Better Medicine Through Machine Learning | Suchi Saria | TEDxBoston 

Machine Learning: Start with linear regression (did it in school) where you try to find a line that fits data. y=mx+c, where you are trying to find the values of m and c. Once you find a line that fits well, you can use it to predict y for a given x. Machine learning is the same basic principle (assuming you use neural networks). A neural network is just an n-dimensional polynomial (n can be billions), and you are tweaking the coefficients of this polynomial until it fits your data well The process of tweaking the parameters with sample data is called 'learning'. So, in essence you are just doing a curve-fit. The problem is, reality cannot always be modeled by a curve (or n-dimensional plane). So all AI hits a wall. There will never be a Fully Self Driving car (ask Elon to lie in front of his Tesla while I set it off on auto pilot, see if he agrees), nor will there ever be a Fully Automated Diagnostic robot (or whatever this woman is peddling). Why? Because the underlying neural network can only model a subset of reality (the subset that can be modeled by n-dimensional curves). Investors, with freshly printed cash from the Fed, splurged on all things AI. Watch them die as Fed stops printing money.

I have same thinking too, lets just revolutionize the healthcare

Don't feel bad and apologetic to your patients. Things can be presented beautifully and perfectly on power point, infact in real life it's hard to implement 0.5 %of what being presented on the power point.

I am a software engineer. AI/Machine Learning is glorified curve-fitting. You probably did linear regression in school, finding the best line to fit your data in a two dimensional plot. Neural networks are simply an extension of this, into n dimensions, resulting in curve (we cannot visualize higher than 3, so we cannot draw it). Just like in your linear regression, if the data is well-behaved (already linear) you get a good fit, and can use the line to predict values, so neural networks can work very well if the n-dimensional polynomial is indeed representative of the data. But if you have an outlier, no amount training will ever get your line to touch that point. So it is with neural networks. It can never model reality, because reality cannot be represented by an n-dimensional surface. So we will never have self-driving cars, or diagnostics by computers. Obama got donations from Silicon Valley and diverted lots of money to 'e-health'. Boston was a cheer leader. They have given a long time ago. We have no idea how animal brains work, so it is naive to claim we can replicate it. As for the state of medicine... my siblings are all MDs in USA. The state of medicine is pathetic because most of the money goes to non-doctors. Meanwhile, doctors are doing fine but there are good doctors and bad doctors. Not much we can do about that.

Read Deep Medicine by Eric Topol. He goes into this more in depth. There are a lot of factors that could be recorded that would involve wearable technology and the like.

Naveed Farrukh I agree we need an anonymous database of medical records, including images, and other information collected by doctors on patients. Only problem is making sure the exact information about the patient is hidden from third parties.

I think This comment is the most relevant and professional to the TREWS performance. It shows TREWS is slightly better than human physicians. The only question is whether TREWS performance is able to be reproducible ubiquitously to any other data sets. If not, it is purely resulted from data cleaning or manipulation.

interesting

....and yet TREWS is taught as a CASE STUDY in graduate "Machine Learning in Health Care" courses, taught to students. It has become a cornerstone case study for teaching.

You made an interesting point about involving the pharmaceutical industry to incentivise the implementation of this system! Something that can be tried and tested to see where it could take us.

I do, to a major degree!

Partial replacement only. Unlike robots, where you can take the risk of them falling, no one can risk a patient from the alert mind of a physician. If we talk of a country like India, there are many socio-economic & cultural factors too that influence treatment decisions. Tailoring treatment to each individual patient is not expected to be within the reach of artificial "intelligence".

Ok ...we aren't able to solve 100%of problem...but from this method at least we can solve 95% problem...isn't this a breakthrough then?...just to nullify a solution because it cannot solve a problem 100% is not right.

I guess deep reinforcement learning sounded cooler to say, also it's easier to implement for large data since there is no need to hand label it.

I am currently 25 thinking of going back to university, but I am undecided to what to choose! Computer science major or pharmacy. I want to be able to improve the healthcare sector. Thank you

@@didodido103 cs

@@didodido103 I will chose pharmacy and minor in computer science

AI

very dead

wtf does that have to do with anything. This talk was about using AI/ML to analyze data and diagnose diseases/conditions, not how AI can be used to advertise essential oils.

Holistic medicines are usually untested, very little data is available regarding their efficacy in treatment of specific symptoms. While it may work for some people, it is not something that can reliably be called 'medicine' , as we use the term now. Many of the drugs are infact active substances isolated from plants, based on the 'remedies' and then studied further in controlled conditions, regarding their effects on human physiology. The reason why 'natural' remedies is discounted is because unless it is studied further in controlled conditions, (like recently curcumin, found in turmeric, has been shown to be beneficial as an anti-inflammatory agent, and turmeric is an integral part of Ayurvedic medicine and everyday ingredient for Indian food), It cannot be touted as a remedy by healthcare professionals who are liable for the advice they impart and the consequences of their advice. There is much to be discovered. We have barely scratched the surface. Its not even a game of validation of previous myths. We need to study and understand what is truly beneficial and what is purely placebo.

"reliance on technology" is a bit too broad of a conclusion. AI could reduce time in hospital, reducing such risks you speak of.

any useful information please (reply on me). I always had melatonin for a good sleep. even I can have a sleep without any medicine but I sitll want to sleep better. but if I use melatonin too much I will lose one or more days of (good -sleep ) I think its ( withdrwal symptons ).so I help AI to help me decide how much I should take.

the audience are not all computer engineers. she had to explain ML to them somehow

Yes. She is an extremely renowned professor at Johns Hopkins University.