Space-time personalized short message service (SMS) for infectious
disease control – Policies for precise
public healthPlease cite this article as: Ling Yin, Applied
Geography, https://doi.org/10.1016/j.apgeog.2019.102103
0143-6228/© 2019 Elsevier Ltd. All rights reserved.
Space-time personalized short message service (SMS) for infectious
disease control – Policies for precise public health
Ling Yin a, Nan Lin a, Xiaoqing Song a,b, Shujiang Mei c, Shih-Lung
Shaw d, Zhixiang Fang e, Qinglan Li a, Ye Li a, Liang Mao f,*
a Shenzhen Institutes of Advanced Technology, Chinese Academy of
Sciences, Shenzhen, Guangdong, China b College of Territorial
Resources and Tourism, Anhui Normal University, Wuhu, Anhui, China
c Shenzhen Center for Disease Control and Prevention, Shenzhen,
Guangdong, China d Department of Geography, University of
Tennessee, Knoxville, TN, USA e State Key Laboratory of Information
Engineering in Surveying, Mapping and Remote Sensing, Wuhan
University, Wuhan, Hubei, China f Department of Geography,
University of Florida, Gainesville, FL, USA
A R T I C L E I N F O
Keywords: Short message service (SMS) Dengue fever Mobile phone
tracking Trajectory analysis Precision health
A B S T R A C T
Mobile phones and short message service (SMS) have been widely used
in disease control and prevention. Personalized SMSs further allows
real-time, precisely targeted interventions that achieve better
cost- effectiveness. Few SMSs are personalized based on
spatiotemporal travel behavior of individuals, which plays an
important role in disease spread. We proposed a set of SMS policies
tailored to individuals’ travel behavior derived from massive
mobile phone tracking records. These policies tend to alter
spatial, temporal, or spatio- temporal patterns of individuals’
daily activities, in order to reduce the risk of disease spread.
Taking Shenzhen city, China, as a study area, we simulated and
evaluated these policies for Dengue Fever intervention. Our
simulation results show that the spatially targeting policy that
discourages discretionary trips produces the highest
cost-effectiveness to control disease spread in areas with high
importation risk. For the entire city, however, the temporally
targeting policy that shifts individuals’ travel schedules achieves
the best cost- effectiveness. Our study contributes to a new ground
of precise public health that calls for individualized, real-time,
and accurately targeted interventions. Utilizing big mobile phone
data, we present a novel approach to design, simulate, and evaluate
space-time precise intervention for disease control.
1. Introduction
Mobile phones and short message service (SMS) that are already a
part of people’s daily life are powerful tools to improve health by
alerting disease risk and assisting disease prevention (Badawy
& Kuhns, 2017; Fjeldsoe, Marshall, & Miller, 2009; Krishna,
Boren, & Balas, 2009). For instance, short messages were sent
to a large number of subscribers with health promotion slogans for
HIV/AIDS in India, to encourage parents to get their children
vaccinated in Pakistan, to pro- mote hand-washing in Nepal for
preventing diarrhea, and to remind the population about bed net
distribution for malaria control in Africa (Deglise, Suggs, &
Odermatt, 2012).
In addition to the ‘one-message-fits-all’ policy, the SMS can be
further personalized to improve intervention effectiveness. For in-
stances, Franklin, Waller, Pagliari, and Greene (2006) reported
an
increased adherence to medicine taking when sending patients daily
text-messages with personalized goal-specific prompts and tailored
to patient’s age, gender, and insulin regimen. Yoon and Kim (2008)
found that weekly optimal advice via SMS based on patients’
medication detail, diet and exercise can rapidly improve and stably
maintain their health measures. Therefore, the personalized SMS is
considered as a promising policy in ‘precision public health’,
which calls for ‘providing the right intervention to the right
population at the right time’ to maximize cost-effectiveness
(Khoury, Iademarco, & Riley, 2016; Patrick, Griswold, Raab,
& Intille, 2008).
Although SMSs can be tailored by a variety of personal information
for disease control, little attention has been paid to individual’s
travel pattern over space and time. The important role of travel
behavior or human mobility in spreading diseases has been well
documented, particularly for communicable diseases, such as flu and
dengue fever (Le
* Corresponding author. 3141 Turlington Hall, Gainesville, FL,
32601, USA. E-mail address:
[email protected] (L. Mao).
Contents lists available at ScienceDirect
Applied Geography
2
Menach et al., 2011; Meloni et al., 2011; Huang, Das, Qiu, &
Tatem, 2012; Wesolowski et al., 2012). Travel restriction has also
been widely considered as an effective intervention policy for
epidemic control (Bajardi et al., 2011; Germann, Kadau, Longini,
& Macken, 2006). The personalized SMSs based on travel
behavior, however, were not well studied in the literature because
of two research challenges: how to obtain detailed individual
travel records for a massive population; and if succeed, how to
tailor the sent message for each individual based on this ‘big’
data.
Mobile phone tracking technology has been recognized as one of
promising solutions toward ‘precision’, due to its popular use,
timeli- ness, and location awareness (Patrick et al., 2008;
Wesolowski, Buckee, Engø-Monsen, & Metcalf, 2016; Lai, Farnham,
Ruktanonchai, & Tatem, 2019). Millions of mobile phone tracking
records make it possible to depict movements and contacts of
individuals in unprecedented details (Blondel, Decuyper, &
Krings, 2015; Deville et al. 2014; Gonzalez, Hi- dalgo, &
Barabasi, 2008; Schneider, Belik, Couronne, Smoreda, &
Gonzalez, 2013; Song, Qu, Blumm, & Barabasi, 2010; Xu et al.
