Ask Good Questions
I start out my Stanford cloud computing class the same way every year. I tell the students the mark of a great student is not measured by what they know, but by the questions they ask. So learn to ask good questions. Recently I was invited to speak to a group of 50+ interns at the Children’s Hospital of Orange County. They range from high school students to recent college graduates. Rather than do a presentation I asked them to listen to this podcast and submit a question. I answered a few of them in my talk but decided to answer all of them online. What follows are their questions in no particular order.
How do you approach someone who’s skeptical about advancement in tech with healthcare, specifically AI because of the sci-fi depictions of it in films and shows?
Adoption of new technology always starts with early adopters who see the vision and engage. An entrepreneur’s challenge is to realize they will have to talk to 10 people to find 1. The trick is not to waste time on the 9. For more check out the classic book — Crossing the Chasm.
Are there any downsides that you can see coming from pediatric Cloud Computing?
I really haven’t thought of any — not that there might not be.
Throughout your work with Dr. Chang, both of you were pioneering a relatively untouched field with the overlap of AI in medicine. Given that there was such little work done in these fields when you started, what challenges did you face in getting support or finding information that was useful and how were you able to overcome these challenges?
Pioneering any new area requires a belief that it’s possible and the drive to make it happen. You see obstacles only as a test of your ideas and something to solve on your way to your goal.
I have 2, sorry: Can you explain some of the pitfalls with extracting only model parameters in federated learning? Can this dilute the interpretations made from the final model? With cloud computing in medicine, do we have the CPUs necessary to analyze the amount of data and supplementary data needed for useful models?
Work in federated learning has been driven by the need for consumer application to learn while preserving privacy and network bandwidth. Google keyboard and Siri area good examples. The application to healthcare is in the very early stages so we’ll find the pitfalls together. Check out some of the work in the federated learning and swarm learning.
You mentioned how one of the obstacles for transformation is change and that the obstacle of change is in itself an obstacle. In an era where science and technology are more than prominent, why do you think that we are so hesitant to accept change? How would you convince someone who is reluctant to change that transformation is happening?
Change is difficult for humans. We know we should exercise more and eat better, but do we change? Per a previous note the job of anyone leading a new science or technology is to find the early adopters, the one in ten and make him or her successful. In time what was once thought of, as radical technology will become Main Street. Many people thought eCommerce would never happen, until Amazon. Many people don’t think we could have autonomous electric cars, until Tesla.
How do you think can cloud computing integrated in health care be brought to areas where even the simple basic human needs is unmet?
Pneumonia is still a killer of kids in Africa, not because there is not low cost scanner technology, but because there is no one to read the scan. The country of Rwanda has one pediatric cardiologist. Rural America is no better. AI holds the promise of bringing the best pediatric cardiologist to every kid on the planet. From an infrastructure perspective BevelCloud’s edge cloud holds to key to building AI4Kids.
What advice would you give to someone who’s in the process of building a startup?
It will take longer than you think and be harder than you think. Patience, persistence and drive are key. You might also consider applying to an accelerator, such as Alchemist. In the spirit of full disclosure I’m the Chairman there.
What are the main problems you foresee in bringing the cloud technology into healthcare?
People & Money. Healthcare is managed in silos, largely because the system has evolved to high degrees of specialization. As a result making systemic change is difficult. Furthermore (and this is largely a US comment) the way we pay for healthcare is broken. Who will pay for the development of AI in pediatric healthcare?
What has been the biggest challenge you’ve faced along the way to applying your background in technology to pediatric medicine?
Speed. Healthcare is slow to move so we’ve had to be patient — but persistent.
Given how necessary collaboration and accessibility are to this “cloud” of data, how do you think this would be applied at an international scale? Would this project fall under the jurisdiction of the WHO or be in the hands of private companies and institutions or be managed some other way? In other words, how would it go about navigating each country’s laws and jurisdictions to be as effective as possible without being bogged down by policy?
BevelCloud engineered the edge cloud with privacy management as core. One of my former Stanford students is working with us. We just published why privacy is not a reason to slow down AI in medicine. Bevelcloud’s approach is to go to each of the 500 children’s hospitals and deliver a scalable, secure infrastructure on which innovation can occur, while preserving privacy and adhering to each countries rules around data flow. The edge cloud is unique in that we can store and process data in each country.
How do you think AI can address distrust and accessibility issues for minorities in healthcare?
First, we need to make sure any AI is trained with diverse data sets. Using data from kids in Palo Alto can’t train an AI algorithm for to diagnose pneumonia for kids in Africa.
Moving forward, what is something you wish you could encourage aspiring healthcare workers to do to assist in advancing pediatric care and technologies?
Get educated on software. Software is unique in that the cost to build is only the cost of imagination.
With cloud data, how can you ensure that you are including data from a wide set of ethnicities and people to ensure that physicians aren’t bombarded with a large repetitive set? How can you collect your data from rural areas to ensure inclusivity?
BevelCloud’s edge cloud is designed to insure data can be shared from a wide set of ethnicities. Today most of the data from healthcare machines (ultrasounds, CT, blood analysis is digital exhaust.
