An analytical tool to support public policies and isolation barriers against SARS-CoV-2 based on mobility patterns and socio-economic aspects

2.1. Visual analytics based on mobility data and networks for urban dynamics

Analysis of massive data produced by different technological sources, such as mobility data, can be performed through visual analysis tools. This approach has been adopted by some researchers to transform data into robust information for the final user [16], [17]. Visual analytics tools may incorporate mobility patterns to achieve a better analysis of urban dynamics.

These visual analytics tools can extend the comprehension about different aspects of mobility [18], by providing visual insights to support decisions relative to urban systems, such as public transportation [19], [20], [21], traffic analysis [22], tax evasion [23] and smart cities applications [17], [24]. The study of mobility data brings challenges due to complex spatiotemporal patterns, which can be analyzed by the visual tools that present adequate mobility representations [25].

Network science has been used in different contexts where visual analytics tools are applied. Networks are widely applied to provide insights in the context of social networks [26], [27], [28]. Real-world situations can be represented more precisely by weighted directed networks, which contain more information about the relationship between pairs of nodes [29]. Thus, it enables us to characterize terrestrial transportation route networks [30], model and visualize scientific datasets [31], and to analyze patient flow in hospitals [32].

Urban mobility data can be combined with networks to study the relationship among neighborhoods. This analysis can include the detection of communities, which are common mobility patterns in an urban environment. Krueger et al. [33] proposed a visual interactive tool to provide insights for urban planning research, using mobility graphs and unsupervised learning to detect mobility patterns and behavior anomalies.

An interactive visual analytics tool was developed in [34] to enable the detection and changes in spatial–temporal patterns, and also provide clutter reduction when comparing flows of people in certain moments, by combining spatial aggregation of flows and temporal clustering. Wang et al. [35] used geo-textual data, which embodies geographical and human activity information, to generate sequential latent graphs and applied the Louvain algorithm for community detection method to identify and analyze different types of region transformations.

The works found in the literature use automated clustering approaches to detect and analyze mobility patterns. Also, these works did not consider interactive approaches to simulate mobility restriction, which makes it impossible to study the impact of relaxation or tightening of social distancing. In this work, an interactive clustering approach is used in a visual analytics tool to obtain constrained and relaxed mobility patterns using networks generated through anonymized mobile phone data.

2.2. Visual analytics for epidemic analysis and the COVID-19 outbreak

Data visualization encompasses essential tools to aid policymakers during an epidemic, offering statistical analysis, surveillance, and indicator monitoring. The tools are developed such that human cognitive effort is reduced and the capacity to analyze big datasets is augmented. Unified global visualization technologies can provide insights to support decisions to manage an epidemic, and consequently minimize its propagation [36], [37].

The devastating power of COVID-19 became more evident as the pandemic advanced [38]. Therefore, several governments and institutions have used visualizations in their reports on the COVID-19 spread, including economic and labor factors [39], [40], [41]. Many challenges for deploying solutions for COVID-19 management and control arose, such as the development of problem-oriented big data acquisition systems and inference of population behavior [42].

Open information about the geographical distribution of COVID-19 is important to support scientific analysis and to provide a better understanding of the spread of COVID-19 [43]. A good example is a dashboard provided by John Hopkins University [44], which includes heat maps concerning the active number of cases, cumulative cases, hospitalizations, and incidence rates. The World Health Organization [45] also provides an interesting dashboard that includes specific visualizations regarding statistics and region comparisons.

Furthermore, articles exploring data visualization have been published. Dey et al. [39] present an exploratory data analysis of the pandemic with visualizations of the spreading of COVID-19 in China. Gao et al. [40] propose the use of cartograms to visualize the expansion and spread of COVID-19. Barone et al. [46] created a surveillance dashboard for COVID-19 containing graphs, figures, and criticality tables with data collected from the European Centre For Disease Prevention and Control. Zhang et al. [47] used a spread index and an extinction index to investigate when the pandemic grows and declines, and also used bubble charts and animations to visualize how the pandemic is changing over time.

