Learning from COVID-19 Data on Transmission, Health Outcomes, Interventions and Vaccination

February 24, 2021

Xihong Lin, PhD
Professor, Department of Biostatistics
Harvard T.H. Chan School of Public Health

Tuesday, February 23, 2021

Information

ID: 6223
To Cite: DCA Citation Guide

Download Transcript

00:03So, so first it is my honor
00:05to today to welcome Doctor.
00:07She only as our summer speaker,
00:10Xihong Lin is a professor from Harvard,
00:13jointly appointed by both as deaths.
00:16Annabelle Slash Department and she has
00:18received broad recognition and many
00:21awards for her great contribution to the
00:23field and her research has covered so
00:26many different topics ranging from Mr
00:28Logical work including hypothesis testing,
00:30had dimension statistics
00:31and color inference tools,
00:33data applications.
00:34And the computational statistics
00:36such as Statical Genetics.
00:38Scalable statical inference
00:39as well as applications with
00:41epidemiological and health data.
00:43So today she will share with us
00:46her work on analyzing large scale
00:49coordinating databases from both
00:50China and the US and provide several
00:53takeaways and discuss priorities.
00:55It's Insp endemic,
00:57so I will not occupy everyones time for more.
01:01I will hand it over to see home from here.
01:05So should we start
01:07alright? Thank you laying so
01:08much and for inviting me and
01:10for the very nice introduction
01:12and so I will share my screen.
01:141st oh, I cannot share the screen.
01:16I have just met
01:18your Co host so you should
01:20be able to get it now.
01:25Can you can you share your
01:26screen I can do now? Thank you thanks.
01:33Alright, can you see my screen?
01:35Yes. Cool excellent alright?
01:38So I'll share with you some of the
01:42work and we have been doing last year
01:46on the Covic 19 so this is the data
01:50just downloaded earlier this this week.
01:52So as you can see that the right now
01:56they're over 110,000,000 cases in
01:59US and also 2.4 two point 4,000,000
02:02deaths were 110,000,000 cases and
02:052.4 million deaths worldwide.
02:07So if you look at the curve on the left,
02:10that is the case curve and for
02:12a few selective countries.
02:14So as you can see for both the UK and
02:16also United States and the number of
02:18cases had been going down in January,
02:21so that is a good sign.
02:23And also Israel as you know that
02:25Israel has been a really leader in
02:27a vaccination and so their cases
02:29have been going down as well.
02:31But if you look at Africa I think
02:33the number this particular country
02:35you can see the number of cases.
02:37Has been going up likely due to the new
02:40violence in Africa and on the right.
02:42That is the case curve with the desk curve.
02:45You can see the patterns pretty similar,
02:47especially you can see for this
02:49African country and the number
02:51of deaths has going up quickly,
02:53so that is really worrisome.
02:55So here's a talk outline,
02:57so I'll start talking about the covid
03:01transmission intervention and using the data,
03:03and then they turn talk about
03:06the USN word data.
03:08Then I'll talk about epidemiological
03:11characteristics of Mackovic.
03:12Then I'll talk about the 221
03:14playbooks and also the defining
03:16challenges in particular about
03:19the vaccine rollout and uptake.
03:21Our focus more on the uptake and
03:24also how can we do scalable testing?
03:28And in particular,
03:29I talk about this support design and
03:32we call the hypergraph we called hyper.
03:35That is based on the hypergraph
03:39factorization.
03:40So I started working on the Covic
03:4219 research mainly by coincidence
03:44and last February.
03:46So my post,
03:47our former postdoc column one
03:49he is currently is professor in
03:51school public Health at Wild on
03:53Science and Technology University,
03:56which is located in Wuhan,
03:58and so last February,
04:00'cause Wuhan was epicenter.
04:01So I wrote him a message asking
04:04how and he and his family were
04:07doing and he told me that.
04:10He and his colleagues were
04:12analyzing the Wuhan data,
04:13then at that time there was already one case,
04:17and in Seattle and one case in Boston.
04:20So I sense that the the the
04:23virus might spread and so.
04:25I decided to join them and working
04:28on analyzing the Wuhan data so we
04:31work the we work for several weeks
04:34and in February especially child
04:36when his Cody worked really hard and
04:39then finish this preprint and we post
04:43the preprint on March 6 and so then
04:46with the hope that the findings we
04:48want we want to share their findings
04:51with the US and also other country
04:54as soon as possible to help them.
04:58The other countries,
04:59and so I I did not expect this
05:02preprint attract lots of attention
05:04that you can see the number of
05:07abstract view and download,
05:09and also there lots of.
05:13A lot of free trade and also the news outlet
05:17coverage and also the policy documentation.
05:20And after I twittered this paper in
05:23the Twitter and so then this paper
05:26preprint cover too much stuff and so we
05:29decided to split the paper into split
05:32the preprint into two paper and one
05:35was published in JAMA and last April.
05:38Another one was published in
05:40Nature last summer and so this work
05:43on the JAMA paper was led by.
05:46I'm Pan and a towel,
05:48and here they both were addressed.
05:50pH alarm and country.
05:52Who is the Dean of school?
05:54Public health as well doing science
05:57and technology University and the
06:00Nature paper was led by Charlo.
06:03And then emerge after the preprint
06:05was posted them in Twitter.
06:07And then I got quite a few interview request.
06:10But as a station,
06:12my first reaction was turned
06:13down all the interviews,
06:15so I turned on all the interviews
06:18and in March, and then April,
06:20I decided that probably is not a good idea.
06:24It's good to talk with the media
06:26and then they can understand the
06:28scientific funding correctly.
06:30Then that will help,
06:31and the combat the Covic.
06:33And so I decided to accept the interviews
06:36and so you can see there's a quite
06:39few coverage of the findings and
06:41also the interviews and in the US medias.
06:44And also in the UK medium and
06:47also in scientific journals
06:48such as Nature and Science.
06:50And the one thing I I learned
06:53was that 'cause when as a faculty
06:56member in Academic Institute
06:58and we were not trained to.
07:01Speak to the media and so to the
07:04scientific communication on the two.
07:05The general public is very important,
07:08so it's important to have more
07:10training and in this area.
07:12Another thing I learned was is
07:14important that speak in simple
07:16language and without the jargon and
07:18so general public could understand.
07:21And then I testified the in the
07:23science and Technology University
07:25$0.10 at Technology Committee
07:27of UK Parliament on April 17.
07:30And so this sense technology Committee
07:33has about 8 to 10 Parliament
07:35members like a senators on it.
07:38And then they later on they also
07:41invited few other witnesses.
07:43And then they wrote a letter to the
07:46Prime Minister Johnson and they make
07:49a 10 recommendations in their letter.
