Data Evaluation for Early Disease Outbreak Detection

点击翻译按钮获取中文摘要

Timely implementation of appropriate disease control measures is facilitated by earlier detection of disease outbreaks whether due to bioterrorism or naturally occurring pathogens. There currently exists a range of automated health systems data concerning e.g. ambulatory care and emergency department visits, hospitalizations, diagnostic tests and pharmaceutical drugs; moreover, the availability of these data is likely to increase with greater use of health information technology. While such information could be invaluable for disease outbreak detection, not enough is known about the relative merits of different data sources for early detection of disease outbreaks. In this project we will evaluate and compare the efficacy of different health services data sources for early disease outbreak detection, including telephone inquiries, ambulatory care visits, emergency department visits, laboratory test requests and results, radiology tests, hospitalizations, drug prescriptions and drug dispensings. As test-beds we will use two large integrated health delivery systems (Harvard Pilgrim Health Care / Harvard Vanguard Medical Associates and Kaiser Permanente Northern California) with comprehensive electronic medical information on over four million persons. This means that we will have information about each health encounter data source for exactly the same well-defined population, which is critical for proper comparison. The data sources will be evaluated using all three statistical signal detection algorithms chosen by the BioSense Initiative, plus the space-time permutation scan statistic. The latter automatically adjusts for any purely temporal and purely spatial variation in the data, so that the data comparison does not depend on our relative success at modeling that noise through statistical regression models for different data sources. The different data sources will be evaluated with respect to the number, timeliness, accuracy and precision of signals in four different ways, (i) Total number of signals compared to expect under the null hypothesis of no outbreaks, (ii) Concordance between signals and known disease outbreaks as defined by e.g. local public health departments, (iii) Confirmation or rejection of signals by boking at subsequent detailed health information for those individuals generating the signals, (iv) Presence or not of signals when the real data is spiked with simulated outbreaks.