Khan et al. [48] designed an ontology-based infrastructure that captures relationships between data streams, visualization functions, and web pages to allow the rapid deployment of web-based interactive dashboards across multiple data streams, in a context with limited resources and low programming expertise. Ipenza et al. [49] developed a public visual analytics tool, named QDSSUS, to interactively explore millions of Brazilian healthcare records through interactive queries. The authors discovered how patterns evolved over one year of COVID-19 in different regions of Brazil, such as the relationship between patient age groups and their corresponding dominant symptoms.

A survey on the use of visual analytics for public health was presented by Preim and Lawonn [50], which included research about computational epidemiology. The survey mentions that visualizations containing spatiotemporal data are often used to describe the disease’s behavior. This is because epidemic spread simulation based on the analysis of connections between different locations or individuals over time can output more realistic results in what–if scenario analysis.

A view of epidemic modeling through networks is given by [51]. The use of social network analysis and visualization [51], [52] is very important as an epidemic spread often depends on mobility patterns among locations in the studied region. Tao et al. [53] present a visual analytics framework, called HoNVis, which can be used to explore epidemic outbreak scenarios through domestic and international travels, identifying airports of interest and tracing the propagation of an epidemic.

Luo [54] proposed a visual analytics tool to evaluate the effect of control measures using face-to-face interaction patterns of a primary school. Geo-social mixing patterns are explored from human interaction network data, which enables the identification of critical individuals, locations, and clusters of locations. Then, a control measure can be designed, considering the interaction patterns, and evaluated through a compartment model that simulates the propagation of an epidemic.

Dynamic Network Visualization (DyNetVis) [55], [56], [57] was proposed to be an open-source visual analytics tool for dynamic network exploratory visualization. DyNetVis also implements dynamic processes, including epidemic compartment models. It is possible to analyze and explore the infection path of simulated epidemics with different parameters, visualizing the transmission tree and different groups of individuals (i.e. susceptible, infected, and recovered).

Some scientific visualization tools that adopted network analysis were proposed in the COVID-19 context, Abel and Gietel-Basten [58] proposed a tool to analyze the global economy interconnectedness, and thus investigate the economic and social impact of COVID-19, based on a chord diagram.

Saraswathi et al. [59] used software tools, Cytoscape and Gephi, to create social network visualizations and explore common transmission patterns. The visualization approach enabled the identification of evolving hotspots, such as those associated with international travel and principal cities. Also and key actors were identified in the transmission, by considering age-sex attributes in the nodes of the transmission network. The exploratory analysis occurred during different phases of the lockdown in India.

Luo et al. [60] used a visualization approach to construct disease transmission network graphs for the COVID-19 epidemic, according to relationships among cases. The proposed visualization tool for transmission networks can identify the transmission source and contacts, assess the current situation of transmission and prevention, and provide an adequate response to control the propagation of COVID-19.

It is possible to see that past studies for COVID-19 dashboards considered exploratory analysis and surveillance of epidemiologic variables and automated development generation of visualization components. Also, works that proposed graph-based visualization tools for epidemics and COVID-19 were mainly focused on exploratory analysis of the transmission network, the visualization of simulation output from compartment models, and analysis of economic impact based on the degree of connection between different locations.

To the best of our knowledge, this is the first tool that uses mobility patterns, generated from millions of anonymized mobile phones, to manage the prioritization of risk mitigation actions to a set of locations. The insights provided to mitigate risk across a set of locations is obtained by crossing mobility, socioeconomic and epidemiologic data. Thus, it is possible to interactively define the restriction or relaxation of common mobility patterns, defined by clusters, given the values of risk contamination and impact on socioeconomic variables, observe the impact of the pandemic in economic arrangements associated with the clusters, and monitor the clusters’ (and its components) contamination risk over time.

3.1. Data used in the experiments

The mobility data was provided by the In Loco company, which collects anonymized location data from about 60 million devices around the world, from which about 2.2 million devices’ data has been collected in Pernambuco, which is considered an entirely reasonable number in light of the entire population of Pernambuco, found about 8.8 million in the last census in 2010 and recently estimated in 9.5 million.