07:52Under so so then I I was honored
07:55that several of my recommendation
07:58we included in their recommendation
08:01in the recommendation they made it
08:04to the Prime Prime Minister Johnson.
08:07And then, uh,
08:08this last month and as a one year
08:12anniversary of the colic outbreak and
08:15nature published their third editorial
08:18and focusing on the major findings,
08:21a key finding from ideology.
08:24And so we will.
08:26We will honor that our paper was
08:29featured on the in this editorial.
08:33So let me first talk about
08:36the Wuhan analysis.
08:37Then we talk about the US and other
08:39countries and the 1st I'll introduce
08:42this effective reproductive number.
08:44So this concept,
08:45our value is right now and everybody
08:48understand what that is and so
08:50that measures average number
08:52of infected people by one case.
08:54So you can see on the right if
08:57Artie called before that means one
08:59person could affect to four people.
09:02That means the virus spread.
09:04So therefore, to in order
09:05to control the pandemic,
09:06the army to be less than one.
09:10And so here's a woman analysis under.
09:13So we found the two major features of Kovik.
09:17The first feature is the virus
09:19is highly transmissible,
09:20and so the before January 23rd,
09:23and there was no intervention.
09:25As you can see,
09:27the number of cases had went up quickly,
09:30exponentially,
09:31and so the first case was reported
09:34on the in this orphan and seafood
09:37market and this seafood market
09:40was closed on January 1st.
09:42Under then on March on January 23rd,
09:45and that was a lunch of the lock down and so.
09:50As you can see that before the lock down
09:54the estimated RTT values is about 3.5,
09:57so that means each purse case could be.
10:00Like 3.5 people and so that is not good.
10:04So that means this is really the
10:06disease is very transmissible
10:08and then after the lock down and
10:11with the social distancing is
10:13really how the RT value dropped to
10:16a little over 1.2 and then.
10:19But it's not good enough and so then
10:22after February 2nd and that was a lunch
10:25of the centralized isolation and quarantine.
10:28So they basically the city build.
10:31Two new hospital field Hospital and
10:33also 16 field hospitals converted
10:36from the stadium and expectation
10:38Center and after that.
10:40And then you can see the number of
10:43cases drop down very quickly and the RT
10:46values and by March 8th was about .27
10:50and then the pandemic was controlled.
10:55And so to estimate those RT values.
10:58So basically the model we use is a person
11:01partial differential equation model
11:03and so this captured by the left here.
11:06Then you can see here is a from them.
11:10So here this is the symptom onset
11:12and between exposure and symptoms
11:14answer discord, incubation period.
11:16So generally this is about 5 days and then
11:19from the exposure to presymptomatic onsite.
11:22This so this part period
11:24called latent period.
11:25So that means a patient is infected
11:28and but the person is not transparent,
11:32doesn't transmit the disease, and so.
11:35Between the pre symptomatic period
11:372 symptomatic period so this period,
11:39the person,
11:40even though a person doesn't have
11:42a symptom and but very load,
11:44is high enough and then the person
11:48could become.
11:49Infectious and so this period is
11:51about two days between presymptomatic
11:53two symptomatic and between
11:55exposure to pre symptomatic period.
11:57This is about 3 days,
11:59and so we built in those components
12:02and in the model,
12:03and so we introduce this
12:05presymptomatic compartment,
12:06and also because at that time the testing
12:09case will not that widely available,
12:12and so therefore there were a lot
12:14of cases which were uncertain,
12:16and so therefore we built in this.
12:20Unstained components,
12:20the observed data here those are the
12:23observed data and then those are.
12:26Then we construct all those components and
12:28drew the deep partial differential equation.
12:31So this is isolation component.
12:34And so so here the we
12:36after you feed this model,
12:39one can construct the
12:41reproductive number or value.
12:43So basically,
12:43as I mentioned,
12:45the datas are here and those are the data.
12:48The rest part are basically coming
12:50from the partial differential
12:51equations and then we fit the
12:54partial differential equation,
12:55Apostle model and using MCMC and
12:57the Gulf Coast parameter estimate.
13:00And so here you can see that before
13:03the intervention on our journey,
13:05then if there was suppose there
13:08were no intervention,
13:09then this will the blue curve will
13:12be the predicted number of infected
13:15cases and so you can see that 170% of
13:18the Wuhan population were infected.
13:20That will reach the natural
13:22herd immunity and so now has a
13:2510 million population size.
13:27Then that means 7,000,000 people need.
13:30Need to be infected in order to
13:32reach the herd immunity and so
13:34this is not a good strategy and
13:37many old people will die and the
13:39sudden after supposed when user,
13:41social distancing and lock down
13:43and so you can see from Wuhan the
13:46number of cases still going up
13:48but not as fast as the without
13:50intervention but after the.
13:52Centralized isolation quality on top
13:54of the social distancing and the
13:56number of cases went down very quickly.
13:59And so this is what what they
14:02did in one hand.
14:04So if the case subject was tested
14:07positive and this person to patient
14:09was admitted to the field hospital,
14:12and so this is different from
14:14the US in the sense that in the
14:17last spring and US also build
14:20multiple field hospital,
14:22but they only admit as a severe diseases,
14:25so amount and diseases cases
14:27they were isolated at home.
14:29And so that these notes are isolation at
14:32home could still infect the family members.
14:35And but in Wuhan all the mild cases and
14:38were admitted to the field hospital,
14:41so they were monitored.
14:43If anybody became severe and the
14:45patient was transferred to the ICU,
14:47the regular hospital and for the
14:49people who had a symptom and but who?
14:53Because at that time there were
14:55not enough testing kits and so
14:57they were quitting the odds of
14:59hotels and will University dorm.
15:02If so, all the children and stay with
15:05the parents so the family were together.
15:07If anybody became a test positive
15:10and the person was transferred to
15:12the field hospital in two weeks,
15:15if a person was tested negative
15:17and the person went home,
15:19and similarly for close contact and they
15:22were quarantined as a hotel as well.
15:25I see how can I ask a very basic question.
15:28If I look at the Group One Group
15:30two Group One is confirmed.
15:32I guess if it means if you
15:34perform some kind of PCR test
15:36you are positive and Group 2 is
15:37with symptom but not confirmed.
15:39Can it go the other way?
15:41Can you first be confirmed
15:43but with no symptoms?
15:45Oh yeah, I think there were cases.
15:49Possible that would have had
15:51no symptoms but test positive,
15:53but at that time because they were
15:55not enough testing kids so many
15:58of the cases and who were able
16:00to be tested at the same time,
16:02so that's why there were lots
16:04of undetected cases.
16:05I said thanks.
16:09And so this strategy worked
16:11quite well in Wuhan.
16:13So in less than two months they
16:17reach 0 confirmed case and then by
16:20March 18 and by April 8 and after.
16:24To confirm the case for three weeks
16:26and then the city was reopened.
