To deliver this we use statistical, machine learning, as well as mechanistic model-based methods. These methods utilize data from the CDC, Google Health Trends, and high-resolution meteorological data (temperature, absolute humidity, precipitation, etc.). The mechanistic model-based methods also use US census data, mobility data, and synthetic population data developed by our lab. 

The pipeline has been successfully run since 2017. It provides forecast data so that public users can query the flu index for any day of the next four weeks at any location in the US. Weekly executive summary reports allow organizations and business users to extract insights from the data analytics on surveillance and forecasts.

Predicting influenza accurately is a difficult task due to the randomness of the influenza dynamics in unstructured populations, the sparsity and noisiness of the surveillance data, and the adaptive nature of human behavior. While high-quality forecasts can help at the public health level, public awareness of an ongoing epidemic and its severeness through the season is also extremely important.  This knowledge allows proper preventive actions to be taken at the individual and household levels, such as vaccination and self-isolation.

1. We have built HIIPP, a High-resolution Influenza Index Prediction Pipeline. It provides weekly predictions of the FluReady flu index, which ranges from level 1 to 10, with 1 being the lowest flu activities and 10 being the highest. The index is a measure of the intensity of the ongoing influenza epidemic, aligned with but distinct from the state activity indicator published weekly by the CDC. We extend the surveillance of this measure at state-level to nowcasts, and forecasts at sub-state levels. The flu index is a measure that allows one to easily compare the intensity of flu activities across different locations and over time, and make informed decisions. Compared with ILI percentage or case count measures, the flu index is easier to convey and more intuitive to comprehend.
2. We have developed models and methods for making high-resolution forecasts. For the general public, surveillance and forecasts of flu activities at aggregate levels (e.g. state)  are less relevant than those at local levels such as counties. We have mapped county-level forecasts to AccuWeather weather stations to allow users to consistently query their local flu activity levels.
3. We have built an end-to-end software pipeline that runs weekly to ingest surveillance data, call statistical and machine learning models, make forecasts, and publish forecasts and summary reports. The software pipeline was designed in a modular manner so that new forecasting methods can be easily added to improve the overall performance of our forecasts.

HIIPP provides insights into how academic research and industry business can work closely to benefit the public effectively.

Figures

The outputs of our forecasting pipeline are provided every week during the flu season (October to April) as data sets of flu index predictions and summary reports to AccuWeather, Inc. AccuWeather makes the forecasts accessible to the general public on their weather forecast platforms.

• The FluReady flu index is an index developed to indicate the relative level influenza-like illness activity level within the population at a particular location. It varies from 1 (minimal) to 10 (very high).
• Flu index predictions are provided both at the weather station level and for each 0.1 degree (latitude) by 0.1 degree (longitude) grid cell of the USA.
• Flu index predictions are provided for the next 28 days.