Project 4

Zi Xin | Rachel | David | Xavier

Predicting presence of West Nile Virus in mosquitos

Introduction

West Nile virus (WNV) is a mosquito-borne infectious disease that flares up during mosquito season from summer to early fall. It first emerged in the eastern U.S. in 1999, and has now become the leading cause of mosquito-borne disease in the continental U.S. [1]

WNV is most commonly spread to humans through infected mosquitos, usually of the genus Culex. [2] In most cases, WNV infection does not cause symptoms, but up to 20% of people who become infected with the virus develop symptoms ranging from a persistent fever, to serious neurological illnesses that can result in death. Currently, there is no human vaccine against WNV and prevention of the illness in humans is based on mosquito control. [3]

In 2002, WNV first hit Chicago with 225 human cases reported that summer. In response, the Chicago Department of Public Health (CDPH) implemented city‐wide surveillance and mosquito control measures, which resulted in a dramatic decline of human cases in subsequent years. Chicago continues to have one of the most ro‐ bust urban mosquito control programs in the country. [4]

Problem Statement

The goal of this project is to assist the City of Chicago a.k.a Windy City and the CDPH in developing a model that can accurately predict when and where different species of mosquitos will test positive for WNV and identify the factors that are the strongest predictors of WNV presence.

If outbreaks of WNV in mosquitos can be reliably predicted, the City of Chicago and CPHD will be able to more efficiently and effectively coordinate resources towards preventing transmission of this potentially deadly virus.

Problem Approach

Predicting the presence of WNV is a binary classification problem. To that end, we tested several classification models listed below using a total of 66 feature variables. The evaluation metrics used will be AUC score and sensitivity (recall).

• Logistic regression
• Support vector machine
• Bagged decision trees
• Extra trees
• Random forest
• AdaBoost
• Gradient boosting
• XGBoost

Conclusion and Recommendations

In this analysis, we explored the performance of several models in predicting the presence of WNV in mosquitos. Using AUC score and sensitivity as our main evaluation metrics, the Random Forest classifier, which obtained an AUC score of 0.846 and sensitivity of 0.819, was selected as the final production model.

As seen from the model's top 20 features, factors related to seasonality, weather, location and mosquito species are highly important in predicting the presence of WNV. In particular, seasonality and weather-related features play the biggest role as they made up the majority of the top 20 features. This aligns with the fact the WNV is a seasonal virus that emerges during summer as hot temperatures favour mosquito-breeding.

In our cost-benefit analysis of mosquito spraying, we found that spraying seems to be inversely correlated with WNV incidence. However, we do not have sufficient data at hand to quantify the effectiveness of spraying and thus cannot form a strong conclusion on whether the benefits of spraying are justified by its costs. Furthermore, it is difficult to isolate the effects of spraying as the spread of WNV is also influenced by other variables such as geographical location and weather.

Whilst mosquito spraying certainly has its benefits, it is best for the CDPH to take a multi-pronged approach in tackling WNV. CDPH could drive educational campaigns to teach the community how they can help to prevent mosquito breeding in their homes. For example, ensuring that they remove all standing water so that mosquitos will not lay their eggs there.

For spraying, since WNV incidence is the highest in August, spraying efforts should be concentrated in August for greatest effect.

Data description

Train/Test Data (split into separate train and validation sets for modelling) – Every week from late spring through the fall, mosquitos in traps across the city are tested for presence of WNV. These test results are summarised by date from May 2007 to Oct 2014, trap, and species. Location data is also provided for all the traps.

Train Data consist of information for 2007, 2009, 2011 and 2013, while Test Data consist of information for 2008, 2010, 2012 and 2014.

Spray Data – Data is provided for the locations where the City of Chicago did mosquito spraying in 2011 and 2013. Spraying can reduce the number of mosquitos in the area, and therefore might eliminate the appearance of WNV.

Weather data – Daily weather records are provided for the period 2007 to 2014 from 2 weather stations. Hot and dry weather conditions are conducive for a seasonal virus like WNV, and it is worth using weather data to improve the prediction of WNV presence.