16:29So it's a whole thing only took
16:31in two months and so the first
16:34take home take home messages,
16:36the social distancing and centralized
16:38isolation quarantine were critical
16:39for controlling the outbreak,
16:41so using the social distancing
16:43alone that help but was not good
16:46enough so that helped make our
16:49reduce around 1:00 and but did
16:51not bend the curve and the reason
16:53is the there were lots of the.
16:57Community transmissions on the
16:59social distancing help block the
17:01community transmission that is
17:03between household transmission but
17:05within family transmission and cost.
17:07Place transmission with common
17:09and so so the social distance.
17:12Distancing does not help block that and so,
17:16especially with how many families are
17:19multi generation families and they
17:22live in apartment so it compared to
17:24US as even harder under to isolate.
17:28At home.
17:29And the idea that centralized isolation,
17:31creating to social distancing that help
17:34bend the curve and stop the pandemic.
17:38And so we validate those findings
17:40and in other countries last spring.
17:42So if you look at the curve on the left,
17:45that is a Italy data.
17:47So you can see that Italy also did
17:50the social distancing that help reduce
17:52the R and the R curve lingered around
17:55one for over a month and did not bend
17:58the curve but and also on the right
18:01that Germany data in the spring.
18:04The same thing under the curve did not band.
18:09And the second feature from
18:11analyzing the Wuhan data is the
18:13Covic is highly converged on the,
18:15so we estimated about 87% of
18:17the cases were undetected.
18:19So in other words,
18:20the detected cases was only
18:22the tip of the iceberg,
18:24and so you can see that on the on
18:28the left and the right bars are
18:31the detected cases and then the.
18:34Yellow bars also uncertain the cases,
18:36so we estimated on the we estimate
18:39entertainment rate and so we estimated
18:42about 87% of the cases were uncertain,
18:45and many of those cases were asymptomatic
18:47or mildly symptomatic cases.
18:49By adding the yellow and red that
18:52can give us a prevalence estimate
18:54that is about 2.5% in one hand,
18:57and so this is similar to the
18:59theological studies based on antibodies,
19:02and that was about 3%.
19:04And then US result very similar.
19:07the CDC did theological study last
19:10year and then the estimated about
19:13862 twenty times the number of
19:15cases were six to twenty time of
19:19the cases which were reported.
19:23And also those undetected tasted
19:25post a high risk of resurgence
19:27if one reopened too early,
19:28lifting the controls, and so we estimate
19:31the probability of the researchers.
19:32Think about this is the first
19:34day and one has a confirmed case.
19:37When has a confirmed case.
19:38It doesn't mean there is no
19:40case at all because there are
19:42still a lot of undetected cases.
19:45And suppose when we open in 14
19:47days by lifting all the control
19:49measures and the first strategy is
19:51after the first day observing the.
19:53Zero confirmed Case No matter
19:55whether the second day has the case
19:58or not and when to reopen info.
20:00This and the second strategy is one has
20:04a confirmed 0 case for 14 consecutive days.
20:08That basically means 000 our
20:11way through and what is.
20:13Research is probability.
20:15So that is what we found that if one
20:19reopen in 14 days after the first day,
20:22observe 0 case.
20:23So that means it can be zero and 120
20:26again in this type of situation then
20:29the researchers probability is 97%
20:31and if one observes the zero case
20:34for 14 consecutive days and then
20:36the resurgence probability is 32%.
20:38So what is tell us is we need to be
20:41management and don't reopen too early.
20:44So this is happened last.
20:46May and many of the state in the
20:49South do it reopened too early.
20:51Then we saw those cases are searched
20:53in the in the summer in the South.
20:56So what's the take home away?
20:59Take a take away message on the
21:02number 2 is to control the pandemic.
21:05A single control measures not enough
21:08money to use multiple control measures,
21:11and including the mask,
21:13wearing social distancing and
21:15massive testing,
21:15contact tracing and also supported
21:18isolation and quarantines and also
21:20effective treatment and also the vaccine.
21:23And so the in the JAMA paper
21:26we call it multi faceted.
21:28Intervention and then later on in
21:31the summer and people give it a nice
21:34name and called the Swiss cheese model.
21:36So that is a nice name and so the
21:40challenge is we we know those
21:42in control measures,
21:43but it's difficult to implement
21:45those control measures and also keep
21:48high compliance in many countries.
21:49So the defining challenges the
21:51public house control,
21:53measure implementation and then
21:54keep up with the compliance and
21:57also the vaccine definitely is.
21:59Really critical and we.
22:01We know that by now there are
22:04two successful ovac seen,
22:06one in US1 is the face by Pfizer,
22:09the other is more donor under
22:11the efficacy is 95%.
22:12This really really amazing
22:14scientific advance.
22:15Developing the vaccine in such a short time.
22:20Under so we also.
22:22Last spring we also developed a
22:25website on that help estimate the
22:27RT value as a different resolution
22:30at the for different countries,
22:33States and counties and so you can see that.
22:38But we copies are key curve and
22:42for a different.
22:44Reach reaches and so this work
22:46was led by Andy Xu, my student,
22:49and she lucky there my poster.
22:52And so this website was featured on
22:55the in Nature Article last summer.
22:59So how do we fit this model?
23:02So because there are lots of data points,
23:05so we want to estimate the curve so
23:08therefore instead of using the partial
23:11differential equation model and we
23:13extended this epidemic model which was
23:15originally proposed by query in 2013,
23:18and so the model in this type of
23:20epidemic model is quite different
23:23from the traditional logistics
23:25traditional statistical model.
23:26So we need to build in the.
23:30If the. An infectious component,
23:33so here is supposed.
23:34Why is the number of cases so think
23:37about the number of cases and for
23:40each day in Connecticut and then
23:42one first need to calculate this
23:44Lambda T and this Lambda T is called.
23:47Basically calculates the number
23:49of people at risk,
23:50so that is calculated using the products
23:52of the serial interval distribution.
23:55Multiply the number of cases in the
23:57previous period said like 7 days and
24:00then the Ark is a parameter one moment.
24:03Estimate so in the original model,
24:05the estimate RT at each time point
24:07that estimate a lot of parameters,
24:09and then when building this
24:11person model and with Lambda T as
24:14offset and RT as a parameter.
24:16But they asked me lots of parameters
24:18and one also account for the
24:20reporting deley by using a lag.
24:22So we what we did here with we try to.
24:27Accommodate on the Covic features
24:29and so we estimate us zero interval
24:32distribution and from this comma
24:34distribution using the paper in
24:36publishing in nature method and
24:38then in order to estimate RT as many
24:41many values and we assume a curve
24:44and the estimate by using a spline.
24:47So there are few angle in work
24:49and so we want to estimate RT as a
24:52function but cover it and also the
24:56in the traditional epidemic model.
