Audioburst Search

How to Ask an Actionable Question

Automatic TRANSCRIPT

Eighty thankful is a tech bootcamp building. The world's next workforce with thankful you can launch a career tech online with the full support of a personal mentor support for this episode of talking machines comes from thankful ready to take the future you've been dreaming about and make it a reality that you want you are listening to talking machines ends right so do you need a Vaso presser do you need a ventilator these are treatments were if we could plan then that would be helpful for the patient it would be helpful for the doctor you've spoken to have said that seems to be a common theme in in machine learning so I actually did computer science and electrical engineering as an Undergrad I double majored there's in biomedical engineering and at the end of my masters applied to PhD programs and one time I t nice excellent and then so now you so what we try to focus on actionable insights for human health I sort of got to this statement through a series of improvements on initial and so we were able to do that using different kinds of neural networks and we had these really great performance numbers but then we were noticing that there were still a few people who are getting wrong during our interview with researcher this week's guest on talking machines is Marzia guests me who is a professor of computer science at the University of Toronto because I liked soldering also likes programming so I didn't WanNa give either either one up and then I went and worked for a couple of years at Intel I worked in there oh health group and in their emerging market platform group I got to travel a lot it was very fun but then I got a Marshall Scholarship so went to Oxford studied for a couple of years hair like mortality and I'll predict it is a good one right like we could we could do that but then what happens is you learned that if you predict mortality within twenty four hours who said I have this crazy high accuracy and it's because the feature was when all the machines get turned off that that patient is probably Bruce in an intensive care unit the doctors probably already knew that for example the dominant feature you find in a note that creates high accuracy predictions is freezes like instantly no matter what what kinds of models were using and so we took a look at it with a doctor we were collaborating with at Beth Israel deaconess medical center and it turned out that Oh end at the vector institute and when we sat down with Marzia a at the symposium for Geoff Hinton that was put on by the Vector Institute a couple of weeks back we asked for the first question that we ask all of our guests how did you get where you are it's been a circuitous path which I'm sure several of the people whole thoughts about what I wanted to do because originally as a PhD student I thought what I wanna do is predict something important in healthcare so let's let's find something important in health going to make it so it's things like that where I said okay maybe we don't want to predict something important we want something actual predicts something actionable and so then we thought okay. Let's look at predicting intervention all the priests you start to think oh man that's that's not what I was going for you've got a you've got an excellent wait for Catholicism at that point I had actually one student you are here and when did you come to the University of Toronto I've only I've only been here for a year I'm very fresh and you're also you're also at the veterans I'm Katherine Gorman and this week we are getting ready beginning to get ready for and so we have a little bit of a special episode for you we're going to be just fee some of the patients that we were miss predicting didn't really look like Miss Predictions and so for example we would say well we think you could win the patient off of the vessel Presser at time fifteen hours but that patient wouldn't actually be weaned until our forty five right so that's thirty hours right we're missing by thirty hours but then we go look at which is also still very new very yes so tell me about your group at at the university you've tried to work problems that you guys focusing on what are you hoping to do sort of like as a unit the issue is we have this data and it's labeled as part of a system if you're in an intensive care unit and there's a person on a Vaso presser that you could potential wean the nurse says maybe you can win this patient and then in the room next door there's a patient whose coating right and they need to be resuscitated you're going to put your attention the person who needs it right and so we don't have labels that lend themselves to complete trust right on so if you're trying to I predict something actionable in healthcare you may still not get exactly what you want and so then we thought of how we're GONNA look for these actionable insights in healthcare I don't just predict something and so we tried looking at whether you could use things like Wasserstein Ganz to say how a particular person might respond the patients notes and an hour fifteen the nurse says you should probably win this patient an hour twenty-five in her says we should really win this patient in an hour so you know the onto a particular drug right which is kind of interesting right and the reason it's interesting technically is usually if you wanna forecast how people respond you need to see an example of them before is because healthcare is this this tiny slice of human health so the thing that I tell all of my students is what does it mean to be healthy so if I asked you drug and an example of them after a drug absolutely which when you have acute care and you need to administer a drug sometimes you don't have that baseline often we don't try it and so we're able that work and then the reason that we've moved onto sort of this this final iteration of we want actionable insights in human health and not in health care with the healthcare system you are healthy when there is no data on you know the negative space into healthy how could I be healthier they'll tell you I don't know because we define health and being healthy as an absence of interactions up for myself and that makes sense right we have this sort of general idea of what it might mean for a single person to be healthy right but if you ask a doctor what does it mean not used again and specifically cycle again right to try and learn what do people normally look like before a- drug what do they normally look like after a dry the amount of problem here it sounds amazing is a huge opportunity because it means we could try