Real-time Big Data Forecasting of the COVID-19 Outbreak

May 22, 2020

Information

Nicholas A. Christakis, M.D., Ph.D., M.P.H. Sterling Professor of Social and Natural Science

ID5239

To CiteDCA Citation Guide

00:00I would now like to
00:02introduce our next Speaker,
00:04Doctor Nicholas Christakis Doctor.
00:05Christakis is a sociologist,
00:07an physician who conducts
00:08research in the areas of social
00:10networks and biosocial science.
00:12His current research is
00:13mainly focused on two topics.
00:15First, the social mathematical,
00:17an biological rules governing,
00:18have social networks form,
00:20and 2nd the social and biological
00:22implications of how these networks
00:24operates to influence thoughts,
00:25feelings, and behaviors.
00:27Doctor Crystal Kiss.
00:28Thank you for being here.
00:31Thank you all.
00:32Of course I can't see any of
00:34you and there are 268 of you.
00:36I see.
00:36It's a very odd experience using
00:38zoom to do this and I lately
00:40have not been using slides in
00:42order to communicate in a way
00:43I think is more effective.
00:44So I'm going to try to cultivate
00:46some visual images in what I'm
00:48going to talk to you about.
00:50I'm going to start with just a
00:51couple of brief remarks or to set the
00:54stage about the coronavirus pandemic.
00:55I suspect everyone on this audience
00:57will know these numbers already.
00:58Then I'm going to talk a little bit
01:00about the issue of waves of pandemics.
01:02And then I'm going to tell you a little
01:06bit about some of the work in my lab,
01:09including an app that we just released
01:12a soft launch earlier this week
01:14called who nala HUNALA that we think
01:16can be quite helpful and it's quite
01:19different than all the other apps
01:21that are available at the moment.
01:23So as you all know,
01:25the are not for a pathogen is felt to
01:28be something intrinsic to the pathogen
01:31or as it relates to the host as well.
01:34Uh,
01:35this is the number of new cases
01:37that can arise from a prior case.
01:39In, uhm,
01:40you know fully susceptible population.
01:41That is to say,
01:42no individual is immune to a pathogen
01:44and the end host hasn't responded yet.
01:47We haven't taken any actions were
01:49not pulling apart and living in
01:51like Hermits for example were
01:52interacting normally and it is
01:54estimated and there was a recent
01:56meta analysis just released by the
01:58land set a couple of days ago or
02:00a couple of weeks ago that they
02:02are not for this condition is.
02:05Probably around 3,
02:05maybe 2.4 somewhere in the range
02:07in the high choose probably,
02:09and that's quite high.
02:11Actually that is a high are not.
02:13The seasonal flu has and are not of
02:16about 1.3 Ebola of about 1.5 to one point 9.
02:19And of course a chicken pox
02:21around 3:00 to 6:00.
02:22And of course measles is the
02:24champion which has one of the
02:26highest arnotts estimated of around
02:2818 new cases for each new case.
02:30This so called are not of course is
02:33different than what is called the RE.
02:36The affective reproductive rate,
02:37which is the number of new cases
02:40that arise in a kind of more
02:42steady state of the epidemic.
02:44For every old case.
02:45So some people are immune.
02:47People are beginning to
02:48take responsive action,
02:49for example, were beginning to
02:51engage in physical distancing,
02:52and this would be known as the
02:55affective reproductive rate and the
02:57ariav courses can fall as an epidemic
02:59proceeds because of what we do or how
03:02we've been affected by the pathogen.
03:04I just a different parameter,
03:06something known as the case fatality rate.
03:08That's the fraction of people
03:10who died conditional on coming
03:11to medical attention course.
03:13Whether you come to medical
03:14attention or not depends on what
03:16kind of health care system you have,
03:19what patients do when they get symptoms,
03:21and so on.
03:22And sometimes people estimate
03:23something instead known as the
03:25symptomatic case fatality rate.
03:26The SCF are,
03:27which is what fraction people die,
03:29conditional on developing a symptoms.
03:31We think this is still much
03:33more debated than they are not.
03:35We think this numbers between 0.5 and
03:371% or perhaps as low as zero point 3%.
03:41Now the case fatality rate for the
03:43seasonal flu is about zero point,
03:451% ignoring other features, just on average.
03:48About one out of 1000 people who
03:50get the seasonal flu will die.
03:53We think that this disease is between
03:555 and 10 times as bad as that and the
03:58Stars to the epidemic we're having
04:00right now is about a 10th as deadly
04:04as the SARS one epidemic from 2003.
04:06The case fatality rate for
04:08SARS of one was about 10%.