2016). When combined with disease incidence or prevalence, the
mobile phone data can be used to predict movements of disease
carriers, and identify sources and sinks of disease dispersion,
then suggesting targeted control and elimination programs to the
right places (Bengtsson et al. 2015; Finger et al. 2016; Mao, Yin,
Song, & Mei, 2016; Meankaew et al. 2010; Le Menach et al.,
2011; Panigutti, Tizzoni, Bajardi, Smoreda, & Colizza, 2017;
Peak et al., 2018; Tatem et al. 2014; Wesolowski 2015; Weso- lowski
et al. 2012, 2014). In the current literature, however, few of
these intervention programs could be labelled ‘space-time
personalized’, even though big data were heavily utilized. First of
all, few of these inter- vention programs are personalized at a
fine temporal scale. A majority of studies have been focused on
spatial risk mapping and intervention, for instance the recent work
in Namibia (Tatem et al. 2014), Kenya (Wesolowski et al. 2012), and
Zanzibar region of Tanzania (Le Menach et al., 2011). The suggested
intervention strategies were spatially ori- ented but did not
consider temporal dynamics of individuals’ travel behavior.
Wesolowski et al. (2017) only used mobile phone tracking data to
explore seasonal travel and then inter-seasonal intervention
strategies within a year. Despite high temporal resolution (e.g.,
hourly) offered by mobile phone tracking technology, current
studies have seldom taken this advantage to investigate temporally
resolved in- terventions, such as the best hours in a day to deploy
intervention. Second, many studies are not spatially personalized
either (Lee et al. 2016). Since most of these studies were oriented
to entire countries, the smallest spatial unit for analysis and
disease intervention is often large in size. For example, the study
in Namibia considered 402 basis units for analysis (Tatem et al.
2014), which makes each spatial unit covering 2000 km2 wide on
average. Interventions targeted to such large spatial units could
be costlier in time and efforts than those to small areas or to
specific individuals. In a nutshell, the current literature has not
made full use of the high spatiotemporal resolution offered by the
mobile phone big data. A major reason is these studies were
conducted at a national level, and the detailed mobile phone
records needed to be aggregated to coarser spatial and temporal
scales (e.g., by prefecture and by season) to ease the analysis. To
the best of our knowledge, there are few attempts to explore
spatiotemporally personalized interventions for disease
outbreaks.
To fill the research gap, we proposed a set of SMS policies
tailored to individuals’ travel behavior with a high spatiotemporal
resolution, which were derived from massive mobile phone tracking
records. Our proposed policies tend to alter spatial, temporal, or
spatiotemporal patterns of individuals’ daily trajectories, in
order to contain the spread of diseases. Taking Shenzhen city,
China, as a study area, we simulated and evaluated these policies
for controlling dengue fever outbreak.
In 2014, the Guangdong province in South China was swept by an
unexpected dengue fever epidemic, which started with 423 cases in
August and reached 42,358 cases by the end of October (Cheng et
al., 2016; X.; Jin, Lee, & Shu, 2015; Lin et al., 2016). Aedes
albopictus is the
major vector of Dengue fever in this study area (L. Q. Jin &
Li, 2008). As the second largest city in the province (Fig. 1), the
study area had also experienced an abrupt increase of dengue
infections in 2014 and is now scaling up its dengue control program
to fight future outbreaks of dengue and Zika fever.
2. Materials and methods
2.1. Mobile phone tracking data
The mobile phone dataset is produced by the China Mobile Tele-
communications Company, and obtained from Shenzhen Transportation
Operation Command Center for research purposes. It tracks millions
of anonymized mobile phone users during a regular weekday (24 h in
total) in 2012 without any major event. The mobile phone tracking
dataset is generated from real-time monitoring of phone signals
through a network of cell towers. Each cell tower gives a cellular
coverage to an area that is often approximated by a non-overlapping
Thiessen polygon, hereinafter referred to as the phone catchment
area. The radius of a catchment area varies from 200 m to 2 km,
dependent on the proximity between towers. A cellphone is connected
to a cell tower when entering the tower coverage area, and
meanwhile obtains a current location as latitude and longitude of
the cell tower. It is noteworthy that this dataset is different
from the caller detail records (CDR) widely used in many other
litera- ture, in that the mobile phone status, including its time
and location, is recorded at a regular basis, i.e., every 0.5–1 h,
no matter whether a telecommunication transaction is made or
not.
The China Mobile Telecommunications Company takes around 75% of
mobile phone user market in Shenzhen, and the original mobile phone
tracking dataset includes 16 million anonymous mobile phone users.
Due to signal loss and power off issues, a portion of mobile users
were not tracked and lost their records during some hours, leading
to gaps in location tracking. We only selected users who had at
least one record in every hour of the 24 h to guarantee continuous
tracking. Since this se- lection procedure may introduce sampling
bias over space, we employed a random resampling strategy to
mitigate such bias. Specifically, we first identified the most
frequently appeared location during the nighttime (0:00–6:00) as an
estimated home location for each mobile phone user, from which we
calculated the user distribution as percentages over 10
administrative regions of Shenzhen city (Fig. 1). Following the
user distribution, we randomly sampled a proportion of mobile phone
users who had continuous tracking records from each of 10
administrative regions, which finally formed a dataset of 3.87
million mobile phone users. For each mobile phone user, the travel
trajectory is composed of 24 hourly records, and each record has
three elements: an anonymous user ID, the latitude and longitude of
currently connected cell tower, and the time period of recording.
Each time period spans for an hour, i.e., 0:00–1:00, 1:00–2:00, …,
23:00–24:00, respectively. Since the mobile phone records are
updated every 0.5–1 h, for users who happen to have more than one
location records within a time period, we only kept the first
recorded location.
