12 Dec Interview with Andrew Carroll of Google AI
The Genomics team in Google AI develops methods to help understand genomic data and combine it with non-genomic biomedical data. Our team builds DeepVariant, a highly accurate deep learning variant caller. The real world applications are in improving clinical genomics, drug discovery, and in non-human biotechnology. Andrew shapes product strategy to drive adoption of the technology and ensure its connection to high-value application. His role has a strong focus on genomics community engagement, collaboration, and partnership. Read his full bio.
Interview with Andrew Carroll of Google AI
Q: Artificial intelligence (AI) techniques have sent vast waves across healthcare, even fueling an active discussion of whether AI doctors will eventually replace human physicians in the future. Do you believe that human physicians will be replaced by machines in the foreseeable future? What are your thoughts?
A: I believe that applying AI technologies in healthcare will make physicians more valuable, and will make their careers more enjoyable and sustainable. Practicing medicine is not only very demanding on physicians’ time, it also involves huge cognitive overheads from multi-tasking and context switching. Recent trends like Electronic Health Records bring benefits of consolidating information, but at the cost of greater time burdens and context switching.
We will apply AI in Healthcare to organize and integrate streams of information; to help summarize it meaningfully for physicians in ways that allow them to manage complexity. This will allow doctors to have more time and focus for the things which they are uniquely specialized for and want to do: understanding the important concepts of their field, connecting with their patients, and observing the subtle facets of their illnesses.
Even in specialty fields like histology and pathology, where the concern around AI replacing doctors is most pronounced, AI is likely to enhance physicians’ usefulness. In a recent publication from Google AI and collaborators “Impact of Deep Learning Assistance on the Histopathologic Review of Lymph Nodes for Metastatic Breast Cancer”, using AI to annotate key areas of slides allows pathologists to reach diagnoses faster and more accurately. This is one power example of humans being stronger with AI as opposed to the idea of competing. We also see this to be true in other domains, including ophthalmology.
In the end, I feel that as we apply AI further into healthcare, it will off-load from doctors the machine components of their job, and ultimately make the practice of medicine more human.
Q: Can you provide some use cases that have already successfully demonstrate the value of AI/Machine Learning in healthcare?
A: AI has already transformed one area of how we relate to health and our healthcare, though I feel few realize it. Right now, AI is an essential technology in organizing and querying information in search engines and phones (such as Google search and Android). These technologies are changing the access that people have to information about how they can stay healthy, how they realize they are sick, and what they can do to recover. These examples show the use of AI in personal settings; as AI is increasingly applied into more technical and specialized domains, it will have similar transformative effects on the information available to health professionals and the simplicity of accessing it.
Q: What areas in healthcare will benefit the most from AI/Machine Learning applications and when will that be?
A: I am uncertain which domains will ultimately receive the greatest impact, but I have a clear idea of the first to benefit: those fields which relate to imaging and natural language processing. These domains have the most mature machine learning infrastructures and community of expert developers. These fields have huge databases of annotated examples from both general and specialized domains, and the AI applications are already of profound value in non-specialist commercial systems. We can expect progress in areas like understanding, summarizing, and annotating speech, text, images, and video to continue to advance rapidly.
This will provide the foundation for applications which organize, summarize, and annotate medical imaging, doctor’s transcriptions, unstructured and structured information in the EHR, and medical billing.
Q: What are some of the challenges to realize AI/Machine learning in healthcare?
A: One of the biggest points of resistance I hear from the community about AI is a perceived lack of explainability and, behind that, a hesitation to trust AI. This is something the AI community will have to address, as no matter how capable we make the technology, it is ultimately applied through the hands of people who must trust it.
The good news is that AI systems are not fundamentally un-explainable. There are many techniques which allow AI systems to show why they are making the decisions they do. One example is a recent Google AI publication “Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning”. This demonstrated that, amazingly, deep learning models can predict from eye images a person’s age, sex, BMI, smoking status, or cardiovascular health with high reliability. The models were also designed to highlight the most informative features for these predictions, allowing the model to specify what parts of the eye relate to these traits, and in turn suggesting links to the underlying biology.
As a field, we must continue to ground our AI systems and predictions in the underlying biology. Because deep learning models act on relatively raw information, they often construct models of problems independent from human characterization of the field. When we work to ground these models in biology, not only do we provide a check on their validity, we can also use them as a check on our understanding. This can even discover new insight we didn’t know was there.
Q: What are the products and/or services Google AI offers/develops in the AI/Machine Learning sector? What makes Google AI unique?
A: My team within Google AI makes a tool called DeepVariant, which processes sequencing data to identify the positions in the genome which make an individual unique. DeepVariant is highly accurate, having won an award in the 2016 PrecisionFDA competition for highest SNP accuracy. DeepVariant is also high extensible, because it can be quickly improved or changed by retraining with new data as opposed to needing extensive coding for a new problem type.
DeepVariant is also an open-source tool. Anyone is free to download the code and modify it. Anyone is free to train their own models for it, whether for a different sequencing technology or tuned for a specific organism. DeepVariant is itself built on top of TensorFlow, which is an open-source library for deep learning also developed by Google in an open-source manner. I am grateful for Google’s commitment to open-source in these foundational technologies.
Q: What is your role at Google AI and what excites you about your work?
A: I have always been fascinated by both biology and AI. I started working in genomics around the time that the cost of sequencing fell dramatically and whole human genomes started to be something you could do at scale. It was amazing to watch the field of genomics grow and change so quickly. Sometimes people would ask about where a given technology would be in ten years and I would have to remind them that ten years ago, none of the technologies we are using had been invented.
More recently, I feel the same is true about Machine Learning and AI. It reminds me of that same time when the first large genomics cohorts were being sequenced. I feel so lucky to be at the intersection of not just one, but two fields that are actively re-inventing themselves.
I also feel the energy of a community which sees the potential in the combination of AI, genomic, and healthcare to change our lives for good. That these are essential tools which help us understand what makes us sick and healthy, so that we steer ourselves toward health and happy lives.