Modeling the Ebola Outbreak in West Africa, 2014

August 11th UpdateBryan Lewis PhD, MPH (blewis@vbi.vt.edu)

Caitlin Rivers MPH, Stephen Eubank PhD, Madhav Marathe PhD, and Chris Barrett PhD

Goals

• Estimate future cases in Africa• Offer any guidance on potential for

transmission in the United States• Explore impact of various countermeasures

Data Sources

• Using case counts from WHO for Model Fitting– Lots of variability from different sources, generally similar– Challenging to estimate what proportion of infections are

captured

• Liberia’s Ministry of Health for Model Selection and geographic resolution

Currently Used WHO DataCases Deaths

Guinea 495 363

Liberia 516 282

Sierra Leone 691 286

Nigeria 13 2

Total 1779 961

● Data reported by WHO on Aug 8 for cases as of Aug 6

● Sierra Leone case counts censored up to 4/30/14.

● Time series was filled in with missing dates, and case counts were interpolated.

Measure of Awareness?

Aug 8Jul 29

Compartmental Model

• Extension of model proposed by Legrand et al.Legrand, J, R F Grais, P Y Boelle, A J Valleron, and A Flahault. “Understanding the Dynamics of Ebola Epidemics” Epidemiology and Infection 135 (4). 2007. Cambridge University Press: 610–21. doi:10.1017/S0950268806007217.

Legrand et al. Model Description

Optimized Fit Process• Parameters to explored selected– Diag_rate, beta_I, beta_H, beta_F, gamma_I, gamma_D,

gamma_F, gamma_H– Initial values based on two historical outbreak

• Optimization routine– Runs model with various

permutations of parameters– Output compared to observed case

count– Algorithm chooses combinations that

minimize the difference between observed case counts and model outputs, selects “best” one

Fitted Model Caveats

• Assumptions:– Behavioral changes effect each transmission route

similarly– Mixing occurs differently for each of the three

compartments but uniformly within• These models are likely “overfitted”– Many combos of parameters will fit the same curve– Guided by knowledge of the outbreak and additional

data sources to keep parameters plausible– Structure of the model is supported

Liberia Fitted Models

Assuming no impact from ongoing responsesand DRC parameter fit is correct:

142 cases in next week182 cases in the following week

Assuming no impact from ongoing responsesand Uganda parameter fit is correct:

178 cases in next week235 cases in the following week

Liberia Fitted ModelsSources of Infections

Currently 14% of Liberian Infections among HCWSupports use of “Uganda” parameter set

Liberia Forecasts over time

1. Model trained on Liberian data, using “Uganda” parameters up to specified date

2. Model projected past “trained to” date

3. Complete case count data provided for reference

Sierra Leone Fitted Models

Assuming no impact from ongoing responsesand DRC parameter fit is correct:

208 cases in next week267 cases in the following week

Assuming no impact from ongoing responsesand Uganda parameter fit is correct:

211 cases in next week273 cases in the following week

Sierra Leone Forecasts over time

Model trained on Sierra Leone data up to specified date, projected into future, Complete case count data provided for reference

Explore Intervention Requirements

Vaccination of large swaths of population required to reduce txm, unless a targeted strategy is used

Explore Intervention Requirements

This does not capture reduction in deaths, but shows nominal interruption of transmission

Notional US estimates

• Under assumption that Ebola case, arrives and doesn’t seek care and avoids detection throughout illness

• CNIMS based simulations– Agent-based models of populations with realistic social

networks, built up from high resolution census, activity, and location data

• Assume:– Reduced transmission Ebola 70% less likely to infect in

home and 95% less likely to infect outside of home than respiratory illness

– Transmission calibrated to R0 of 3.5 if transmission is like flu

Notional US estimates Approach

• Get disease parameters from fitted model in West Africa

• Put into CNIMS platform– ISIS simulation GUI– Modify to represent US

• Example Experiment:– 100 replicates – One case introduction into Washington DC– Simulate for 3 weeks

Notional US estimates Example

100 replicatesMean of 1.8 casesMax of 6 casesMajority only one initial case

Conclusions

• Still need more information (though more is becoming available) to remove uncertainty in estimates

• From available data and in the absence of significant mitigation outbreak in Africa looks to continue to produce significant numbers of cases in the coming weeks

• Under current assumptions, Ebola transmission hard to interrupt in Africa with “therapeutics” alone

• Expert opinion and preliminary simulations support limited spread in US context

Next Steps

• Gather further data from news media and reports to support model parameter selection

• Build patch model framework to incorporate more geographic location information

• Build more detailed population of area to support agent based simulations

ADDITIONAL SLIDES FOR MORE DETAILS

Liberia Fitted Models

Model Parameters

No behavioral Changes included

Liberia Disease Parameters for Model Fitting UgandaOut Uganda_in DRCOut DRC_in

beta_F 0.858 1.093 0.081 0.066beta_H 0.091 0.113 0.003 0.002beta_I 0.123 0.084 0.204 0.505dx 0.585 0.650 0.867 0.670gamma_I 0.050 0.100 0.079 0.100gamma_d 0.084 0.125 0.050 0.104gamma_f 0.665 0.500 0.512 0.500gamma_h 0.335 0.238 0.153 0.200Score 62370 NA 103596 NA

Sierra Leone Fitted Models

Model Parameters

No behavioral Changes included

Sierra LeoneDisease Parameters for Model Fitting UgandaOut Uganda_in DRCOut DRC_in

beta_F 1.752 1.093 0.045 0.066beta_H 0.260 0.113 0.001 0.002beta_I 0.083 0.084 0.296 0.505dx 0.323 0.650 0.300 0.670gamma_I 0.247 0.100 0.149 0.100gamma_d 0.211 0.125 0.159 0.104gamma_f 0.330 0.500 0.814 0.500gamma_h 0.247 0.238 0.333 0.200Score 140931 NA 114419 NA

Legrand et al. Approach

• Behavioral changes to reduce transmissibilities at specified days

• Stochastic implementation fit to two historical outbreaks – Kikwit, DRC, 1995 – Gulu, Uganda, 2000

• Finds two different “types” of outbreaks– Community vs. Funeral driven

outbreaks

Parameters of two historical outbreaks

NDSSL Extensions to Legrand Model

• Multiple stages of behavioral change possible during this prolonged outbreak

• Optimization of fit through automated method

• Experiment:– Explore “degree” of fit using the two different

outbreak types for each country in current outbreak