# Human Mobility, Networks and Disease Dynamics on a Global Scale

- 1 Citations
- 1.3k Downloads

## Abstract

Disease dynamics is a complex phenomenon and in order to address these questions expertises from many disciplines need to be integrated. One method that has become particularly important during the past few years is the development of computational models and computer simulations that help addressing these questions. In the focus of this chapter are emergent infectious diseases that bear the potential of spreading across the globe, exemplifying how connectivity in a globalized world has changed the way human-mediated processes evolve in the 21st century. The examples of most successful predictions of disease dynamics given in the chapter illustrate that just feeding better and faster computers with more and more data may not necessarily help understanding the relevant phenomena. It might rather be much more useful to change the conventional way of looking at the patterns and to assume a correspondingly modified viewpoint—as most impressively shown with the examples given in this chapter.

## 19.1 Introduction

In early 2009, news accumulated in major media outlets about a novel strain of influenza circulating in major cities in Mexico [1]. This novel H1N1 strain was quickly termed “swine flu”, in reference to its alleged origin in pig populations before jumping the species border to humans. Very quickly public health institutions were alerted and saw the risk of this local influenza epidemic becoming a major public health problem globally. The concerns were serious because this influenza strain was of the H1N1 subtype, the same virus family that caused one of the biggest pandemics in history, the Spanish flu that killed up to 40 million people in the beginning of the 20th century [2]. The swine flu epidemic did indeed develop into a pandemic, spreading across the globe in matters of months. Luckily, the strain turned out to be comparatively mild in terms of symptoms and as a health hazard. Nevertheless, the concept of emergent infectious diseases, novel diseases that may have dramatic public health, societal and economic consequences reached a new level of public awareness. Even Hollywood picked up the topic in a number of blockbuster movies in the following years [3]. Only a few years later, MERS hit the news, the Middle East Respiratory Syndrome, a new type of virus that infected people in the Middle East [4]. MERS was caused by a new species of corona virus of the same family of viruses that the 2003 SARS virus belonged to. And finally, the 2013 Ebola crisis in West African countries Liberia, Sierra Leone and Guinea that although it did not develop into a global crisis killed more than 10000 people in West Africa [5].

Population density is only one side of the coin. In addition to increasing face-to-face contacts within populations we also witness a change of global connectivity [9]. Most large cities are connected by means of an intricate, multi-scale web of transportation links, see Fig. 19.1. On a global scale worldwide air-transportation dominates this connectivity. Approx. 4,000 airports and 50,000 direct connections span the globe. More than three billion passengers travel on this network each year. Every day the passengers that travel this network accumulate a total of more than 14 billion kilometers, which is three times the radius of our solar system [10, 11]. Clearly this amount of global traffic shapes the way emergent infectious diseases can spread across the globe. One of the key challenges in epidemiology is preparing for eventual outbreaks and designing effective control measures. Evidence based control measures, however, require a good understanding of the fundamental features and characteristics of spreading behavior that all emergent infectious diseases share. In this context this means addressing questions such as: If there is an outbreak at location *X* when should one expect the first case at a distant location *Y*? How many cases should one expect there? Given a local outbreak, what is the risk that a case will be imported in some distant country. How does this risk change over time? Also, emergent infectious diseases often spread in a covert fashion during the onset of an epidemic. Only after a certain number of cases are reported, public health scientists, epidemiologist and other professionals are confronted with cases that are scattered across a map and it is difficult to determine the actual outbreak origin. Therefore, a key question is also: Where is the geographic epicenter of an ongoing epidemic?

Disease dynamics is a complex phenomenon and in order to address these questions expertises from many disciplines need to be integrated, such as epidemiolgy, spatial statistics, mobility and medical research in this context. One method that has become particularly important during the past few years is the development of computational models and computer simulations that help address these questions. These are often derived and developed using techniques from theoretical physics and more recently complex network science.