Data dictionary

Feature	Variable type	Datatype	Dataset	Description
Id	Norminal	int64	test	the id of the record
Date	datetime	datetime	train and test	date that the WNV test is performed
Address	Norminal	object	train and test	approximate address of the location of trap
Species	Norminal	object	train and test	species of mosquito
Block	Norminal	int64	train and test	block number of address
Street	Norminal	object	train and test	Street name
Trap	Norminal	object	train and test	Id of the trap
AddressNumberAndStreet	Norminal	object	train and test	approximate address returned from GeoCoder
Latitude	continuous	float64	train and test	Latitude returned from GeoCoder
Longitude	continuous	float64	train and test	Longitude returned from GeoCoder
AddressAccuracy	Norminal	int64	train and test	accuracy returned from GeoCoder
NumMosquitos	discrete	int64	train and test	number of mosquitoes caught in this trap
WnvPresent	discrete	int64	train and test	1 means WNV is present, and 0 means not present.
Station	discrete	int64	weather	Station 1 or 2 where weather data is collected
Date	datetime	datetime	weather	Date of weather record
Tmax	discrete	int64	weather	Max temperature in Fahrenheit
Tmin	discrete	int64	weather	Min temperature in Fahrenheit
Tavg	continuous	float64	weather	Avg temperature in Fahrenheit
Depart	discrete	float64	weather	The difference from normal temperatures for the last 30yrs
DewPoint	discrete	int64	weather	Dew Point temperature
WetBulb	discrete	int64	weather	Web Bulb temperature
Heat	continuous	float64	weather	Difference between daily avg temperature and 65 Deg F
Cool	continuous	float64	weather	Difference between daily avg temperature and 65 Deg F
Sunrise	datetime	datetime	weather	Sunrise timing
Sunset	datetime	datetime	weather	Sunset timing
CodeSum	norminal	object	weather	Weather Phenomena
Depth	continuous	float64	weather	Snow inches
Water1	continuous	float64	weather	Snow/Ice on ground
SnowFall	continuous	float64	weather	Snowfall in inches
PrecipTotal	continuous	float64	weather	Rainfall and melted snow (inches)
StnPressure	continuous	float64	weather	Avg Station Pressure
SeaLevel	continuous	float64	weather	Avg Sea level Presure
ResultSpeed	continuous	float64	weather	Resultant Wind Speed
ResultDir	continuous	int64	weather	Resultant Wind Direction
AvgSpeed	continuous	float64	weather	Avg WInd Speed
Date	datetime	datetime	spray	Date of the spray
Time	datetime	datetime	spray	Time of the spray
Latitude	continuous	float64	spray	Latitude of the spray
Longitude	continuous	float64	spray	Latitude of the spray
year	discrete	int64	engineered feature	year corresponding in train data
month	discrete	int64	engineered feature	month corresponding in train data
weekofyear	discrete	int64	engineered feature	week of the year corresponding in train data
yearmonth	norminal	object	engineered feature	yearmonth string corresponding in train data
weekday	discrete	int64	engineered feature	weekday corresponding in train data
closest_station	discrete	int64	engineered feature	the closer weather station to each trap
row_count	discrete	int64	engineered feature	number of "duplicate" rowsto get an estimate for nummosquitos
intensity_acc	continuous	float64	engineered feature	calculates the "intensity" of a spray for each observation in the train dataset
daytime	discrete	int32	engineered feature	total number of minutes from Sunrise to Sunset
trap_weight	continuous	float64	engineered feature	weightage of the trap based on Wnvpresent/no. of times sampled
rhumidity	continuous	float64	engineered feature	relative humidity derived from dewpoint and tavg
tavg_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling temp averages for 7, 10, 14, 30, 60, 90 days
resultspeed_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling wind speed averages for 7, 10, 14, 30, 60, 90 days
dewpoint_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling dewpoint averages for 7, 10, 14, 30, 60, 90 days
rhumidity_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling relative humidity averages for 7, 10, 14, 30, 60, 90 days
tmax_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling max temp averages for 7, 10, 14, 30, 60, 90 days
tmin_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling min temp averages for 7, 10, 14, 30, 60, 90 days
preciptotal_7/10/14/30/60/90	continuous	float64	engineered feature	Rolling preciptotal averages for 7, 10, 14, 30, 60, 90 days