to learn what it means to be indicator of when someone who's actually dead to like how do we understand the difference between being healthy and being not sick like those fundamental definitions seem so earth shifting I think for me some of the the realizations I had as a technical student really came from the biggest to other kinds of data and so what we've been thinking about recently as can we combined things like primary care records with other kinds of passes doing well and we interact with it most in fact when we're doing really poorly then we need to move beyond these electronic healthcare records that capture us at ours ugh and so try to forecast a little bit better how a person who we haven't seen respond to a drug might respond to a drug and that was really cool so I really liked that doctors that mentored me so I would say things like can you give me a label for this condition or what is an appropriate gap time they're in the hospital that sounds amazing what a transformation of just understanding the like very basic things that you're interested in like from predicting like okay this is a really great have mobile data with these more acute electronic healthcare records and get a sense of how do we keep people out of the hospital not figure out one they'll die once right but if we're going to do that we probably need to move beyond healthcare because if we only interact with the healthcare system when we're not realizations made me feel like we have a maybe more primal issue to deal with in healthcare we could use machine learning to have depression maybe we can help figure that out and I think there are many spaces in which we haven't explored the full extent of definition to create evidence we already sort of agree on what makes a chair cherry it probably needs to allow you to sit on it probably should be for forecasting in this recurrent model and they would they would say things like let's let's take it back even one step how do you think this label is yeah absolutely oh my God what a huge thing to tackle so are there particular areas you mentioned you mentioned depression you mentioned thinking about sort of constructed what sample size do you think it's based on does it apply to the population you're now generalizing it to and I think those kinds of realizations realize the things that I'm excited about there are some that are more methodology based questions in some that are more science questions right so some methodological questions I'm really thing that a lot of the knowledge we have in healthcare is very biased right so it's often based on small sample sizes that don't reflect a general population sort of comfortable unless it's you know there are some exceptions to that but if we don't really know what it means for example what it means to be using ambient data digital exhaust are there particular projects that you are really excited about so I think the that when trained on one data set that experiences many different kinds of shifts so maybe it's different electronic health care system that you have maybe you know what does it mean to you to be healthy you would say probably getting enough exercise sleeping well eating well sort of thinking about my body is a thing I need to take care of that sort of like how I define a lot of what I've been trying to understand from a privacy or federated learning or generalization perspective is how do we create models turns in health of you know disease ideologies of Endo types that could be created more understanding is what's really necessary sites the one that we all use as mimic and there's a huge void that's left in the wake of this one shining example of healthcare data that is is de identified and protected behind a data usage agreement and training agreement we need more things like that right so when we have people with private data sets maybe locally interested in has to do with reproducibility of the results that we get so there are very few data sets that exist for doing machine learning and health the the one that we all John there are often exclusion criteria that mean it may not generalize to even the people it's used to treat for example talking about a drug I think those kinds of a different population that you have maybe it's a similar demographic population but they suffer from different conditions how can we make sure that our model results are generalized simple so it's an interesting technical question some of the clinical areas that I've been focusing in on our I'd like to move beyond acute care and that's because acute care we've you know as community the people who do machine learning health we tend to focus on acute care conditions because that's where there's data and that's where honestly there are labels we have but it requires a significant amount of thought so understanding how normal conditions like pregnancy can lead to like what does it mean to have a healthy pregnancy that's a really big question and so there are there are technical ways that we can ask that question a set number of actions that we can perform that might benefit that broken arm so there's there's slightly more that we feel like we can do but if we're thinking about questions elsie pregnancies and so you can try to take a look at what that might mean another thing I'm really interested in is mental healthcare and that's because the labels there slightly more confidence and so when people are septic were all sort of septic in the same way a broken arm is broken arm and we feel like there are the outcomes for mom and baby is one thing I've been thinking about and the Nice thing about Ontario is because it's a single payer system we have data on many many eight to decide that we want to explore these kinds of questions with enough urgency that we are willing to hello I'm very sad that's meaningful but it may be more meaningful that every morning you texture mom I am sad enough that I can't get out of bed right that we need to be thinking about are gathering data in different ways and sort of like encouraging more gathering of Ambien data or Lake Virginia's data and combine it into something meaningful because for example if I look at your electronic health record you see a therapist once a month and believe that a qualified set of vetted researchers can look at data that has some privacy concerns but is important enough that we want to how for people even within a single diagnosis like major depressive disorder and also there that seems there's a ton of opportunity for machine learning to do this the experience matters more than a single measurement by a health professional for machine learning to really perform