Supporting COVID-19 Hospital Planning and State Reopening Using Model Projections

May 22, 2020

Information

Forrest W. Crawford, PhD Associate Professor of Biostatistics, Associate Professor of Ecology and Evolutionary Biology, Associate Professor of Management, and Associate Professor of Statistics and Data Science

ID5234

To CiteDCA Citation Guide

00:00I would like to now introduce
00:03our next speaker, a Doctor,
00:05Forrest Crawford, now to Crossville,
00:07is an associate professor of high statistics,
00:10associate professor of ecology
00:11and evolutionary biology,
00:12associated professor of Management
00:14and associate professor of
00:16statistics and data science.
00:18Out of Crawford's work focuses on
00:20mathematical and statistical problems
00:22related to discrete structures
00:24in stochastic processes, mapping,
00:26genealogy, public health, bio,
00:28medison and Evolutionary Science.
00:30Doctor Crawford thank you for being here.
00:35Great, thank you very much.
00:37I'm very happy to be here.
00:39Very honored to be among
00:40these amazing presenters.
00:42I would like to present for you
00:442 recent projects and I won't go
00:47into a lot of technical detail.
00:49There's some mathematics and
00:50statistics behind this work,
00:52and I'm not going to talk about any of that.
00:55I'll just try to talk about.
00:58Uh, the need that we were trying
01:00to respond to when we worked on
01:03these projects and what the research
01:05product square and where to find them.
01:08So this is joint work with post
01:11doc solely omarova Richard Lee.
01:13So hey Lexi and PhD students
01:16Margaret Earline's,
01:17daughter Jinhao Son and also the
01:19COVID-19 statistics policy modeling
01:21and Epidemiology collective.
01:23So the first thing that happened,
01:25I think this was in in late
01:28March was that we heard.
01:31That there was an acute needed to Yale.
01:33New Haven health system for
01:35help with capacity planning.
01:37Trying to prepare the hospital
01:39and health system.
01:40For what was then believed to be a
01:44coming onslaught of new patients,
01:46which had the potential to overwhelm
01:49the health system to overwhelm the
01:51supply of ICU beds and Ventilators?
01:54So we we tried to respond to this challenge,
01:58which came,
01:59I think,
02:00from directly from senior hospital
02:03leadership to build a model,
02:05an idealized representation of
02:07the dynamics of patient flow
02:09through the hospital COVID-19
02:11patients who presented to the Ed.
02:14And then we moved to the floor,
02:17possibly released their move,
02:19possibly to the ICU,
02:21and then received care in the hospital.
02:25And we are especially interested in
02:27helping the health system helping Yale,
02:30New Haven and also other health
02:32systems to plan their expansion
02:34in capacity to plan the ability
02:37to accommodate patients who are
02:39coming in every day so that the
02:42systems would not be overwhelmed.
02:43And we ended up in a very short
02:46amount of time writing software for
02:49web application that implemented
02:51in mathematical model whose
02:52structure I I'm not going to show.
02:55I guess beyond beyond this last this diagram.
02:59And the idea here is that if you are
03:01helping to manage the health system,
03:03then you can dial in a lot of the
03:05features of your health system,
03:07the capacity,
03:08the number of beds you have in the floor,
03:10and I see you.
03:12How you expect the patterns of
03:14change of patient presentations
03:15to the D to change overtime,
03:18you can dial in your expected or
03:21planned capacity increases in
03:22terms of beds into the future,
03:24and you can look to see how how
03:27patients will end up flowing
03:29through the hospital.
03:30So I think this was this was useful
03:33in augmenting some of the existing
03:35capacity planning tools and software
03:37that Yale New Haven Health System had,
03:40and we did receive feedback from
03:42health systems throughout the country.
03:44That they were using this and
03:47other tools to help plan for.
03:50For a very rapidly increasing number
03:53of patients presenting to the D,
03:55so this was this is a project
03:57that was done
03:59very quickly in late March in anticipation
04:02of a very fast increase in the number
04:05of cases were very fortunate in
04:08Connecticut that hospital systems were
04:10able to expand capacity quite rapidly
04:13and at the state level at least.