From the provided Origin–Destination (OD) Matrices that estimate the flow amongst the cities of Pernambuco, Metropolitan Region of Recife, and Recife’s neighborhoods from 2020-05-06 to 2020-05-26. The entries of the OD Matrix represent outflows from each city $i$ to a city $j$. Thus it was calculated the percentage of people that remained in the city $i$ and the percentage that left the city $i$ to each of the other $N-1$ cities at each date of the considered period. Given an OD matrix A, if $a_{ij}$ = 0.2, then a 20% flow was estimated of population moving from city $i$ to city $j$.

In this study, we considered the worst-case scenarios of the COVID-19 spread by estimating the worst-case mobility matrix for the studied period. Thus, given that a matrix $A^t$ is the OD matrix for period $t$, the non-normalized representative matrix of the worst-case mobility for a period $t=1,2,\dots ,T$ scenario is given by $W_{i,j}^{\prime }=\underset{t}{\max }\left(A_{i,j}^t\right)$, where $T$ is the final day available in the dataset. The worst-case mobility matrix is given by $W_{i,j}=W_{i,j}^{\prime }/{\sum }_{j=1}^N{W}_{i,j}^{\prime }$.

Furthermore, data concerning the COVID-19 spread in Pernambuco was collected from the Brazilian Ministry of Heathy reports [64], and the SEPLAG of Pernambuco [65]. Also, the socioeconomic variables of the cities were collected from IBGE [66]. A summary of the spatial variables for Recife neighborhoods and the group from which they belong is presented in Table 1

3.2. Networks and clustering analysis

This section presents the necessary concepts from network analysis used to obtain the results in this paper. Also, we provided some sources that can be used by the interested reader to explore more measures and definitions concerning the topics presented.

Let $D=(V,A)$, be a directed graph, or digraph, where $V$ represents a set of vertices and $A$ (disjoint from $V$) consists of a set of directed arcs together with an incidence function ${\varphi }_D$ that associates each arc of D with an ordered pair of vertices [67].

In our analysis, we use directed graphs to represent mobility patterns where each node, $v_i$, represents a geographical location (e.g., a district, a city, etc.) and each arc, $(i,j)$, represents aggregated movement of one or more people from location $v_i$ to location $v_j$.

Furthermore, for each arc $a=(i,j)$, we define the weight of an arc, $w(i,j)$, as the proportion of the population of each location $i$ that travels region $j$. When considering the mobile location data, we collected the origin–destination matrix for each day from January to the date after social distance policies. Thus, we considered two possibilities for building the network scenarios: (i) the maximum flow observed before social distance policies and (ii) the maximum flow after social distance policies.

The maximum value was used to ensure a conservative risk approach, although other values such as the average could be used. These weights shall be considered according to the current social distance state. Then, the digraph, together with its arcs’ weights, is called a weighted directed graph $(D,w)$.

Our analysis is based on a threshold, $l$, used to filter the arcs in a weighted directed graph to build derived undirected graphs as described in the following. An arc between two vertices $v_i$, $v_j\in V$ exists (i.e., passes the filter) if and only if, $w_{i,j}\ge l$ or $w_{j,i}\ge l$.

In other words, this threshold represents the fractional volume of people, which implies a well-established communication amongst different locations, therefore indicating that there is a proportion of $w_{i,j}\ge l$ of people that travel from city $i$ to $j$.

Fig. 2 presents the directed graphs obtained by choosing three different threshold values in a graph representing the mobility in the city of Recife, where the arc weights were obtained from the anonymized mobile phone location data described in Section 3.1. The algorithm to generate clusters with a threshold value $l$ is shown in Algorithm 1.

The inputs are an origin–destination matrix $W$, a set of Nodes, and the mobility threshold $l$. The first step is to create a graph filtered by the mobility threshold $l$, as shown in lines 1–4. In line 5 we create a list of nodes to be covered by at least one cluster. The partition is initialized in line 6. The loop from lines 7–29 evaluates the cluster from which a node $v$ and its neighbors belong.