best so how how do you think we can is it that you search for what can I do I'm so depressed so it feels like there are opportunities in some of these chronic conditions wear your day to day you look at the state or to try to address what might be reasonable options moving forward so you can imagine that just looking at de identified health records her have less definition and if you if you talk to many people in the mental health space they'll tell you we don't quite know what treatments work like like mimic you can address a small set of questions you can address hospital questions but if we want to address these larger set of questions we do need to ask people to do that it's protected by a an entity with scientific rigor that has a merit based review I think that would be incredibly valuable because right now it's hard to imagine how we would collect data in any other way you can't run a randomized controlled trial for this because we don't want to recruit people who are ill and people that would be really valuable yeah a holistic a holistic view take on the on the questions that you may want to be asking or approaching so that you you are not over fitting for your problem but does that does that require some level of literacy about what the questions think that's a huge resource and it's also fantastic because it combines these different sorts of data sources Ryan you have some wearable data you have expert coded data you have balance between maybe person's own desires for privacy and society's need for more information so that we have fewer women dying in child am not literate about how an airplane works I really am not like I can say this confidently I don't really know how the airplane works about is the FDA trying to solicit comments from both patient groups and from scientists to ask what does regulation look like broadening the pools that we already have to include more how how do we accomplish the bringing in of that other ice cream clean need as a as a society going to be asking are on the part of the person who's donating their data so this is a question that different countries have dealt with in different ways so for example in Ontario if you of how different societies have made individuals fuel empowered right by having a representative council that gets to make decisions about is this in the best interest of people donate their data the same way that they donate blood or Daytona an Oregon right if you can donate this data towards science right and believed hey with you using their genetic data to explore issues of tribe ancestry it's not something they want and so I think that there are there are great examples I want to ask a question using first nations data you have to present the tribal council with the research question you're going to ask because maybe they would not be okay and to be frank I don't care to understand it because I have a lot of confidence that the people operating that airplane are doing so unethical birth fewer people committing suicide right these are these are problems and we'd like to be able to address them and I think the question of literacy is a really hard one in question and I think we all have these internal review boards right ethical review boards that we pass so I think we we have tools where we know how to address the manner and there's a federal aviation agency that is going to oversee those individuals and there's regulation one of the things that I've been really happy or the NIH saying we want to know what scientists think transparent machine learning should be what does that mean that's that's sort of a square to have this medication or that medication we want to figure out what seems to work for people and that requires a broad sample and so things like the UK by self reported data and so I think having data sets like that creating national data set that are representative of different countries and the problems that we're facing us societies or are you just trying to sort of answer maybe a personal bitch here I think that that can be done responsibly right so scientists often are trying to get at a very these processes and I think people will want some amount of information about what kinds of questions can you ask what sort of review will be undergone in order unders because what's transparent to me may not be transparent to somebody else so I think that we do have good tools start investigating how we might want to go about the you know a question in a lecture about will I have this machine learning model and it's going to revolutionize career coaches and a network based in your local area your time is now visit thankful dot com slash machines to learn about our courses and start building the future she term can we make that a little bit more solid and what what would a patient like to see what would a provider like to see what are these things mean and what do they mean two different states is and disrupt you know healthcare condition a with technology be and it's going to help people see who ensure that this is an ethical study I think we can address this issue yeah do you think we are thinking about them as fast as we need to be I'd like to think about them right in those concern me so so the fact that people who don't have a a depth and yeah I think so so here's what worries me about once a week I get an email or edit funded deploy it though that's concerning to me because we know that any modification in any process has I at the VEX petits in the field believe like honestly believe that they could come in Traina model make a start up supervised deep learning for this problem and that you talk about privacy concerns and how you alleviate them but eighteen months that's going to be five years and to NIH grants and seven graduate students and I think that there's not a literacy about how hard I just don't understand why you're worried about these old fashioned methods and these old fashioned problems because I know that now Oester to talk about the situations in which the supply and when it won't work but right now what happens is these lectures that normally in sections I suppose I think that's concerning so I am worried that as a community we need to many wonderful computer science talks that start with something like you know this thing that you think is impossible it's not achievable the data that would be great yeah can you tell me when that happens because I would love yeah and then and then he said that gets the strengths of our papers and show good results and then you you wait for people to read the limitation section of your paper or come and speak to you the new unsupervised deep learning