04:10That was a much deadlier.
04:12Condition and the 1918 influenza A
04:14pandemic had a case fatality rate.
04:17We think about four to 5%.
04:20So these two parameters there are
04:21not or the RE something about the
04:23transmissibility of the disease,
04:25and the fatality of the disease
04:27could be put on two axes,
04:29and you could plot every respiratory
04:31pandemic for the last 100 years
04:33on these two axes,
04:34and then you could see well, how?
04:37How does this pandemic compared
04:38to previous ones,
04:39and when you do this exercise,
04:41it's actually kind of alarming.
04:44Of the worst pandemic we've had
04:46in terms of how transmissible it
04:48was and how deadly it was is 1918
04:51in the upper right corner.
04:53The second worst we had was in 1957,
04:56which had sort of Intermediate
04:58Intermediate Lethality,
04:59an intermediate transmissibility,
05:00and this disease is probably slightly
05:02more transmissible and slightly
05:03more deadly than the 1957 pandemic,
05:05so it's getting up there.
05:07It's above the 1957 pandemic and
05:09not as bad as the 1918 pandemic.
05:12However, the point is, it's bad.
05:15This is bad and I think what we
05:17have to accept is that this moment
05:20in historical time.
05:21That we all happen to live in is
05:24a moment when a new species of
05:26pathogen has entered our species.
05:28There's a new germ out there that is
05:31is going to have its way with us.
05:34It's going to spread in our species
05:36and and affect us and it is bad and
05:39without action many people would
05:41have died even with the actions
05:43we have taken about.
05:45100,000 people have already died.
05:47Now as a nation and as different districts,
05:50we have taken different sorts of
05:52actions that people have been sort
05:54of locked down in various ways.
05:56We've closed our schools,
05:58we've pathetically done some contact
06:00tracing and some testing we have had
06:02work from home orders or stay at home
06:04orders in most states in the union,
06:06and the point of this,
06:08of course, was to flatten.
06:09The curve was to reduce the
06:11intensity at any given moment of
06:13the number of cases that we have.
06:15But every single respiratory pandemic.
06:17The last century has come in waves.
06:19All we have done by flattening the curve
06:21is we've not eradicated the pathogen.
06:23We just stopped the transmission.
06:25The germ is still out there.
06:26It will come back to China.
06:28It will come back to us.
06:30It's going to come back.
06:31All of the rest atory pandemics,
06:33even the mild ones of the last 100
06:35years have come back and typically
06:37they come back in the fall.
06:39And this has to do with a variety of things.
06:41It has to do,
06:43and they typically come back
06:44every fall for two or three years
06:46before they kind of damp down.
06:48Uh,
06:48and and eventually end and if time permits,
06:51we can talk a little bit
06:53about why pandemics end.
06:55And this has to do partly
06:57with human behavior.
06:58You know, when the fall comes,
07:00the students return to school,
07:02adults return to work.
07:04We move indoors.
07:05So are are dense interactions
07:06the proximity which we interact
07:08with other people increases,
07:10which enhances transmissibility.
07:11There may be some environmental factors
07:13which affect the pathogen heat or humidity,
07:16or our responsiveness to the pathogen.
07:18So we might do better in the sunny weather.
07:21Our bodies might be more able to resist
07:24the pathogen is not sunny weather.
07:27And so on.
07:28And ultimately,
07:29one of the factors of parameters
07:31that we could think about is it
07:33bikini ologists In addition to.
07:35Well, there are a number of parameters,
07:37but In addition to the transmissibility,
07:39the R and the lethality that
07:41case fatality ratio is something
07:42known as the attack rate,
07:44which is the fraction of people
07:46who actually get the disease in
07:48the end and in the end for this
07:50pathogen it'll probably be above 50%,
07:52maybe a bit higher if we overshoot
07:54in ways we can discuss.
07:56If if there's time now in the 1957 pandemic.
08:00Nationally,
08:00about 25% of people got the
08:02disease were infected,
08:03but in some hard hit areas
08:05it was as high as 40%.
08:07I got the disease now.
08:09Our best estimates of how many
08:11people have gotten it already
08:13in the United States is low.
08:15So for example,
08:16if you look at Sweden,
08:17they just released a quite good study.
08:20Sweden has had less severe sort of
08:22social or physical distancing than
08:24we've had about 4% of Swedes using a
08:26national Sero prevalence study just
08:28released this week have been infected.
08:30And they have been mixing
08:32more than we've been mixing.