In line 8 a node that was not evaluated is selected. In lines 9–16 we evaluate if the neighbors of $v$ belong to at least one existing cluster. If it is true that the neighbors of $v$ belong to at least one existing cluster, we identify the indexes of the clusters with a list $S$. Lines 17–24 merge the clusters that contains some neighbors of $v$. Note that the list $S$, which stores the index of the clusters to be merged, is reversed in line 18 to guarantee that the highest indexes from the list $U$ will be removed first.

The clusters saved in $S$ are merged into a new cluster $C$ and are also deleted from $U$ in Lines 18–22. In line 24, a $v$ and its neighbors will be added to the merged cluster $C$ to constitute a new cluster in $U$. As can be seen in Lines 25–27, if the neighbors of $v$ do not belong to any cluster, a new cluster is created in $U$, containing $v$ and its neighbors. In line 28, $v$ is deleted from the set of nodes that are being evaluated. It can be observed that the sensitivity of the clustering method depends on the filtered graph, which depends only on the mobility threshold $l$.

We can see that the connection between pairs of locations in the considered region is sensitive to the threshold $l$. As the threshold is reduced, fewer locations will be connected, and clusters will begin to appear. These clusters are unconnected graphs, and each cluster is associated with a set of cities that commute at least a proportion $l$ of people between them.

Common approaches that automatically cluster graphs mainly optimize a quality function, such as modularity [28], [29], [68]. This was not adequate for our context, since it is interesting for the policymaker to understand how isolating and relaxing isolation measures may affect the community and the contamination risk within a community.

The proposed clustering approach is the most adequate for our study since the only criteria needed to define connections between locations is the proportion of people commuting from one location to another, respecting the most probable connections, when $l$ is increased or decreased, according to real mobility data. It is possible to obtain the most probable mobility patterns of a certain region and the impact of constraining and relaxing mobility in the risk metrics through this interactive clustering approach.

The adopted clustering approach represents mobility patterns that are more probable to occur according to the value of $l$. These mobility patterns are well-established connections between locations associated with frequent travels relative to family, work, and expenditure of services and goods.

Although singletons will exist as a result of the clustering, they will not contain representative mobility patterns relative to the region under study, since it is possible to define $l$ to cover a significant part of the population of the region (state or city) under study. Thus, one can consider generating a set of clusters that cover at least 90% of the population of the region under study.

Different mobility patterns imply different cluster configurations which will impact the contact between people. Thus, contamination risk metrics will be affected depending on the clusters’ configuration. The division of the region into clusters is important to allocate resources according to the risk of each cluster.

Therefore, this threshold implies a risk exposure considering the connectedness among cities. It is possible to see that this clustering approach makes it possible to configure clusters such that they represent real mobility patterns between locations in a region and to constrain or relax these connections to reduce contamination risk.

To make the importance of this approach clear, we can compare it with other common clustering methods. Clustering methods such as K-means generate a partition from a set of training examples evaluated in a set of observable features. The clustering method presented in Algorithm 1 generates a partition from a graph. It tries to agglomerate locations in different groups that share a well-established connection according to $l$.

K-means will cluster training examples according to their similarity in a feature set. Thus, if anonymized data from cell phones are not available, K-means can be used to generate clusters of locations when it is possible to collect other datasets and extract features that can be processed with this clustering method. Of course, K-means will not provide the same level of interactiveness as our approach, since the user is requested to choose the number of clusters, which will be automatically obtained through the considered feature set and the quality function.

We can now define risk measures for each cluster that will summarize the infection’s possible spread based on the current state of the pandemic and the mobility data. These measures can help policymakers to identify and prioritize those regions that are critical. There are two types of regions to be prioritized: infected and healthy. First, we will present metrics that support plans for economic recovery. We defined the Risk Exposure Measure of a region, $R_{v_j}$ by the following equation:

where $p_i$ is the likelihood of one individual from $v_i$ having the infection. This equation quantifies the risk of infection in a healthy region. In other words, it represents the risk of a healthy region becoming an infected region. In Eq. (1), it was considered that the contamination within a region and its people moving to distinct regions are independent events.