doesn't use data it's like other they said it's like a bait nuys that when we when we speak at technical conferences what do we tell our students what have we been trained to do we've been trained to present in recent results and a person stood up the back of the room and asked a question they said I see that you are doing this modal hyperbolic press about about machine learning about scientific articles right the come out about machine learning and these papers limitation sections but nobody reads the reckon with the fact that we're having a moment and in this moment there is a disproportional amount of interest and cle- intrigue in what we're doing in this field and so to be responsible we need to be overly transparent about where we are as a community and where we're going into your privacy concerns so I think that the the big opportunity here is for us to see Sion's but I I do feel like some people hype it and I worry that that hype creates a perception that we can use unsupervised deep learning oh you know humans could never do this I can train a model in in an hour to do that better than any human this thing that an infant can do at be it learns with no yeah exactly with no data and you know so like they kept going kept going and then they talked about how phones would be made out of paper in a decade and you know is it got it got even more out there as a the question finished but some problems are and how progress is being made at an astonishing rate right gets the community is doing a fantastic job of pushing research in good direct my favorite so far claimed that I've had is I was I this is this is a true story true story I was giving a lecture and presenting nobody who who is outside of the community would see are being seen by millions of people who think people on Youtube who the last thing that they encountered that deal with artificial intell- think about ourselves as humans intelligence transparency fairness these things there has one technical meaning and there is another conflated sort of like very invite scientifically because so many of the words that we use to mean things in this field technically also means something else and carry great weight for the way that we like percents disembarkation is really hard I think that communities that work for example in fairness agents was a movie or you know like a work of fiction right and so I think people make connections between words you might say in a lecture and a work of fiction they might have read recently and I I do worry about that yeah because we find we find ourselves in a unique position scientifically or the field finds itself in unique of made huge efforts recently to try and say what does fairness mean and and this is technical fairness right this is not we're not trying to for making sure that sixteen other groups can hopefully plan their own little tiny conferences which seems huge every year I spoke name real fairness in a social justice sense for example right there's there's no epidemic claim here I also feel like when we've worked on fields like transparency the community there I've been very happy recently to see several talks where people have said like let's let's not say do the one of the prior our workshop chairs about this before I started and she gave me the fantastic advice of human meaning and like I don't even want to get into when you're trying to speak a different language right like this is just English it's challenging I do think that the the word and so I think either distancing ourselves from certain terms and making a collective decision about what we're going to use or redefining terms and saying this is a technical things that that is challenging being in the workshops is that the workshops I think spiritually are intended to be a place where early work I think you'll do a very good job at it but I don't think you'll be happy about I think she was right it's so I think one of the so I'd love to talk to you a little bit about your your work with the nerves conference you are on the organizing committee for the workshops this year which is this definition and it is this I think those are ways to deal with it but I mean intelligence were never we're never going to get past that one that's that's just going to be a problem forever like very exciting Yeoman's work thank you for doing it I know it's like planning conference sort of goes on and then the the workshop chairs sort of charged with planning it can be shown right so I in the past really benefited from showing a poster at a workshop and having people come in and say transparency anymore because that means too many things to too many people and it's it's a squishy technical term it could mean any of a number of things the thing that's been really impressive to see is the workshop organizers deal with the massive amount of interest that they've been presented with an get that great feedback and I think because there's so much pressure now on Europe's as a conference the workshops have had to raise the bar because workshop organizers have done this fantastic job of this Herculean task of notifying all of these workshop presenters and getting for later and so I I've had several papers that have been made by doing an early work as a poster and workshop and then later on that's going to be morphed because and also you know as as cherish this year we instituted this earlier deadline this global notification deadline to ensure that people who need extra time for visa acquisition a few is a workshop get a thousand submissions. How are how are you supposed to deal with the volume and so I think that the the thing that's been really in could actually potentially get visas and it's significantly earlier in the calendar year than than it used to be and the meany conference can operate the way that it does but I think the Nice thing about the workshops is they let in a younger crowd of people and they bring in this group that maybe might be intimidated by the main Europe's conference and often as a Pi or as a their ducks in a row and allocating tickets? So I've been very impressed that this two day you know fifty you know parallel session Jordan has not been terrible because therefore people doing it and so if you have four very responsible very responsive people who are working towards objective you'll be able to make it so that's been really fantastic and tell me a little bit this was the first year that the workshops included a call around ideas into to include explicitly a call for diversity is because there have been several studies that have shown that if you don't have for example a diverse organizing group Prophet what you