24:58One assumed answer.
24:59Him and trade is a constant
25:02overtime and so the,
25:03but in practice the entertainment
25:05rate is not constant,
25:07especially when the number
25:08of tests are goes up.
25:10What number of positive test
25:12rate goes up and then uncertain
25:15manner it will get better and.
25:17So we want to answer payment way to
25:20be a function of the coverage and
25:23also we want to instead of fixing the
25:26reporting deal if we want to use the
25:29data to model the reporting deley
25:31and using all those met component,
25:33we can estimate the prevalence.
25:35So here are some preliminary result and so.
25:39The the code so you this is for the US
25:42data you can see right now and many
25:45countries a number of cases between
25:47being going down really nicely and
25:49so the current USRT value is about .78,
25:52and so we hear when you can see
25:54we have this arty curve that
25:57expanded so below 1 now.
25:59And also you can see the number of new
26:01cases have been going down and also
26:04the number of deaths has been going down.
26:07But there is a lag between the best.
26:09Under the case. And also this is
26:13the state level are key value.
26:15So just give example like for California
26:18you can see that the art in California is
26:21about .67 and so does this very nice banded
26:24curve for the cases and also for that.
26:28So now let me talk about the what are the
26:31factors associated with Covic infection.
26:34So as we start from the Wuhan data,
26:37then I'll move to the US data.
26:40So the data we estimated the
26:43attack rate on the my age.
26:45So you can see that the each
26:48of the period separately,
26:50and so you can see that for the
26:53older people that purple and yellow,
26:55and then the tax rate was.
26:58Much higher than the younger people,
27:00and so this is a good lesson.
27:03And in the spring last spring,
27:05then later on, as you know,
27:07like in US and there were more cases.
27:11Elderly cases in the spring and but
27:14then the elderly is become very careful
27:17and try to protect themselves and
27:20most of the cases in the summer and
27:22also in the fall were younger people.
27:25And then on the right that shows that
27:28the male and female that you from one
27:31day to the attack rate was similar.
27:34But health care worker,
27:35the purple bar has much higher
27:38infection rate,
27:39especially before the intervention,
27:40and then after interventions.
27:42Acrid among the health care worker,
27:45and was better and so that calls
27:47for the importance of the PPS
27:50and before the intervention.
27:52People were not aware of the Covic,
27:55and so, therefore, is this not many people,
27:58not many health care workers.
28:00Hard to pee pees.
28:02So I give a talk on the Wuhan finding on
28:06the March just before the school public.
28:09Just before Harvard started the spring
28:12break and throw in one of the slides,
28:14I showed that the on the day
28:17before the ABC News,
28:19there's one picture of
28:20the health care workers.
28:22And so I showed up there so the
28:24the the health care worker will
28:26not properly protected in US and
28:29so they had no protection suit
28:31and no face shell for example.
28:34And then the infection could
28:36be go through eyes.
28:38I did not realize that those three
28:40slides on the showing the health care
28:42workers not properly protecting US were
28:45widely distributed during the weekend.
28:48So the March 13 was a Friday.
28:50Then on March 16, that was a Monday,
28:53and so there was a national campaign on
28:57the protection of health care worker,
28:59which comprehensive PP is.
29:01And so the in short time and there were.
29:04Over 1.7 million Xan signatures
29:07and sold in this.
29:09No, the this picture was taken from my talk.
29:12So during that period I got to
29:14know a lot of health care workers
29:17and many of them and wrote to me
29:20and so it's it's kind of like a.
29:22Nice to see,
29:23like a little statistical
29:25analysis and could
29:26help the community.
29:27And also in the spring on the so
29:30I did something that station are
29:33supposed to do and so that we spend
29:37quite a bit time how working with
29:41the state of Massachusetts and also
29:43with abroad and so helping shifting
29:46the PPE under swap on from China.
29:49And so I was on the state, Massachusetts.
29:53The task force in the spring and then like.
29:57One thing I was really touched
29:59last spring was.
30:00Many, many peoples and step in to help
30:03without asking expecting any credit.
30:05So they really a wonderful experience
30:08and by working with so many peoples
30:11and who stepped in to help and so
30:13like in the screen 'cause there
30:16were not many flight from China to
30:18US so was difficult to shift under
30:21those medical supplies and two US
30:23and then so was really wonderful.
30:25Many people help out and so you can
30:29see that there were four flight.
30:32Shifting the usapyon swap watered
30:34by the state of Massachusetts
30:36from Shanghai to Boston under the
30:38first was the flight of the first
30:41flight leaving could own,
30:42and because there were not many
30:45commercial flight available
30:46and travel flight available,
30:48so this flight was converted from the
30:50from the Air Canada Flight and the
30:53passenger flight to a charter flight
30:56and then the picture on the right is
30:59the first flight arriving Boston.
31:03And also I the innerspring or we
31:06launch how we feel up and so this
31:10app collects the information about
31:12the Covic 19 symptoms and behaviors
31:16and also testing a result.
31:18And so this was in collaboration
31:21with some junk.
31:23Many of you probably know fun by
31:26his work in CRISPR editing and
31:29gene editing and CRISPR and also
31:32also banned Superman.
31:34Who is the CEO of country?
31:36So this is really a great.
31:39Collaboration between academia
31:40and industry 'cause we are not
31:43very good at developing up,
31:45but people in industry.
31:47They're much better developing app.
31:49So so many volunteer helping with
31:51this how we feel project and we
31:54build a nonprofit organization and
31:56with so many volunteers and then
31:59this app has over 750,000 users
32:02and also 50 million responses.
32:04And so I'll present some of those
32:07results and this is the first paper.
32:10Out of this,
32:11how we feel project was published in Nature.
32:14Human behavior last summer.
32:17So here last spring,
32:18who were more likely to be tested?
32:21And it turns out that people
32:23who had symptoms,
32:24CDC symptoms or health care workers and
32:27essential workers and people of color,
32:29they were more likely to be tested.
32:31So that makes sense,
32:32because in the spring the testing
32:35kids were not as widely available,
32:37so the vulnerable group should
32:39have priority to be tested.
32:40And so this also present analysis challenge,
32:43because the people who were
32:45tested or likely to be sicker.
32:47And so therefore this is not a random sample,
32:50so when we studies Association between the
32:53factors associated with the infection,
32:55we have taken into account
32:57that people who were tested
32:59was not a random sample and so therefore
33:01in the analysis we use the inverse
33:04probability weighted procedures and
33:06to account for the selection bias.
33:08So we found that male with a higher
33:11risk of infection than females.
33:13And also we found that people of color
33:16were at higher risk of infection.
33:19And also the essential workers
33:21and health care worker and these
33:24were at higher risk of infection.
