Spatial health surveillance systems typically identify high risk regions based only on the residence address of diseased individuals. Geo-located social media data o ers a unique opportunity to obtain information on the spatial movements of individuals as well as their disease status. This can be a rich source of information to identify high risk places even in regions where no one lives such as parks and entertainment zones. We develop two models and their respective algorithms to deal with this challenging problem. We demonstrate the applicability and e ectiveness of our proposed methods by applying them to a collection of geotagged Twitter messages coming from Brazil. In particular, we target the identi cation of spatial clusters associated with Dengue, a vector-borne disease that a ects millions of people in Brazil annually, and billions worldwide, to show the usefulness of our methods for disease surveillance.
The current fast increasing popularity of devices equipped with location sensors o ers unprecedented possibilities for data mining research. The deluge of geo-located data daily generated across several sources (e.g. mobile phones, connected vehicles) allows us to answer many questions regarding the behavior of targeted populations with a timing and precision not possible in the last decade.
For instance, health surveillance systems usually identify high risk places based only on the residence address or the working place of diseased individuals. This approach ignores a multitude of exposures the individuals are daily subject to and therefore provides little information about the actual places where people are infected, the truly important information for disease control. The increasing availability of geo-located data in online platforms o ers a unique opportunity: in addition to identifying diseased individuals, we can also follow them in time and space as they move on the map. Incorporating the mobility of individuals into spatial analysis requires the development of new models that can Proceedings of the VLDB 2018 PhD Workshop, August 27, 2018. Rio de Janeiro, Brazil.Copyright (C) 2018 for this paper by its authors. Copying permitted for private and academic purposes.. cope with this type of data in a principled way and e cient algorithms to deal with the ever growing amount of data.
In this research, we give a contribution in such direction. We exploit geo-located social media data to detect spatial clusters of disease infection. Our main goal is to contrast observed mobility patterns for diseased and non-diseased individuals in order to detect localized regions where likelihood of being infected by a given disease is higher than in the rest of the map. Identifying places where people have higher risk of being infected may be key to surveillance actions. In summary, our contributions are as follows: (i) We introduce the problem of detecting spatial clusters of infection risk from mobility data; (ii) We propose two novel models, and their respective algorithms, for the discovery of spatial clusters of infection risk; (iii) We propose an extraction and modeling strategy of geo-located social media data to the proposed problem; (iv) We present experimental results to illustrate our approach in action by applying our algorithms to a collection of geo-located Twitter data from Brazil. 2.
We use Figure 1 to explain our problem. Each individual is indexed by a number i and has a set of ni spatial positions. The spatial points are given by each lat-long coordinate pair embedded in a geo-located social media post. The positions from a single person are connected by line segments in Figure 1, representing her movements. The individuals are additionally labeled by two colors according to their status: disease cases (in red) or controls (in blue). The cases are those individuals who mentioned a personal experience with the disease in at least one message (further details in Section 4). The messages which have mentioned personal experience with the disease are marked with a hatched shadow in Figure 1. The gure also shows a spatial region Z where the risk of becoming a case might be higher than in the rest of the region.
Our main goal is then to scan the map varying the position, shape, and size of the candidate regions, looking for the region Z^ that most likely is a higher risk area. After nding this most likely hot spot Z^, we want to calculate its probability of occurrence to evaluate whether there is enough evidence to call it a real cluster.
Several challenges emerge in this problem: (i) Each i-th
individual is not associated with a single location, as in the
usual spatial cluster detection task [
We adopted a case-control framework, where the data consist of locations, within a speci ed geographical region, of all known cases of a particular disease, and of a random sample of controls drawn from the population at risk. We labeled the individuals such that the rst N of them are the cases and the last M are the controls.
In our analysis, the key innovation is that the input is a series of locations rather than a single location for each individual. Let xi = (xi;1; : : : ; xi;ni ) be the point events associated with the ni messages issued by the i-th individual, i = 1; : : : ; N + M . Each xi;k represents the geographical message location such as a lat-long coordinate pair. For the cases i = 1; : : : ; N , at least one message in xi refers to a personal experience with the disease and it is denoted disease-labeled message. Typically, there will be a small percentage of disease-labeled messages for each individual. None of the control individual messages are disease-labeled.
