Are global epidemics predictable ? V. Colizza School of Informatics, Indiana University, USA M. Barthélemy School of Informatics, Indiana University, USA A. Barrat Universite Paris-Sud, France A. Vespignani School of Informatics, Indiana University, USA “Networks and Complex Systems” talk series
Epidemic spread: 14 th century Dec June 1348 Dec June 1349 Dec June 1350 Dec Dec Dec Black Death
Nov Mar Epidemic spread: nowadays SARS
SARS
Modeling of global epidemics Modeling of global epidemics Ravchev, Longini. Mathematical Biosciences (1985) multi-level description : intra-city epidemics inter-city travel
World-wide airport network complete 2002 IATA database V = 3880 airports E = weighted edges w ij #seats / year N j urban area population (UN census, …) V = 3100 airports E = weighted edges >99% of total traffic Barrat, Barthélemy, Pastor-Satorras, Vespignani. PNAS (2004)
World-wide airport network = 9.75 k max = 318 = w min = 4 w max = e +06 Frankfurt Sapporo - Tokyo
Broad distributions strong heterogeneities 3 different levels: degree weight population World-wide airport network summary
Epidemics: Stochastic Model compartmental model + air transportation data N1N1N1N1 N2N2N2N2 N0N0N0N0 N5N5N5N5 N4N4N4N4 N3N3N3N3 w 54 w 45 SIR model Susceptible Infected Recovered
Stochastic Model Travel term j lw jl Travel probability from j to l # passengers in class X from j to l multinomial distr.
Stochastic Model Travel term j lw jl Transport operator: other source of noise: two-legs travel: outgoing ingoing
Stochastic Model Intra-city S I R Independent Gaussian noises Homogeneous assumption rate of transmission -1 average infectious period
compartmental model + air transportation data SIR model Intra-cities Inter-cities Epidemics: Stochastic Model summary Does it work ?
Case study: SARS Susceptible Latent Infected Hospitalized R Hospitalized D Recovered Dead dd (1-d) DD RR Infected Hospitalized
Case study: SARS data: WHO reported cases final report: 28 infected countries 8095 infected cases 774 deaths refined compartmentalization parameter estimation: literature best fit initial condition: t=0 Feb. 21 st seed: Hong Kong I 0 =1, L 0 estimated, S 0 =N
Case study: SARS results
statistical properties epidemic pattern ? strong heterogeneity in no. infected cases: large fluctuations Full scale computational study of global epidemics: statistical properties epidemic pattern effect of complexity of transportation network forecast reliability SIR model
Results: Geographic spread Epidemics starting in Hong Kong
Results: Geographic spread Results: Geographic spread Epidemics starting in Hong Kong Gastner, Newman. PNAS (2004)
Results: Geographic spread Epidemics starting in Hong Kong t=24 dayst=48 dayst=56 days t=66 days t=160 days
maps heterogeneity epidemic spread appropriate measure role of specific structural properties: topology, traffic, population comparison with null hypothesis 1 st PART: Heterogeneity
Epidemic heterogeneity and Network structure HOM HOM P(k) P(N) P(w) WANWAN P(w) P(k) k P(N) HETk HETk P(k)P(w) w P(N) N HETw HETw
Epidemic heterogeneity Entropy: prevalence in city j at time t normalized prevalence H [0,1] H=0 most het. H=1 most hom.
Results: Epidemic heterogeneity global properties average over initial seed central zone: H>0.9 HETk WAN importance of P(k)
Results: Epidemic heterogeneity epidemics starting from a given city average entropy profile + maximal dispersion noise: small effect
Results: Epidemic heterogeneity epidemics starting from a given city percentage of infected cities
2 nd PART: Predictability t=24 days t=48 days t=56 days t=66 days t=160 days time One outbreak realization: Another outbreak realization ? t=24 days t=48 days t=56 days t=66 days t=160 days ? ???? epidemic forecast containment strategies
Predictability normalized probability Similarity between 2 outbreak realizations: Hellinger affinity Overlap function
Predictability 2 identical outbreaks 2 distinct outbreaks time t
Results: Predictability left: seed = airport hubs right: seed = poorly connected airports HOM & HETw high overlap HETk low overlap WAN increased overlap !!
Results: Predictability j lw jl j l HOM: few channels high overlap HETk: broad P(k) lots of channels! low overlap WAN: broad P(k),P(w) lots of channels! emergence of preferred channels increased overlap !!! + degree heterog. + weight heterog.
Conclusions air transportation network properties global pattern of emerging disease spatio-temporal heterogeneity of epidemic pattern quantitative measurement of the predictability of epidemic pattern epidemic forecast, risk analysis of containment strategies Ref.: qbio/