ARTICLE
OPEN
Modeling the effect of exposure notification and
non-pharmaceutical interventions on COVID-19
transmission in Washington state
Matthew Abueg
1,4
, Robert Hinch
2,4
, Neo Wu
1
, Luyang Liu
1
, William Probert
2
, Austin Wu
1
, Paul Eastham
1
, Yusef Shafi
1
,
Matt Rosencrantz
1
, Michael Dikovsky
1
, Zhao Cheng
1
, Anel Nurtay
2
, Lucie Abeler-Dörner
2
, David Bonsall
2
, Michael V. McConnell
1,3
,
Shawn O’Banion
1
✉
and Christophe Fraser
2
✉
Contact tracing is increasingly used to combat COVID-19, and digital implementations are now being deployed, many based on
Apple and Google’s Exposure Notification System. These systems utilize non-traditional smartphone-based technology, presenting
challenges in understanding possible outcomes. In this work, we create individual-based models of three Washington state
counties to explore how digital exposure notifications combined with other non-pharmaceutical interventions influence COVID-19
disease spread under various adoption, compliance, and mobility scenarios. In a model with 15% participation, we found that
exposure notification could reduce infections and deaths by approximately 8% and 6% and could effectively complement
traditional contact tracing. We believe this can provide health authorities in Washington state and beyond with guidance on how
exposure notification can complement traditional interventions to suppress the spread of COVID-19.
npj Digital Medicine (2021) 4:49 ; https://doi.org/10.1038/s41746-021-00422-7
INTRODUCTION
The COVID-19 pandemic has brought about tremendous societal
and economic consequences across the globe, and many areas
remain deeply affected. Due to the urgency and severity of the
crisis, the poorly understood long-term consequences of the virus,
and the lack of certainty about which control measures will be
effective, many approaches to stopping or slowing the virus are
being explored. In seeking solutions to this problem, many
technology-based non-pharmaceutical interventions have been
considered and deployed
1
, including data aggregation to track
the spread of the disease, GPS-enabled quarantine enforcement,
AI-based clinical management, and many others.
Contact tracing, driven by interviews of infected persons to
reveal their interactions with others, has been a staple of
epidemiology and public health for the past two centuries
2
.
These human-driven methods have been brought to bear against
COVID-19 since its emergence, with some success
3
. Unfortunately,
owing in part to the rapid and often asymptomatic spread of the
virus, these efforts have not been successful in preventing a global
pandemic. Further, as infections have reached into the millions,
traditional contact tracing resources have been overwhelmed in
many areas
4,5
. Given these major challenges to traditional contact
tracing, it has been suggested that apps that make use of
Bluetooth technology can assist in detecting exposures to those
carrying the virus, and serve as a complementary tool to human
contact tracing initiatives
6
.
Technological solutions in this space have never been deployed
at scale before, and their effectiveness is unknown. There is an
acute need to understand their potential impact, to establish and
optimize their behavior as they are deployed, and to harmonize
them with traditional contact tracing efforts. Specifically, we will
examine these issues in the context of the Exposure Notification
System (ENS), developed by Apple and Google, which is currently
being adopted by many states and countries
7
. In this system, GPS
and location data are not used—instead, Bluetooth alone is
utilized to exchange anonymous, randomly-generated IDs which
can later be checked against a list of positive cases. In order to
protect user privacy and build user trust, ENS does not require
users and their contacts to be identified or located, and
recognition of each user’s exposure risk level can take place only
on the user’s smartphone
8
.
To improve our understanding of this new approach, we
employ individual-based computational models, also known as
agent-based models, which allow the exploration of disease
dynamics in the presence of complex human interactions, social
networks, and interventions
9,10
. This technique has been used to
successfully model the spread of Ebola in Africa
11
, malaria in
Kenya
12
, and influenza-like illness in several regions
13,14
, among
many others. In the case of COVID-19, the OpenABM-Covid19
model by Hinch et al.
15
has been used to explore smartphone-
based interventions in the United Kingdom. Individual-based
models simulate individuals and their interactions in home, work,
and community contexts, using epidemiological parameters to
guide the compartments in an expanded SEIR model
16
and
demographic parameters to simulate individuals and their
interactions. Although past work has studied disease transmis-
sion
17–19
, progression
20
, and social distancing interventions
10,21,22
,
we seek to understand the combined effect of exposure
notifications and non-pharmaceutical interventions in an environ-
ment calibrated to the demographics
23
, occupational structure,
and epidemic trend of that location.