The Cluster Risk Exposure Measure, $R_{C_k^0}$, is given by:

A conservative approach has been assumed for both risk exposure metrics $R_{v_j}$ and $R_{C_k^0}$, thus including the redundant intersection of events, i.e., being contaminated by more than one city simultaneously. These risk metrics support the prioritization of regions where progressive economic recovery is possible. Thus, one may prioritize locations or clusters with small risk exposure to develop economic recovery plans.

The final risk measure is calculated for singletons or clusters that are infectious (contain at least one active case). This metric is important to rank clusters according to their infection rate. If a cluster’s infection rate is high, more efforts are required to mitigate the risk of contamination by defining mobility restrictions. The name of the measure is the Force of Infection (FOI), and it represents the rate at which a single individual contracts a disease [69]. FOI is defined according to Eq. (3).

where $C_k^1$ is an infected location, $Y_{v_j}$ is the number of infected individuals from a city $v_j\in C_k^1$ and $N_{C_k^1}$ is the population size of $C_k^1$. $\beta =c\times t$, where $t$ is the transmission probability and $c$ models the contact rate. Similarly, a more conservative analysis was considered and, therefore, the transmission probability was settled to 100%.

The proxy adopted to represent the contact rate was the mean strength of the cluster [70]. We can see that the risk is reduced when the contact rate is reduced, where this reduction is directly associated with mobility restriction measures, considering the limits of isolation barriers.

The proposed approach enables a way to protect and isolate uncontaminated areas for reducing local economic losses since it allows policymakers to choose which weak mobility links shall be interrupted with risk-informed decision support. Using a simple approach for clustering the neighborhoods brings the advantage of easy understanding and usability by policymakers and epidemiologists. This feature is also an advantage for enhancing communication with society to better support isolation plans.

3.3. System requirements and analytical tasks

In our discussions with the domain experts in our team, we were able to identify four main requirements for the visualization tool developed:

R1 — Visualization of multiple data sources: The analysis of isolation barriers requires the investigation of multiple data sources: mobility networks, COVID-19, and socioeconomic indicators. One particularly important aspect of the analysis is how these sources interact to define the properties of the clusters (defined in Section 3.2).

R2 — Consider multiple scenarios: Mobility restriction policies, associated with the defined isolation barriers have big consequences on the daily activities of the society. For this reason, it is essential to consider the multiple levels of restriction. It will also enable decision-makers to adapt this policy for different locations and also to different moments of the pandemic development.

R3 — Simplicity: Designing policies for such an urgent pandemic situation requires the consideration and discussion with many stakeholders with a wide variety of backgrounds (outside the technical team). Thus, the visualizations designed should be simple and easy to discuss with all the people involved in the decision-making process.

R4 — Easy to share: To quickly promote discussion of policies associated with isolation barriers, it is important that the tool can be easily shared with the stakeholders and the city administration. Given that this is a general theme in the current pandemic situation, this also promotes a more extensive and quicker idea propagation and possible replication in other locations.

5.1. Pernambuco networks: State scale

After setting up the desired threshold using the mobility threshold slider, the user can monitor the clusters’ states (infected or not infected) over time. Fig. 4 illustrates how clusters generated for $l=0.025$ behaved in the period from 2020-05-12 until 2020-05-26 and included cities from other states bordering with Pernambuco that have relevant people flow.

We highlighted some clusters on the first day of analysis, manually associating a letter for each of them. Cluster $A$ did not become infected during the considered period. Therefore, the policymaker may judge if this cluster is a candidate for a specific isolation barrier policy associated with an economic recovery plan or how to relax the isolation policy after observing a tolerance time.

On the other hand, the states of the clusters $B$, $C$, $D$, and $E$ changed in the considered period. Cluster $C$ became infected while Cluster $B$ became not infected on 2020-05-14. Cluster $D$ became not infected on 2020-05-22. Cluster $E$ became infected on 2020-05-14 and then became not infected on 2020-05-22. It is important to highlight that during this period since the tool presented in this paper was not available, no condition-based policy was planned for these clusters of cities, so we expect that this proposed graphic tool may be useful for preserving the population of areas that are not infected.

