Deep Learning on Healthcare (1)

Deep Learning on Healthcare (1)

Hello, everyone. My name is Jihoon Jeong. Creator of this youtube channel “Information
and Intelligence”. Nowadays, more and more people from outside
Korea is visiting my youtube channel, so that I decided try to start english lecture. This is the first try … so that I’m a
little bit nervous. For the Korean listeners, I’ll post Korean
subtitle for this lecture and I want you to know that I’ll lecture in Korean, too. Don’t be too disappointed. I want to introduce myself first. I graduated Hanyang University Medical College
and I’ve got Medical Doctor degree there. And I started my master degree for public
health at Seoul National University. At that time, my master thesis title was “Patient
Satisfaction Analysis and Quality Improvement in Hospital Management”. I used Case Based Reasoning and Decision Trees
such as C4.5 and CART, association rule called a priori for analyzing the question & answers. Yes, it is Good Old Fashioned AI technologies. But at that time that was working quite well. After that, I started my PhD degree at USC
on biomedical engineering. I majored in optics and my PhD research theme
was “Multi-Mode Optical Imaging”. The key issues of my research was managing
N-dimensional big image data to make human perceptible efficiently. So most of my research was focused on dimensionality
reduction using PCA, ICA, etc and unsupervised learning such as K-means clustering algorithm
and smart signature selection processes with embedded database. After my PhD period, I came back to Korea
and worked for the hospitals and universities. Currently, I’m senior teaching fellow of
Kyung Hee Cyber University, Media and Communication major. I also had a position at KAIST Graduate School
of Culture Technology as adjust professor. At that time, my major interests was VR and
AR, smart object and space and Human-Computer Interaction. I wrote many books in Korea on various themes
such as computer programming, trends and future scenarios and nowadays I’m writing SF novel
at CoinDesk Korea. I’ve also worked for many Korean companies
as an advisor. Samsung Electronics, NHN, LG Electronics,
Hyundai Auto Company are some of them. As an angel investor, I’ve invested more
than 50 startup companies so far. Now, I co-founded two angel investing and
startup accelerator companies BigBang Angels and Digital Healthcare Partners and also working
for those companies as managing partner and partner respectively. But, maybe this is much more popular than
who I am in English speaking people. Have you seen this picture? When there is a ranking introducing laziest
guy page this is almost always introduced. Actually this is my son. I uploaded this picture at twitter at that
time and that was popular. So, I am the father of world laziest boy. OK. now we need to start lecture. I’ve picked up my first lecture theme on
“deep learning on healthcare”. Actually, this lecture was given to Google,
Sloan School at MIT, Center for Data Science at NYU and Massachusetts General Hospital. And also I spoke at Nvidia GTC Korea last
year on this themes. This is more like MBA style lecture based
on my experiences especially on my investment portfolio companies. I’ll try to explain what is the reality
and myths when you guys try to do something using deep learning on healthcare. That is totally different stories from just
research. This is quite long lecture for just an online
lecture, so that I’ll divide this lecture into several episodes time about 10 to 15
min or slightly longer. Let’s start. Firstly, I want to explain “What Deep Learning
can do for Healthcare?”. There are 4 different stake holders, patient,
medicine (hospital or doctors), public and pharmaceutical companies. Each of them has different needs and restrictions. For patient and medicine, machine learning
can help to understand physiological changes over time. We can estimate what was going on based on
the health record and lab, medical image data. And, we can forecast progression or onset
of the disease. This is called prognosis and estimating prognosis
after the diagnosis and/or treatment intervention is very important in medicine and also for
the patient. Finally, we expect personalising treatment
strategies in near future based on genetic, health and life-style data. Oh, I forgot to thank Danielle Belgrave for
adopting this amazing slides and items. I’ll use her material in next slides, too. For population health or for the public, machine
learning can do more complex job. It can help to elucidate average effects and
deviations from average effects. For the public people, it is very important
to know what’s the average, since it is related with cost/efficiency, but it is also
important to know when we can find out deviations from average effects. Based on the machine learning results, decision
makers can do better decision. So that policy recommendation is another benefit
using deep learning. Health education. Nowadays, may mobile app or internet service
can do health education for the patient. If the app understand patient situations and
data better, their education will be much more personalized and effective. Outreach. As you know, medical service is very expensive. So that many patients in the world to visit
doctor’s office. But, if machine learning agent can help the
people to decide proper decision with no cost or very little cost, it will help to outreach