04:15The number of covered patients
04:18did not outpaced the hospital's
04:20ability to accommodate them.
04:22So I think the the need for this particular
04:25application has waned a little bit
04:28since mid April when hospitalization,
04:30census covert hospitalization census
04:32began to decline in Connecticut.
04:34If there is a second wave of
04:36infections in Connecticut,
04:38we anticipate this tool becoming
04:40very useful and relevant again.
04:43But the main thing that I'd like
04:45to talk to you about today is work
04:48in support of the Connecticut
04:50governor's plans to reopen the state.
04:53Governor Lamont convened a panel of experts
04:56that reopened Connecticut advisory panel,
04:58including many people from Yale,
05:00and I was asked to support the work of that
05:03panel by providing modeling projections,
05:06transmission, modeling,
05:07projections of COVID-19 incidents,
05:09hospitalizations,
05:09and deaths under reopening scenarios.
05:11Articulated at the time in a very
05:13general way by the governor to plan
05:16for interventions like testing,
05:18contact tracing and to assess the
05:21risk of a second wave of infections
05:24occurring over the summer or in the fall.
05:28Following reopen and release of
05:30contact that had been suppressed
05:33during the state lockdown.
05:35As you probably know,
05:38Connecticut began its reopening
05:40phases yesterday on May 20th.
05:43And the work of this this committee
05:45to assist in that process may
05:48be coming to a close.
05:50But I think that there is a very
05:52important ongoing need for projections
05:54to inform decision making and
05:57epidemiological study design at
05:58the Department of Public health
06:00and at the state level overall,
06:02as the state considers how to move
06:05forward in its reopening phases,
06:07whether there is a need to revert to
06:10a previous more restrictive phase and
06:12how this process should play out.
06:15In particular,
06:16I think policymakers are very
06:18interested in having an early
06:19warning system that could tell them
06:21if there is a coming but hidden
06:24wave of new infections that will
06:26become hospitalizations and deaths
06:27in the near future.
06:29I think that it is fair to say
06:32that Connecticut policymakers,
06:33along with a lot of decision
06:35makers throughout the world,
06:37have access to very high quality data
06:40streams that describe the current
06:41state of the pandemic in their area.
06:44Here in Connecticut,
06:45the governor has access to various
06:47dashboards and reports daily reports
06:49from the Department of Public Health
06:51on the number of tests administered,
06:53the number of positive tests the
06:56Connecticut Hospital Association
06:57reports daily.
06:57The hospitalization census from
06:59the previous night.
07:00The number of beds that are theoretically
07:02available for kovid patients,
07:04including search capacity and beds that
07:07have been added on a temporary basis.
07:10Decision makers have access to near
07:12real time information about case counts
07:14and deaths and possibly excess deaths
07:16that are occurring in major Health Systems.
07:19An outside.
07:20And this is very good policy.
07:22Makers have access to this
07:24real time information.
07:25But that information alone
07:26may not be enough to tell them
07:29when a second wave is building,
07:31and about two occur, and if that is
07:34going to occur sooner this summer.
07:36The model projections that
07:38my group has been developing.
07:40Have the ability to tell us about
07:43possible futures instead of the
07:44current state of the metrics that
07:46the state has chosen to track.
07:48What we're really interested in is
07:50what might occur in the future.
07:51What are the things that we can't
07:54see today that will become observable
07:56two or three weeks from now?
07:58So in particular,
07:59we want these projections to inform
08:02reopening phases in the state.
08:03The decision about how and even whether
08:06to open schools for for young people,
08:09and also colleges and universities.
08:11How to inform efforts to expand
08:14testing and contact tracing in a way
08:17that is equitable and also targets
08:20the areas that are highest need.
08:23And how to develop continued or
08:25modified distancing guidelines
08:27into the future and possibly
08:29change those guidelines as needed.
08:34And in doing this work we asked ourselves
08:36and probably other people ask themselves,
08:39does the world really need another
08:42COVID-19 transmission model?
08:43and I think you know at the worldwide level,
08:47even at the national level,
08:49the answer is probably no.
08:51But locally at least I
08:52think that Connecticut does.
