Dr. Craig Mermel, Research Lead in Pathology at Google AI on the future of AI in healthcare, his work developing the Apple Watch and winning a personal battle with childhood cancer. Part one of a fascinating two-part interview…

 

Craig Mermel is currently the Research Lead in Pathology at Google Health where he leads a research team focused on accelerating the application of machine learning for improved diagnosis of important human diseases, especially cancer.

Prior to joining Google, Craig worked at Apple on the Apple Watch and related health initiatives. He completed joint MD/PhD training at Harvard Medical School, where his PhD dissertation focused on developing novel statistical methods for mining the cancer genome. He conducted residency training in Clinical Pathology at Massachusetts General Hospital and is board-certified in Clinical Pathology.

 

 

What initially sparked your interest in medicine and subsequently, AI in medicine?

As a child, I was always interested in mathematics. I thought it was a lot cooler than science so I dreamt of becoming a mathematician or a code breaker. When I was a high school freshman, we had a chemistry teacher who was this amazing human being, a cancer survivor who taught me to appreciate science. He was the one who opened my door to this amazing yet mysterious subject. He introduced us to cutting-edge discoveries and to make connections to things that happen in everyday surroundings. He passed on his knowledge and passion for science and had a major influence on my career.

I said to myself, I could do mathematics but it’s unlikely to have any impact on people’s lives. I got to focus on science because that’s where I could make a difference. So, I became interested in science when I was in high school and in college I double majored in mathematics and biochemistry. At that time, I was trying to find a balance between the two rather than thinking about medical school. I thought I would be a scientist as I had been working in the laboratory as an undergraduate student.

In the lab, I worked with medical doctors who were also scientists in the laboratory. What they did amazed me, and seemed like fantastic way to work on practical problems. So I went into medical school with the mind of being a scientist and not a practitioner. I wanted to be the kind of doctor who helps develop new insights into diseases and new therapies. But all along, my interest in mathematics never died down. I went onto an MD-PhD program to be a physician-scientist as I felt this would be a natural way to blend medicine, science, and mathematics.

It wasn’t smooth sailing because I did have trouble figuring out where and how to blend and make sense of these subjects and to do the kind of scientific work I wanted to do. I did a few rotations in different laboratories and realized most research work is manual, non-quantitative and required only limited quantitative skills. I was eager to look for new angles. It was the early 2000s and I learned about the explosion of genomics data in my first-year genetics course in medical school. I thought perhaps I could leverage those data and use analytical tools to solve some important clinical problems.

My PhD research work was in cancer genomics, applying some of the latest quantitative techniques to large datasets to uncover new cancer genes, generate new targeted therapies and shorten the time between bringing discoveries in the lab to treating patients. It was an amazing and exciting time, with an explosion of data being driven by rapidly developing technologies bringing new, practical insights into human disease. That was where I first found a practical intersection between data and medicine and I’ve been following that passion ever since.

You were involved in the development of the Apple Watch and other health initiatives before joining Google AI. What was that experience like?

After I finished my MD-PhD program, I entered residency training in Pathology to learn how new genomic technologies were going to make their way into actual clinical practice. My initial plan was to stay within healthcare and work as an academic pathologist. At the same time, I became personally quite interested in all the new types of data coming from wearable devices. I was convinced that one day these devices would be used to drive greater insights into people’s health.

This was in 2011 and 2012. The earliest devices like Fitbit and others could only count your steps and not give you insights into your physiology. At the end of my residency, I was offered a an opportunity to join Apple as part of the Apple Watch team. Apple is a secretive company and I was just a small part of a much bigger team, so I can’t talk about everything I worked on, but I was part of the team that developed some early fitness applications for Apple Watch. I also got to observe the technology as it evolved from supporting healthy people into thinking about how it could help monitor and even diagnose diseases like atrial fibrillation.

Personally, it was an amazing experience. As a physician-scientist, my prior training focused around thinking at the level of an individual patient or at most a single hospital. Whereas working for a company like Apple, I was working on a product that’s potentially going to be used by millions of people. We also had to think about developing technology for people of all ages, backgrounds, fitness levels and health conditions. Thinking about how people are going to use and interact with the product you had worked on, as an engineer, that’s an amazing thought.

I spent almost four years at Apple. I loved the interesting and exciting work but I missed the connection to pathology and the prior clinical training I received. This compelled me to look for a new opportunity. I already had the tech bug in me and now I wanted to apply what I’d learned to my clinical training. It was then that I came across a team in Google, which was doing a lot of work on the intersection of AI and imaging. They were looking for someone to help them in developing that sort of product, so I moved to Google at the beginning of 2018.

What are some of your responsibilities as the Research Lead for Pathology at Google AI?

My job title and responsibilities change every year! Some of the things I was working on in the early days, when we were just starting on the research path, was to outline the problems we should focus on, assess their impact, and think about the feasibility of the technologies involved. Once those were defined, we would have to come up with a road map to accomplish it.