Another analysis involving the selected cluster set is to investigate the situation of some local productive arrangements (LPAs), which constitute the state’s strategic economic regions. These types of economic clusters are supported by government programs that promote regional development by stimulating the enterprises’ technological development, competitiveness, and sustainability. The analysis is shown in Fig. 5.

Areas in green are not contained in any cluster. Purple areas represent dairy’s LPA, pale orange areas represent plaster’s LPA, yellow areas represent clothing production’s LPA, and blue areas represent viticulture’s LPA. On 2020-05-26, pale’s LPA was the only one not infected. This shows that the majority of the considered strategic economic regions are infected, and policymakers should give special attention and develop a plan of action and monitoring for these areas.

5.2. Recife metropolitan area

An analysis including only the Recife metropolitan area gives more details about this region. In this scenario, instead of considering Recife as a single vertex, each of its neighborhoods was considered in the graph as a vertex to reveal these neighborhoods’ relations with the cities of the Recife metropolitan area. Fig. 6 presents the clusters constructed with a threshold of $l=0.05$, including singletons.

A more detailed analysis of Recife’s neighborhoods is included in Section 5.3. Nevertheless, having a view of the relations between the neighborhoods and cities within the region is interesting. Since the flow of people in the metropolitan area is intense during working hours, it is essential to understand the networks before relaxing social distancing and working policies related to the pandemic. As shown in Fig. 6, Recife’s neighborhoods are within clusters that contain cities such as Olinda, Paulista, Jaboatão dos Guararapes, and Itapissuma.

5.3. Neighborhoods of Recife

This section presents the analysis that considers only the neighborhoods of Recife. Fig. 7 presents the clusters construction considering three different thresholds. The lowest threshold ($l=0.01$ on the left) results in the identification of a single cluster, considering all Recife’s neighborhoods. Increasing the threshold ($l=0.05$ on the center and $l=0.1$ on the right), more clusters are identified, this being interesting for policymakers when deciding over social isolation policies.

Of course, if the threshold is large enough, only singletons will be shown. The dispersion plots in Fig. 7 show the relationship between the Force of Infection of the clusters and their active cases per capita at two dates, considering the same thresholds of $l=0.01$, $l=0.05$, and $l=0.1$.

In Fig. 8, bar plots illustrate the most infected Recife’s neighborhoods in terms of the total number of cases and the number of cases per capita. While Boa Viagem has a higher number of cases on both dates, other neighborhoods (Paissandu and Cidade Universitária) have a higher number of cases per capita. Several hospitals and clinics of Recife are located in Paissandu, which can indicate that this neighborhood receives people from other locations that represent potential contamination cases.

Cidade Universitária is a neighborhood that is mostly centered on activities related to Universidade Federal de Pernambuco, which has a large Clinic Hospital that serves for teaching and provides general public health services. Although the university has stopped its teaching activities since March 15, 2020, and deployed the home office for all non-essential activities, it was not expected that this neighborhood would have such a high incidence of cases.

It is possible that this could be revealing a similar pattern that was found by [71] around two Wuhan hospitals related to a high concentration of SARS-CoV-2 RNA in aerosols, although the infectivity of the aerosolized virus still requires further studies.

Thus, while further studies are not carried out, authorities should reinforce rigorous sanitation procedures and protection for the surrounding population of these areas, which could include extra care for the elderly that lives in such a neighborhood.

Further analysis with the clusters generated with $l=0.05$ and socioeconomic variables can be performed to highlight specific aspects that might be important considering individual behavior or risk factors that may influence public policy success or suggest additional preventive actions.

There are several socioeconomic composite indexes [72], [73], [74] that may reveal specific conditions of the population. Thus, besides considering HDI, other aspects that can provide insights for policymakers have been included due to its correlation with faster spread of disease, severe symptoms of SARS-CoV-2 and lethality.

Fig. 9 shows clusters and colored regions for different socioeconomic indicators and variables related to SARS-CoV-2. The five visualizations enable the investigation of different characteristics that may drive specific public policies for a cluster or even part of a cluster.