## 19.2 Modeling Disease Dynamics

*S*uceptible-

*I*nfected-

*R*ecovered” (SIR) model, a parsimoneous model for the description of a large class of infectious diseases that is also still in use today [13]. The SIR model considers a host population in which individuals can be susceptible (

*S*), infectious (

*I*) or recovered (

*R*). Susceptible individuals can aquire a disease and become infectious themselves and transmit the disease to other susceptible individuals. After an infectious period individuals recover, acquire immunity, and no longer infect others. The SIR model is an abstract model that reduces a real world situation to the basic dynamic ingredients that are believed to shape the time course of a typical epidemic. Structurally, the SIR model treats individuals in a population in much the same way as chemicals that react in a well-mixed container. Chemical reactions between reactants occur at rates that depend on what chemicals are involved. It is assumed that all individuals can be represented only by their infectious state and are otherwise identical. Each pair of individuals has the same likelihood of interacting. Schematically, the SIR model is described by the following reactions

*I*/

*N*of infected individuals. Because we have

*S*susceptibles the expected change of the number susceptibles due to infection is

*N*of individuals in a population. Depending on the magnitude of

*N*a model in which reactions occur randomly at rates \(\alpha \) and \(\beta \) a stochastic system generally exhibits solutions that fluctuate around the solutions to the deterministic system of Eq. (19.4).

Both, the deterministic SIR model and the more general particle kinetic stochastic model are designed to model disease dynamics in a single population, spatial dynamics or movement patterns of the host population are not accounted for. These systems are thus known as well-mixed systems in which the analogy is one of chemical reactants that are well-stirred in a chemical reaction container as mentioned above.

### 19.2.1 Spatial Models

When a spatial component is expected to be important in natural scenario, several methodological approaches exist to account for space. Essentially the inclusion of a spatial component is required when the host is mobile and can transport the state of infection from one location to another. The combination of local proliferation of an infection and the disperal of infected host individuals then yields a spread along the spatial dimension [13, 14].

*D*is the diffusion coefficient. The reasoning behind this approach is that the net flux of individuals of one type from one location to a neighboring location is proportional to the gradient or the difference in concentration of that type of individuals between neighboring locations. The key feature of diffusive dispersal is that it is local, in a discretized version the Laplacian permits movements only within a limited distance.

*S*,

*I*or

*R*and reactions occur only with nearest neighbors on the lattice. These models account for stochasticity and spatial extent. Given a state of the system, defined by the state of each lattice site, and a small time interval \(\Delta t\), infected sites can transmit the disease to neighboring sites that are susceptible with a probability rate \(\alpha \). Infected sites also recover to the

*R*state and become immune with probability \(\beta \Delta t\). Figure 19.3a illustrates the time course of the lattice-SIR model. Seeded with a localized patch of infected sites, the system exhibits an asymptotic concentric wave front that progresses at an overall constant speed if the ratio of transmission and recovery rate is sufficiently large. Without the stochastic effects that yield the irregular interface at the infection front, this system exhibits similar properties to the reaction diffusion system of Eq. (19.7). In both systems transmission of the disease in space is spatially restricted per unit time.

### 19.2.2 The Impact of Long-Distance Transmissions

*r*decreases as an inverse power-law as explained in the caption to Fig. 19.3. The possibility of transmitting to distant locations yields new epidemic seeds far away that subsequently turn into new outbreak waves and that in turn seed second, third, etc. generation outbreaks, even if the overall rate at which long-distance transmission occur is very small. The consequence of this is that the spatially coherent, concerntric pattern observed in the reaction diffusion system is lost, and a complex spatially incoherent, fractal pattern emerges [16, 17, 18]. Practically, this implies that the distance from an initial outbreak location can no longer be used as a measure for estimating or computing the time that it takes for an epidemic to arrive at a certain location. Also, given a snapshot of a spreading pattern, it is much more difficult to reconstruct the outbreak location from the geometry of the pattern alone, unlike in the concentric system where the outbreak location is typically near the center of mass of the pattern.

