Infectious diseases are a major burden to global health. Understanding their mechanisms and being able to predict and intervene epidemic outbreaks is an important challenge for researchers and decision makers alike. It should not be too hard either―if we include human contact patterns, the mechanisms of contagion and the typical features of the disease, we could model most infectious-disease related phenomena. Of these three components, the network epidemiology of the last decade has shown that our limited understanding of human contact patterns is probably the most important focus are for advancing infectious disease epidemiology. We will discuss what is known about human contact patterns and how to include this knowledge in epidemic modeling. First, we discuss recent work on what the epidemiologically most important temporal structures of human contacts are. We use about 80 empirical temporal network datasets, several arguably important for disease spreading, and scan the entire parameter space of disease-spreading models. By comparing to null-models, we identify important, simple temporal patterns that affect disease spreading stronger than the bursty interevent time distributions. Furthermore, we investigate how to eliminate the temporal information to make an as relevant static network as possible. After all, static network epidemiology has more methods and results than temporal network epidemiology and it for some purposes it is necessary. We find that an “exponential threshold” representation almost always the best performance, but time-sliced network (with a carefully chosen window, usually considerably different than the sampling time of the data) works almost as good. In contrast, networks of concurrent contacts do not seem to carry so important information.
6. Sociopatterns gallery
P H Y S I C A L P R O X I M I T Y
Prostitution
Sociopatterns conference
Hospital system
N = 16,730, L = 50,632, T = 6.0y
N = 113, L = 20,818, T = 59h
N = 159(8), L = 6,027(350), T = 7.3(1)h
N = 293,878, L = 64,625,283, T = 3,570dReality mining
N = 63, L = 26,260, T = 8.6h
7. ELECTRONIC COMMUNICATION
N = 57,189, L = 444,162, T = 112.0d
Bornholdt’s e-mail
Eckmann’s e-mail
N = 3,188, L = 115,684, T = 81.6d
Filmtipset forum
N = 7,084, L = 1,412,401, T = 8.61y
Filmtipset messages
Pussokram dating
N = 28,972, L = 529,890, T = 512.0d
QX dating
N = 80,683, L = 4,337,203, T = 63.7d
N = 35,624, L = 472,496, T = 8.27y
Facebook wall posts
N = 293,878, L = 876,993, T = 1591d
12. GOOD REPRESENTATION:
RANKING OF IMPORTANT
VERTICES CONSERVED
FOR ALL PARAMETER VALUES:
MEASURE AVG OUTBREAK SIZE
WHEN SPREADING STARTS AT i
FOR ALL PARAMETER VALUES:
MEASURE DEGREE OF i
FOR ALL PARAMETER VALUES:
MEASURE CORENESS OF i
degree 4
coreness 0
coreness 2
coreness 3
coreness 4
static importance
optimal params.
dynamic
importance
Spearman
rank correlation
coefficent
=
Quality of
representation
20. STEP 1 Assign stubs to vertices from a
random number distribution.
1
2
3
4
5
6
21. STEP 2 Connect random pairs of stubs
to form a simple graph.
1
2
3
4
5
6
22. STEP 3 Create active intervals for each
edge.
(1,2)
(2,3)
(2,4)
(2,5)
(3,4)
(3,5)
(4,5)
(5,6)
time
23. STEP 4 Create a time series of contacts
from some interevent-time
distribution.
time
24. STEP 5 Split the time series into
segments proportional to the
intervals and impose the
contacts of the segments to the
intervals.
(1,2)
(2,3)
(2,4)
(2,5)
(3,4)
(3,5)
(4,5)
(5,6)
time
25. STEP 6 Forget the active intervals.
(1,2)
(2,3)
(2,4)
(2,5)
(3,4)
(3,5)
(4,5)
(5,6)
time
36. Compensate for the size bias on intervals because of finite
T0
t’
t
sampling time (t’ would only be recorded if it starts within [0,T–t’])
37. Compensate for the size bias on intervals because of finite
T0
t’
t
sampling time (t’ would only be recorded if it starts within [0,T–t’])
Compensate for the chance an interevent time t is active
0
t
at the start of the sampling is proportional to t
38. Compensate for the size bias on intervals because of finite
T0
t’
t
sampling time (t’ would only be recorded if it starts within [0,T–t’])
Compensate for the chance an interevent time t is active
0
t
at the start of the sampling is proportional to t
ti
T–tii: ti≥t
∑ /
ti
T–tii
∑
Sum up and normalize
39. 0
0.2
0.4
0.6
0.8
1
0 500 1000 1500 2000
time t (days)
predicted from
interevent times
end times
beginning times
PROSTITUTION
P(t)B
45. 0
0.1
0.2
0.3
0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
fractionofinfectives
O r i g i n a l d a t a S I R
46. 0
0.1
0.2
0.3
0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
fractionofinfectives
I n t e r e v e n t t i m e s S I R
47. 0
0.1
0.2
0.3
0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
fractionofinfectives
B e g i n n i n g t i m e s S I R
48. 0
0.1
0.2
0.3
0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
fractionofinfectives
E n d t i m e s S I R
50. 0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
O r i g i n a l d a t a
0
0.2
0.3
0.4
averagenumberofinfections
S I S
0.1
51. 0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
I n t e r e v e n t t i m e s
0
0.2
0.3
0.4
averagenumberofinfections
S I S
0.1
52. 0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
B e g i n n i n g t i m e s
0
0.2
0.3
0.4
averagenumberofinfections
S I S
0.1
53. 0
0.2
0.3
0.4
0.1 0.2 0.90.8 10.70.60.50.40.3
0.1
1
0.01
0.001
per-contact transmission probability
durationofinfectivestage
averagenumberofinfections
E n d t i m e s S I S
0.1
56. Science by: Illustrations by:
Petter Holme Fredrik Liljeros Mi Jin Lee
P Holme, 2013, PLoS Comp. Biol. 9:e1003142.
P Holme, F Liljeros, 2013, arxiv:1307.6436.