33:26Also, we found another household
33:29exposures and and also community
33:31exposure are significant risk factor
33:33for infection and so you can see that
33:37for the household exposures after
33:38the show is almost 17 for Community
33:41exposures as we show almost three.
33:44So So what that mean is we need
33:47to break the within household
33:50and close place transmission and.
33:53Cluding, the nursing home,
33:55homeless shelters and prisons,
33:57and so and also we need to control
34:00the community transmission and so
34:02this finding was supported by the
34:05Massachusetts data that Massachusetts
34:08last year reported that almost 90%
34:11of covid cluster were household.
34:13So what that mean is household
34:16transmission is dominant is prevalent.
34:18Dominant lots of transmissions.
34:20And also we found the most important
34:23symptoms and was not the fever and
34:26cough was lots of peace and smell.
34:29So in particular we found out
34:31ratio is almost 33 associated with
34:34loss of taste and smell.
34:36About 40% of those who were past
34:39positive had lost of taste buds,
34:41taste and smell.
34:42Among those who are not testing about
34:466.6% among those who are test negative.
34:49That was about 5%.
34:50So this is an important symptom is also
34:54distinguished from the flu symptom.
34:56Then we also build a prediction
34:59model giving there were not enough
35:01tests available and then can we
35:04use the screening on two and two?
35:07Predict whether a person is
35:09likely to be infected or not.
35:12So by using the CDC symptom,
35:14you can see that the RC
35:16curve is AOC is about 70%.
35:19Using all the variables and it's
35:22about 80% if we use a simpler
35:25model only used for variable,
35:27including the three exposure
35:28variable and also the loss of taste,
35:31smell,
35:32the symptom variables and then
35:34you can see the AOC is also 80%.
35:37And so this is very simple model but has
35:40very good predictability for infection.
35:42And when we build this model we use.
35:45This actually proves a boost that is a
35:48scalable gradient tree boosting method.
35:51So now let me talk about the
35:54defending challenge on the in 2021.
35:57So first is the vaccine rollout and optic,
36:00so the science was really
36:02wonderful last year,
36:04so developing the vaccine such as
36:06short time with such high efficacy,
36:09that's really amazing and so.
36:12So the challenge is the vaccination program.
36:15So basically,
36:16how can we get the vaccine into people's arm?
36:20And so so that basically includes the
36:23distribution and also the administration.
36:26Also,
36:27it's important to have equitable
36:29and scalable vaccination,
36:31and also is important to
36:33overcome vaccine hesitancy.
36:34I'm going to focus on this one.
36:39And the second defining challenge is
36:43the massive scalable testing and so.
36:47PCR test yes, a gold standard,
36:50but it is expensive and to do
36:53the massive regular testing.
36:55So I'm going to talk about
36:58efficient testing strategy using
37:00the pooled testing and also the
37:03other strategies rapid testing.
37:05And the third component is the
37:07implementation and compliance of
37:09public health control measures.
37:10So if you look at a quick job
37:13of the cases in January is not
37:16likely to do to the vaccine,
37:19because only less than 10% of the
37:22US population had been vaccinated.
37:24I seem like the last last month and
37:27the the implementation and compliance
37:29and control measures and became
37:32better and people pay more attention
37:35to the behavior changes so that.
37:37Definitely is an important message.
37:41So let's look at the vaccine rate and so.
37:45Overlap and this is from the one word data.
37:48You can see.
37:49Israel is definitely the role model
37:51and so the right now they have an
37:54average 70 doses and per 100 people.
37:57And so after we account that some people
38:00have two doses on average about 40%
38:03people in Israel had been vaccinated
38:05and with so that's really amazing.
38:07And you have this less than 10%
38:10if on the right you can see we
38:12have a serious equity issue.
38:15And in particular,
38:16you can see basically nobody
38:18in Africa has been vaccinated,
38:20so that's really not good.
38:23So the another defining challenges
38:26vaccine hesitation.
38:27So in order to achieve the
38:29vaccine induced herd immunity,
38:31we need to overcome vaccine hesitation.
38:34So I'm going to present the
38:36findings and from the how we feel
38:39data show McCabe is my Postal.
38:42He take a lead in this work in
38:45collaboration with many colleagues.
38:48So here is a way, a lunch,
38:51the Maxim question in how we
38:53fill up in early December.
38:55So with here the result of analyzing the
38:58first month data about 30,000 people.
39:01So we develop a partnership with
39:04Kinetica and last spring and so
39:06that's why you can see we have more
39:09respondents and in the kinetica,
39:11and also because the countries
39:13is located in California.
39:15So we had more respondent respondent.
39:18In California,
39:19so if you look at overall vaccine hesitancy,
39:23hesitancy read,
39:24you can see like thoughts are more hesitant,
39:27and so overall the vaccine hasn't
39:30hesitancy rate is about 1818% from the
39:33hallway field data and 82% on the.
39:37What said they are likely were
39:39more likely to take the vaccine.
39:42So if you look hard, um,
39:44vaccine hesitancy rate by race and ethnicity,
39:47then you can see that people of color
39:50are much more likely to be vaccine hesitant.
39:53So in particular,
39:55if you look at a black for example,
39:58the vaccine hesitancy is is all.
40:0046%, almost 50% so so compared
40:03to white is about 15%,
40:06but compared to Hispanic,
40:08about 30% you can see a large
40:11fraction of them are undecided group.
40:14So what that mean is that a
40:18community engagement through
40:19the education of outreach is important.
40:23To overcome vaccine hesitancy,
40:25so here are the results.
40:27Who are more likely to be vaccine
40:30hesitant and so we found the
40:32younger people are more likely to
40:34be a vaccine hesitant and females,
40:37and also health care worker essential
40:40workers and also the people of color.
40:43In particular,
40:44black people are 3.5 times more likely
40:47to be vaccine hesitant than white,
40:49and people with pre existing
40:52conditions and low income.
40:54And also rural areas and also the
40:57thoughts and also places with high kufic
41:01burden and also the people who they are.
41:05So those are more likely to be vaccine
41:09hesitant people who wear masks and
41:12also use the protective measures.
41:15They are less likely to be vaccine hesitant.
41:20Talk to us in summary.
41:22So the the vulnerable group are
41:24more likely to be vaccine hesitant,
41:27and so they include people of
41:30color health care worker,
41:32essential worker and the young people female
41:35and the regions with high kovik burdens.
41:38And also the people with
41:40pre existing conditions,
41:41parents and low income.
41:44And also people not using
41:46the protective measures.
41:48So an Irish last late last year,
41:51the lunch, a community engagement alliance.
41:54And so this is,
41:56uh,
41:56involved multiple centers and the one
41:59of the goal is to do the Community
42:03engagement to help with participation
42:06in clinical trial and also.
42:09Overcome the vaccine hesitancy.