A visual inspection of the air-transportation system depicted in Fig. 19.1 is sufficiently convincing that the significant fraction of long-range connections in global mobility will not only increase the speed at which infectious diseases spread but, more importantly, also cause the patterns of spread to exhibit high spatial incoherence and complexity caused by the intricate connectivity of the air-transportation network. As a consequence we can no longer use geographic distance to an emergent epidemic epicenter as an indicator or measure of “how far away” that epicenter is and how long it will take to travel to a given location on the globe. This type of decorrelation is shown in Fig. 19.4 for two examples: The 2003 SARS epidemic and the 2009 influenza H1N1 pandemic. On a spatial resolution of countries, the figure depicts scatter plots of the epidemic arrival time as a function of geodesic (shortest distance on the surface of the Earth) distance from the initial outbreak location. As expected, the correlation between distance and arrival time is weak.

## 19.3 Modeling Disease Dynamics on a Global Scale

*m*to population

*n*in a given unit of time [19, 20]. For example \(N_{n}\) could correspond to the size of city

*n*and \(F_{nm}\) the amount of passengers the travel by air from

*m*to

*n*. One of earliest and most employed models for disease dynamics using the meta-population approach is a generalization of Eq. (19.4) in which each population’s dynamics is governed by the ordinary SIR model, e.g.

*n*is a parameter. In addition to this, the exchange of individuals between populations is modeled in such a way that hosts of each class move from location

*m*to location

*n*with a probability rate \(\omega _{nm}\) which yields

*n*from all other locations, the second term the flux in the opposite direction. Combining Eqs. (19.9) and (19.10) yields:

*m*to

*n*per unit time. However, based on very plausible assumptions [11], the system can be simplified in such a way that all parameters can be gauged against data that is readily available, e.g. the actual passenger flux \(F_{nm}\) (the amount of passengers that travel from

*m*to

*n*per day) that defines the air-transportation network, without having to specify the absolute population sizes \(N_{n}\).

*n*obeys

*m*with destination

*n*. Because passengers must arrive somewhere we have \(\sum _{n}P_{nm}=1\).

*n*with respect to the entire population size \(\mathscr {N}\). As expected the time course of a global epidemic in terms of the epicurve and duration depends substantially on the initial outbreak location.

A more important aspect is the spatiotemporal pattern generated by the model. Figure 19.6 depicts temporal snapshots of simulations initialized in London and Chicago, respectively. Analogous to the qualitative patterns observed in Fig. 19.3b, we see that the presence of long-range connections in the worldwide air-transportation network yields incoherent spatial patterns much unlike the regular, concentric wavefronts observed in systems without long-range mobility. Figure 19.7 shows that also the model epidemic depicts only a weak correlation between geographic distance to the outbreak location and arrival time. For a fixed geographic distance arrival times at different airports can vary substantially and thus the traditional geographic distance is useless as a predictor.

## 19.4 Issues with Computational Models

The system defined by Eq. (19.12) is one of the most parsimoneous models that accounts for strongly heterogeneous population distributions that are coupled by traffic flux between them and that can be gauged against actual population size distributions and traffic data. Surprisingly, despite its structural simplicity this type of model has been quite successful in accounting for actual spatial spreads of past epi- and pandemics [19]. Based on early models of this type and aided by the exponential increase of computational power, very sophisticated models have been developed that account for factors that are ignored by the deterministic metapopulation SIR model. In the most sophisticated approaches, e.g. GLEAM [21], the global epidemic and mobility computational tool, not only traffic by air but other means of transportation are considered, more complex infectious dynamics is considered and in hybrid dynamical systems stochastic effects caused by random reactions and mobility events are taken into account. Household structure, available hospital beds, seasonality have been incorporated as well as disease specific features, all in order to make predictions more and more precise. The philosophy of this type of research line heavily relies on the increasing advancement of both computational power as well as more accurate and pervasive data often collected in natural experiments and webbased techniques [21, 22, 23, 24, 25].

