92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Wednesday, 25 January 2012: 5:15 PM
An Application of a Newly Developed Machine Learning Technique for Predicting Climate-Meningitis Seasonal Outlook Over West Africa
Room 242 (New Orleans Convention Center )
Isaac K. Tetteh, North Carolina State Univ., Raleigh, NC; and D. L. Gonzalez II, Z. Chen, N. F. Samatova, and F. H. M. Semazzi

Our motivation is primarily based on the important role surface relative humidity (sfc RH) plays in meningitis epidemics over tropical West Africa. The disease, caused by a gram-negative, diplococcal Neisseria meningitidis, is associated with several environmental factors such as dusty Harmattan winds and low humidity. It is a highly pervasive, devastating, and debilitating disease responsible for up to about 250,000 deaths annually (with case fatality rate of 10-50%) in the sub-Saharan African “Meningitis Belt”. It imparts several degrees of disabilities that include, but not limited to, auditory impairment, cognitive impairment, necrosis, septicemia and paralysis, to surviving victims. Due to limited vaccine supplies and logistical constraints, the current intervention strategy relies on sfc RH information, on intraseasonal time scale, to channel the vaccines and logistics to the most vulnerable communities. However, on seasonal to interannual time scale, sfc RH variability and predictability associated with the disease, and in particular the roles of the tropical Atlantic and other large-scale teleconnection patterns actively engaged in the modulation have not been fully explored. In this study, we have investigated some of these phenomena. Based on a priori selection of global (sea surface temperature: SST; U and V winds) and regional (precipitable water: PW) scale predictors, we successfully applied our recently developed hierarchical classification and supervised feature selection algorithms, with the overall goal of selecting multivariate features and locations of the predictors anticipated to optimize the prediction of sfc RH. Following this, we implemented a least absolute deviation (LAD)-based ensemble agglomeration with varying ensemble size, to generate a climate-meningitis outlook for West Africa, in one year ahead of time for 2009 Jan-Feb-Mar (JFM) meningitis season. The results demonstrate that the top three cross-validation (CV) correlation skills for sfc RH (i.e., meningitis-proxy), based on leave-one-out CV (LOOCV), ranged between 0.85 and 0.89, at 5-9 months lead times. This study has, therefore, the potential of contributing to the development of a well-orchestrated risk management strategy, to complement the efforts done at intraseasonal time scales, in saving lives.

Supplementary URL: