The UCAR meningitis effort: Results and next steps.

The UCAR meningitis effort: Results and next steps

Key Questions, relative to this meeting

• What have we learned from our research?• How can this new knowledge be integrated

into the outbreak response strategy? • How, generally, can the integration of research

into public health actions be facilitated?• What can public health needs tell us about

where to go next with research?

Paraphrased from meeting agenda

Our foci

Research that will improve the impact of reactive mass vaccination campaigns

1. Can can we predict the end of an epidemic with sufficient certainty that we avoid implementing a campaign that would have little impact?1

2. What criteria (epidemiological, population, climate, environment, vaccine, etc.) should be factored into a decision tree to decrease the time between epidemic onset and vaccination campaign while maintaining the same level of specificity?1

1. These questions come from the WHO Issue Paper, July 2011

Objectives and Strategies

1. Predict epidemic end by– Verifying Greenwood hypothesis linking season end and

humidity– Leveraging existing meteorological forecasts

2. Characterize risk factors by – Surveying 222 households for knowledge, attitudes and practices– Testing disease models against atmospheric, demographic, and

epidemiological data

3. Characterize economic impact by – Surveying 74 households for Cost of Illness

4. Inform reactive vaccination campaigns by– Developing a useful Decision Support System that includes archived

and real-time data and analysis tools

The Greenwood Hypothesis…

Greenwood, 1999

First Line of Evidence – Weather Simulations vs. Meningitis Cases

• Mera examined link between hindcasted weather variables and cases in regions with Meningitis outbreaks

• In April 2009, the Kano epidemic ended after relative humidity crossed above a 40% threshold

• Attack rates fell in D’jamena and Gaya when average relative humidity for the week rose above 40%.

Slide from Roberto Mera

Second line of evidence – Local Data

• Dukic analyzed monthly meningitis cases and environmental variables from 1998 to 2008 for Navrongo, Ghana– Data collected by A. Adams-Forgor, Navrongo Health

Research Centre• Found that RH showed strongest correlation with

number of cases– Couldn’t investigate lags because data were monthly– Dukic also found an inflection point in RH vs. cases

curves at RH of 40%

Pairwise correlations between meteorological and health variables January 1998 - December 2008 for months that reported

at least one case of meningitis.

Third line of evidence –Disease model tested against regional data

• Start with a differential equation disease model based on known transmission dynamics

• Use the model (trained with data) to determine the probability of a meningitis epidemic based only on a two-year, spatial incidence of epidemic across the region.

• Ask if you can better predict an epidemic if you use other variables (over 90 were tried, including RH, wind speed, temperature, …) in your modeling coefficients Data came from Clément Lingani, and covered

the entire belt for the 2008 and 2009 meningitis season.

Relative humidity improves prediction

• Hopson and Dukic found that knowing the RH two weeks ago improves accuracy in predicting an epidemic by ~25%1

• Coupled with a two week forecast, this indicates an improved ability to anticipate a roll-off in epidemic 4 weeks in advance

1It turns out other variables (air temp, winds, NE winds) also help, but less than relative humidity

Without RH

Using RH

Leveraging Forecasts

• Several meteorological centers produce global forecasts that can be used to estimate future humidity

• These forecasts need to be corrected, slightly– we can do this using past forecasts and

observations via “quantile regression”• The many forecasts themselves can also be

used to estimate the uncertainty in the prediction (ensembles)

Forecasting: Thorpex-Tigge “grand ensemble” -

Using ‘Quantile Regression’ can better calibrate ensembles to actual data

Without Quantile Regression:Observations outside range of ensembles

With Quantile Regression: Ensembles bracket observations

From Tom Hopson

Knowledge, Attitudes, and Practices Survey• A survey of 74 Households who have had meningitis and 158 age/gender/region matched

controls

• Part I: KAP– Knowledge of meningitis– Personal and household experience with meningitis– Customs and practices– Attitudes about diseases

• Part II: Socio-demographics– Education-literacy– Occupation (travel)– Housing (ventilation, sleeping arrangements)– Cooking, water, waste, etc. – Household assets; food security

• Part III: Cost of Illness– Costs of the case

• Medicine, transport to and from hospital, provision of meals, treatment– Costs in terms of missed work (either directly or for caregiver)– Costs due to sequelae– Limitations – recall bias in earlier cases

Preliminary Survey Results• In northern Ghana, a case of

meningitis costs, on average, more than a year’s wages

• Seasonal migration can protect from meningitis

• Early meningitis symptoms are often misidentified, delaying treatment – suggests an intervention

• Airborne matter (especially from local and regional burning) may increase risk

Using Generalized Additive Models

• GAM for meningitis developed by Dukic • Adjusts for time-varying confounding processes (e.g.

seasonal migration) that co-vary with the weather variables

• Analyzed 11 years of Navrongo data from Adams-Forgor

• Reaffirms importance of temperature and RH, and shows that amount of CO in the air due to burning biomass (fires).

Developing a Decision Support System

• Need to allow regional view of humidity to support regional decisions

• Use differential-equation based model to predict cases and end-of-season

• Allow users to look at district-level predictions• Incorporate current epidemiological data to

improve prediction– Can this start to create an archive of cases for

future research?

• DSS Walkthrough Slides go here, please….

View of West Africa looking from the northeast, with precipitable water in blue and meningitis reports by district -- no report (gray), alert (yellow), and epidemic (red).

Research Result Potential Public Health Action

Next steps or Hurdles

Humidity can be used to predict the end of an epidimic with 2-4 weeks lead time

Allocating vaccine to areas with persistent low relative humidity

Decision Support System for regional-scale decision makersHurdles include need for current epidemiological data

CO, Humidity, Wind, current cases can predict future cases by district

Earlier vaccination because of knowledge of trends in disease rate

Decision Support System for local scale. Same hurdle as above

Early cases of meningitis are often mistaken for less serious disease

Earlier self-diagnosis

Public Health Outreach

Migration and socioeconomic status influence disease transmission in identifiable, if surprising, ways

Targeted vaccination, education about travel

Public Health Outreach

Abudulai Adams-Forgor, Patricia Akweongo, Timothy Akwine, Dominic Anaseba, Anaïs Columbini, Maxwell Dalaba, Vanja Dukic, Mary Hayden, Abraham Hodgson, Thomas Hopson, Benjamin Lamptey, Jeff Lazo, Roberto Mera, Gertrude Nyaaba Nsormah, Raj Pandya, Jennie Rice, Fred Semazzi, Madeleine Thomson, Sylwia Trazka, Tom Warner, Christine Wiedinmyer Tom Yoksas

21

Funder

CollaboratingInstitutions

Collaborating Individuals

Department of Earth and Atmospheric Sciences, North Carolina State University, USAIRI, Columbia University, USANavrongo Health Research Centre, GhanaUniversity Corporation for Atmospheric Research, USARegional Maritime University, Ghana

Probability of exceeding risk threshold in any given year across belt

>15 cases per 105>10 cases per 105

>5 cases per 105>0 cases per 105

End-of-season date (risk less than “climatological” risk)

Day of Year

>15 cases per 105>10 cases per 105

>5 cases per 105>0 cases per 105

Yearly variation of End-of-Season dates across the belt

• Extra slides below

Probability of exceeding risk threshold in any given year across belt

>0 cases per 105 >5 cases per 105 >10 cases per 105

Probability