Despite the success of these quantitative approaches, this strategy bears a number of problems some of which are fundamental. First, with increasing computational methods it has become possible to implement extremely complex dynamical systems with decreasing effort and also without substantial knowledge of the dynamical properties that often nonlinear dynamical systems can possess. Implementing a lot of dynamical detail, it is difficult to identify which factors are essential for an observed phenomenon and which factors are marginal. Because of the complexity that is often incorporated even at the beginning of the design of a sophisticated model in combination with the lack of data modelers often have to make assumptions about the numerical values of parameters that are required for running a computer simulation [26]. Generically many dozens of unknown parameters exist for which plausible and often not evidence-based values have to be assumed. Because complex computational models, especially those that account for stochasticity, have to be run multiple times in order to make statistical assessments, systematic parameter scans are impossible even with the most sophisticated supercomputers.

Finally, all dynamical models, irrespective of their complexity, require two ingredients to be numerically integrated: (1) fixed values for parameters and (2) initial conditions. Although some computational models have been quite successful in describing and reproducing the spreading behavior of past epidemics and in situations where disease specific parameters and outbreak locations have been assessed, they are difficult to apply in situations when novel pathogens emerge. In these situations, when computational models from a practical point of view are needed most, little is known about these parameters and running even the most sophisticated models “in the dark” is problematic. The same is true for fixing the right initial conditions. In many cases, an emergent infectious disease initially spreads unnoticed and the public becomes aware of a new event after numerous cases occur in clusters at different locations. Reconstructing the correct initial condition often takes time, more time than is usually available for making accurate and valueable predictions that can be used by public health workers and policy makers to devise containment strategies.

## 19.5 Effective Distance

Given the issues discussed above one can ask if alternative approaches exist that can inform about the spread without having to rely on the most sophisticated highly detailed computer models. In this context one may ask whether the complexity of the observed patterns that are solutions to models like the SIR metapopulation model of Eq. (19.12) are genuinely complex because of the underlying complexity of the mobility network that intricately spans the globe, or whether a simple pattern is really underlying the dynamics that is masked by this complexity and our traditional ways of using conventional maps for displaying dynamical features and our traditional ways of thinking in terms of geographic distances.

*effective distance*derived from the topological structure of the global air-transportation network. In essence the idea is very simple: If two locations in the air-transportation network exchange a large number of passengers they should be effectively close because a larger number of passengers implies that the probability of an infectious disease to be transmitted from

*A*to

*B*is comparatively larger than if these two locations were coupled only by a small number of traveling passengers. Effective distance should therefore decrease with traffic flux. What is the appropriate mathematical relation and a plausible ansatz to relate traffic flux to effective distance? To answer this question one can go back to the metapopulation SIR model, i.e. Eq. (19.12). Dispersal in this equation is governed by the flux fraction \(P_{nm}\). Recall that this quantity is the fraction of all passengers that leave node

*m*and arrive at node

*n*. Therefore \(P_{nm}\) can be operationally defined as the probability of a randomly chosen passenger departing node

*m*arriving at node

*n*. If, in a thought experiment, we assume that the randomly selected person is infectious, \(P_{nm}\) is proportional to the probability of transmitting a disease from airport

*m*to airport

*n*. We can now make the following ansatz for the effective distance:

*m*arrives at

*n*and thus \(P_{nm}=1\) the effective distance is \(d_{nm}=d_{0}\) which is the smallest possible value. If, on the other hand \(P_{nm}\) becomes very small, \(d_{nm}\) becomes larger as required. The definition (19.13) applies to nodes

*m*and

*n*that are connected by a link in the network. What about pairs of nodes that are not directly connected but only by paths that require intermediate steps? Given two arbitrary nodes, an origin

*m*and a destination

*n*, an infinite amount of paths (sequence of steps) exist that connect the two nodes. We can define the shortest effective route as the one for which the accumulation of effective distances along the legs is minimal. So for any path we sum the effective distance along the legs according to Eq. (19.13) adding up to an effective distance \(D_{nm}\). This approach also explains the use of the logarithm in the definition of effective distance. Adding effective distances along a route implies the multiplication of the probabilities \(P_{nm}\) along the involved steps. Therefore the shortest effective distance \(D_{nm}\) is equivalent to the most probable path that connect origin and destination. The parameter \(d_{0}\) is a free parameter in the definition and quantifies the influence of the number of steps involved in a path. Typically it is chosen to be either 0 or 1 depending on the application.