Evaluating the clusters and neighborhoods according to HDI, it is possible to verify in Fig. 9(a) that there is one specific cluster, $C_1$, which shows the highest social differences located in the north area of Recife. This reflects the mobility pattern of this low-income population who work, in general, to provide services for this wealthy area of the city.

As shown in Fig. 9, Fig. 9, longevity is related with quality of life conditions. However, the elderly population can be an important aspect when analyzing FOI to minimize fatalities. Similarly, the number of persons per bedroom can bring insights and be an important aspect of policymakers. As shown in Fig. 9(c), there are specific neighborhoods associated with low HDI where several persons live in the same habitation, and therefore more than two persons share the same bedroom.

This is an important piece of information for policymakers as it can be impossible for such people to keep social distancing or even isolate the family member who develops COVID-19 symptoms. This means that if one family member is infected, all persons living in this house might be infected at the same time, increasing the disease contagion rate.

Concerning the infection-related variables, the absolute number of active cases can hide the severity of epidemic spread over neighborhoods with a lower population, as shown in Fig. 9, Fig. 9. For instance, although the neighborhood of Boa Viagem has the highest number of active cases in Fig. 9(d), it is one of the less critical neighborhoods according to Fig. 9(e) because this neighborhood has about 120 thousand inhabitants.

Fig. 10 presents three scatter plots concerning FOI against socioeconomic indicators. From Fig. 10(a), we can see that until the date of this analysis, the majority of neighborhoods that contain the above-average value of the indicator “more than two persons per bedroom” are concentrated in lower FOI values. This reveals that there should be preventive or educational actions to keep these values under control. Otherwise, there will be fast growth in infected cases that originated in these areas.

Paissandu is a neighborhood with one of the highest FOI and a percentage of the population with more than two persons per bedroom. Therefore public services should plan on how to support this population by aiding social distancing and isolation of infected persons since the virus can infect more people per domicile.

In Fig. 10(b), it can be observed that most of the neighborhoods with a low-educated population still have a low FOI. Therefore behavioral aspects should be reinforced to avoid epidemic spread amongst this population. The Totó neighborhood has the highest FOI amongst these neighborhoods with a low-educated population.

Fig. 10(c) shows that neighborhoods with a higher elderly population and higher FOI, such as Derby, are in dangerous situations. Similarly, public services can seek to offer specialized support for these neighborhoods to reduce the number of fatalities.

5.4. Expert feedback

We presented the developed system and case studies above to our project team, agents from the city administration, and researchers working on similar initiatives in other states of Brazil. All of them gave positive feedback concerning the usefulness and easiness of usage of the tool. In general, they were particularly pleased with the functionality of testing multiple isolation scenarios interactively by changing the mobility threshold.

Important feedback was about using socioeconomic variables such as the service level of water supply per neighborhood, homes with more than two persons per bedroom, HDI, and education. Depending on how these factors are associated with a neighborhood and the epidemic advance, there are possible actions to be deployed in terms of enhancing water supply and educational campaigns to stress the population about hygiene and protective measures. Furthermore, the ability to inspect different urban data sources in the context of the pandemic spread data was also appreciated.

Finally, the domain experts made many suggestions for extensions to the tool. For example, they wanted to see different spatial data aggregation levels to investigate in more depth the isolation strategy that could be used. While meeting with experts from the city administration’s mobility intelligence department, we had access to the mobility survey carried out by experts of the Secretariat for Urban Planning.

They were very interested in this tool and posed that it could be interesting to integrate our tool with their mobility survey dataset to perform a similar analysis. This survey was performed in 2018 with a sample of two hundred thousand questionnaires that include the modes of transportation used (bus, bike, private cars, etc.).

Thus, one direction for future work is to include this piece of information for statistical Bayesian models to enhance new case estimates according to different scenarios. Another constructive feedback was related to the integration of this tool with the mobility survey to optimize the schedule of opening hours in different economic activities, which shall be a new project to be developed after the COVID-19 outbreak.

Finally, one important concern raised by the experts is that the sub-notification of COVID-19 cases could very much influence the results due to low testing rates in the region. While this is out of the scope of our tool, we acknowledge that this is an important issue, and strategies to compensate for this problem could be used. We intend to investigate this in the future.