|
Abstract:
|
Semi-regular seasonal epidemics of respiratory and vector-born diseases such as influenza and dengue impose a heavy economic and mortality burden on society. Seasonal (non-pandemic) influenza, for example, is associated with ~250k--500k deaths worldwide each year (WHO estimate). Case counts follow a pattern most years: a sharp peak or interval of high activity with relatively low and flat surroundings. Accurate and reliable forecasts of disease prevalence should enable more effective countermeasures and preparation. However, variations in the timing and intensity of the epidemics, nonlinear dynamics, hidden states, and holiday effects impair the performance of popular time-series modeling tools such as SARIMA. We present some of our group's forecasting systems, which have performed well in several government initiatives. They include an ensemble of data-driven techniques incorporating Bayesian reasoning about timing and intensity; Epicast, a wisdom-of-crowds approach leveraging real-time human predictions; and Archefilter, a filtering approach which incorporates digital surveillance data such as Twitter.
|