2.2. Baseline scenario: local and importation risk of dengue
fever
To form a baseline scenario of no SMS interventions, two types of
dengue transmission risk were estimated, namely the risk of local
infection and that of importation through individual travels. The
local infection risk of an area indicates the likelihood of
acquiring dengue fever virus from existing cases within the same
area. Since dengue fever is vectored by mosquitoes, the local risk
map also implies spatial risk of local mosquito bites that cause
dengue infection. On the other hand, the importation risk of an
area measures the repeated introduction of dengue fever into the
area, via human mobility, that can initiate or sustain the
transmission there. Detailed procedures for estimating both risks
are elaborated in the previous study by Mao et al. (2016), and are
briefly described below.
L. Yin et al.
3
To map the local infection risk, de-identified data for dengue
fever cases from February 2013 to December 2014 were collected by
the Shenzhen Disease Prevention and Control Center. A total of 489
confirmed cases were georeferenced onto the street map as points
using their reported home addresses. Among these dengue fever
cases, there were 350 local infection cases and 139 imported cases
based on their travel history records. Spatial covariate datasets
were assembled to represent weather (the minimum, maximum and mean
of temperature and rainfall from June to December), topography,
land cover, land use, and population at 100 m resolution. A
scalable grid resolution of 100 m was chosen so that each phone
catchment area (with a radius from 200 m to 2 km) can have at least
four grid cells and thus the heteroge- neity of infection risk
within it can be represented. A 100-m local risk grid map was then
created following the random forest classification procedures
outlined by Cohen et al. (2013). Note that the resulting local risk
map did not vary over time during the epidemic. Instead, this
research focused on spatial heterogeneity of infection risk,
meanwhile simplifying the temporal dynamics as an average day of
the epidemic.
By combining the local risk map and the mobile phone tracking
dataset, the importation risk into a phone catchment area was
further calculated as a sum of infection risk of every visitor,
formulated as Equation (1):
IRL¼ XNL
i¼1 Pri (1)
Where Pri ¼ 1 Q24
t¼1ð1 RLðtÞÞ AðtÞ IRL denotes the importation risk to a
phone catchment area L, i.e., the number of imported infections per
day. i represents an individual and NL is the total number of
individuals who visited L during a day. Pri is the total
probability of acquiring dengue fever virus for individual i over
24 h in a day. RL(t) is the local infection risk at a phone
catchment L derived from the local risk map. t (¼1, 2 …, 24 h)
indicates the time of visit in hour and the exponent A(t) is the
standardized activity level of mosquitoes in blood meal hunting
during time t derived from the observational study (Li, Li, &
He, 2004) (Fig. 2). Details can be found in the supplementary file
and a published work (citation blinded for review).
Using Eq. (1), we identified top 50 phone catchment areas with
remarkably high importation risk (Fig. 3), hereinafter referred to
as high-importation-risk areas (HIR_Areas). A total of 310,000
individuals who visited HIR_Areas were referred to as
high-importation-risk visitors
(HIR_Visitors), including those who stay at this area for certain
time or just travel through this area. These HIR_Areas and
HIR_Visitors were set as a baseline scenario for exploring
intervention polices.
2.3. Space-time personalized SMS policies
We proposed a set of travel intervention policies that precisely
target HIR_Visitors with particular spatial, temporal, or
spatiotemporal travel patterns by sending them text messages
regarding recommendations against dengue infection. We assumed
these messages could alter their daily travel behavior over space
and time, and consequently reduced the importation risk.
1) Spatially targeting policies (Sp) aim to reduce the number of
HIR_Visitors by changing their travel destinations, but not the
time sequence. In other words, a text message is sent to all
HIR_Visitors and suggests them to avoid visiting a list of
HIR_Areas. We designed two types of messages: the first type (S1p)
says “Due to dengue fever outbreak recently, please consider
cancelling all your trips today to following HIR_- Areas: XXX, …“;
the second type of messages (S2p) states, “Please cancel your
discretionary trips to following HIR_Areas: XXX …“, aiming to
discourage non-obligatory trips destined to HIR_Areas. Compared to
S1p,
Fig. 1. Geographic location of Shenzhen city in Guangdong province,
China and its administrative districts.
Fig. 2. Hourly activity levels of Aedes albopictus and human
movements during a day in the study area.
L. Yin et al.
4
the policy S2p is less restrictive on travel behavior, because it
does not affect obligatory trips (e.g., work trips) to
HIR_Areas.
As Fig. 4 illustrates, to simulate S1p, we identified trips with
desti- nations at HIR-Areas and trips passing by HIR-Areas from
trajectories of HIR-Visitors. First, for trips with destinations at
HIR-Areas, we further divided them into two types. One type was
commuting trips such as going to work and school, and the other
type referred to non-commuting trips such as shopping and
entertainment. For commuting trips with destinations in HIR-Areas,
we reset their destinations from workplaces and schools back to
homes. This is because, for each mobile phone user, we did not have
information regarding alternatives for these must-visit places. For
the non-commuting trips with destinations in HIR-Areas, we
re-routed their destinations to a randomly selected phone catchment
area not in HIR-Areas but with a similar travel distance. Here the
travel distance was measured as the Euclidean distance between a
mobile phone user’s departure cell tower and arrival cell tower.
Second, for a trip passing by HIR-Areas, regardless of its trip
purpose, commuting or non-commuting, we created a detour route to
replace its original tra- jectory. Specifically, we first
identified a passing point Pk locating at HIR-Areas from a trip
trajectory. Then we replaced Pk with its nearest cell tower outside
HIR-Areas. This step was repeated until all passing point Pk in
HIR-Areas were replaced.
