1 of 47

Joint work with Lars Lorch, William Trouleau, Efstratios Tsirtsis, Heiner Kremer, Aron Szanto and Bernhard Schölkopf

Manuel Gomez Rodriguez

A Spatiotemporal Epidemic Model �to Quantify the Effects of �Testing, Contact Tracing and Containment

2 of 47

Contact tracing technology for Covid-19

2

Location-based technologies�(e.g., GPS, Wifi triangulation, LTE, QR)

Proximity-based technologies�(e.g., bluetooth)

Use smartphones to track �who you have met in the �recent past

If a person is tested �positive, notify contacts

3 of 47

The promise of contact tracing technologies

3

Predict the spread of COVID-19 at an unprecedented spatiotemporal resolution

When and where new individual infections may happen

4 of 47

The promise of contact tracing technologies

4

Design more effective containment & mitigation strategies

Predict the spread of COVID-19 at an unprecedented spatiotemporal resolution

When and where new individual infections may happen

This would allow authorities to lift the most restrictive measures

5 of 47

The promise of contact tracing technologies

5

Design more effective containment & mitigation strategies

Predict the spread of COVID-19 at an unprecedented spatiotemporal resolution

When and where new individual infections may happen

This would allow authorities to lift the most restrictive measures

Data-driven insights into disease parameters

Relative importance of different modalities of disease transmission

6 of 47

The promise of contact tracing technologies

6

Design more effective containment & mitigation strategies

Predict the spread of COVID-19 at an unprecedented spatiotemporal resolution

When and where new individual infections may happen

This would allow authorities to lift the most restrictive measures

Data-driven insights into disease parameters

Relative importance of different modalities of disease transmission

For this to happen, we need �data-driven models designed to make use of individual contact tracing data

7 of 47

Most models: country or state level predictions

7

https://arxiv.org/pdf/2004.01105.pdf

Li et al., Science 2020

Most models make country or state level predictions & they �do not characterize the state of each individual over time

8 of 47

What about agent-based epidemiological models?

8

Agent-based models proceed in �sequential discrete epochs

Epoch 1

Epoch 2

Epoch 3

9 of 47

What about agent-based epidemiological models?

9

Agent-based models proceed in �sequential discrete epochs

How long is each epoch?

How do we faithfully go beyond memoryless distributions?

How do we aggregate events within a epoch?

Epoch 1

Epoch 2

Epoch 3

However, epidemiological state, �mobility, testing, tracing & containment measures spans �wide range of temporal scales:

10 of 47

What about agent-based epidemiological models?

10

Agent-based models proceed in �sequential discrete epochs

How long is each epoch?

How do we faithfully go beyond memoryless distributions?

How do we aggregate events within a epoch?

Epoch 1

Epoch 2

Epoch 3

However, epidemiological state, �mobility, testing, tracing & containment measures spans �wide range of temporal scales:

In addition, they cannot characterize individual mobility & contacts patterns within specific towns or cities

11 of 47

Our spatiotemporal epidemic model

11

Our model uses marked temporal point processes (TPP) to jointly represent…

Epidemiological state[contact-based variation of SEIR]

Mobility pattern[Check-in’s or encounters]

Testing�[A variety of testing policies]

Social distancing & business restrictions

[per demographic groups & business type, dynamic]

Each �individual

Health authority & government

Tracing�[A variety of testing policies]

Want to know more? http://learning.mpi-sws.org/tpp-icml18/

12 of 47

How does the TPP representation work?

12

Temporally sorted �queue of events

Social distancing �business restrictions

[thinning]

[superposition principle]

Epidemiological & �mobility events

Testing & contact tracing events

13 of 47

Individual mobility patterns

13

Bob

Time

Time

Time

Time

14 of 47

Individual mobility pattern

14

Bob

Time

Time

Time

Time

Intensity (or rate) at which Bob goes to a supermarket

Duration distribution for supermarket visits

15 of 47

Epidemiological state of each individual

15

Time

Susceptible

Exposed

Asymptomatic

Resistant

Presymptomatic

Symptomatic

Hospitalized

Deceased

[1] Lauer et al., Annals of Internal medicine, 2020

[1]

[2] He et al., Nature Medicine, 2020

[2]

[1]

[2, 3, 4]

[2, 3, 4]

[3] Woelfel et al., medRxiv, 2020

[4] WHO, 2020

[5] Wang et al., JAMA, 2020

[6] Linton et al., JCM, 2020

[5]

[6]

[6]

Time to event distributions:

Exposure intensity

[2, 3, 4]

16 of 47

Exposure intensity depends on mobility patterns

16

Christine

Time

Time

Time

Time

17 of 47

Exposure intensity depends on mobility patterns

17

Christine

Time

Time

Time

Time

18 of 47

Exposure intensity depends on mobility patterns

18

overlap

Christine

Time

Time

Time

Time

Environmental �transmission

19 of 47

Exposure intensity depends on mobility patterns

19

overlap

Christine

Time

Time

Time

Time

Environmental �transmission

20 of 47

Exposure intensity depends on mobility patterns

20

overlap

Christine

overlap

Environmental �transmission

Time

Time

Time

Time

21 of 47

Exposure intensity depends on mobility patterns

21

overlap

Christine

Environmental �transmission

overlap

Time

Time

Time

Time

Environmental �transmission

22 of 47

Social distancing & business restrictions

22

Before social distancing & business restrictions:

3 times a week

twice a week

once a week

Time

Time

Time

23 of 47

Social distancing & business restrictions

23

During social distancing & business restrictions:

3 times a week

twice a week

once a week

once a week

university is closed

once every two weeks

Closed

Time

Time

Time

24 of 47

Social distancing & business restrictions

24

Time

Time

Time

During social distancing & business restrictions:

3 times a week

twice a week

once a week

once a week

university is closed

once every two weeks

Closed

Given historical mobility/contact patterns, we can use thinning to simulate different social �distancing measures & business restrictions

25 of 47

Testing

25

Individuals are added to a testing queue according to a testing policy (e.g., add all symptomatic individuals)

Time

Time

Queue additions are �event-driven

n individual from the queue �are tested

time period T

Each individual receives test outcome after 𝝙 units of time

testing capacity

reporting delay

26 of 47

Contact tracing for testing and/or isolation

26

Identify contacts the last N days before testing positive

Bob

27 of 47

Contact tracing for testing and/or isolation

27

Pick the M contacts with the highest empirical probability �of infection due to Bob

Identify contacts the last N days before testing positive

Bob

28 of 47

Contact tracing for testing and/or isolation

28

Pick the M contacts with the highest empirical probability �of infection due to Bob

Those contacts who overlap the longest with Bob in locations k with high

Identify contacts the last N days before testing positive

Bob

t - N

t

Christine

Bob

overlap

overlap

Environmental �transmission

Environmental �transmission

29 of 47

Contact tracing for testing and/or isolation

29

Pick the M contacts with the highest empirical probability �of infection due to Bob

Those contacts who overlap the longest with Bob in locations k with high

Test them and/or isolate them for T days

Identify contacts the last N days before testing positive

Bob

t - N

t

Christine

Bob

overlap

overlap

Environmental �transmission

Environmental �transmission

30 of 47

Proximity-based vs. spatial-based technology

30

overlap

Environmental �transmission

Time

To identify overlaps, we only �need to record

To identify environmental transmissions, �we need to record

Christine

Bob

Proximity �Spatial︎

Proximity* �Spatial

*Proximity with beacons

31 of 47

31

How did we test our model?

Case study using data from

Switzerland

&

Germany

32 of 47

Publicly available data sources: mobility model

32

High resolution �population density maps

[tile size: 600m x 600m]

Open street maps (OSM) data

[bars/restaurants, schools/unis, offices, bus stops, supermarkets]

+

assumptions

Individuals’ mobility patterns

[sites check-in’s]

Census data

[age groups, household size]

+

33 of 47

Publicly available data sources: epidemiology

33

New daily positive cases and fatalities per age group for each…

Landkreis �( )

Canton�( )

34 of 47

Publicly available data sources: testing

34

Maximum testing capacity per day for both Germany and Switzerland, as reported by authorities

Rescaled according to �the population of the

town, city, landkreis or canton

35 of 47

Model calibration using Bayesian Optimization

35

+

Individuals’ �mobility patterns

[sites check-in’s]

Bayesian �optimization�[BoTorch]

Testing settings

[testing capacity, delay,�testing policy]

Model simulator

True and simulated daily positive cases �per age group

Transmission rates per site type

36 of 47

36

What kind of questions can our model answer in Tübingen (Germany)?

37 of 47

What would have happened without “lockdown”?

37

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

Model simulation with �social distancing & business closures from “lockdown”

Model simulation without �social distancing & business closures from “lockdown”

38 of 47

What would have happened without “lockdown”?

38

Model simulation with �social distancing & business closures from “lockdown”

Model simulation without �social distancing & business closures from “lockdown”

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

39 of 47

When should we lift the ”lockdown”?

39

It resurges and there is �a second wave

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

40 of 47

What about alternating curfews for K random subgroups?

40

Monday morning

Monday afternoon

Tuesday morning

Group 1

Group 2

Rationale: alternating curfews may break chain of infections

41 of 47

What about alternating curfews for K random subgroups?

41

Enough to keep the �epidemic under control

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

42 of 47

Contact tracing for highly granular social distancing

42

Flattens the curve & decrease �the peak significantly

N = 3 days, M = 25 contacts, T = 7 days

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

43 of 47

Contact tracing requires high compliance

43

It gradually �worsens performance

N = 3 days, M = 25 contacts, T = 7 days

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

44 of 47

Contact tracing for narrowcasting

44

Probability that a person who visited this location in a given time period is exposed

The ranking of sites in terms of exposure probability does not change in expectation when adjusting for adoption levels across sites

It does not suffer from the x2 adoption problem

It requires location-based contact tracing or proximity-based with (bluetooth) beacons

*Population/sites are downsampled by 10x, results in the (updated) preprint will be without downsampling

45 of 47

45

Open source implementation: https://github.com/covid19-model

It includes a collection of auxiliary functions and notebooks that can be used to mimic our experimental setup from publicly available data at any desired location

Open source implementation

46 of 47

46

Open source implementation: https://github.com/covid19-model

It includes a collection of auxiliary functions and notebooks that can be used to mimic our experimental setup from publicly available data at any desired location

Open source implementation

Already used by a growing set of researchers, e.g.,

Teams from Stanford, Uni Bonn, Uni Vienna, Facebook, Google...

We are happy to help anybody get started [Slack channel]

47 of 47

47

Arxiv preprint [to be updated later this week]:

https://arxiv.org/abs/2004.07641

Details on the model and results