Format

Send to

Choose Destination
Environ Res. 2017 Nov;159:427-434. doi: 10.1016/j.envres.2017.08.039. Epub 2017 Sep 18.

A spatio-temporal prediction model based on support vector machine regression: Ambient Black Carbon in three New England States.

Author information

1
Department of Environmental Health, Harvard T.H. Chan School of Public Health, 401 Park Drive, Boston, MA 02215, USA. Electronic address: yara.abuawad@mail.harvard.edu.
2
Department of Environmental Health, Harvard T.H. Chan School of Public Health, 401 Park Drive, Boston, MA 02215, USA.
3
Department of Environmental Health, Harvard T.H. Chan School of Public Health, 401 Park Drive, Boston, MA 02215, USA; Department of Biostatistics, Harvard T.H. Chan School of Public Health, 677 Huntington Avenue, Boston, MA 02215, USA.
4
Department of Environmental Health, Harvard T.H. Chan School of Public Health, 401 Park Drive, Boston, MA 02215, USA; Department of Epidemiology, Harvard T.H. Chan School of Public Health, 677 Huntington Avenue, Boston, MA 02215, USA.

Abstract

Fine ambient particulate matter has been widely associated with multiple health effects. Mitigation hinges on understanding which sources are contributing to its toxicity. Black Carbon (BC), an indicator of particles generated from traffic sources, has been associated with a number of health effects however due to its high spatial variability, its concentration is difficult to estimate. We previously fit a model estimating BC concentrations in the greater Boston area; however this model was built using limited monitoring data and could not capture the complex spatio-temporal patterns of ambient BC. In order to improve our predictive ability, we obtained more data for a total of 24,301 measurements from 368 monitors over a 12 year period in Massachusetts, Rhode Island and New Hampshire. We also used Nu-Support Vector Regression (nu-SVR) - a machine learning technique which incorporates nonlinear terms and higher order interactions, with appropriate regularization of parameter estimates. We then used a generalized additive model to refit the residuals from the nu-SVR and added the residual predictions to our earlier estimates. Both spatial and temporal predictors were included in the model which allowed us to capture the change in spatial patterns of BC over time. The 10 fold cross validated (CV) R2 of the model was good in both cold (10-fold CV R2 = 0.87) and warm seasons (CV R2 = 0.79). We have successfully built a model that can be used to estimate short and long-term exposures to BC and will be useful for studies looking at various health outcomes in MA, RI and Southern NH.

KEYWORDS:

Air pollution; Black Carbon; Machine learning; Prediction; Support Vector Regression

PMID:
28858756
PMCID:
PMC5623647
DOI:
10.1016/j.envres.2017.08.039
[Indexed for MEDLINE]
Free PMC Article

Supplemental Content

Full text links

Icon for Elsevier Science Icon for PubMed Central
Loading ...
Support Center