Alternative screening method for potential airborne disease : using severe acute respiratory syndrome data as an example

Abstract

Introduction: Infectious diseases are alarming. They are contagious and detrimental as the illnesses concern the human respiratory systems which are the means that the infections are transmitted swiftly from one person to another within seconds. Nowadays, close contacts among people may result in chaotic situations attribute to the potential wide spread of infections when traveling abroad is becoming more common and frequent. The disastrous incident of Severe Acute Respiratory Syndrome ("SARS" thereafter) in 2003 caused more than eight thousands people worldwide infected with an initial of only ten infected people who once stayed in the same hotel, and those ten people were suspected to have shared the single infected origin. This case is the best illustration of how powerful a deadly infection can be. It awakened most countries to re-examine their existing public health delivery systems and to figure out what instant and remedial actions should be taken when infectious disease starts to spread, for instance, to determine the efforts needed for containment and mitigation. Historical incidents have already proved that prompt detections and comprehensive surveillances are vital to the eradication of infectious diseases, but the most appropriate way for effective and efficient screening has not yet been confirmed as each of them carries its own advantages and disadvantages. In order to ensure the isolation precaution starts at the right time for effective containment, a sensitive and reliable screening approach to trigger off the whole process is crucial. Purpose: To explore the feasibility of using data mining as an effective screening method to predict the occurrence of airborne disease based on the pre-hospitalized clinical presentations using the data of SARS in Hong Kong as an example. Method: This study is an observational retrospective case record review study. All patients aged 18 or above, attending the Accident & Emergency Department (AED) of a major hospital during the period from 1st February 2003 to 30th June 2003 with provisional diagnosis of SARS, were recruited. Data collected for analysis included patient particulars, clinical presentations, co-morbidities, and laboratory results for confirmation of SARS based on the World Health Organisation (WHO) guidelines. There are four stages in this study. The first stage is the preparation of a comprehensive database for further analysis, followed by an evaluation of the existing prediction rules reported by others in stage two. The third stage is the attribute identification stage and the last stage is the model testing. Results: A total of 549 adult case records were examined. Eighty percents of them were randomly selected to form the training dataset and the remaining cases were used as testing dataset. The testing data was fitted into the existing prediction model reported by Chen et al. (2004), Wang et al. (2004) and Leung et al. (2004) that all the studies were carried out in the most similar situation or inclusion criteria as the current study. The testing data was then classified into SARS and non-SARS based on each prediction rule and counterchecked with the laboratory diagnostic results. The sensitivity and specificity of each prediction rule were calculated and compared with the quoted value. The poor agreement of the calculated sensitivity (ranged from 0.17 to 0.95) and the specificity (ranged from 0 to 0.67) with the quoted values showed a strong need to have a new prediction model with better prediction power. Data mining technique was employed to see if it can be an alternative prediction method for airborne disease. Association rule mining could not find any sequential /affinity relationship between the clinical variables and the disease status. Classification rule mining showed that malaise, sore throat, fever and shortness of breath were critical clinical predictors where clustering method identified chills, malaise, sore throat and shortness of breath as critical clinical predictors. The testing data was fitted into the mined rules again and another set of calculated sensitivity (0.86) and specificity (0.71) values were obtained for comparison. The results were further tested under different circumstances and similar findings were obtained. Conclusion: Data mining can be a better and an efficient option with higher specificity and sensitivity for predicting airborne disease in AED in the future.