In this work, we adapt the OpenABM-Covid19 model to
simulate the ENS approach, and apply it to data from Washington
state in the United States to explore possible outcomes. We use
data at the county level to match the population, demographic,
and occupational structure of the region, and calibrate the model
1
Google Research, Mountain View, CA, USA.
2
Nuffield Department of Medicine, University of Oxford, Oxford, UK.
3
Department of Medicine, Stanford University School of
Medicine, Stanford, CA, USA.
4
These authors contributed equally: Matthew Abueg, Robert Hinch.
✉
email: obanion@google.com; christophe.fraser@bdi.ox.ac.uk
www.nature.com/npjdigitalmed
Published in partnership with Seoul National University Bundang Hospital
1234567890():,;
with epidemiological data from Washington state and Google’s
Community Mobility Reports for a time-varying infection rate
24
.
Similar to Hinch et al., we find that digital exposure notification
can effectively reduce infections, hospitalizations, and deaths from
COVID-19, even if just roughly 15% of the overall population
participates. We extend the findings by Hinch et al. to show how
digital exposure notification can be deployed concurrently with
traditional contact tracing and social distancing to suppress the
current epidemic and aid in various “reopening” scenarios. We
believe the demographic and occupational realism of the model
and its results have important implications for the public health of
Washington state and other health authorities around the world
working to combat COVID-19.
RESULTS
Digital exposure notification
We present forward-looking simulations for Washington state
counties by comparing multiple hypothetical scenarios with
combinations of digital exposure notification, manual contact
tracing, and social distancing. Each simulation uses the same
calibrated model up to July 11, 2020, at which point the
hypothetical interventions are implemented. Beyond this date,
each simulation uses the final calibration model parameters,
except where explicitly specified as part of the intervention. For
each simulated intervention we report the number of infections
(daily and cumulative), cumulative number of deaths, number of
hospitalizations, number of tests per day, and fraction of the
population in quarantine. The simulation runs for a consistent
300 days from the beginning of our mobility data, March 1, 2020,
through Dec 25, 2020, plus the additional calibrated seeding
period before March 1. Unless otherwise stated, the reported
result is the mean value over 10 runs with different random seeds
of infection.
Results may be affected by the end date of the simulation
because of the time it takes some interventions to have their full
effect. We believe that a time horizon of approximately 5 and a
half months is long enough to be practically useful for public
health agencies who are considering deploying such interven-
tions, but short enough to minimize the long-term uncertainty
and effects of externalities such as a vaccine becoming available.
We first study the effect of a digital exposure notification app at
different levels of app adoption (15%, 30%, 45%, 60%, and 75%) of
the population in each county. As a baseline, we compare those
results to the “default” scenario assuming no change in behavior
or interventions beyond July 11, 2020. The results show an overall
benefit of digital exposure notification at every level of app
adoption (Figs. 1, 2). When compared to the default scenario of
only self-isolation due to symptoms, each scenario results in lower
overall incidence, mortality, and hospitalizations. Unsurprisingly,
the effect on the epidemic is more significant at higher levels of
app adoption. An app with 75% adoption reduces the total
number of infections by 56–73%, 73–79%, and 67–81% and
deaths by 52–70%, 69–78%, and 63–78% for King, Pierce, and
Snohomish counties, respectively. However, even at a relatively
low level of adoption of 15%, there are meaningful reductions in
total infections of 3.9–5.8%, 8.1–9.6%, and 6.3–11.8% and total
deaths of 2.2–6.6%, 11.2–11.3%, and 8.2–15.0% for King, Pierce,
and Snohomish counties, respectively.
In addition to its ability to suppress the epidemic, we also
evaluate the effects of exposure notification adoption on the total
number of quarantine events. There is an incentive to minimize
the quarantine rate because of the perceived economic and social
consequences of stay-at-home orders. At 15% exposure notifica-
tion adoption the total number of quarantine events increases by
4.6–6.4%, 6.6–6.8%, and 5.8–10.2% for King, Pierce, and Snohom-
ish counties (Fig. 3). In general, the higher the level of exposure
notification adoption the greater the number of total quarantine
events, with the exception of very high levels of adoption (60%
and 75%) where this number plateaus or even decreases, likely
due to the significant effect of the intervention in suppressing the
overall epidemic in those scenarios.
