Ensemble Forecasting for Disease Outbreak Detection

We describe a method to improve detection of disease outbreaks in pre-diagnostic time series data. The method uses multiple forecasters and learns the linear combination to minimize the expected squared error of the next day's forecast. This combination adaptively changes over time. This adaptive ensemble combination is used to generate a disease alert score for each day, using a separate multi-day combination method learned from examples of different disease outbreak patterns. These scores are used to generate an alert for the epidemiologist practitioner. Several variants are also proposed and compared. Results from the International Society for Disease Surveillance (ISDS) technical contest are given, evaluating this method on three syndromic series with representative outbreaks. Problem Description In modern biosurveillance, time series of pre-diagnostic health data are monitored for disease outbreaks. Prediagnostic time series typically consist of daily counts of regional emer...

• Type: Conference
• Year: 2008
• Where: AAAI
• Authors: Thomas H. Lotze, Galit Shmueli
• Keywords: AAAI 2008, Disease Outbreak, Intelligent Agents, Pre-diagnostic Health Data, Time Series