08:54We saw a very acute need,
08:56especially at the state level right
08:58now to develop a scenario analysis
09:00tool that is specifically responsive
09:03to the needs of Connecticut leadership
09:05as they plan to re inform to inform
09:08reopening strategies to reopen the state
09:10and to design interventions that are
09:12appropriate for Connecticut specifically.
09:14And to do that we have access to a
09:16lot of data streams that essentially
09:19none of the national level.
09:21Transmission modeling efforts have access to.
09:24In particular,
09:25we have a connection to the
09:27Connecticut Hospital Association,
09:29so we know exactly how many
09:32patients are hospitalized.
09:33Throughout the state and what the
09:35bed capacity is as a dynamically
09:38changes overtime.
09:39We can calibrate transmission models
09:41in particularly clinical models,
09:42of what happens to patients after
09:44they enter the health system using
09:46patient trajectory data from Yale.
09:48New Haven health system.
09:50We've accessed at Yale here.
09:53Fortunately to the ale emerging
09:55infections program,
09:56surveillance data from DPH and
09:58close connection to people who are
10:01planning and conducting testing
10:03and seroprevalence surveys to
10:05inform further scientific efforts.
10:08I hope that in the future we will
10:10continue to have access to colleagues
10:12at the Department of Public health
10:14who are actually implementing
10:16the intervention strategies.
10:17Contact tracing and testing,
10:19and encouraging individuals who
10:21test positive to isolate themselves
10:22and we want to be able to help
10:25them design those interventions.
10:26So we built a model.
10:28I'm not going to show the structure.
10:31It is a generalization of the sci,
10:34our class of transmission models
10:35that has been described previously.
10:37Today we fit that model along with
10:40the information that we have about
10:42when the governor closed schools
10:44and when the state lockdown occurred.
10:47To produce projections,
10:48and here I'm showing projections that
10:51begin in early March and we have real data,
10:54actual observed data up to,
10:56I think yesterday overlaid as dots.
10:58So on the left we have hospitalizations.
11:02Reported an projected and we have
11:04cumulative deaths on the right
11:06and the model overall recovers.
11:08Historical dynamics of hospitalizations
11:09and deaths very, very accurately,
11:11and I think this is partly because
11:14we have very specific information
11:16about what the governor did and when.
11:19And how those interventions affected
11:23transmission and these downstream outcomes?
11:27Here are some projections that the
11:29group just finished working on this.
11:31I should have said earlier this is
11:33specifically joint work with oleum,
11:34rozafa and and Richard Lee,
11:36who have worked tirelessly over the
11:38last couple of days to put all of
11:41this together and also to write 2
11:43reports which I'll tell you about
11:45in the moment.
11:46So in the upper left hand corner we have.
11:50A representation of the amount of
11:53interpersonal contact that occurs
11:54in Connecticut,
11:55historically prior to March 20th.
11:58Sorry, May 20th the first drop is
12:01due to the governor's closure of
12:04schools in the second drop is due
12:07to the state stay at home order.
12:10And the changes in that contact
12:13curve that occur after May 20th.
12:15Our guess is this is a scenario
12:18that we developed based on
12:20ideas about a slow reopening,
12:22in which contact between
12:24individuals returns to baseline
12:26or returns to normal very slowly,
12:28and by slowly I mean that 10% of
12:31this latent suppressed contact is
12:34released roughly once per month.
12:36And so the time series of contact
12:39going forward is just the step
12:42function that increases by 10% of
12:44the suppressed amount every month.
12:47So this is what we imagine.
12:49This is not necessarily what
12:50will occur in real life.
12:51It could be better, could be worse,
12:53but this is one scenario that we
12:55want to present to the governor.
12:57Um?
12:57And here we look at the implications
12:59of this scenario in terms of new
13:03infections or daily incidents.
13:04In Connecticut we see a small
13:07spike after reopening,
13:08but daily incidence remains low and
13:10begins to rise only into late August.
13:13In the lower left hand corner.
13:15We see hospitalizations.
13:17The dotted line above is the overalls
13:21hospital bed capacity in Connecticut,
13:24including temporary or search beds.
