| Description: |
Having the ability to assess the risk of an outbreak becomes crucial when considering the effectiveness of responses to public health issues. Failing to notice changes in how an outbreak is transmitted can cause delays in the strategic responses for containment. To counter this, we propose Outbreak Sentry, a real-time forecasting pipeline constructed in R. Outbreak Sentry can take in data from various sources, such as mobility information, daily epidemiological surveillance, temperature and rainfall, environment-related data, and geo-location data. Outbreak Sentry employs feature construction, lag feature addition, predictive modelling, and predictive model re-calibration. It performs rigorous automatic data quality inspection and inline data forecasting for time series, machine learning, and Bayesian models, combining the results using ensemble forecasting. Outbreak Sentry also supports time series, machine learning, and Bayesian models which it combines using ensemble forecasting. Over six years of detecting dengue instances in a tropical urban area, Outbreak Sentry was more accurate and better calibrated than baseline models like naive carry forward, ARIMA, gradient boosting, and Bayesian regression. The system was able to issue alerts with a 1-2 week delay prior to the outbreak occurring. We focus on the system’s architecture alongside model performance, alert systems, and system deployment. Outbreak Sentry was constructed to be used as a public health tool in the lower management levels, providing real-time responses to data-driven health issues. |