To simulate S2p, we identified non-commuting trips with destina-
tions at HIR-Areas and trips passing by HIR-Areas from trajectories
of HIR-Visitors. Using the same simulation as S1p, we reset the
destinations of non-commuting trips and create detour routes for
trips passing by
HIR-Areas. 2) Temporally targeting policy (Tp) does not reroute
HIR_Visitors to
other destinations but aims to reduce their stay time in HIR_Areas
by shifting departure/arrival time of their trips. According to
hourly ac- tivities of Aedes albopictus in the study area,
5:00–8:00 and 18:00–20:00 are two peak periods of blood meal
hunting (Li et al., 2004), overlapping with the two peak periods of
human travels (7:00–10:00 and 17:00–20:00) (Fig. 2). To reduce such
overlap, the Tp policy recom- mends mobile phone users to adjust
their travel time to avoid staying at areas with higher risk during
peak hours of mosquito biting. For simu- lation, we examined each
HIR_Visitor’s temporal pattern of traveling, and designed a
personalized message, such as “Please adjust your departure/arrival
time of next trip to XX: XX to avoid mosquito biting peak
time”.
Specifically, as Fig. 5 illustrates, Home_RL denoted the local
infection risk RL of a user’s home location. FirstStay_RL was
referred to as the local infection risk RL of the first destination
after leaving home in the morning (5:00–10:00). LastStay_RL is the
local infection risk RL of the last destination before returning
home in the evening (17:00–23:00). The simulation of Tp strategy
for a mobile phone user is described as the follows.
To adjust the departure time from home in the morning: a) If First-
Stay_RL Home_RL and the user left home before 8:00, then we sent a
message as “Please leave home after 8:00”. We simulated this
strategy by postponing the user’s departure time to 8:00. b) If
First- Stay_RL <Home_RL and the user left home after 6:00, then
we sent a
Fig. 3. The identified high-importation-risk areas and
high-importation-risk visitors to control Dengue fever in Shenzhen
city. The visitors of high-importation-risk areas mainly come from
neighboring areas.
L. Yin et al.
5
Fig. 4. Flowchart of simulating the effect of spatial targeting
policies.
Fig. 5. Simulating the effect of temporal targeting policy.
L. Yin et al.
6
message as “Please leave home before 6:00”. We simulated this
strategy by moving up the user’s departure time to 6:00. Note that
although the morning peak of mosquitos’ activity level starts from
5:00, it was not usual for regular working people to leave home
before 5:00, and therefore, we chose 6:00 as the start of the
intervention policy.
To adjust the arrival time at home in the evening: a) If Last-
Stay_RL Home_RL and the user return home after 18:00, then we sent
a message as “Please return home before 18:00”. This was simulated
by moving up the user’s arrival time to 18:00. b) If LastStay_RL
<Home_RL and the user returned home before 20:00, then we sent a
message as “Please return home after 20:00”. This was simulated by
postponing the user’s arrival time to 20:00. Note that users might
leave home again at night, but these night-outgoing trips were
beyond our intervention.
Both of the above messages for departure time in the morning and
arrival time in the evening were sent to users if their travel
pattern met the criteria.
3) Spatiotemporally targeting policies (STp) aim to reduce total
number of trips of HIR_Visitors through cancelling trips for a
certain time period, which changes both spatial and temporal
sequence of a trajec- tory. We designed two types of messages: the
first type (ST1p) stated “Please stay at home from 18:00 to 8:00 to
reduce the risk of dengue infection”, aiming to discourage all
trips during peak hours of mosquito bites. We simulated this policy
by anchoring HIR_Vistors’ trajectories at home between 18:00 and
8:00. The second type (ST2p) of messages is an extreme travel
restriction, stating that “Please stay at home all day today”. We
simulated this policy by anchoring the whole trajectory of a HIR_-
Vistor at home.
To further consider people’s compliance to SMS, we examined ten
compliance levels of mobile phone users to each intervention
policy, from 0% to 100% with an increment of 10%. For each
compliance level we randomly select a fraction of HIR_Vistors to
form a compliant population.
2.4. Evaluation of intervention effectiveness and
cost-effectiveness
2.4.1. Local and regional effectiveness To evaluate the
intervention effectiveness in each phone catchment
area, we computed a local reduction ratio of importation risk
(RR_IRL), formulated as Equation (2):
RR IRL ¼ IR’
L IRL
IRL (2)
where IRL and IR’ L denote the importation risk to a phone
catchment area
L before (the baseline scenario) and after an intervention,
respectively. This local measure was further aggregated to a
regional level as an averaged reduction ratio (ARR_IR) in Equation
(3) and a total reduction of importation risk (TR_IR) in Equation
(4):
ARR IR¼ 1 M XM
L¼1 RR IRL (3)
TR IR¼ XM
(4)
where M is the total number of phone catchment areas in a region.
The region can be either the HIR_Areas or the entire city.
2.4.2. Cost-effectiveness analysis To further evaluate the
feasibility, we investigated the trade-off be-
tween the effectiveness and costs of each intervention policy. The
costs, here, refer to negative impacts on people’s daily travels
and activities, due to cancelling trips and being grounded at
homes. For each inter- vention policy, the resulting costs were
gauged by the Total Trips Cancelled and the Total Increased Hours
at Home for all HIR_Visitors. The cost-effectiveness was then
measured as ratios of effectiveness over
costs, i.e., TR_IR per Trip Cancelled and TR_IR per Increased Hour
at Home. To account for randomness, we simulated each intervention
policy
100 times at each compliance level, and calculated the mean value,
as well as the coefficient of variance (CV), for those
effectiveness and cost- effectiveness indicators.
