IMED 2018: Predicting spatiotemporal risk of yellow fever using a machine learning approach.
- 1. Predicting Spatiotemporal Risk of Yellow Fever Using a Machine Learning Approach RajReni B. Kaul, MichelleV. Evans, Courtney Murdock, and John M. Drake bit.ly/YF-Kaul
- 2. Yellow fever (YF) Endemic to 44 countries in Africa and Latin America Estimated 60,000 deaths annually
- 3. 2017 spike in Brazilian YF cases Year ConfirmedCases 780 data stops Jan 2018 Reported byTABNET
- 4. 2017 spike in Brazilian YF cases Year ConfirmedCases 780 data stops Jan 2018 Reported byTABNET
- 5. Previous years had fewer cases Year ConfirmedCases Reported byTABNET
- 6. 2017 spike on the range edge Many of the cases were outside the historical range
- 7. Hg. leucocelaenus Hg. janthinomys Sa. chloropterus Ae. aegypti Ae. albopictus Sylvatic cycle dominates YF transmission
- 8. Hg. leucocelaenus Hg. janthinomys Sa. chloropterus Ae. aegypti Ae. albopictus Sylvatic cycle dominates YF transmission
- 9. Drivers of spillover Adapted from Gortazar et al 2014
- 10. Connecting drivers and data for spillover NHP and vector range maps human density
- 12. Connecting drivers and data Fire density NHP and agriculture land overlap
- 13. Connecting drivers and data for spillover Fire density NHP and agriculture land overlap NHP and vector range maps human density Rainfall Temperature NDVI
- 14. Connecting drivers and data for spillover Fire density NHP and agriculture land overlap NHP and vector range maps human density Rainfall Temperature NDVI
- 15. Data aggregation: municipality Human Population Density Rainfall NDVI YF case(s) reported Temperature
- 16. Data aggregation: monthly* Human Population Density Rainfall Temperature NDVI month 1 month 2 month 3
- 17. Dataset units: municipality-months spillovers were very rare ~0.01% 106 spillover events between Jan 2001 – Dec 2013 month 1 month 2 month 3
- 20. Idealized data set Covariates Response Single observation with positive response value
- 21. Idealized data set Covariates Response Single observation with negative response value Single observation with positive response value
- 22. Subsample = bootstrapping Covariates Response Subsample dataset by selecting cases at random Sparse dataset Subsample dataset
- 23. Subsample = bootstrapping Covariates Response Sparse dataset Subsample dataset 0.01% positive observations 10% positive observations
- 24. Subsample many times Model1 Model2 Modeln Final product: Aggregated Model
- 25. Model1 Model2 Modeln Final product: Aggregated Model Advantages of Bagging Aggregating limits variance-bias trade off
- 26. Model1 Model2 Modeln Final product: Aggregated Model Advantages of Bagging Aggregating limits variance-bias trade off Increasing number of positive observations in subsample reduces over fitting
- 28. Model methods Training data subsample logistic regression bagged logistic regression average of 500 models 30% data withheld
- 29. Exploring model predictions Jan 2001
- 30. Exploring model predictions Jan 2001 Feb 2001
- 31. Exploring model predictions Jan 2001 Feb 2001 Mar 2001
- 32. Exploring model predictions Jan 2001 Feb 2001 Mar 2001 Month MeanIntensity Mean spillover intensity by month
- 33. Exploring model predictions Jan 2001 Feb 2001 Mar 2001 Variance in spillover intensity by municipality
- 36. YF spillover intensity Month MeanIntensity Variance in spillover intensity
- 38. Natural break in data led to regional models NHP species richness No.Municipalities High Richness Low Richness
- 39. High Low Spillover intensity by region Low 86 Spillover events High 20 Total 106 AUC 0.79 AUC 0.88
- 40. High Low Mean intensity by region Month MeanIntensity
- 41. High Low Mean intensity by region Variance in spillover intensity
- 42. Comparing national and regional models MeanIntensity Month Variance in intensity
- 43. VARIABLE IMPORTANCE BY PERMUTATIONMachine Learning Method
- 45. Idealized data set Response Covariates C1 C2 C3 C4
- 46. Covariates are linked to a observation Response Covariates C1 C2 C3 C4
- 47. Randomize a single covariate Response Covariates C1 C2 C3 C4
- 48. Randomize a single covariate Original data set Model Model accuracy predicting the training data
- 49. Randomize a single covariate Original data set C1 permutated data set Model Model accuracy predicting the training data Model Model accuracy predicting the training data
- 50. Randomize a single covariate Original data set C1 permutated data set Model Model accuracy predicting the training data Model Model accuracy predicting the training data Performance decline due to permutated covariate
- 53. Assessing drivers High Low RankedVariableImportance National Model Low NHP High NHP Policy implication: Same interventions can be used for whole country
- 54. Assessing drivers High Low RankedVariableImportance National Model Low NHP High NHP Policy implication: Different interventions are needed across the country
- 57. RankedVariableImportance National Low NHP High NHP
- 58. RankedVariableImportance National Low NHP High NHP
- 59. RankedVariableImportance National Low NHP High NHP
- 60. RankedVariableImportance National Low NHP High NHP
- 61. Yellow fever in Brazil Model drivers of spillover are regional requiring regional interventions
- 62. Yellow fever in Brazil Model drivers of spillover are regional At times coastal cities are at higher risk than the Amazon River basin
- 63. Yellow fever in Brazil Model drivers of spillover are regional At times coastal cities are at higher risk than the Amazon River basin Month MeanIntensity
- 64. Acknowledgements Eric Marty JP Schmidt Spencer Hall Annakate Schatz @renikaul renikaul/YF_Brazil reni@uga.edu
- 68. Assessing drivers High Low RankedVariableImportance National Model Low NHP Low 5350 Municipalities High 210 Total 5560