Lyme disease is the fastest-growing vector borne disease in North America, and therefore a growing public health concern. TickTock is a predictive model that incorporates historical data from the CDC, historical weather and climate data, national land-cover data, and climate change models, to create predictions for the number of Lyme cases in USA on a per-county basis. It uses Facebook Prophet for the time-series prediction, resulting in a an improvement of 25% on the mean absolute percentage error. It is hosted at https://ticktocklyme.herokuapp.com/
- CDC historical data
- NLCD data
- CMIP5 projections
- Historical climate and weather data
I validated the model using 2016,2017 data as test data. Note, another possible option is to leave out counties - fit the models without 10% of the counties in them, and see how it performs on the new ones.
How to most effectively combine time varying and stationary data?
Some of our data is annual eg. the cases per year, while other, such as the land-cover database, is not. To use the time-stationary data effectively, and to reduce noise in the data at the same time, I clustered the counties based on the forest-cover, latitude etc.
- http://danielcameronmd.com/lyme-disease-can-cost-billions/
- https://www.lymedisease.org/lymepolicywonk-costs-75billion/
- https://www.hhs.gov/ash/advisory-committees/tickbornedisease/meetings/2018-07-24/written-public-comment/index.html
- https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5872223/
- https://www.jhsph.edu/news/news-releases/2015/lyme-disease-costs-more-than-one-billion-dollars-per-year-to-treat-study-finds.html
- https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0116767
- Incorporate hierarchical time-series along with the additional regressors.
- Incorporate individual models for the different classes
- Include the individual time-series models for each county