3. Results and discussion
3.1. SMS intervention effectiveness for HIR_Areas only
We first focused our analysis on HIR_Areas, the top 50 phone catch-
ment areas with the highest importation risk. The effectiveness of
all policies is linearly proportional to the compliance level (Fig.
6A). This is because the importation risk was calculated as an
arithmetic sum of infection risk of every visitor in Equation (1).
The maximum variability was limited to 1.33% around the mean value
(Fig. 6B). indicating that the simulated effectiveness was well
stabilized after 100 realizations. The stability of simulation
results can be attributed to the similarity and regularity of local
residents’ movements. That is, the HIR_Visitors mainly moved
locally (Fig. 3), which increased the similarity of their visited
places. Previous studies have also revealed that most of urban
residents have less than four frequent activity locations and often
have similar movement patterns among their frequent activity
locations (Schneider et al., 2013; Yin et al., 2015).
The most effective SMS policy was the spatiotemporally targeting
policy ST2p that urged HIR_Visitors to stay at home all day in a
day. It achieved the highest ARR_IR in HIR_Areas across all
compliance levels, given that the spread of dengue fever was highly
localized. The policy that follows was the spatially targeting
policy S1p, which asked HIR_- Visitors to avoid all trips to
HIR_Areas. This policy was effective because a majority of
HIR_Visitors lived around HIR_Areas (Fig. 3), and most of their
trips were related to HIR_Areas, either destined to or through
HIR_Areas. For the same reason, it was not surprising that another
spatially targeting policy (S2p) ranked the third in effectiveness,
since it suggested HIR_Visitors to only avoid discretionary trips
to HIR_Areas.
The fourth effective policy was the spatio-temporally targeting
pol- icy (ST1p) that recommended targeted users staying at home
from 18:00 to 8:00. This intervention policy reduced trips of
HIR_Visitors during the peak hours of mosquito activity, and
resulted in two offsetting processes regarding importation risk. On
one hand, staying at home decreased the number of visits to
HIR_Areas, and lowered the importation risk to HIR_Areas. On the
other hand, if the local infection risk at home was higher than the
places they were supposed to go, staying at home during peak hours
would otherwise increase the chance of acquiring dengue fever,
which in turn raised the importation risk when they traveled in
non-peak hours. The simulated effectiveness of ST1p arose from
these two offsetting processes and hence was not as effective as
the previous three policies. Lastly, the temporally targeting
policy (Tp) was the least effective one for HIR_Areas. It only
shifted HIR_Visitors’ travel time, not their destinations. The
control effect was not as direct as other four policies that
reduced visitor volume into HIR_Areas.
3.2. SMS intervention effectiveness for entire city
Beyond HIR_Areas, SMS intervention policies also affected other
areas of the entire city. Heat maps in Fig. 7 show the estimated
RR_IR for each mobile phone catchment area given an 80% compliance
level. The proposed policies produced positive effects at some
areas (RR_IR > 0 as warm colored), but also negative effects at
some other places (RR_IR < 0 as cold colored).
With regard to the spatially targeting policies (Fig. 7A and B),
posi- tive effect areas were primarily HIR_Areas, where mobile
users were suggested not entering. Around these positive areas,
there were negative effect areas as results of trip substitution
and detour. For trip substitu- tion, HIR_Visitors, possibly with
infection, were simulated to visit other alternative places with
similar travel distance instead, thus affecting
L. Yin et al.
7
Fig. 6. The intervention effectiveness for HIR_Areas and the
variability of intervention simulation estimates at different
compliance levels. (A) The simulated average reduction ratios of
importation risk (ARR_IR) for HIR_Areas. (B) The coefficient of
variance (CV) of simulation estimates.
Fig. 7. Spatial distribution of the reduction ratios of importation
risk (RR_IR) for the entire study area by: (A) the spatial
targeting policy (S1p); (B) the spatial targeting policy (S2p); (C)
the temporal targeting policy (Tp); (D) the spatiotemporal
targeting policy (ST1p); (E) the spatiotemporal targeting policy
(ST2p) with compliance level of 80%. Warm colors indicate positive
effect (reduced importation risk) and cold colors indicate negative
effect (increased importation risk). Simulation results at other
compliance levels can be found in the supplementary file. (For
interpretation of the references to color in this figure legend,
the reader is referred to the Web version of this article.)
L. Yin et al.
8
other areas negatively, especially neighboring areas due to the
travel distance constraint. For detour, HIR_Visitors who should
have passed by HIR_Areas were simulated to choose nearby detour
routes, thus importing infection risk into neighbor areas.
The temporally targeting policy (Fig. 7C) generated relatively
small change of importation risk for the entire city. In other
words, it did not reduce importation risk as effectively as the
spatially targeting policies, but it did not negatively affect
other places either. The spatiotemporally targeting policies (Fig.
7D and E) produced the largest positive effect areas among all
proposed policies. Their positive effect centered at the HIR_Areas
and radiated outward. Since individuals were limited at home, this
policy had no severe negative effect for other places.
Fig. 8 compares the overall performance of policies for the entire
city and for the HIR_Areas only. The total importation risk
reduction (TR_IR) of spatially targeting policies (S1p and S2p) for
the entire city was even less than that for HIR_Areas. Therefore,
the spatially targeting policies had strong local effectiveness,
but were not optimal for global disease control. On the other hand,
the temporally targeting policy produced significantly higher total
risk reduction for the entire city than that for the HIR_Areas. For
city-wide intervention, the temporally targeting policy
outperformed the spatially targeting policies from a perspective of
global optimization. Furthermore, the spatiotemporally targeting
policies remain the most effective for the entire city.