08:33As you know, in New York it was about
08:3621% of New Yorkers in a rather good
08:38sort of representative sample of
08:40New Yorkers have become infected,
08:41and other studies around the United States
08:43of high quality that have been done
08:45show relatively low fractions of people
08:47have yet been exposed to the epidemic,
08:49so we have quite a way to go still.
08:54You know, before we before we reach the
08:57final attack rate that we're likely to get.
09:00So the disease is going to come back
09:03more people are going to get infected.
09:06Unfortunately, we're going to have
09:08more deaths with this condition.
09:10What can we do to predict the course of this?
09:13Oh, and I should say that I think
09:15that I've been flip flopping on my
09:17opinion as to the likelihood of
09:19successful development of a vaccine.
09:21So some days I'm optimistic some
09:23days I'm pessimistic.
09:24I'm not an expert on vaccines,
09:26but what I suspect is that no matter
09:28how fast we go on the vaccine,
09:30it's likely that plus or minus six months
09:33the vaccine will be widely available
09:34around the same time we otherwise
09:36would have gotten herd immunity anyway,
09:38so I don't think the vaccine
09:40is going to change the story.
09:42Very much unfortunate.
09:45How can we predict the course
09:47of this epidemic?
09:48Can we develop some tools that help us
09:50to confront how we might emerge from
09:52the lockdowns that we're currently
09:54engaged in and might anticipate
09:56the course of the epidemic in the
09:58fall when it comes back my love.
10:00Has been doing quite a few projects
10:02in this regard.
10:03We have a in the midst of developing
10:05New Haven wide sero prevalence study
10:07that will follow people longitudinally.
10:09Our work if we launch it,
10:11will have some different features
10:12than some of the other studies that
10:14have been done around the world.
10:16Some features that we think offer some
10:18interesting research opportunities,
10:19but I'm not going to talk about that today.
10:22Instead,
10:22I want to talk about two other things.
10:24One is a work that exploits the
10:26use of human movement.
10:28We had a paper just published in nature.
10:30About two weeks ago that took advantage
10:33of a big data that track the flow
10:35of people through Wuhan in China
10:38throughout the whole of China are
10:40up through sort of late February,
10:42so we had data on 11.5 million
10:44transits using phone data.
10:46So people paying the tower when
10:48they were in Wuhan,
10:49and then they relocated to another
10:51part of China and such data can be used
10:54to track the flow of human beings,
10:57even if you don't know who's
10:59infected or who is not.
11:01The movement of people, which,
11:03depending on data availability,
11:05could be tracked in basically in real time.
11:08Can be used.
11:09We showed using a certain model
11:11to predict the intensity,
11:13location and timing of the pandemic.
11:15So there are tools you can use that
11:17rely on other sorts of information.
11:20For example human movement,
11:21and it doesn't have to be
11:23fun data. It could be tolling
11:25data on highways as cars,
11:27a shift from place to place.
11:31Sort of other kinds of Geo.
11:34Location data, air travel data, and so forth.
11:38So human movement can be used.
11:40Another kind of thing that can be
11:42done is is using searches and probably
11:45many of you are remember the so called
11:48Google flu trends that was proposed.
11:50You know 10 or 15 years ago.
11:52Now the idea there was the following idea.
11:55So right now what the CDC does or
11:58other monitoring agencies do and what?
12:01What doctor Weinberger's talk just
12:03talked about as well is you wait
12:05in a central location for data to
12:08accumulate and be reported to you.
12:09For example,
12:10testing data for people doing influenza
12:12testing or people showing up in an
12:14emergency room or death counts for example.
12:16And what that means is that you know
12:19some period of time distant from
12:21now two to three weeks from now.
12:23You might know where the epidemic is today.
12:26Well, that's frustrating because
12:27you're always behind the epidemic.
12:29You can never get out ahead of it.
12:32Well, Google flu trends was an idea.
12:34That's it.
12:35Well,
12:35maybe we can use something about
12:36people's behavior today like there
12:38searching behavior for flu symptoms.
12:39For example.
12:40Maybe that can tell us where the
12:42epidemic is today and their first
12:43paper by Larry brilliant group
12:45showed that that could be affected.
12:46Then there was a whole literature that
12:48emerged that sort of debunk that and said,
12:50well, no,
12:51there it won't be effective and so on.
12:53But that's an illustration of a set
12:55of tools like the movement of data
12:57that I just described you from.
12:58The other project we had done.
13:00It's an illustration of a set of tools that.
13:03Allow you to a survey or note
13:05where is the epidemic today based
13:08on what I'm seeing today?
13:10But we have another idea that I'm about
13:12to tell you about that allows you to tell
13:15where the epidemic will be in the future.
13:18So it's not just rapid notification,
13:20it's advanced warning of the epidemic.
13:22How does this work?
13:24Well, imagine you're the network of people.
13:26There may be many of you can
13:28cultivate in your mind's eye,
13:29a kind of image of such a network.
13:31Since I'm doing this talk without slides,
13:33I have to kind of try to do that.
13:36Their little dots that are people
13:37in lines that connect the people.
13:39Many of you have seen these images and you
13:42have this sense that there's a middle of it,
13:44which is a very densely
13:45interconnected group of people.
13:47And then it feathers out
13:48to the social periphery,
13:49where there are people who,
13:51let's say,
13:51only have very few friends and whose
13:53friends have very few friends.
13:55So in the middle of the network you
13:57have people that are very popular and
13:59whose friends are very popular and as
14:01you get to the edge of the network,
14:02you don't have those qualities.
14:04That sense of centrality in the
14:05network can be quantified in a
14:07variety of mathematical ways.
14:08In fact, the mathematics of that lies
14:10at the core of how Google you know
14:12the billions of dollars that were
14:14made by the founding of Google using
14:15the so called page rank algorithm.
14:17So you can figure out what
14:19is the central website.
14:20Or you can figure out who's this
14:22central person in a network.
14:24Now imagine is such a network that
14:26a pathogen begins strikes someone
14:28at random in the population and
14:29then begins moving across the
14:31ties through the social graph.
14:33You should have the intuition that it
14:35should reach central people in the
14:37network sooner in the course of the epidemic,
14:40popular people should be more
14:42likely to get infected.
14:44And popular people should get
14:45infected sooner in the course of
14:47the epidemic than unpopular people.
14:49That means if we can identify this,
14:51incidentally,
14:52is the same reason that popular
14:53people get better stock tips
14:55or more information sooner.
14:56'cause if information flows
14:57through the network,
14:58there are more central in the network
15:00they can acquire this knowledge just
15:02like they acquired germs sooner
15:03in the course of the epidemic.
15:05Actually,
15:05there's a side light on some work
15:07we've done in the lab on the
15:09evolutionary biology of friendship,
15:11where we argue that the spread
15:12of germs is the price we pay
15:14for the spread of information.
15:16That's a whole other topic anyway,
15:18so central people can be like
15:20Canaries in a coal mine.
15:22If we can find them and monitor them,
15:24they will tell us those people
15:26should get the epidemic should
15:28strike them earlier in the course.
15:29Then it strikes a random person,
15:31so identifying such people and
15:33monitoring them gives us a tool to
15:35forecast the future course of the epidemic.
15:37My lab about 10 years ago for the H1N1
15:40Pandemic showed that this was possible.
15:42It could be done,
15:44and now we've developed new tools
15:45in combination with a mean car.
15:47Posse's group in the at Yale
15:49Electrical Engineering using
15:50certain machine learning tricks,
15:52which I'll describe in just a moment.
15:54That allow us to deploy these
15:56ideas in the form of an app that is
15:58sort of like ways for coronavirus,
16:00where everyone anonymously and
16:02privately contributes a little information.
16:03This information is aggregated
16:04and then fed back to the users,
16:06just like when you use when you drive
16:09on the highway and you use ways,
16:11you report that there's a traffic
16:13accident or that there's a traffic jam,
16:15and then this informs people that
16:16are behind you on the highway and
16:18gives them something of value.
16:20You get something of value and
16:22you share with others.
16:23It's like a crowdsourced way.
16:25Of tracking traffic,
16:26but we have like ways for coronavirus.
16:28So for example,
16:28when we saw all the politicians
16:30and celebrities that were getting
16:32that were in the news with getting
16:34sick from coronavirus early on,
16:36it was not just that they were
16:38rich and famous,
16:39so they were able to get tests and
16:41people cared what happened to them.
16:43They actually were getting sick more
16:45so Boris Johnson was out there shaking
16:47hands with all these other people.
16:49He was also spreading the germs exactly,
16:51which was irresponsible or Tom Hanks
16:53and his wife are all of these people.
16:55There are more connected
16:57so they get stricken.
16:58Earlier they were Canaries in a coal
17:00mine so our new app which were just
17:03released this past week on Monday
17:05were doing a soft launch this week.
17:08If you would like to use it
17:10you can go to who nala HUNAL,
17:13a.yale.edu and download it.
17:14We are asking that you not
17:16broadly advertised it.
17:17You can invite your friends but
17:19please don't broadly advertise it yet.
17:21We're still debugging it.
17:23If you find any bugs please email me or.
17:26Let us know and then we will.