poorer people. One of the key issues in public health is
finding the source of disease and cause. Prevention is the most important issues. We expect machine learning also can help these
issues. Finally, machine learning also can decrease
the healthcare inequalities. For pharmaceutical industry, machine learning
also can do many things. Typically, this is the pipeline of developing
new drugs. Start from ideas, basic research including
animal studies are performed and there are typically 3 phases of clinical trials. After that we need to get approval for this
new drug from regulatory agency. And this is not the end. we need to know and monitor this drug is working
well and what will be the side effects. We believe that machine learning can be helpful
for every pipeline processes. This is the diagram from CBInsight. Every year they are publishing quite good
report on AI and healthcare industries. I want to show you this diagram because we
need to know really many areas exist in healthcare industry. This is the field that many companies and
startups need to collaborate to make some progresses. It is totally different from consumer internet
business. this is not covering every details AI healthcare. But as you can see, many medical imaging & diagnosis
startups were created and many of them are doing very well. One of them, Lunit here, is my investment
portfolio. And there are several startups dealing with
mental health and virtual assistants. Some of the companies are using wearables
to promote health using AI technologies. In the hospital, already many companies are
actively working for patient care and hospital management solutions based on machine learning. Some of them are more focused on emergency
room and surgery. Nutrition and lifestyle management, monitoring
is also very important for wellness. And nowadays, more research based deep learning
companies and patient data & risk analytics are growing. This is the diagram from Lunit’s oncology
brochure. I like this diagram very much. We are experiencing another paradigm shift
of the medicine. Before 1990s, medicine was just like art. Every doctor has their own opinion and since
they have their license and authority, it is very difficult to challenge their decision. Education was done just like master-apprentice
relationships. It changed in 1990s, more and more insurance
system govern the medicine, they requested to make some guidelines for payment and right
decision. To deal with this request, evidence-based
medicine is now de facto standard. Every medical decision should be made with
evidence and such a guideline needs to be made based on very strict scientific scrutiny. It made medicine totally different animal. Many inexperience doctor was gone and average
quality of the medicine was much better than before. But, there are also trade-offs. Making evidence is very high cost and time-consuming
job. Even though you’ve got some evidence when
you are practicing, it takes very long time and cost to change previous guideline. It is easily take 10 years to 20 years. Why this process is taking so long time and
paying big cost? Because it depends on many doctors agreements
and they are get together once a year and always conservative position are majority
without clear evidence. This has very big problem. Let’s say one doctor find out “this is
wrong treatment and that is better” and it takes 20 years to agree everybody and it
is now guideline for payment from insurance system. Is it OK? Well, through those 20 years … Every patient
got deprived of better treatment options. I think this is not ethically acceptable. We need to overcome this disadvantage for
evidence based medicine. Fortunately, I think machine learning and
deep learning can alleviate this problem. We can now access to large-scale digital data
in hospitals and using deep learning technologies we can make many evidences with the support
by deep neural networks. This kind of accomplishment need to be supported
by medical data. So we can say it as “Data-Driven Medicine”. I think we are now entering totally new medical
paradigm, data-driven medicine. Under the graph, this is the bar that representing
the degree of deep learning software’s accomplishment. We can now detect the abnormalities 10 out
of 10 doctors can also do. It is called ‘Easy’. If 50% of experts can miss and AI can detect,
it can be called ‘Intermediate’ task. If 1 out of 10 doctors can detect, but AI
can do, then it is ‘Hard’ task. If no human experts can detect the abnormalities
and AI can, it will be very “Challenging” task. But, already, some of the project are entering
this tasks. Finally, AI can do more than detection. It will translate clinical implications and
provide unprecedented insight. Then, we will be at the medicine with totally
new paradigm. Still long way to go, but we can make the
goal like this. OK. This is the 1st lecture. I hope many researchers and entrepreneurs
to innovate medicine with deep learning to like this lecture. I’ll be back in next week with a subject
for the types of medical data and different approaches for them. Thank you very much for listening this lecture.

4 thoughts on “Deep Learning on Healthcare (1)

  1. 잘들었습니다. 고맙습니다~
    그런데 왼쪽 이어폰으로는 소리가 안들리네요. 녹화할 때 뭔가 잘못된 것 같아요ㅋㅋ

Leave a Reply

Your email address will not be published. Required fields are marked *