development. There's a gamut there, running from early discovery, basic research and development, clinical trial management and going on to commercialization.  There are various applications.  On the other side - the clinical care side - applications right now are focused on providing support in the decision-making process and not being the sole source of the decision. They are basically augmenting human expertise and medical expertise.

JA: Could you explain to us what are patient avatars?  These appear to be a very useful concept in discovering new compounds and facilitating and expediting the drug discovery process and allowing us to build processes where we don't have to experiment, so to speak, on humans, but can somewhat virtualize the process.

SC: Patient avatars are patient profiles based on the available big data of both episodic and life care instances of patients. They can be made richer by including information about social determinants of health, demographic information, and so on.  These avatars are being used to understand how drugs are impacting patients outside of the controlled environment of a clinical trial.  They are being used to provide information around which patients are experiencing maximum efficacy and what kinds of adverse events are being experienced.  

All of this information is opening, I believe, a new area of clinical trials called the “the virtual control arm.” You are using a set of data to understand what the baseline in a demographic is and using that information to understand how effective a new molecule or a new drug target modality is likely to be.

Continued in part 2 …