Table 4. Model applications.
Model | Study | Prediction Target | Sample Size | Spatial Domain | Time Domain | Testing Method1 | Metric Score2 |
---|---|---|---|---|---|---|---|
B. Spatial Risk Random Forest | [29] | Mean annual incidence per 100,000 population | 43,512 county-years | Conterminous US (3,108 counties) | 2005–2018, averaged | Bootstrapping | R2pred = 0.59 [0.44–0.70], RMSE = 3.7 |
D. Spatial Risk High Resolution BRT | [30] | Ranked relative risk (0–1) | 1,378 human cases | South Dakota | 2004–2017 | Out of sample data | AUC = 0.727 |
E. RF1 | [28] | Annual human cases | 882 county-years | New York and Connecticut | 2000–2015 | LOYOCV | R2pred = 0.72, RMSE = 1.6 |
E. RF1 | [28] | Seasonal mosquito MLE | 218 county-years | New York and Connecticut | 2000–2015 | LOYOCV | R2pred = 0.45, RMSE = 2.3 |
E. RF1 | [28] | Seasonal mosquito MLE | 2,596 trap-years | New York and Connecticut by trap | 2000–2015 | LOYOCV | R2pred = 0.53, RMSE = 1.0 |
F. NE_WNV County-years | [34] | 2018 human cases3 | 1,472 county-years | Nebraska | 2002–2017 | Out of sample data | CRPS = 1.90 |
F. NE_WNV County-years | [34] | 2018 WNV positive counties3 | 1,472 county-years | Nebraska | 2002–2017 | Out of sample data | Accuracy = 0.717 |
G. GLMER Ensemble | [20] | MLE mosquito infection rate | 225 grid-years | Suffolk County, New York | 2001–2015 | LOYOCV | RMSE = 4.27 |
H. Harris County | [21] | MLE mosquito infection rate (1-month lead) | 130,567 trap-nights | Harris County, Texas | 2002–2016 | Out of sample data |
R2pred = 0.8 |
H. Harris County | [21] | Mosquito abundance (1-month lead) | 10,533,033 mosquitoes | Harris County, Texas | 2002–2016 | Out of sample data | R2pred = 0.2 |
I. ArboMAP | [35] | Positive county-weeks | Approximately 9,504 county-weeks (training) Approximately 792 county-weeks (testing) |
South Dakota | 2004–2015 (training) 2016 (testing) | Out of sample data | AUC = 0.836–0.8564 |
I. ArboMAP | [36] | Positive county-weeks | Approximately 11,088 county-weeks | South Dakota | 2004–20175 | Fit to training data only | AUC = 0.876, Rs = 0.84 |
J. Chicago Ultra-Fine Scale | [22–24] | Human case probability (by hexagon) | 1,346,940 hexagon-weeks | Variable, up to 5,345 1-km hexagons | 2005–20166 | Fit to training data only | R2 > 0.85; RMSE < 0.02; AUC > 0.90 |
K. Model-EAKF System | [25] | Annual human cases; peak mosquito infection rates; peak timing of infectious mosquitoes; annual infectious mosquitoes | 21 county-years | 2 counties (Suffolk, New York and Cook, Illinois) | Weekly, Varied by location | Retrospective data assimilation | Threshold-based accuracy 7 |
K. Model-EAKF System | [26] | Multiple8 | 110 outbreak-years | 12 counties | Weekly, Varied by location | Retrospective data assimilation | Threshold-based accuracy 7 |
K. Model-EAKF System | [39] | Multiple8 | 4 county-years | 4 counties | Weekly, 2017 | Real-time data assimilation | Threshold-based accuracy 7 |
L. Temperature-forced Model-EAKF System | [26] | Multiple8 | 110 outbreak- years | 12 counties | Weekly, Varied by location | Retrospective data assimilation |
Threshold-based accuracy 7 |
M. California Risk Assessment | [31] | Historical outbreaks of western equine encephalomyelitis and St. Louis encephalitis as proxy for WNV | 14 agency-years | California | Half-months | Temporal correspondence | Early detection of arbovirus risk prior to outbreaks |
M. California Risk Assessment | [32] | Onset and peak of human cases by geographic region | 12 half-months in 3 regions | California | Half-months | Retrospective data assimilation | Early detection of WNV risk prior to onset and peak of human cases |
M. California Risk Assessment | [33] | Emergency planning threshold (risk ≥ 2.6) | 11,476 trap-nights | Los Angeles Country, California | 2004–2010 | Retrospective data assimilation | AUC = 0.982 |
1LOYOCV: leave-one-year-out cross-validation; Out of sample data: accuracy based on data not used to develop the model; Fit to training data only: accuracy based on the same data used to develop the model; Retrospective data assimilation: finalized data until the time of forecast; Real-time data assimilation: data processed and available at the time of forecast.
2R2pred: predictive R2, i.e., an R2 calculated on data outside the sample, Rs: Spearman correlation coefficient, AUC: area under the curve, Threshold-based accuracy: +/−25% of peak week, human cases, total infections over the season; +/−25% or 1 human case, RMSE: Root Mean Squared Error, CRPS: Continuous Ranked Probability Score.
3Results for 2018 reported here, validation was also performed separately for 2012–2017, see [34] for details.
4Three analyses presented: short-term: AUC = 0.856, annual made on July 5: AUC = 0.836, annual made on July 39: AUC = 0.855.
5Restricted to July–September for each year.
6Restricted to 21 epi weeks per year.
7Varied by analysis and lead time.
8Prediction targets: human cases in next 3 weeks; annual human cases; week with highest percentage of infectious mosquitoes; peak mosquito infection rate; annual infectious mosquitoes.
AUC, area under the curve; CRPS, Continuous Ranked Probability Score; LOYOCV: leave-one-year-out cross-validation; RMSE, Root Mean Squared Error; WNV, West Nile virus.