Google is not a pathology lab, so we don’t have access to tissue samples or images. So a big part of my early work was forming research partnerships with medical institutions and other laboratories to provide us with the de-identified data to help train and validate different machine learning models. As we finished these projects, we prioritized publishing our work for the scientific and clinical community, so that was also part of my responsibility: helping decide what papers we should write, what journals we should send them to, and ensuring the work was accessible to the right audience. We also want to share our progress with the non-clinical audience, so I got to share our work more broadly by writing blog posts or at various meetings.

For the past year and a half, my job has been mostly focused on translation. Now that we have demonstrated what AI is capable of, we have to take the research and translate it into an actual setting to measure its impact. As mentioned, Google is not a pathology lab and it doesn’t read your biopsies,so we’ve formed partnerships with others in the Alphabet organization to try and get these tools into the healthcare system. Ultimately, we wish to have a full lifecycle from demonstrating that AI can solve certain clinical challenges to helping deploy or translate these solutions to the real world.

How challenging is your work at Google to improve diagnoses of diseases through machine learning?

Google has been working on healthcare AI applications for several years now. One of our earliest papers on the use of AI for diagnosing diabetic eye disease was published in 2016, And the field has rapidly advanced during the last six years. It’s no longer all that novel to show that AI can achieve certain clinical tasks, but the real challenge has moved to getting the technology into real clinical practice. AI will only help a patient if it’s deployed into actual clinical care.

Training and validating a machine learning model are only the first steps. Healthcare sets a very high bar for safety and efficacy, and it takes time to show that AI, as with any other cutting-edge technology, is really safe and effective. At the same time, the market for adopting these tools is also very fragmented, which makes it difficult for any one tool to get widespread adoption. There’s also a need to think about reimbursement; to convince the payors that AI is safe, cost-effective, and should be considered part of the standard care. That’s another challenge so there’s a lot of work and all of them will proceed on a different timescale.

In terms of pathology, there is a specific challenge in that practicing pathologists today still largely rely on glass slides and traditional microscopes. That makes it particularly challenging to collect data and to deploy AI solutions that can be used in clinical practice.

The COVID-19 pandemic has accelerated the adoption of telehealth, has it also accelerated the digitization of pathology?

Yes, I think it has. It’s still too early to say how sustainable it will be in the long run but COVID-19 has increased the appreciation for the value of being able to do remote diagnosis when you can’t have everyone in one place. But, in reality, changing the workflow in the lab requires big capital and time investments. It’s not possible to snap a finger and make a lab go digital. COVID-19 has put a lot of pressure financially on hospitals so it’s also not a good time to pour lots money into a new investment.

I think there are two pieces of tension. On the one hand, I do believe there’s a benefit for pathology to go digital. Being able to read slides on the computers enables us to remain robust even if labs have to close due to unforeseen circumstances, and take advantage of AI solutions that can further improve diagnostic accuracy and efficiency. But on the other hand, it’s challenging to make the capital investment to buy or upgrade all the equipment needed for full digitization and train all pathologists to use technology. Transformation takes time and it’s not realistic to assume all labs can go digital tomorrow.

It sounds a little gloomy. Are we asking too much to want to see AI in pathology?

I am just pointing out the reality. I believe AI will be a major driver for digital pathology, and that the advantages that AI brings over time will motivate more labs to go digital. Globally, there’s also a demographic shift. In the US, the average age for pathologists is 55 to 60-years old. Pathologists are fairly senior and we are not replacing new pathologists at the rate at which they are retiring. Yet, the number and complexity of cases are increasing every year. So there’s a mismatch in supply and demand and we desperately need AI tools to fill up the gap and make pathologists more efficient at what we do.

What’s the greatest challenge you’ve overcome?

Cancer is a disease that has affected me personally. I was diagnosed with acute leukemia when I was 12 years old and underwent three years of chemotherapy. I was incredibly fortunate to have a curable disease – I’ve been in remission now for over 25 years – but every day in the hospital I met kids who weren’t as fortunate, which I think has always left me with an appreciation for the role that chance can play in your life.

People often ask me if I became a doctor because of my experience. I do think that’s too simplistic of an explanation, but I can’t discount that it played an important role in guiding my choices in ways both large and small. Mostly, it left me with both a tremendous desire to give meaning to my personal suffering and a feeling of obligation to use whatever talents I do have to help others who are less fortunate that I was.

While winning my own battle against cancer is certainly the greatest challenge I have ever overcome, I’m also constantly reminded that more times than not, cancer is still a battle we lose (I lost my own father to multiple myeloma 7 years ago). With what time I have, I want to have as much of an impact on this disease as possible.

My greatest fear is that this is a tough disease, and the challenges we face are real. Cancer is not an abstract problem but it’s something that affects anyone and everyone. I am sure many of us have known someone or lost someone through cancer. It’s personal for everybody. That’s why I always ask myself if I have done enough or if we are on the right track in our fight against cancer.

 

Part two of this wide-ranging interview, where Craig reveals what it’s really like to work at Google and Apple, can be read here

Dr. Craig Mermel will be speaking at AIMed’s virtual multi-track CME-accredited event, ‘Imaging’ on 29th and 30th June.

View the full, exciting two day agenda and book here