13:27And you can see that under this
13:29very slow reopening scenario,
13:31hospitalization continues its slow decline,
13:33becomes very flat in July,
13:35and part of August,
13:37and begins to rise very slowly
13:39as we get towards September.
13:41But overall hospitalization remains
13:43well below the census peak which
13:46occurred in mid April and likewise
13:49deaths begin to flatten out and.
13:51And we end up with almost 6000
13:53deaths in our simulations.
13:55In this scenario,
13:57under slow reopening,
13:58I think this is an optimistic scenario.
14:01Here's a more pessimistic scenario in
14:04which contact for returns much more
14:06quickly to the pre lock down baseline.
14:09Here we release 10% of this latent
14:12suppressed contact every two weeks.
14:14This is a much more rapid rise in contact.
14:18Again,
14:18we don't know what exactly will happen when.
14:22People return to work and maybe
14:24children return to summer camps in
14:26day cares and things like that,
14:28but this is perhaps a more pessimistic
14:31scenario in which people experience much
14:33more interpersonal contact than they did,
14:35say,
14:36a week ago.
14:37Here we see a really dramatic rise in
14:40daily incidents into August and September.
14:43Uh,
14:43with very large numbers of individuals
14:46getting infected per day in Connecticut.
14:49Likewise,
14:50hospitalizations rise very dramatically
14:51in August under this scenario,
14:53and we are looking at the possibility
14:56of possibly exceeding hospital capacity.
14:59Even the surge capacity by mid
15:01August or early September,
15:03and this is very bad because people
15:06who need hospitalization but don't
15:09get it are very likely to die much
15:12faster than they would otherwise.
15:14Likewise,
15:15here we see a dramatic increase
15:17in deaths in August,
15:18and it just gets worse into
15:21September under this scenario.
15:22So I think in reality,
15:24what will occur in Connecticut is
15:26probably something between these
15:28two extreme scenarios, but these,
15:30I think might be benchmarks against
15:32which we measure the governments
15:35true response and the response of
15:37the people in terms of their contact.
15:40We're not just interested in looking into
15:43a crystal ball an predicting the future.
15:46We also want to be able to inform
15:50concrete intervention efforts,
15:51including scientific intervention,
15:53with scientific efforts to learn
15:55more about the Epidemiology of
15:57COVID-19 specifically in Connecticut.
15:59In particular,
15:59the design and planning and
16:01implementation of future seroprevalence
16:03studies will require accurate
16:05estimates of cumulative incidence.
16:07That is,
16:08the number of people in Connecticut
16:11who have evidence of prior infection.
16:14And so these are things that
16:16actually will come out of the
16:19model projections if you plan to
16:20run so prevalent study in a month,
16:23we can tell you under different scenarios,
16:26roughly how many people are likely
16:28to have evidence of prior infections
16:30at that moment under the assumptions
16:33articulated in the model.
16:34So we hope that this tool will
16:37be useful prospectively for study
16:39planning and design of testing
16:41and other interventions,
16:43In addition to just predicting the future.
16:46So, uh, so going forward?
16:50We want to be able to share this
16:53information in the form of reports
16:56with policymakers, policymakers,
16:58in the state government,
17:00and decision makers throughout the state.
17:02So we put together a website
17:05along with the code for software
17:08and two reports so far.
17:10One policy report in one technical
17:13report on how the model works,
17:15this website just went live about
17:18an hour ago and now now these?
17:21Reports are posted publicly for
17:23anyone to see as we update these
17:26reports in real time.
17:27We will document the updates and
17:29post new versions on the website.
17:32If we ever change anything,
17:33we will provide a note saying
17:36what has changed so that you can
17:38follow our progress as we go.
17:40We will post these reports roughly
17:43once every four to six weeks to
17:46coincide with the governor's stated
17:48reopening phase plans and so I will.
17:50Paste a link here in the web and
17:53our chat window if you'd like
17:55to check out this website,
17:57you don't have to copy down the URL.
18:00Basically,
18:00over the next few months will try to
18:03provide actionable intelligence to
18:05state decision makers so that they can
18:07better plan the states response an reopening.
18:10In this crisis.
18:11And that's all I have for you.
18:14Thank you very much.
18:17Thank you very much, I'd like.