3.3. Cost-effectiveness analysis
Feasible policies are often trade-off between effectiveness and
costs. Given the compliance level of 30% and 80% as examples, Table
1 lists the indicators representing effectiveness, costs, and
cost-effectiveness, respectively. For the HIR_Areas only, the
spatially targeting policy (S2p) did not produce the highest
effectiveness, but had the least costs, thus offering the best
cost-effectiveness for HIR_Areas. This policy can reduce the total
importation risk by 9.52 for every trip cancelled, and by 2.76 for
every increased hour at home.
With regard to the entire city, the temporally targeting policy
(Tp) that only adjusted people’s travel time produced the highest
cost- effectiveness. It reduced the total importation risk by 4.43
for every trip cancelled, and by 1.65 for every increased hour at
home. Thus, the temporally targeting policy can be considered as
the optimal choice for dengue intervention over the entire
city.
4. Discussions
4.1. The issues of data bias
The issue of data bias always comes with studies using mobile phone
data (Lai et al., 2019). There are two major types of data bias in
mobile phone data related to this study. First, mobile phone data
only come from one mobile phone carrier company, which may not
represent all mobile phone users. Second, even if the user market
is large enough to represent all mobile phone users, mobile phone
users may not represent all population, since children, elderly and
socioeconomically disad- vantaged groups are likely to be
underrepresented.
As for the first type of bias, the selected mobile phone carrier
com- pany dominates the market in our study area with 75% share of
total mobile phone users. For this reason, we believe our data from
this company can well represented mobile phone users’ travel
behavior, and well support our exploration on personalized
intervention policies.
We are fully aware of the second type of bias in the mobile phone
data. Children often live with and are supervised by adult family
members. When adult family members receive intervention messages,
they will guide their children’s travel behavior as instructed. For
elderly and socioeconomically disadvantaged groups who do not use
mobile phones, many of them are highly vulnerable to dengue, and
should be informed about intervention via traditional offline
strategies, for example, wall-posting of intervention information
at where those people are most likely to gather.
4.2. Generalizability of the approach
Besides of dengue, the proposed approach can be extended to con-
trolling other mosquito-borne diseases, such as malaria and Zika,
in an urban context. There are three reasons for its
generalizability. First, the estimation model for local infection
risk includes spatial covariates representing weather, topography,
land cover and land use, which are commonly related to many
mosquito vectored diseases. Second, the design of temporally and
spatiotemporally targeted policies depends on peak hours of Aedes
albopictus, but can be easily substituted by other mosquito
species. Third, the concept of importation risk and personal- ized
intervention rely on a key assumption that human movements play a
critical role in disease spreading, which is true for many
mosquito- borne diseases.
Moreover, a main innovation of this study lies in the space-time
personalized service based on individual’s daily travel pattern.
This idea can be applied to broader infectious diseases that are
spatially spread by human movements such as diseases by
human-to-human transmission. In that case, the disease risk models
and intervention models need to be tailored to the specific type of
diseases.
4.3. Limitations of the approach
The estimation models of dengue transmission risk used in this
approach cannot represent the dynamic process of a disease spread
in both spatial and temporal dimensions. Therefore, it is difficult
to inte- grate space-time precise intervention policies with the
different stages of an epidemic. This can be solved by developing
spatially explicit agent- based models to simulate disease dynamics
(Fras-Mart;nez et al., 2011; Bian et al., 2012; Merler et al.,
2015). In such a way, the power of mobile phone data to capture
human spatiotemporal activities can be utilized to design staged
space-time precise intervention. In the future studies,
interactions among mobile phone users need to be considered, for
example social contagion of their travel behavior and their compli-
ance with interventions. We simply assumed that mobile phone users
behaved independently, and thus the importation risk could be esti-
mated as a linear sum of individual infection risks. A more
realistic but non-linear scenario warrant a further investigation.
Second, due to the data accessibility issue, we only used one day
of mobile phone tracking
Fig. 8. The total importation risk reduction (TR_IR) of space-time
intervention polices with compliance level of 80%.
L. Yin et al.
9
data to represent individuals’ daily activity-travel patterns. A
mobile phone tracking dataset with a longer time span such as one
week and one month, can offer more details in individual’s travel
pattern. Besides, if longer periods of trajectory data are
available, the prediction model of travel behavior can be developed
along with external factors such as weather, holidays, major events
and so on. The future importation risk can be estimated with the
proposed model. Then, based on the predicted high-importation risk
areas, future targeted people can be identified to tailor their
space-time interventions accordingly. It would be more helpful if
the personalized travel interventions are proposed based on the
projected future travel behavior. Third, this study only focused on
precise intervention regrading individual travel behavior. Other
types of intervention such as vector control are not involved. In
the future study, the combined intervention can be further explored
for precise health.
5. Conclusions
Precision public health is promising, but more work lies ahead to
develop a robust scientific foundation for use. Using dengue fever
as an example, this study proposed a new approach that utilizes
high- resolution massive mobile phone tracking data to design,
simulate, and evaluate space-time precise intervention for disease
control. This study reveals that each space-time personalized SMS
policy has its own strengths and weakness. An appropriate policy
choice depends on a comprehensive evaluation based on multiple
factors including the focused area (high importation risk area or
entire city), disease control effectiveness, and resulting costs.
Particularly, to be cost-effective in practice, our simulation
results show that the spatially targeting SMS policy to discourage
discretionary travels could be a wise choice for controlling
diseases within the high importation risk areas, while the
temporally targeting SMS policy of shifting travel time is a
cost-effective intervention policy for the entire city.
