Project 4 - West Nile Virus Prediction

---

Chenyze | Elaine | Kenrick | Raphael

---

West Nile virus (WNV) is the leading mosquito-borne disease in the United States (CDC, 2009). It is a single-stranded RNA virus from the family Flaviviridae, which also contains the Zika, dengue, and yellow fever virus. It is primarily transmitted by mosquitoes (mainly Culex spp.), which become infected when they feed on birds, WNV's primary hosts. Infected mosquitos then spread WNV to people and another animals by biting them. Cases of WNV occur during mosquito season, which starts in the summer and continues through fall (CDC, 2009).

While WNV is not extremely virulent and only about 1 in 5 people who are infected develop West Nile fever and other symptoms, about 1 out of every 150 infected develop a serious and sometimes fatal illness (CDC, 2009). There is currently no vaccine against WNV.

In view of the recent outbreak of WNV in Chicago, the Chicago Department of Public Health (CDPH) has set up a surveillance and control system to trap mosquitos and test for the presence of WNV. The goal of this project is to use these surveillance data to predict the occurrence of WNV given time, location, and mosquito species. Findings from this project will guide and inform decisions on where and when to deploy pesticides throughout the city, to maximise pesticide effectiveness and minimise spending.

Folder Organisation:

---

|__ code
|   |__ 01_eda_and_feature_engineering.ipynb   
|   |__ 02_model_tuning_and_insights.ipynb     
|__ data
|   |__ train.csv
|   |__ test.csv
|   |__ processed.zip    
|   |__ weather.csv
|   |__ spray.csv
|   |__ mapdata_copyright_openstreetmap_contributors.txt
|__ images
|   |__ heatmaps.png
|   |__ species_count.png
|   |__ wnv_by_year.png
|   |__ wnv_distribution.png
|__ planning_documentation.xlsx
|__ group_presentation.ppt
|__ README.md

Summary of Findings & Recommendations

---

(Occurrences of WNV and mosquito spraying efforts in Chicago from May - Sept in 2007, 2009, 2011, and 2013.)

	predicted WNV absent	predicted WNV present
actual WNV absent	2031	458
actual WNV present	50	88

Using XGBoost (our best performing model), we achieved an ROC_AUC of 0.834 with the confusion matrix above. Using the model, we found that WNV is more prevalent under certain conditions. Week of year was the top predictor by far for our model, followed by day of year, 18-days rolling sum of daylight hours, Culex restuans, and 18-days rolling mean of average temperature. This means that WNV is most likely to occur during certain weeks of the year (in August, as shown in the figure above), and therefore spray efforts should be concentrated during this period. Location was not a strong predictor in our model, suggesting that WNV clusters may be transient, occurring where best conditions emerge.

After conducting a cost-benefit analysis, we found that the money saved from reducing WNV cases would at most fund about 300 - 500 sprays. However, as the current datasets do not substantially point to a significant impact from spraying, more evidence (from a better designed spraying regime) are needed for a more concrete recommendation. For example, spraying efforts could be concentrated at the beginning of August so that there would be enough time to observe if mosquito numbers decline, in the relative absence of other confounding factors (such as temperature).

Data Dictionary

---

Feature	Type	Dataset	Description
Id	int	train.csv/test.csv	ID number of the record
Date	datetime	train.csv/test.csv	date the WNV test is performed
Address	datetime	train.csv/test.csv	approximate trap address; sent to GeoCoder
Species	str	train.csv/test.csv	mosquito species in trap
Block	str	train.csv/test.csv	block number of address
Street	str	train.csv/test.csv	street of address
Trap	str	train.csv/test.csv	ID number of the trap
AddressNumberAndStreet	str	train.csv/test.csv	approximate address retrieved from GeoCoder
Latitude	float	train.csv/test.csv	latitude retrieved from GeoCoder
Longitude	float	train.csv/test.csv	longitude retrieved from GeoCoder
AddressAccuracy	int	train.csv/test.csv	accuracy of information returned from GeoCoder
NumMosquitos	int	train.csv/test.csv	number of mosquitoes in a sample
WnvPresent	int	train.csv/test.csv	whether or not WNV is present in a sample (1 = present; 0 = absent)
Date	datetime	spray.csv	date of spray
Time	datetime	spray.csv	time of spray
Latitude	float	spray.csv	latitude of spray
Longitude	float	spray.csv	longitude of spray
Station	int	weather.csv	weather station (1 or 2)
Date	datetime	weather.csv	date of measurement
Tmax	int	weather.csv	maximum daily temperature (F)
Tmin	int	weather.csv	minimum daily temperature (F)
Tavg	int	weather.csv	average daily temperature (F)
Depart	int	weather.csv	departure from normal temperature (F)
Dewpoint	int	weather.csv	average dewpoint (F)
WetBulb	int	weather.csv	average wet bulb
Heat	int	weather.csv	heating degree days
Cool	int	weather.csv	cooling degree days
Sunrise	int	weather.csv	time of sunrise (calculated)
Sunset	int	weather.csv	time of sunset (calculated)
CodeSum	str	weather.csv	code of weather phenomena
Depth	int	weather.csv	unknown
Water1	int	weather.csv	unknown
SnowFall	int	weather.csv	snowfall (inch)
PrecipTotal	str	intweather.csv	total daily rainfall (inch)
StnPressure	int	weather.csv	average atmospheric pressure (inch Hg)
SeaLevel	int	weather.csv	average sea level pressure (inch Hg)
ResultSpeed	float	weather.csv	resultant wind speed (mph)
ResultDir	int	weather.csv	resultant wind direction (degrees)
AvgSpeed	int	weather.csv	average wind speed (mph)