Let Z be a (large) set of geographical regions that are candidates to be spatial clusters. There are potentially innite regions in Z and they cover the entire region under analysis. By varying Z 2 Z we scan the map looking for the zone Z^ that most likely is a higher risk area. After nding this most likely hot spot Z^, we calculate its likelihood to evaluate whether there is enough evidence to identify it as a real cluster.
The multiple number of locations associated with each
individual, rather than the usual single location (such as
their place of residence), leads us to consider two di erent
models, which we call Visit Model and Infection Model [
Let Vi;z be the random number of messages in Z among the ni total number of messages issued by the i-th individual. Use 1[A] to represent the indicator random variable that the event A occurs. Hence 1[Vi;z 1] is the binary random variable indicating whether the i-th individual ever issued a message inside the candidate zone Z. These random variables can be assumed independent, but they are not identically distributed as the success probability depends on the number ni of messages issued by each individual. Denote by p = p(Z) the probability that, giving that a case individual is tweeting, she does it from within Z. Let p = p(Z) be the similar probability for a control individual. We are interested in zones where p(Z) > p(Z).
For a user who is a case, we have P (Vi;z 1) equals to 1 (1 p)ni and, for a control user, it is equal to 1 (1 p)ni . Considering a xed zone Z, the visit model likelihood is given by L1(Z; p; p) = (1 p)PiN=1 ni1[Vi;z=0] (1 p)PiN=+NM+1 ni1[Vi;z=0] N Y h(1 i=1
N+M
Y h(1 i=N+1 (1 p)ni )1[Vi;z 1]i (1 p)ni )1[Vi;z 1]i where we simpli ed the expression by dropping the zone Z from p(Z) and p(Z) writing simply p and p. 3.2
We will estimate the probability that someone issues a disease-labeled message (becomes a case) given that she visited k times the region Z. Let r = r(Z) be the infection risk inside the candidate cluster and r = r(Z) the infection risk in Z, the region outside Z. We are interested in zones Z where r(Z) > r(Z).
Let Ii be the binary indicator that the individual i is a case. We assume that these binary random variables are independent. They are not identically distributed since their probability of Ii = 1 depends on the number of visits Vi;z by the i-th individual to the zone Z. We have P(Ii = 1jVi;z = ki) = 1 P(Ii = 0jVi;z = ki) = 1 (1 r)ki (1 r)ni ki = (ki; r; r). Therefore, the likelihood of the pattern of cases and controls is given by
N+M L2(Z; r; r) = Y ( (ki; r; r))Ii (1 i=1 (ki; r; r))1 Ii 3.3
Recall that Z is the set of candidate zones to be scanned. The test statistic we adopt for the Visit Model is T1 = L1(Z^; p^; p^) =
L1(Z^; p^(Z^); p^(Z^))
(1) sup
Z2Z p^(Z^)>p^(Z^) and an analogous formula de nes T2 for the Infection Model. In order to verify its statistical signi cance, we must use Monte Carlo simulation to obtain the null hypothesis distribution of T1 and T2 as the exact or asymptotic analytic calculation is not feasible. The null hypothesis is given by either H0 : p = p or H0 : r = r for all Z 2 Z for the Visit Model and the Infection Model, respectively.
The Monte Carlo distribution is determined by randomly permuting the labels of cases and controls among all individuals. Using this pseudo dataset, we proceed the entire scan over all Z 2 Z to obtain a pseudo value for T1 and T2. As this will be replicated several times, we call these values T1(1) and T2(1). We then select another random permutation of the labels, scan the zones and nd T1(2) and T2(2). Independently, we repeat this procedure a large number B 1 of times generating a set of pseudo values plus the values calculated with the actually observed dataset: T1; T1(1); T1(2); : : : ; T1(B 1) and T2; T2(1); T2(2); : : : ; T2(B 1). Under the null hypothesis, these values are independent and identically distributed. Therefore, the rank of the real observed statistics T1 and T2 are uniformly distributed on the integers 1; : : : ; B. This implies that an exact p-value for the null hypothesis of visit model is given by and an analogous formula de nes p2 for the Infection Model. The test is signi cant at the level 2 (0; 1) if pm < . When either test is signi cant, the most likely zone is given by the corresponding maximizing argument Z^ in (1).