Funding sources
Appendix A. Supplementary data
Supplementary data to this article can be found online at
https://doi. org/10.1016/j.apgeog.2019.102103.
References
Badawy, S. M., & Kuhns, L. M. (2017). Texting and mobile phone
app interventions for improving adherence to preventive behavior in
adolescents: A systematic review. JMIR mHealth and uHealth,
5(4).
Bajardi, P., Poletto, C., Ramasco, J. J., Tizzoni, M., Colizza, V.,
& Vespignani, A. (2011). Human mobility networks, travel
restrictions, and the global spread of 2009 H1N1 pandemic. PLoS
One, 6(1), e16591.
Bengtsson, L., Gaudart, J., Lu, X., Moore, S., Wetter, E., Sallah,
K., et al. (2015). Using mobile phone data to predict the spatial
spread of cholera. Scientific Reports, 5, 8923.
Bian, L., Huang, Y., Mao, L., Lim, E., Lee, G., Yang, Y., … Wilson,
D. (2012). Modeling individual vulnerability to communicable
diseases: A framework and design. Annals of the Association of
American Geographers, 102(5), 1016–1025.
Blondel, V. D., Decuyper, A., & Krings, G. (2015). A survey of
results on mobile phone datasets analysis. EPJ Data Science, 4(1),
1–55.
Cheng, Q., Jing, Q., Spear, R. C., Marshall, J. M., Yang, Z., &
Gong, P. (2016). Climate and the timing of imported cases as
determinants of the dengue outbreak in Guangzhou, 2014: Evidence
from a mathematical model. PLoS Neglected Tropical Diseases, 10(2),
e0004417.
Cohen, J. M., Dlamini, S., Novotny, J. M., Kandula, D., Kunene, S.,
& Tatem, A. J. (2013). Rapid case-based mapping of seasonal
malaria transmission risk for strategic elimination planning in
Swaziland. Malaria Journal, 12(1), 61.
Deglise, C., Suggs, L. S., & Odermatt, P. (2012). Short message
service (SMS) applications for disease prevention in developing
countries. Journal of Medical Internet Research, 14(1).
Table 1 The cost-effectiveness analysis for space-time intervention
policies with compliance level 30% and 80%.
L. Yin et al.
10
Deville, P., Linard, C., Martin, S., Gilbert, M., Stevens, F. R.,
Gaughan, A. E., et al. (2014). Dynamic population mapping using
mobile phone data. Proceedings of the National Academy of Sciences,
111(45), 15888–15893.
Finger, F., Genolet, T., Mari, L., de Magny, G. C., Manga, N. M.,
Rinaldo, A., et al. (2016). Mobile phone data highlights the role
of mass gatherings in the spreading of cholera outbreaks.
Proceedings of the National Academy of Sciences, 113(23),
6421–6426.
Fjeldsoe, B. S., Marshall, A. L., & Miller, Y. D. (2009).
Behavior change interventions delivered by mobile telephone
short-message service. American Journal of Preventive Medicine,
36(2), 165–173.
Franklin, V. L., Waller, A., Pagliari, C., & Greene, S. A.
(2006). A randomized controlled trial of Sweet Talk, a
text-messaging system to support young people with diabetes.
Diabetic Medicine, 23(12), 1332–1338.
Fras-Martnez, E., Williamson, G., & Fras-Martnez, V. (2011,
October). An agent-based model of epidemic spread using human
mobility and social network information. In 2011 IEEE third
international Conference on privacy, security, Risk and Trust and
2011 IEEE third international Conference on social computing (pp.
57–64). IEEE.
Germann, T. C., Kadau, K., Longini, I. M., & Macken, C. A.
(2006). Mitigation strategies for pandemic influenza in the United
States. Proceedings of the National Academy of Sciences, 103(15),
5935–5940.
Gonzalez, M. C., Hidalgo, C. A., & Barabasi, A.-L. (2008).
Understanding individual human mobility patterns. Nature,
453(7196), 779.
Huang, Z., Das, A., Qiu, Y., & Tatem, A. J. (2012). Web-based
GIS: The vector-borne disease airline importation risk (VBD-AIR)
tool. International Journal of Health Geographics, 11(1), 33.
Jin, X., Lee, M., & Shu, J. (2015). Dengue fever in China: An
emerging problem demands attention. Emerging Microbes &
Infections, 4(1), e3.
Jin, L. Q., & Li, D. (2008). A recent survey of mosquito fauna
in Guandong Province, southern China, with a review of past
records. Medical and Veterinary Entomology, 22 (4), 359–363.
Khoury, M. J., Iademarco, M. F., & Riley, W. T. (2016).
Precision public health for the era of precision medicine. American
Journal of Preventive Medicine, 50(3), 398–401.
Krishna, S., Boren, S. A., & Balas, E. A. (2009). Healthcare
via cell phones: A systematic review. Telemedicine and e-Health,
15(3), 231–240.
Lai, S., Farnham, A., Ruktanonchai, N. W., & Tatem, A. J.
(2019). Measuring mobility, disease connectivity and individual
risk: A review of using mobile phone data and mHealth for travel
medicine. Journal of Travel Medicine, 26(3), taz019.
Le Menach, A., Tatem, A. J., Cohen, J. M., Hay, S. I., Randell, H.,
Patil, A. P., et al. (2011). Travel risk, malaria importation and
malaria transmission in Zanzibar. Scientific Reports, 1.
Lee, E. C., Asher, J. M., Goldlust, S., Kraemer, J. D., Lawson, A.
B., & Bansal, S. (2016). Mind the scales: Harnessing spatial
big data for infectious disease surveillance and inference. The
Journal of Infectious Diseases, 214, S409–S413.
