Evaluating a Seasonal ARIMA Model for Event Detection in New York City

Seasonal autoregressive integrated moving average (ARIMA) models can generate future forecasts, making it a potential method for modeling syndromic data for aberration detection. We built ARIMA models for five routinely monitored syndromes in New York City and tested the models' ability to prospectively detect outbreaks using datasets spiked with simulated outbreaks. Less than 10% of all outbreaks were detected at a fixed alert threshold of 1 signal per 100 days. These models did not perform well in detecting outbreaks and may require frequent monitoring and readjustment of model parameters.