We also identify secondary clusters, zones with highly signi cant p-values, which do not intersect with the most likely zone Z^. The non-intersecting restriction is necessary because, if one zone Z^ is the most anomalous in Z, many other sets in Z that are only slightly di erent from Z^ will produce very similar likelihood numbers. These zones should be ignored since the most anomalous among them has already been pinpointed. Among the non-intersection zones, we look for those whose p-value pm is smaller than where the p-values are calculated as described above.
In this section, we apply both models to search for spatial
clusters of Dengue infection using Twitter data.
Dengue Overview: Dengue is an emerging mosquito-borne
viral disease with estimated 100 million global infections per
year [
Dataset: Our geo-located data were collected through the Twitter Streaming API2. The collection period goes from January 1st, 2015 to December 31th, 2015 during which we were able to crawl a total of 106,784,441 Twitter messages. We set a geographic boundary box covering the Brazilian territory and consequently all collected tweets are geo-tagged with lat/long GPS coordinates. Based on the geographic coordinates, we assigned each tweet to a valid municipality to perform a city-level analysis.
Content Filtering and Analysis: Individuals are labeled
as cases or controls based on the content of their tweets.
In order to nd individuals presenting a dengue infection
episode, we follow [
In order to run the algorithms, the zones Z are de ned by overlaying di erent grids on the map and each grid cell corresponds to a zone to be scanned. The size of the grid cells vary in order to accommodate risk zones that present di erent characteristics. We set the number of Monte Carlo replicas to B 1 = 999 and de ne the signi cance level as = 0:05. We present the results for 4 Brazilian cities in Table 1, Goia^nia (GOI), Limeira (LIM), Sa~o Jose dos Campos (SJC) e Sorocaba (SOR).
Notice that our models were able to nd spatial clusters in 3 cities that faced strong surges of Dengue during the year of 2015. This is a good indication of the usefulness of the algorithms for disease surveillance. Figure 2 depicts the zones found by each model in the corresponding cities. Notice that in the city of Limeira the models identi ed different regions within the same city. These results also point out the complementarity of the models, so that they may be used together towards establishing two di erent levels of surveillance.
The spatial cluster detection task aims at detecting
localized spatial regions or zones, called spatial clusters, where
the probability of some event occurrence is higher than in the
rest of the map. Spatial cluster detection methods, such as
the spatial and subset scan statistics [
In all this large body of work there has been one invariant aspect of the spatial characterization of the individual input data: there is only one spatial position associated with each one of them. Let it be a pixel (as in a medical image) or a random spatial event (such as an accident location), they have one single spatial location associated, either it be a case or a control individual. In spatial epidemiology, searching for environmental putative sources of infection or disease, in a few cases there have been two positions associated with each individual, his residential and working place addresses. The recent availability of spatial data o ers a unique opportunity and the existing data mining techniques for spatial cluster detection fail to address this new setting as they require a single location to each individual under analysis. Our proposal generalizes this approach by considering the individuals' mobility patterns instead of a single point.
Geo-located social media data o ers a unique opportunity
to obtain information on the spatial movements of people.
These data are easily available, in large amount and with
almost no delay. Furthermore, we can extract the disease
status as cases and controls of the individuals from the textual
content. The stochasticity of location data is not
appropriate for the usual spatial cluster detection tools such as the
traditional spatial scan statistic approach [
One limitation of our approach is the self-selected sample nature of our data. A random sample of social network users is not a random sample of the risk population. There are several biases involved in such a sample. However, we feel that there is merit in developing and using these methods for two reasons. First, in poor regions with lack of information and resources, the suggestion of potential regions of high risk may target a higher proportion of the available resources toward regions with larger probability of being true risk clusters. Second, the population coverage of social networks is expected to continue to expand, resulting in a larger and less biased sample of the population. Additionally, we could imagine using these methods not just on geo-tagged social media data but on user location data more frequently collected from devices such as cell phones. For example, new initiatives have sampled individuals and, upon their consent, tracked their movement 24/7 as well as measured their disease status (case or control) after some time.
We would like to thank FAPEMIG, CNPq and CAPES for their nancial support. This work was also partially funded by projects InWeb, MASWeb, EUBra-BIGSEA, INCT-Cyber, ATMOSPHERE and by the Google Research Awards for Latin America program.