Li, R., Li, J., & He, Y. (2004). Daily observations on blood
meal hunting behavior of Aedes albopictu. J. Med. Pest Control, 20,
27.
Lin, H., Liu, T., Song, T., Lin, L., Xiao, J., Lin, J., et al.
(2016). Community involvement in dengue outbreak control: An
integrated rigorous intervention strategy. PLoS Neglected Tropical
Diseases, 10(8), e0004919.
Mao, L., Yin, L., Song, X., & Mei, S. (2016). Mapping
intra-urban transmission risk of dengue fever with big hourly
cellphone data. Acta Tropica, 162, 188–195.
Meankaew, P., Kaewkungwal, J., Khamsiriwatchara, A., Khunthong, P.,
Singhasivanon, P., & Satimai, W. (2010). Application of
mobile-technology for
disease and treatment monitoring of malaria in the" Better Border
Healthcare Programme. Malaria Journal, 9(1), 237.
Meloni, S., Perra, N., Arenas, A., Gomez, S., Moreno, Y., &
Vespignani, A. (2011). Modeling human mobility responses to the
large-scale spreading of infectious diseases. Scientific Reports,
1, 62.
Merler, S., Ajelli, M., Fumanelli, L., Gomes, M. F., y Piontti, A.
P., Rossi, L., … Vespignani, A. (2015). Spatiotemporal spread of
the 2014 outbreak of Ebola virus disease in Liberia and the
effectiveness of non-pharmaceutical interventions: A computational
modelling analysis. The Lancet Infectious Diseases, 15(2),
204–211.
Panigutti, C., Tizzoni, M., Bajardi, P., Smoreda, Z., &
Colizza, V. (2017). Assessing the use of mobile phone data to
describe recurrent mobility patterns in spatial epidemic models.
Royal Society Open Science, 4(5), 160950.
Patrick, K., Griswold, W. G., Raab, F., & Intille, S. S.
(2008). Health and the mobile phone. American Journal of Preventive
Medicine, 35(2), 177.
Peak, C. M., Wesolowski, A., zu Erbach-Schoenberg, E., Tatem, A.
J., Wetter, E., Lu, X., … Buckee, C. O. (2018). Population mobility
reductions associated with travel restrictions during the Ebola
epidemic in Sierra Leone: Use of mobile phone data. International
Journal of Epidemiology, 47(5), 1562–1570.
Schneider, C. M., Belik, V., Couronne, T., Smoreda, Z., &
Gonzalez, M. C. (2013). Unravelling daily human mobility motifs.
Journal of The Royal Society Interface, 10 (84), 20130246.
Song, C., Qu, Z., Blumm, N., & Barabasi, A.-L. (2010). Limits
of predictability in human mobility. Science, 327(5968),
1018–1021.
Tatem, A. J., Huang, Z., Narib, C., Kumar, U., Kandula, D.,
Pindolia, D. K., et al. (2014). Integrating rapid risk mapping and
mobile phone call record data for strategic malaria elimination
planning. Malaria Journal, 13(1), 52.
Wesolowski, A., Buckee, C. O., Bengtsson, L., Wetter, E., Lu, X.,
& Tatem, A. J. (2014). Commentary: Containing the Ebola
outbreak-the potential and challenge of mobile network data. PLoS
currents, 6.
Wesolowski, A., Buckee, C. O., Engø-Monsen, K., & Metcalf, C.
J. E. (2016). Connecting mobility to infectious diseases: The
promise and limits of mobile phone data. The Journal of Infectious
Diseases, 214(suppl_4), S414–S420.
Wesolowski, A., Eagle, N., Tatem, A. J., Smith, D. L., Noor, A. M.,
Snow, R. W., et al. (2012). Quantifying the impact of human
mobility on malaria. Science, 338(6104), 267–270.
Wesolowski, A., Qureshi, T., Boni, M. F., Sundsøy, P. R.,
Johansson, M. A., Rasheed, S. B., et al. (2015). Impact of human
mobility on the emergence of dengue epidemics in Pakistan.
Proceedings of the National Academy of Sciences. 112(38),
11887–11892.
Wesolowski, A., zu Erbach-Schoenberg, E., Tatem, A. J., Lourenço,
C., Viboud, C., Charu, V., … Metcalf, C. J. E. (2017).
Multinational patterns of seasonal asymmetry in human movement
influence infectious disease dynamics. Nature Communications, 8(1),
2069.
Xu, Yang, Shaw, Shih-Lung, Zhao, Ziliang, Yin, Ling, Lu, Feng,
Chen, Jie, et al. (2016). Another tale of two cities: Understanding
human activity space using actively tracked cellphone location
data. Annals of the Association of American Geographers, 106(2),
489–502.
Yin, L., Wang, Q., Shaw, S.-L., Fang, Z., Hu, J., Tao, Y., et al.
(2015). Re-identification risk versus data utility for aggregated
mobility research using mobile phone location data. PLoS One,
10(10), e0140589.
Yoon, K.-H., & Kim, H.-S. (2008). A short message service by
cellular phone in type 2 diabetic patients for 12 months. Diabetes
Research and Clinical Practice, 79(2), 256–261.
L. Yin et al.
1 Introduction
2.2 Baseline scenario: local and importation risk of dengue
fever
2.3 Space-time personalized SMS policies
2.4 Evaluation of intervention effectiveness and
cost-effectiveness
2.4.1 Local and regional effectiveness
2.4.2 Cost-effectiveness analysis
3.3 Cost-effectiveness analysis
4.2 Generalizability of the approach
4.3 Limitations of the approach
5 Conclusions
Funding sources