Presented by Jodi K. Haponski (GSSP Summer Program) Mentors Radina P. Soebiyanto (USRA/NASA) Richard K. Kiang(NASA)
Most common disease Annually Infects 5-15% of global population Annually in the U.S: ▪ Up to 200,000 hospitalizations, at least 30,000 deaths ▪ Estimated economic burden is ~87.1 billion TypeLocationYearDeaths H1N1Worldwide million H2N2 (ie. Asian)U.S ,000 H3N2 (ie. Hong Kong)U.S ,000 SARS37 Countries200210% fatality
Many new influenza strain first appear in tropical regions Influenza spread varies with latitude Seasonal in temperate climate Year-round outbreaks in tropical climates Influenza Process FactorsRelationship Virus Survivorship TemperatureInverse HumidityInverse Solar irradiance Inverse Transmission Efficiency TemperatureInverse HumidityInverse Vapor pressure Inverse RainfallProportional ENSOProportional Air travels and holidays Proportional Host susceptibility SunlightInverse NutritionVaries
Study area: Hong Kong Sub-tropical climate ▪ Comfortable temperatures in the winter months ▪ Temperature low of 6 C (~43 F) ▪ Summers are hot and humid with occasional showers and thunderstorms ▪ Temperatures often exceed 31 C (~88 F) DATA All reported flu cases in Hong Kong Obtained weekly data from Satellite Derived Data Land Surface Temperature (LST) from MODIS data set Precipitation from TRMM 3B42 GOALS Determine a relationship between environmental factors and number of disease cases Model and forecast with the above results
1. Hilbert Huang Transform, using EEMD (Ensemble Empirical Mode Decomposition) 2. Stepwise Fit Includes only significant variables in model 3. Remove dependent variables in model Original Data: Nonlinear & Nonstationary Time Series HHT: Series of Decomposed Signals. Uses empirical mode decomposition (EMD) to finitely decompose the original into a well-defined Hilbert transform. Decomposed Signal: Linear and Stationary
Original Signal Decomposed Signals
First let all decomposed environmental signals make up the model Fit each signal with the flu signal (using univariate regression) Remove any signal with a p-value greater than.05
Step Variables with Corresponding Decompositions in Model InitialAll Univariate regression on each decomposed environmental signal (fitted with the original flu signal) Those with p-values less than.05 Step-wise Fit (multivariate regression) Signal is removed from model if it’s p-value is greater than.1 Signal added back into the model if it’s p-value is less than.05 Eliminate Dependency: Compute correlation coefficients between each decomposed signal. Labeled dependent if correlation is greater than.5 Signal with the smaller p-value (from univariate regression)
Decomposed signals of each variable were fitted with the original pos flu signal Selected the variable and corresponding decomposed signal with p-value greater than.05
ENV. VAR.ENV. SIGNALCORR. COEF.P-VALUE RAD EVAP DPMEAN RHMAX RHMEAN RHMIN TRMM EVAP RHMIN TRMM SUN RAD TRMM RHMAX E-06 RHMEAN E-05 RHMIN E-07 TRMM E-05 CLOUD RAD RHMAX RHMEAN RHMIN WSPDMEAN
2. Step-wise Fit: ‘In’ Variables ‘In’ Signals ENV VARENV SIGNAL TRMM4 RAD4 TRMM5 RHMAX5 RHMIN5 CLOUD6 WSPDMEAN6 Performed multivariate regression Signal is removed from model if it’s p- value is greater than.1 Signal added back into the model if it’s p-value is less than.05
The correlation coefficients were computed between each signal decomposition Two signals are labeled dependent if their correlation is greater than.5 Signals were eliminated based on p-values Dependent Signals: ENV VARENV SIGNAL RHMAX5 RHMIN5 CLOUD6 WSPDMEAN6
Resulting Model: First Figure: Second Figure: TRMM, 5 th signal removed from model Independent Signals ENV VARENV SIGNAL TRMM4 RAD4 TRMM5 RHMIN5 WSPDMEAN6
The EEMD method was able to give insight into the seasonal relationship between the influenza dynamics with the environmental factors With only two years of training data, we were able to obtain relatively good prediction results