*A*and

*B*. Let’s assume

*A*is a large hub that is strongly connected to many other airports in the network, including

*B*. Airport

*B*, however, is only a small airport with only as a single connection leading to

*A*. The effective distance \(B\rightarrow A\) is much smaller (equal to \(d_{0}\)) than the effective distance from the hub

*A*to the small airport

*B*. This accounts for the fact that if, again in a thought experiment, a randomly chosen passenger at airport

*B*is most definitely going to

*A*whereas a randomly chosen passenger at the hub

*A*is arriving at

*B*only with a small probability.

*m*thus has a set of shortest paths \(\mathscr {P}_{m}\) that connect

*m*to all other airports. This set forms the shortest path tree \(T_{m}\) of airport

*m*. Together with the effective distance matrix \(D_{nm}\) the tree defines the perspective of node

*m*. This is illustrated qualitatively in the Fig. 19.8 that depicts a planar random triangular weighted network.

## 19.6 Recovery of Concentric Patterns

The use of effective distance and representing the air-transportation network from the perspective of chosen reference nodes and making use of the more plausible notion of distance that better reflects how strongly different locations are coupled in a networked system is helpful for “looking at” the world. Yet, this representation is more than a mere intuitive and plausible spatial representation. What are the dynamic consequences of effective distance? The true advantage of effective distance is illustrated in Fig. 19.10. This figure depicts the identical computer-simulated hypothetical pandemic diseases as Fig. 19.6. Unlike the latter, that is based on the traditional geographic representation, Fig. 19.10 employs the effective distance and shortest path tree representation from the perspective of the outbreak location as discussed above. Using this method, the spatially incoherent patterns in the traditional representation are transformed into concentric spreading patterns, similar to those expected for simple reaction diffusion systems.

## 19.7 Reconstruction of Outbreaks

## 19.8 Conclusions

Emergent infectious diseases that bear the potential of spreading across the globe are an illustrative example of how connectivity in a globalized world has changed the way human mediated processes evolve in the 21st century. We are connected by complex networks of interaction, mobility being only one of them. With the onset of social media, the internet and mobile devices we share information that proliferates and spreads on information networks in much the same way (see also Chap. 20). In all of these systems the scientific challenge is understanding what topological and statistical features of the underlying network shape particular dynamic features observed in natural systems. The examples addressed above focus on a particular scale, defined by a single mobility network, the air-transportation network that is relevant for this scale. As more and more data accumulates, computational models developed in the future will be able to integrate mobility patterns at an individual resolution, potentially making use of pervasive data collected on mobile devices and paving the way towards predictive models that can account very accurately for observed contagion patterns. The examples above also illustrate that just feeding better and faster computers with more and more data may not necessarily help understanding the fundamental processes and properties of the processes that underly a specific dynamic phenomenon. Sometimes we only need to change the conventional and traditional ways of looking at patterns and adapt our viewpoint appropriately.