What advice could you guide us with as aspiring clinician-innovators? How can young people as you said “rally the troops” effectively and make lasting change in pediatric medicine?
Get educated, build teams, and build software. You all have seen it in the consumer space (Instagram, Uber, Netflix, etc.). None of this existed when you were born and now it’s commonplace.
Although Cloud Computing is the future of sharing data in regards to healthcare in pediatrics, what limits or concerns do you have regarding successful implementation?
This is not a matter of whether it will happen, but when. Our focus has been to create a community of the willing. There are of course obstacles, but nothing that cannot be overcome with time, money and patience.
Where do you see the cloud and how it is used in medicine in 5 years and in 10 years?
Predicting the future is difficult, but I do think the pandemic has and will continue to accelerate the development and adoption of technology in healthcare. One thing I’m certain of is it won’t be built on EMRs, which are just $100M billing engines.
You mentioned that there are a plethora of machines and scanners being utilized in pediatrics but “nothing is connected, fundamentally.” Why is the case? Evidently, amalgamating pediatric data for universal access and usage is beneficial; hence, what are some reasons why this has not been possible until your project? On a similar note, what are some challenges you have faced throughout this initiative to create a digital infrastructure, and what steps are you taking to address/overcome these challenges?
I’m not sure how to answer why it hasn’t happened yet. I do think it takes a special team with skills.
How can young students get involved in cloud computing early on? Does knowing this skill benefit professionals who are not in the data industry?
You can get an AWS, GCP or Azure instance to play with. Maybe write a simple application or work with a friend to build one. Many many enterprise and consumer applications are running on the cloud so there are benefits to getting educated beyond the world of healthcare.
My one question is: with trying to adapt and use AI to be able to read and diagnose scans in impoverished countries, what to do you think could be done to prevent the AI tech from having a slight misdiagnosis or reading a certain scan wrong, considering that the AI is being programmed by humans and humans themselves are imperfect?
Deep learning/AI algorithms need to be trained with lots of data. Building a self-driving car based on image data in Palo Alto might work well in Palo Alto, but take it to Irvine and it won’t. That all being said, AI is not deterministic, you’ll have to decide what false positive rate to accept (cry wolf) and what false negative rate (missing something) will work. If you’re interested in a more detailed understanding Google “ROC curve”. Clinicians also have a false positive and a false negative rate and are trained on relatively small data sets. The advantage of building AI algorithms is the computer can be trained on 1,000,000s of images — from around the world — and the results can be measured, and improved.
What are some challenges when implementing cloud computing in the world of medicine?
Center cloud computing has several challenges. As the servers from AWS, GCP and Azure run in about 10 data centers in the world, applications, which require low network latency, are challenged to run in the center cloud. Applications which have a large amount of local data (every CT, MRI, gene sequence, etc.) are challenged both to transmit the data, but also in the cost to retrieve it. Applications which require autonomy are challenged when there can be network outages which make the 10 data centers in accessible. Finally applications, which require greater degrees of security, privacy and compliance, can be challenging. BevelCloud’s work on developing an edge cloud is designed to address many of these challenges.
Could you go over more of the basics of cloud computing? It sounds like it’s just computers that can connect to the cloud and allow us to send data back and forth. How is this different from what we currently have?
Compute & storage cloud computing from Amazon, Microsoft or Azure have two basic differences from compute & storage (servers) managed by say the hospital. One, AWS, GCP, Azure spend a lot of money developing automation to manage the security, availability, performance and change of the servers. The traditional on-premises model puts that responsibility with the hospital and typically means staff/people are doing these functions. The second big difference is the business model. In AWS, GCP, Azure, I can decide I need 32 servers for a day, use them, and only pay for a days worth of use. This isn’t possible on-premises.
How difficult is it to get other children’s hospitals to agree to be a part of this system? Do they have any hesitations, and what might those be?
It’s always a challenge doing something new. Objections come from network engineering, security, legal, etc. But we learn something from each objection to perfect the product and how we educate people from all the disciplines about it.
As a teacher at Stanford, what would you say are some outstanding qualities of the specific school/students and do you have any inside information on what they look for exactly among college applicants?
Let’s start with, one of the jobs I’ve wanted to have for one year is to work in admissions because I’m curious what it looks like on the inside. So I have no idea what the university is looking for. We all know good grades, and once upon a time good test scores are ante, but with so many applications and so few slots the admissions office must have a tough time saying no to most people. I’ve taught at Stanford for over thirty years and while the average student is better and the top students are outstanding it is a bell curve. You might guess I talk to a lot of parents. While of course the top schools get all the attention, getting into Stanford, Cal, MIT or Tsinghua is no guarantee of success in your career or more importantly your life. By the way I graduated from North Carolina State University, and it all worked out for me.
As cloud computing becomes used more in field like healthcare, do you think we should be trusting of having our information stored and processed in data centers owned by a small number of companies?
