Data mining techniques for the estimation of variables in health-related noisy data

Abstract

Public health in developed countries is heavily affected by pollution specially in highly populated areas. Amongst the pollutants with greatest impact in health, ozone is particularly addressed in this paper due to importance of its effect on cardiovascular and respiratory problems and their prevalence on developed societies. Local authorities are compelled to provide satisfactory predictions of ozone levels and thus the need of proper estimation tools rises. A data driven approach to prediction demands high quality data but those observations collected by weather stations usually fail to meet this requirement. This paper reports a new approach to robust ozone levels prediction by using an outlier detection technique in an innovative way. The aim is to assess the feasibility of using raw data without preprocessing in order to obtain similar or better results than with traditional outlier removal techniques. An experimental dataset from a location in Spain, Ponferrada, is used through an experimental stage in which such approach provides satisfactory results in a difficult case.