17:28We're working on it and then next
17:30week we're going to do a kind of a
17:32big goof and try to get a lot of
17:35attention and try to get if we can.
17:37Hundreds of thousands of users.
17:38The more people that use the app,
17:40the better it can monitor what's happening
17:42in terms of the flu in your area.
17:44The app when you upload it,
17:46it only takes a on the first time you use it.
17:49You tell us some basic information
17:51about yourself and then.
17:53Every day you are pinned.
17:54If you, uh, if nothing is happening,
17:56you've had no symptoms.
17:58You haven't seen the doctor.
17:59You say no, no,
18:01you're done.
18:01If something is happening,
18:03you have some symptoms,
18:04or you've seen the doctor or you
18:06been out and about in some way.
18:08You might take a minute to answer
18:10and then you immediately get
18:11feedback in the form of your risk.
18:13Your told how much respiratory diseases
18:15there where you live based on a
18:17machine learning algorithm that takes
18:19advantage of lots of information not
18:21only from the CDC and other sources,
18:22but from our users.
18:24And then you're also told your
18:25individual risk based on where
18:27you are in the social network.
18:29For example, if your friends,
18:30friends,
18:30friends have the flu three weeks ago.
18:33This means your risk is different
18:35than some of his friends, friends,
18:37friends did not have the flu three weeks ago,
18:40or if your friends friends are very popular.
18:42Your risk is different than if my
18:45friends friends are not very popular.
18:47This information,
18:47which we validated mathematically.
18:49Another work is then combined and
18:51processed and fed back to you and
18:53you can monitor your risk everyday.
18:55Like I said,
18:56it's anonymous and private and it is
18:58like ways for respiratory disease.
19:01Now and we and and we are also
19:03building from this a dashboard.
19:05So imagine that the state police in
19:07a in a state wanted to know which
19:10parts of the highway are dangerous.
19:12In principle they could take a years
19:14worth of reports or months worth
19:16of reports by citizens traveling
19:18on the highways,
19:19saying where are there traffic
19:21accidents and they could say,
19:22Oh my goodness,
19:23this part of the highway is very dangerous.
19:26Maybe we should put a redesign.
19:28That part of the highlight.
19:30So this is this.
19:31Our app will work the same way.
19:33We're building a dashboard
19:34that could be used,
19:36for instance by people running a
19:37hospital system that want to monitor
19:39what's happening in their area,
19:41or that could be used to to detect
19:43where is are the hot spots,
19:45and to see it's coming.
19:46We're seeing a spike in people
19:48who are central in the network
19:50and not a spike in average.
19:52People that difference between those things?
19:54The difference between the at risk
19:55people and the less at risk people when
19:58there's divergence in those curves,
19:59that's a harbinger.
20:00But the epidemic is going to spike
20:03that the 2nd wave has begun,
20:04for example,
20:05or that the lockdowns are beginning
20:07to foster a spread of the pathogen,
20:09and so of course the individuals
20:11who get this information can act
20:13accordingly on their own benefit,
20:14but collective decision makers can
20:16also now have some vision into
20:18what's happening in the system.
20:21And I'll just stop.
20:22I have two more minutes.
20:24I'll shut up what I want to say is.
20:28That to my eye,
20:29there is no escape from this pathogen.
20:32It will become endemic among us.
20:34There are to our knowledge 7
20:37coronavirus species that that afflict
20:39us for that cause the common cold.
20:41I think that those pathogens that
20:44cause the common cold right now,
20:46the Corona viruses are probably distant
20:49echoes of previous introduction of
20:51coronavirus pandemics into our species.
20:52In the distant past.
20:54I think what happens is these pathogens
20:57tend to mutate to become milder.
20:59Remember from the point of view of the
21:02pathogen, it doesn't want to kill us.
21:04It's better if it doesn't kill us.
21:06So if we if we die too fast
21:08we don't spread it.
21:10So as time goes by,
21:11in general,
21:12pathogens mutate to become milder,
21:13but unfortunately and then there
21:15are two other coronaviruses,
21:16the SARS 1 from 2003 and Murs,
21:18the Middle Eastern respiratory syndrome,
21:19which has a very low are not,
21:22which is one of the reasons
21:23this has not become pandemic and
21:25then the one we're in right now,
21:27which unfortunately is awful for us and.
21:29It will spread and it will
21:31kill many of us I think is the
21:34sad truth until such time is.
21:36Either we invent an effective vaccine
21:38or we get herd immunity so it will
21:41become endemic and we have to use
21:43the best tools that we have to
21:45cope with its existence among us.
21:47Thank you.
21:52Thank you very much doctor Chris Nagus.