Just so you’re thinking about it today your information is stored at 100s of companies, some of it is at AWS, Azure and GCP but there is plenty at Facebook, Wells Fargo and Uber. Managing privacy and security are now and will continue to be central themes. We as individuals need to build trust thru understanding what is being done to secure data and how privacy is being managed. I have always encouraged the Stanford Law students to take my class. Mirena Taskova, was one of those students. She co-authored this article on why privacy is not a reason to slow down AI in medicine. Future public policy needs to be informed by an understanding of technology, which few of lawmakers have.
Throughout your podcast, you continuously emphasize that there is a need who want to change the course of pediatric medicine and who have the vision to implement this change and take action. A lot of us are undergraduates or high school students who are not yet heavily involved in the medical field and the way that medicine is carried out. What are the ways that you recommend to us interns so that we can share those visions, take action and make these changes?
First, get educated. The transition we’re in to a more connected, software-driven world is just beginning. With the Internet today you can learn about anything. Take federated learning as an example you might
Watch Andy Trask on Youtube
Read articles like https://www.tensorflow.org/federated/tutorials/federated_learning_for_image_classification and https://arxiv.org/abs/2003.00295 which were recommended by our friends at Google.
Join a community like https://github.com/FedML-AI/FedML
Connect with interesting people working on federated learning in LinkedIn.
Continue to ask good questions
The cloud computing seems very beneficial for the hospital setting. I am curious if you think it could have any use within the public health/population health setting? If so, what would those look like?
The work BevelCloud is doing to connect healthcare machines to the edge cloud and enable safe secure data sharing from all 1,000,000 machines in all 500 children’s hospitals in the world is a beginning to using global data for public health. Just for you to think about today there is no way to access all the PCR data generated by Corona virus testing globally. Early variant detection is possible, but only thru connected machines.
Hi, you mentioned that people want to “do change,” so one must raise the problem by putting it out there to gather people to help. In terms of implementing artificial intelligence in pediatrics you said that partnering with the top 25 children’s hospitals in the world would lead to the rest of the 500 children hospitals to join; so, do you think that if one would like to implement any innovation in pediatrics it has to start with the top children hospitals first to really have an impact, or is does this only apply to artificial intelligence since it’s so abstract and underfunded?
Change can and often occurs bottom up, not top down. Netflix rise occurred one subscriber at a time. This is particularly true in consumer computing and with applications. The Moonshot Mission to enable safe and secure data sharing from all the machines in all the children’s hospitals is an infrastructure project which requires the cooperation and engagement by the hospitals. When the edge cloud is live then the potential for innovation bottom up using data will happen. Apple had to put a lot of iPhones in people’s hands before software developers wanted to build an IoS application.
Based on the importance and promise of AI, do you see medical school curriculum changing in the future to incorporate more AI?
Curriculum change is difficult. I looked at the CS curriculum 20 years ago at Stanford and it hasn’t changed that much. Part of the challenge is to add some class means something has to go away. If you were interested in learning about AI in medicine I wouldn’t wait for medical school. You can start now, maybe with Dr. Chang’s book.
Will implementing cloud into all medical site databases affect countries with low access to these types of resources?
Software and high quality information democratizes. Rwanda has a single pediatric cardiologist in the country. Rural America is no better. What if we could take Dr. Chang’s brain and attach it to every ultrasound machine in the world so every cardiac scan whether in Rwanda or rural America could have the benefit of his expertise?
What are some things that a mid-level provider can do to foster an environment of innovative thinking at a hospital that is resistant to changes?
Change begins with a community of the willing. In the world today that could be people in the hospital, or like-minded people in your state, country or across the world.
You mention that one of the largest problems that prevented pediatric medicine from growing was the lack of connection between medical devices/computing devices at different children’s hospitals, while other connections have been made rapidly to create the consumer network. Why do you believe it has taken so long for us to connect devices for the sake of improving pediatric medicine as opposed for consumer purposes?
Once upon a time in computing there were silos of data. Some machines ran Netware, others Unix, others Windows. Sharing data between the machines was very difficult. The United States government funded a project in the late sixties (https://en.wikipedia.org/wiki/ARPANET) to connect computing machines to share data in a handful of research sites in the country. By 1994, that grew to 1,000,000 machines, which was enough to give birth to early Internet companies like eBay and Netscape. And of course you know the rest of the story. We see the same thing in hospitals today. There are lots of silos of data, most of the data generated by the healthcare machines is just digital exhaust. Each silo is happy with the current state so it’s going to take a 3rd party to change the status quo.
What inspired you to make your podcast?
The pediatric healthcare community has embraced our work so I was invited to do the podcast by David Cole.
How often do doctors treat kids as little adults and what ramifications does this have on the quality of care young patients receive?
The podcast creaters called it Not Mini Adults, so they would be happy with your question. What I’ve learned is most of medicine is directed at adults. We have money and we’re getting older. As a result those in the pediatric community have to be a MacGyver and adapt the adult technology for kids, whether that is surgical equipment or pharmaceuticals. Thankfully, most kids are healthy, but as I’ve learned the conditions presented by children can be very different than adults. Ask Dr. Chang how pediatric cardiology differs from adult cardiology and you’ll understand this even more.
Perfecting the art of asking a good question is something we all can get better at.