42:13So what this tells us is community
42:17engagement for vaccination,
42:18of which an education is important,
42:21so that Pic home number 5 is important,
42:24remained bigil and to scale up scale
42:27up the control measure and vaccination
42:30by protecting the vulnerable group,
42:33including the health care workers
42:35and essential workers and elderly.
42:37And also it's important
42:39to reach the zero kovik.
42:42So what that mean is.
42:44We need to be careful and reopen
42:46slowly when the number of cases are
42:49sufficiently small and also with
42:51the control measures are so if when
42:54we opened too early and we slipped
42:57in the control measure like what
42:59happened last summer and in the South,
43:02and is likely to see the researchers
43:04and also is important to pay
43:06attention to the long color,
43:09the long term effect especially among the
43:12young people and then also the to a build.
43:15I've seen uptick and it's important
43:18to have community engagement and
43:20outreach and build public trust.
43:23So basically, how can we implement the?
43:27Control measures and also implement
43:30vaccination and ensure high compliance
43:33is the defining challenge this year.
43:36And the truth,
43:38the other component is for this year is
43:41how can we boost the testing capacity
43:44and buy a cover by doing more test.
43:47And so because it's uh if one needs
43:50to do the test frequently and to do
43:53the PCR test is difficult to to do
43:56that for many institution because
43:59it's costly and so they put the
44:01testing provide an alternative.
44:03So I'm going to talk about this
44:06efficient put testing.
44:07A design using the hyper
44:09graph factorization first.
44:11What is the protesting?
44:12The goal is that would put testing is
44:16to screen a large population with a few
44:19tests and giving the limited resources.
44:22So this will help reopen the school
44:25safely and the simple idea is used.
44:28This uh document design sofa.
44:30Suppose we have 100 people and we
44:33do 20 tests and then so we pulled
44:37the people sample into different.
44:39Pools and suppose there's only one case,
44:42and then we test each pool support.
44:44Each pool has a 10 people.
44:46And then we tested each put do 10 pull
44:49test and how we found this cool is costing.
44:53Then we test every individual
44:55in this pool so in.
44:57Therefore instead of doing 100
44:58test you only do 20 tests and
45:01so this is the basic idea.
45:03Put testing.
45:04So what is the limitation of this
45:07simple of protesting design?
45:09And so this document design
45:11allow one individual go to
45:13one pool that is Q equal to 1,
45:16then cycle through the pool until
45:18all individuals are assigned.
45:20So if you look at this example
45:23with eight subjects and six pool,
45:25then you can see that we assign
45:27the face first six subject to the
45:306 four ABCDEF and then recycle
45:33the segments and each subject.
45:35And do the pull A&B so only one person
45:38per pool, so this is not optimal,
45:41only one pool,
45:42only one pool per person and this
45:44could lead to a non redundancy
45:46that also reduce the sensitivity.
45:48So the question is can we do better?
45:51Can we assign each individual
45:53to more than one pool?
45:54That basically makes a Q equal to two.
45:57So let's start from something
45:59like if there's a safe assign
46:01one person to two pools.
46:03So for example I assigned
46:05the first person to.
46:062A B second person to put a C and
46:09third person to pull busy and so on
46:12and then cycle through the order.
46:15So that basically this idea
46:17assign each person to two pools.
46:19What is the problem?
46:20The problem is by doing this simple
46:22way the design is not balanced.
46:25You can see that pull it has
46:27five subjects and puppy has four
46:29and pull up as only one subject.
46:32Because when when does them the
46:34pulling and by assigning one
46:36person to more than one pool
46:38while need to dilute the sample.
46:40So if one has a different solution for
46:43different pools that will affect the.
46:45Accuracy under then the sensitivity.
46:47So can we do better so that is
46:50a basic idea of a more balanced
46:52design we call the hyper design.
46:55So this using the hyper graph factorization.
46:58So the basic idea is we want to
47:01make the spell is that possible?
47:03So for example like here you can see
47:07that assigned person A to pull a BE
47:10person B person to pull CD person
47:123 two pull ENF person four to pull.
47:15PNC Person 5 to pull the D&F and so
47:19on and so this. This idea is after you.
47:22Each pool has four samples and
47:25so you can see for the 1st pool
47:28and the test is negative.
47:30The second pull the test is positive.
47:323rd pool passes positive and so on.
47:35Then afterwards we do the pool testing
47:38and then we can decode to see that
47:41which person is likely to be a case.
47:44And here you can see that.
47:46After do the decoding person 3, four,
47:49and seven are likely to be a positive,
47:53and then we test each of them
47:55individually and find out persons.
47:581 Seven are the cases and so why
48:00it is called hyper graph design.
48:04And that because this is related
48:06to the hypergraph,
48:07and in complete awe metrics
48:10and so you can think about this
48:13as the six pools are the six.
48:16Vertex is under the edges are the
48:19people and soap example like a person.
48:22One will assign this.
48:23This is edge person one and so
48:26that's assigned the pool at A&B
48:29and then person two assigned to
48:31C&D and so so this is so that's
48:34why it's called a hypergraph.
48:36So basically what we do is we
48:38need to assign the individuals
48:40and in the right sequence to make
48:43them as balanced as possible and
48:45not overlap as much as possible.
48:47And so by doing this design,
48:50when we kill equal to 216 pool,
48:53we have 5 factorizations and
48:55so you can see for
48:57each factorization there's no overlap and
49:00and also between every two consecutive.
49:03Assignment under then there
49:05is no overlap as well.
49:08And by doing this hypergraph designs
49:10and so you can see that we can
49:13have a balanced pool and also is
49:16very easy to implement and so this
49:19Calculator and so and also very easy
49:22to decode and so this is based on
49:26the company company Atomic Comics.
49:29Population so we can do this calculations
49:32and for Q equal two and three,
49:34but for Q equals greater than three,
49:37the calculations much more challenging.
49:39And so by doing that then you can see that.
49:43And here we plot out the.
49:46Efficiency against the prevalence,
49:47so only if the prevalence is
49:50low is worthwhile to do.
49:51Put testing if the prevalence is high,
49:53there's no need to do put testing so you can
49:57see that doing the hyper design and it is.
50:01Efficient and then the efficiency is
50:03almost 6 compared to individual design
50:06and also expect her than a redesign.
50:09That efficiency is 4.6 and when the
50:12preference become higher and then you
50:14can see that the efficiency goes down
50:17and then also comparing the hyper
50:20design with a radius and efficient,
50:22the sensitivity is pretty similar and
50:25also when we have 384 subject per batch
50:28and you can see that the hyper design
50:31still outperformed the other design.
50:34And for the Pytest peoples design.
50:36And so it.
50:39Especially when the prevalence become
50:41higher and then you can see that uh,
50:44sensitivity almost reach to 0.