## References

- 1.M.P. Girard, J.S. Tam, O.M. Assossou, M.P. Kieny, Vaccine
**28**, 4895 (2010)CrossRefGoogle Scholar - 2.P.R. Saunders-Hastings, D. Krewski, Pathogens (Basel, Switzerland)
**5**, 66 (2016)Google Scholar - 3.Note 1, examples are
*Contagion*, a surprisingly accurate depiction of the consequences of a severe pandemic, and*Rise of the planet of the apes*that concludes with ficticious explanation for the extinction of mankind due to a man made virus in the futureGoogle Scholar - 4.M. Cotten, S.J. Watson, A.I. Zumla, H.Q. Makhdoom, A.L. Palser, S.H. Ong, A.A. Al Rabeeah, R.F. Alhakeem, A. Assiri, J.A. Al-Tawfiq, et al., mBio
**5**(2014)Google Scholar - 5.J.A. Backer, J. Wallinga, PLoS Comput. Biol.
**12**, e1005210 (2016)ADSCrossRefGoogle Scholar - 6.J. Diamond,
*Guns, Germs, and Steel: The Fates of Human Societies*(W. W. Norton & Company, 1999). ISBN 0-393-03891-2Google Scholar - 7.Current world population, http://www.worldometers.info/world-population/. Accessed 15 Jan 2017
- 8.United Nations (2014)Google Scholar
- 9.J.J. Lee, H. Seebens, B. Blasius, D. Brockmann, Eur. Phys. J. B
**84**, 589 (2011)ADSCrossRefGoogle Scholar - 10.A. Barrat, M. Barthélemy, R. Pastor-Satorras, A. Vespignani, Proc. Natl. Acad. Sci. USA
**101**, 3747 (2004)ADSCrossRefGoogle Scholar - 11.D. Brockmann, D. Helbing, Science (New York, N.Y.)
**342**, 1337 (2013)ADSCrossRefGoogle Scholar - 12.W.O. Kermack, A.G. McKendrick, Proc. R. Soc. Lond. A: Math. Phys. Eng. Sci.
**115**, 700 (1927), http://rspa.royalsocietypublishing.org/content/115/772/700.full.pdf, http://rspa.royalsocietypublishing.org/content/115/772/700. ISSN 0950-1207 - 13.R.M. Anderson, R.M. May,
*Infectious Diseases of Humans: Dynamics and Control*(Oxford University Press, Oxford, 1992)Google Scholar - 14.M.J. Keeling, P. Rohani,
*Modeling Infectious Diseases in Humans and Animals*(Princeton University Press, 2008)Google Scholar - 15.J.D. Murray,
*Mathematical Biology. II Spatial Models and Biomedical Applications Interdisciplinary Applied Mathematics*, vol. 18 (Springer New York Incorporated, 2001)Google Scholar - 16.D. Brockmann,
*Eur*(Phys. J. Spec, Top, 2008)Google Scholar - 17.D. Brockmann, L. Hufnagel,
*Phys*(Rev, Lett, 2007)Google Scholar - 18.D. Brockmann,
*Human Mobility and Spatial Disease Dynamics*(Wiley-VCH, 2009), pp. 1–24Google Scholar - 19.L. Hufnagel, D. Brockmann, T. Geisel, Proc. Natl. Acad. Sci. USA
**101**, 15124 (2004)ADSCrossRefGoogle Scholar - 20.O. Ovaskainen, I. Hanski, Am. Nat.
**164**, 364 (2004)CrossRefGoogle Scholar - 21.W. Van den Broeck, C. Gioannini, B. Gonçalves, M. Quaggiotto, V. Colizza, A. Vespignani, B.M.C. Infect, Dis.
**11**, 37 (2011)Google Scholar - 22.V. Colizza, A. Barrat, M. Barthelemy, A.-J. Valleron, A. Vespignani, PLoS Med.
**4**, e13 (2007)CrossRefGoogle Scholar - 23.D. Balcan, V. Colizza, B. Gonçalves, H. Hu, J.J. Ramasco, A. Vespignani, Proc. Natl. Acad. Sci. USA
**106**, 21484 (2009)ADSCrossRefGoogle Scholar - 24.M. Ajelli, B. Gonçalves, D. Balcan, V. Colizza, H. Hu, J.J. Ramasco, S. Merler, A. Vespignani, B.M.C. Infect, Dis.
**10**, 190 (2010)Google Scholar - 25.P. Bajardi, C. Poletto, J.J. Ramasco, M. Tizzoni, V. Colizza, A. Vespignani, PLoS One
**6**, e16591 (2011)ADSCrossRefGoogle Scholar - 26.R.M. May, Science (New York, N.Y.)
**303**, 790 (2004)ADSCrossRefGoogle Scholar - 27.
*Effective Distance and Shortest Paths in Global Mobility*, http://rocs.hu-berlin.de/D3/waneff/. Accessed 15 Jan 2017 - 28.J. Manitz, T. Kneib, M. Schlather, D. Helbing, D. Brockmann, PLoS Curr.
**6**(2014)Google Scholar