 So machine learning can solve three fundamental problems in cancer care and the first one is access to care today Patient survival rate is a function of their geographical location so what you see here is a map of the United States and you see that there is a very weaker relation between the number of occurrences of cancer and The mortality in this state the second question is a question of cost quality care cost a lot By most conservative estimate we are paying 80 billion dollar per year only for care And as you can see year after year this cost continue to increase make it in unsustainable So the question is what can we get for all this cost and this brings us to the second question Which is how effective is our care? So the good news you can see that mortality is actually goes down But it's very slow For pancreatic cancer five years survival rate is only 7% breast cancer is considered to be success 90% half a million women every year die from this disease So I can hardly call it a success despite all the amazing progress that we made in biology in medicine in the way How we can record information and all the measurements that we are taking about the patients What I would say is that we are really in a primitive state in terms of machine learning and data science in Processing this data imagine that your breast cancer patient There will be thousands of pixels that record your mammogram your pathology slide may be digitized again Lots and lots of very specific information Then the doctor reads it writes you a story in a page Then it's father distilled into a few variables and those are the variables that give you the treatment So this whole pipe up line you are losing information specific to a patient and to make the matters worse This is a statistic. I want you to remember from this talk according to American Society of Clinical Oncology All the decisions today are based on 3% of the population that participated in clinical trial The rest is just not used Imagine what we can do if we utilize all these information and that's exactly what we are trying to do at MIT We're taking information about millions of patients all their own measurements and try to predict their outcomes The sensitivity to treatment the le diagnosis and models of disease progression and the first step here is to take all the data that currently is in language and Extract specific things. So for instance, you want to know what are the features of specific cancer? What happened to the patient machine cannot read it directly So what we build at MIT is a system that takes this data and Translate it into a table into a database where every patient you record all the pertinent features of their cancer and their treatment We can do it with a very high accuracy We did it for the patients in the partner system 96% accuracy and let's say you want to study particular cancerous condition Let's say atypia you can now write one or two queries and get all the patients that had this condition and see what happened to them Over the years Previously to that if you wanted to do it you had to do it by hand You had to go through all these patients and identify which one this condition and you can see in your England Journal of Medicine You publish about breast cancer with 600 cases with one query We can get it to 6,000 cases for a single hospital alone So we can really study disease in a big scale and study it better Another important case where this kind of structured approach can help us is to prevent over treatment For instance if a woman today diagnosed with high-risk breast lesion, it will be excised Only 13% of them are cancer which means that the rest of them did the surgery for nothing disfiguring surgery So in this case with machine learning we can predict 30% of the patients that are actually benign and spare them from the surgery So now it brings me to the most exciting thing that happens in my lab Which is deep learning models that can read the wrong measurements in this case mammograms We can now do it rivaling human performance both in terms of predicting cancer and assessing density And what you will see is a movie that we made in Massachusetts General Hospital where exactly two weeks ago We installed these tools and you will see a radiologist which before they look at the mammogram They have machine reading of the mammogram. This is my collaborator Coni Lehman She just looks at it and approves the reading of the machine So this is really exciting the part that personally excites me is to be able to do stuff better than humans can do What you see here is a breast cancer patient me in 2014 What you would see that my cancer was in the breast since 2012 and it progressed So the question is how early machine can predict something that human cannot do and that's exactly what we are training our deep learning Models to do so I would like to conclude my talk by showing you the map of the cancer And this is mortality and you can see there is a lot of red in this map And what I firmly believe that with advancement in machine learning and big data We can really democratize cancer care and wipe out the red. Thank you