50:46And so this also thought we look at a
50:49different design in different scenarios,
50:52and we showed that is hyper design
50:54is optimal and in terms of allocating
50:57resources and so here we plot out
51:00the X axis is the total number
51:02of sample collect each day.
51:04Suppose each day we collect 3000 samples.
51:07Suppose we only have the resources
51:09to do 12 foot tests.
51:11Then you can see that efficiency screening
51:14capacity using this Q equal to two is 122.
51:17That is much.
51:18A better and so so then also
51:21if one has a Q equal to three,
51:24that means a law allowing assigning
51:26one person to three pools.
51:28Then in that situation we need to
51:32use the hypergraph and with the.
51:35Those kind of. 20 different hyperedges.
51:43So in summary,
51:44to scale up widespread testing,
51:47hyper this is based on hypergraph
51:49factor factorization design provide
51:51efficient pool design to maximize
51:53the balance and efficiency,
51:55and the protesting is useful when the
51:57prevalence is low when the preferences highs,
52:01there's no need for protesting,
52:03just do the individual testing and we build
52:06a website that allows the investigator
52:09and the two design their own study.
52:13And so to combat kovik and so
52:15we are really in this together.
52:18And so we have to be together
52:21and be stronger.
52:22And so it's important to let the data speak
52:26and also develop evidence based strategy.
52:29And we show that there are
52:31two feature of the Covic.
52:33One is is highly transmissible,
52:36second is highly convert.
52:37And also it's important to remain vigilant
52:40and to use the multifaceted interventions.
52:43And to combat Covid,
52:44and so the there are multiple
52:47defining challenges this year.
52:49One is a Black Max magazine,
52:51distribution, uptake and education.
52:53The other is a scalable testing,
52:55so we talk about the put testing and
52:58so I want to thank the many of the
53:02collaborators and so who made many
53:04contributions to help with the project.
53:07And also there's a quick announcement
53:10and the cops and less at lunch.
53:13This Covic 19 data.
53:14Lisa Weaponer last December and
53:17so this is every two weeks and
53:19on Thursday from 12:00 to 1:00.
53:22And so these are we have the
53:24last few two month. Last month.
53:27We have a wonderful speaker.
53:29Great turn out.
53:30So those are the speaker in the
53:32coming weeks and from Denmark
53:34Mukherjee who will talk about Covic
53:37in Indian and Harvey Fineberg.
53:39Many of you know and he's a
53:42former president of National
53:43Academy of Medicine and also.
53:46Jim Young Kim is a former president
53:48of World Bank and so they're going to
53:51give the next week talks and thank you.
53:55Thanks young for this
53:57wonderful talk is very useful.
53:59I want to weather the audience
54:02have any questions for she home.
54:05Yeah I have a question.
54:07Yes, please song.
54:08So I'm wondering,
54:09will people who are willing to
54:11respond to the how we feel study be
54:14more likely to have lower hesitancy?
54:20Um? I would think the how we feel.
54:24We study people probably.
54:27I would think that
54:29probably likely to be true,
54:32and so the how we feel samples
54:35the because of people use the app,
54:38so at least that they are coping
54:41aware and they think a quickly is
54:44problem and so it's possible that
54:47in the national samples when we have
54:51a more representative samples and
54:53the hesitancy rate may be higher.
55:01Donna has a question.
55:03Yeah, I see her name was
55:05incredible work you've done.
55:07It's just absolutely
55:09phenomenal and breathtaking.
55:10How you've addressed
55:11each issue arising in the kobid epidemic,
55:15one by one and come up with such clarity.
55:20To guide us. So my question
55:22is about the hyper designs.
55:23I've been aware of pool testing,
55:25which I you know.
55:26We all know it's been around for awhile,
55:29but I'm just wondering, you know,
55:31is there like a rule of thumb like
55:34safe the prevalence rate is like 5%?
55:36How many digit
55:37in your graph like how many fewer tests
55:40would you have to use using a hyper
55:42design versus like the standard approach
55:44that you know people would tend to use
55:47which is to just test everybody. Yeah,
55:49so that is if you can see that from here.
55:53Yeah, it's a little hard to
55:56see it's a little small.
55:57Oh this hyper yeah I can the the
56:00so you can see the efficiency
56:02that is about hyper design.
56:04Yes almost six so that means that we
56:07can each task and have 6 people and
56:10by individual design so you can see
56:13that suppose you have 100 people.
56:15This is 96 so 96 / 6 and then then
56:19you can see that that is. How many?
56:23How many fewer tasks it less than 20?
56:26Yes, I think about it.
56:28If you do individual test that is 100.
56:32What is the?
56:33What is the?
56:34I didn't understand really what
56:35the 96 and the 3:50.
56:40Yeah batch, so there are 96 so in the
56:42so if you think about when you win
56:45you do test and then basically you
56:47need to layout the sample in a batch.
56:50If you think about it already
56:52then think about that.
56:53Basically they have eight.
56:54You have to think about.
56:56The Matrix is 8 by 12 matrix.
56:58You put all the samples and
57:00in this 8 by 12 array.
57:05OK, thank you yeah.
57:08And also if you look at the capacity
57:11here you can see the capacity is
57:13much better so you can see that.
57:15Suppose I need to test 3000 people a day.
57:19Execution can only afford half 12 tests
57:22and then you can see the efficiency.
57:25Screening capacity is almost 120.
57:30So that is really good.
57:31That's very, very good, yeah?
57:39So I also have.
57:43You still have a question
57:44or say thank you very much.
57:47OK, so I also have a related question.
57:50So she how you mentioned that before.
57:53A future work you want to perform a
57:57regarding the reproduction number
57:59estimation and this intervention
58:01work is to consider different
58:04other covariates when you are
58:06modeling the reproduction rate.
58:08So I wonder,
58:09have you also considered like trying
58:12to take into consideration different
58:14type of various the mutation of
58:17different various and then maybe certain
58:21various various high reproduction rate?
58:23And perhaps others.
58:25This process?
58:25Yeah, that
58:26is excellent suggestions on the so yeah,
58:29if we could have those data will be great
58:32that we could include those in the model
58:36besides the different type of variance.
58:38And also like the vaccination rate.
58:41That would be a very good variable included.
58:44And so the challenge for us right now is,
58:48as you know UK has been doing
58:50a great job in the sequencing,
58:53viral sequencing and so in other words the
58:56surveillance and sequencing surveillance.
58:58But not US, and so we have not doing a great
59:02job in sequencing and so so therefore the
59:06UK could monitor the new virus and well,
59:09but I think with the one of the things
59:12we need to do this year is to increase
59:16the various viral sequencing capacity
59:18so we could monitor the new variants.
59:21So then also make the data available and to
59:25the public and then that can be included.
59:28In the analysis.