Manual contact tracing
Next, we study the potential impact of manual contact tracing in
suppressing the epidemic as a function of the contact tracing
workforce size. We refer to guidance by the Office of the Governor
of WA State with a minimum recommendation of 15 tracers per
100,000 people as well as the staffing rates for King County
including all available staffers (105 FTE for 2.253 million people, or
4.7 per 100,000) and the National Association of County & City
Health Officials (NACCHO) recommended staffing levels during
epidemics of 30 staff per 100,000 people
25
. We set the tracing
delay to one day, which is the optimistic estimate for time to
contact trace, as the goal for Washington state is to notify 80% of
contacts within 48 hours
26
. We similarly use the King County
Phase 2 Application to compute the expected number of initial
contact tracing interviews as well as follow-up notifications. Over a
two-week period, 22 staff members contacted 336 individuals for
initial interviews and 941 for close contact notifications, or
approximately 1 initial interview and 3 notifications per day per
staff member.
Manual tracing with the full desired staffing levels of 15 workers
per 100,000 people is able to affect the epidemic trend in all three
counties, but has a significantly smaller effect at current staffing
levels (Fig. 4). Unsurprisingly, the impact for a given level of
staffing is dependent upon the current epidemic trend, reinforcing
the need for concurrent interventions to effectively manage the
epidemic.
Additionally, we compare the performance of exposure
notification to manual contact tracing to (1) establish similarities
between relative staffing level and exposure notification adoption
and (2) to verify an additive effect of concurrent manual tracing
and exposure notification.
We see improvements in all cases when combining interven-
tions (Fig. 5). In all three counties, exposure notification has a
stronger effect at the given staffing and adoption levels, but
adding either intervention to the other results in reduced
infections, albeit to different extents based on the trend of the
epidemic. This suggests that both methods are useful separately
and combined, and that the trend affects the relative utility of the
interventions.
Concurrent interventions under behavioral changes
While the results shown above suggest that the interventions are
effective in suppressing the COVID-19 epidemic to various
degrees, in practice, health organizations will implement multiple
intervention strategies simultaneously to try to curb the spread of
the virus while also allowing controlled reopenings. Therefore, we
also study the combined effect of concurrent interventions
including digital exposure notification, manual contact tracing,
and network interaction changes. We explore changes to social
distancing in Supplementary Figs. 1, 2.
We examine the effects of combined NPIs under various
“reopening” scenarios by increasing the number of interactions in
every interaction network, including households, workplaces,
schools, and random networks. Specifically, we increase these
interactions by a given percentage from the levels as of July 11,
2020 (0% reopen) up to the levels at the very start of the epidemic
before broad-based mobility reductions (100% reopen). Given the
average number of interactions i for network n at the end of the
baseline as i
b,n
and before the lockdown as i
0,n
, the network
reopening percentage p (in 0–100%) defines the current relative
M. Abueg et al.
2
npj Digital Medicine (2021) 49 Published in partnership with Seoul National University Bundang Hospital
1234567890():,;
Fig. 1 Simulation time series. Simulation results for various levels of exposure notification app uptake (among the total population) during
2020, with the app being implemented on July 11, 2020 in King County (a–f), Pierce County (g–l), and Snohomish County (m–r). The shaded
areas represent one standard deviation. Plots (a, g, m) are the new infections on the given day, which consistently decrease with EN adoption,
although the largest variances occur in the mid-range of adoption changes. Plots (b, h, n) are the total infected percentage, Plots (c, i, o) are
the total deaths, Plots (d, j, p) are the total in hospital, all of which naturally vary by new infections given the natural progression through the
compartmental model. Plots (e, k, q) are the total number of tests performed per day, which also varies by new infections as tests are only
performed on symptoms, not on trace, so they correlate with the proportion of newly infected individuals who eventually show symptoms.
Plots (f, l, r) are the percent of people in quarantine at that time step, which varies non-linearly with the infection rates due to increased
quarantines from increased contact tracing but decreased quarantines with decreasing infection rates.
M. Abueg et al.
3
Published in partnership with Seoul National University Bundang Hospital npj Digital Medicine (2021) 49
1234567890():,;
interactions under reopening i
c,n
as
i
c;n
¼
p
100
1 i
b;n
þ i
b;n
:
(1)
The infectious rate increases due to a 10–20% reopening are
balanced by decreases due to a 22-37% exposure notification app
adoption, although the effect varies by county (Fig. 6). This shows
that limited additional reopenings may be possible after introdu-
cing exposure notification alongside existing fully staffed manual
tracing (15 staff per 100,000 people), but that social distancing
remains an important measure under these circumstances.
Additionally, there is an increased effect to adding exposure
notification under greater reopening scenarios.