WanNa do with your junior students send them to a workshop they get to present their early work and then oh you're there anyway go to the main conference network with some of the other going to a workshop so it was it was meaningful for me antastic that's awesome well yeah I mean the lineup this year looks amazing and I can't imagine what it was like to have to sort through all of that people should be able to come together and present the best ideas and I was very impressed with The diversity plans for several workshops around diversity and making sure that you are accounting for that tell me a little bit about that and did it did it do what you were hoping for so I actually think it did one of the reasons that we want ideas if a diverse group of people are able to contribute and actually one of the things that was also really nice as the workshop chairs you won't get suggestions for diverse speakers and many many others right these are all social science studies right nothing to do with computer science it's just how stuff I have three other fantastic co chairs who who have done a very very good job of helping out with this so I feel like the the women's thing how do they interact do they organize and so we wanted to be very explicit and state that we value diversity many many different the extreme interest specifically around using machine learning tools artificial intelligence tools methods on health health care data thinking about these there's were very responsive to feedback so we noticed for example that there was one workshop where they said you know we're going to try to recruit these speakers that thing that you're doing that's really interesting maybe think about doing it this way add this on and it's not fully baked so you've got that flexibility you can bring it in and it's fantastic because that'll turn into a officiant way would you like to do that and they said Oh wow that's great you know we're we're very happy so people were very responsive and they were very understanding of the the intent the and is seeing all of this press and all of these papers all of this this fire about machine learning is think about the population what would you say to someone who is coming in from the human centered health side of things and perhaps his expertise there a clinician or a healthcare practitioner or care staff anybody who deals with health on a day to day basis hasn't been a significant shift in for many years right there are very few of those in healthcare and so I would say as you come in let's say you're starting so I think it did do what we intended and I hope that what this will do is allow for the community to problem right because in the end we're going to have the best roles that a group of clinicians would not disagree on where there is a standard that there is not much disagreement on that is there talk to the other graduate students at this sister group at another institution so the really valuable and I think it's one of the favorite things that I had as a graduate student was goal the aim and the same thing with early global notification deadline. We tried to be very clear that we're not doing this to torture you it's not just collaborations with a computer scientist you need to bring all of your knowledge and all of your context around a condition and data set to that engagement and also why we upped the global notification deadline because we have these goals we have these aims so let's work towards them to move back to your to your work a little bit given Tiriac it'll mean something different did you realize that it's coded differently by these kinds of doctors and these kinds of doctors did you know that if you and so I think everybody has been really fantastic about understanding why we've included the diversity statement and ask them for a diversity plan wants to work with someone who has expertise in machine learning or these tools what advice would you give them on communicating or collaborating the the advice that I would give a of this minority group in this condition rate so this kind of context that you're aware of and I'm not because I'm not a healthcare provider you need to tell me did you realize that this condition was relabeled two years ago and so any data you get from before two years ago won't meet this human system that creates this data right it's not a visual capture of an object we all agree as humans that object recognition is important the examples you've seen where machine learning has been very successful they are well defined tasks with Labor like spirals and these window box things that's interesting and meaningful right if we're talking about healthcare data even if we're thinking about some are going to try to have this diversity plan and we said Oh you know a sister workshop has a a diversity plan that looks kind of like yours but they were able to implement it in this really when we visualize neurons or we have networks hallucinate often it's for visual data and so you can say wow out at not as a time series but as just let's let's take just snapshots of humans what does it mean if for example I tell you and questions and in your particular path from understanding like how can I get a better label on this data to drilling down on what what does it actually mean when you're looking at Oh payer does that mean that I should make sure to stratified those who get really intervention and don't get an early intervention is it okay to say things is incredibly valuable because these are things I can only attractive I'm aware of them and the the real challenge here is that when we audit models because humans do it and so we can say a ha this is an important task and we would like to understand how the brain works better and how Mike we should over-focus on certain minority groups because we know that they experienced disproportionately more discrimination in society we know that causes stress suicide does that mean that I should be looking at primary care records for children does that mean that I should only look in places that have let's poor people that die there needs to be some sort of our reckoning with the fact that we're dealing with a system to have machines work in a similar way and so it feels like we have an aim that even if we disagree slightly and how you technically constructed feel all up actual type two diabetes later on what are the most effective treatments you know right now even guidelines for that hallucination of the the most doggy dogg grade or or the fact that the filters that activate for car include these wheel tell me the reason that this person died of Sepsis is because here they

Coming up next