59:30So what I've found that last year and
59:33during the Covic people were much more
59:35willing to share the data computer for an,
59:39though this is really,
59:40really wonderful and also the much
59:43more preprint and compared to before,
59:45and that were posted in about
59:47archive and made archive,
59:49and people were really willing
59:51and to share their findings to
59:54the Community as soon as possible.
59:56So these are really wonderful
59:58spirit about open science and.
60:00And is.
01:00:01Fired on the by many researchers last year.
01:00:05Thanks, that's really informed him.
01:00:10Sorry, go ahead who is
01:00:11trying to ask a question.
01:00:14Me, I mean yeah, I see who I have a
01:00:19question regarding to this pulling pulling
01:00:23test. You said the pulling test and
01:00:26compare with the individual test.
01:00:29The sensitivities are similar
01:00:31and right now I'm thinking
01:00:34if each individual does sensitivity they
01:00:37can be test by individual test one this
01:00:42individual mix with five other. Cure
01:00:45lung disease samples.
01:00:47Basically, the concentration is diluted,
01:00:50so how does sensitivity will be keep
01:00:54the same and how the next ways how
01:00:59to compare if there's one positive
01:01:02case with five individual their
01:01:05normal cases an A normal situation
01:01:08controls and compare with all six.
01:01:12There just get exposure
01:01:14with low concentration.
01:01:16So there will be probably have
01:01:19some sensitivity issue if
01:01:22pulling together compared to
01:01:24individual tests, then the error for
01:01:27measurement error testing error for
01:01:29the two different types of tests.
01:01:33How did you consider them additional
01:01:36to the hyper structured testing?
01:01:39Yes, I think this is a great question.
01:01:42Sorry I did not make that clear.
01:01:44What I meant was that hyper design and are
01:01:47ready that they had a similar sensitivity,
01:01:49but the sensitivity is lower
01:01:51than the individual tests.
01:01:52If you look at the curve in about I see.
01:01:57At the green and red they have
01:01:59a similar sensitivity by the
01:02:01compared to the individual test.
01:02:03That is, this black line and it has
01:02:05higher sensitivity and so then they
01:02:08as you are definitely right when one
01:02:10do one month does the pooled testing
01:02:12because the sample needs to be diluted,
01:02:15so therefore we need to pay a price
01:02:17and then sensitivity will be lower.
01:02:20Yeah, so the overall one the population.
01:02:25Pilots large scale testing.
01:02:27We may have more undetectable test.
01:02:30Think about if six samples always happen
01:02:33is 1 sample has a positive positive case,
01:02:37so we may have some testing error
01:02:40here. Yeah yeah, so yes.
01:02:43Yes, I'm sorry.
01:02:44Can I just jump in for a second?
01:02:47So 'cause I think the comparison as
01:02:50you could compare this hyper design
01:02:52to just testing everybody or the
01:02:54hyper design to the traditional
01:02:56pool testing approach where you
01:02:58just divide 100 people in each one
01:03:00is in a single batch and I think
01:03:03she hung what you're saying and it
01:03:05makes sense to me intuitively is by
01:03:08repeating people in multiple batches
01:03:10were increasing the chances of having
01:03:12doubles and triples in the same batch.
01:03:14And then lowering the chance of having
01:03:17false negatives as opposed to the
01:03:20traditional design where you take the
01:03:23100 people and they're only in one match.
01:03:27Is that right?
01:03:28Yeah, that's so that's right.
01:03:30So you don't want to put a
01:03:32hundred 100 people in one batch,
01:03:34because if you do that,
01:03:36then the sample need to diluted a lot and
01:03:39then you will sacrifice the sensitivity.
01:03:42So that's why when want to
01:03:44do the optimal design,
01:03:46want to account for both the
01:03:48balance and also sensitivity and
01:03:50including both of them and so then.
01:03:52So that's why when we build this when we
01:03:55define this efficient screening capacity,
01:03:57this calculation.
01:03:58That incorporated sensitivity
01:04:00in the calculation as well.
01:04:03I see thanks yeah.
01:04:05So the the pool design.
01:04:07So if you look at the traditional design,
01:04:10so here you can see that each person,
01:04:13the traditional design.
01:04:14Basically each person is assigned
01:04:16to a single pool and so this
01:04:19is this called document design,
01:04:21and so this design so you can
01:04:24see that the six for six people.
01:04:28In this example, like the person, one and.
01:04:32Assigned to a person to assign to P,
01:04:35and so this is not efficient design.
01:04:38And so if we assign each person to
01:04:40multiple pools and after decoding
01:04:42that will improve the efficiency.
01:04:47So generally the cute does
01:04:48should not be too big.
01:04:50So here you can secure equal to 1.
01:04:53That means one person assigned
01:04:54to one pool to equal to 2 means
01:04:57a person sent to two pools and
01:04:59just think about this is very
01:05:01interesting and so you can see the
01:05:03basically using using the graph and
01:05:05the ABCD basically means the pool
01:05:07and each edge indicated person.
01:05:09So you can see this person one
01:05:11is assigned to pull A&B and then
01:05:13says that's why there's edge here
01:05:15and person to assign to C&D.
01:05:17So this person too.
01:05:19And then so on.
01:05:24Thanks young, I have one.
01:05:26I have one last question.
01:05:28If other people do not have
01:05:30more question so I wonder is
01:05:32also related to the sensitivity.
01:05:34I wonder how we considered too instead
01:05:36of using on the testing directly but
01:05:39construct some posterior for each person.
01:05:42Use other covariates including
01:05:43your past history.
01:05:44Whether you have higher risk an.
01:05:48I wonder if we use such
01:05:50personalized information combined
01:05:52with this testing results,
01:05:53can we have better sensitivity?
01:05:57Very good question.
01:05:59Yeah, I can see the potential.
01:06:02I can see the potential for doing that.
01:06:05Yeah, I think they're right now in
01:06:10the screening program and the no.
01:06:13Demographic information is collected,
01:06:15and so only the sample collected.
01:06:18So for example, like abroad they.
01:06:22But bro, the dead on.
01:06:25The spring when we first started
01:06:28it was about maybe a 3000 of
01:06:32sample per day and so right now.
01:06:34As you know,
01:06:36broad cover almost 9025% of the testing,
01:06:39and in the New England areas almost
01:06:43cover like 3,000,000 and test and so
01:06:46the event of data were lots of data.
01:06:50Sand were collected and.
01:06:53In the testing,
01:06:53but those data cannot be used for research.
01:07:00Thanks, I don't know if the
01:07:02audience have further question.
01:07:04Maybe you can also email see how afterwards.
01:07:08So we're running a little bit overtime,
01:07:11but it's very. This wonderful talk.
01:07:14Can we have learned so much
01:07:16from Seahawks services talk?
01:07:18Thank you again. Thank
01:07:20you very much.