As part of the Washington State Department of Health’s “Safe
Start” plan, a key target metric to reopen Washington is to reach
fewer than 25 new cases per 100,000 inhabitants over the prior
two weeks
26
. We examine how many days it would take to reach
that target under the combined NPIs. With the spike in cases at
the end of the baseline, the trajectory for reaching these targets
without renewed lockdowns is out of the range of the simulations.
Therefore, to show the relative benefits of the NPIs, we introduce
an artificial renewed lockdown at the mobility levels averaged
over the month before the Phase 2 reopenings (Phase 1.5 for King
County) that occurred on June 5th. Using this averaged mobility
from May 6 to June 5, we model the relative effects of manual
tracing and exposure notification on the Washington Safe Start
key metric.
We find that, for all three counties, the recommended staffing
levels of manual tracing and moderate exposure notification
adoption rates can provide a meaningful reduction in the amount
of time it takes to achieve this metric (Fig. 7). Under the
recommended standard for manual tracing, adding exposure
notification at 30% adoption results in reaching the target in 92%,
87%, and 85% of the time versus no exposure notification for King,
Pierce, and Snohomish counties respectively. At the reduced levels
of 4.7 tracers per 100,000 population, the target is reached in less
than 83% and 88% of the time for King and Snohomish,
respectively, although the exact ratio can not be calculated as
the metric is not achieved in the baseline simulation.
DISCUSSION
We conducted a model-based estimation of the potential impact
of a digital ENS in Washington state by extending the OpenABM-
Covid19 simulation framework. We calibrated our model using
real-world data on human mobility and accurately matched
epidemiological data in Washington state’s three largest counties.
Similar to Hinch et al.’s report on digital contact tracing in the
UK
15
, we found that exposure notification can meaningfully
reduce infections, deaths, and hospitalizations in these Washing-
ton state counties at all levels of app uptake, even if a small
Fig. 2 Peak metrics vs exposure notification adoption rates. Estimated total infected percentage (a–c), total deaths (d–f), and peak in
hospital (g–i)(y-axes) of King (a, d, g), Pierce (b, e, h ), and Snohomish (c, f, i) counties for various levels of exposure noti fication (EN) app uptake
among the population (x-axis) between July 11, 2020 and December 25, 2020. The boxes represent the Q1 to Q3 quartile values with a line at
the median. The whiskers show the range of the data (1.5 * (Q3–Q1)) and any outlier points are past the end of the whiskers.
M. Abueg et al.
4
npj Digital Medicine (2021) 49 Published in partnership with Seoul National University Bundang Hospital
fraction of the population participates. We showed how exposure
notification can be combined with manual contact tracing to
further suppress the epidemic, even if the two interventions do
not explicitly coordinate. Our simulations showed that the
simultaneous deployment of both interventions can help these
counties meet a key incidence metric defined by the Safe Start
Washington plan. The potential overall effect of exposure
notification seems to be greater than even optimal levels of
manual contact tracing, likely because of its ability to scale and
better identify random interactions.
We found that quarantine rates, which contribute to the social
and economic cost of these interventions, do not strictly increase
with exposure notification adoption. In some cases, fewer people
are quarantined even though a greater fraction of the population
participates in the app, which we attribute to successful
suppression of the epidemic at high levels of exposure notification
Fig. 3 Quarantine events vs exposure notification adoption. Estimated total quarantine events of King (a), Pierce (b), and Snohomish (c)
counties for various levels of exposure notification app uptake among the population from July 11, 2020 to December 25, 2020. Note that
even for the “default” (0% exposure notification app uptake) scenario there is a non-zero number of quarantine events because this assumes
that symptomatic and confirmed COVID-19 positive individuals will self-quarantine at a rate of 80%, even in the absence of an app.
Fig. 4 Infections under manual contact tracing. Estimated effect of manual contact tracing on new infections (a–c) and total infected
percentage (d–f) at various staffing levels per 100k people in King (a, d), Pierce (b, e ), and Snohomish (c, f) counties between July 11, 2020 and
December 25, 2020.
Fig. 5 Combined effects of manual contact tracing and exposure notification system. Comparison between manual contact tracing (CT) at
the recommended staffing level and exposure notification (EN) at 30% adoption in King (a), Pierce (b), and Snohomish (c) counties. In all three
counties, the combined effect is greater than individual contributions by either system.
M. Abueg et al.
5
Published in partnership with Seoul National University Bundang Hospital npj Digital Medicine (2021) 49
