RESUMO
The link between particulate matter (PM) air pollution and negative health effects is well-established. Air pollution was estimated to cause 4.9 million deaths in 2017 and PM was responsible for 94% of these deaths. In order to inform effective mitigation strategies in the future, further study of PM and its health effects is important. Here, we present a method for identifying sources of combustion generated PM using excitation-emission matrix (EEM) fluorescence spectroscopy and machine learning (ML) algorithms. PM samples were collected during a health effects exposure assessment panel study in Seattle. We use archived field samples from the exposure study and the associated positive matrix factorization (PMF) source apportionment based on X-ray fluorescence and light absorbing carbon measurements to train convolutional neural network and principal component regression algorithms. We show EEM spectra from cyclohexane extracts of the archived filter samples can be used to accurately apportion mobile and vegetative burning sources but were unable to detect crustal dust, Cl-rich, secondary sulfate and fuel oil sources. The use of this EEM-ML approach may be used to conduct PM exposure studies that include source apportionment of combustion sources.
RESUMO
The inhalation of particulate matter (PM) is a significant health risk associated with reduced life expectancy due to increased cardio-pulmonary disease and exacerbation of respiratory diseases such as asthma and pneumonia. PM originates from natural and anthropogenic sources including combustion engines, cigarettes, agricultural burning, and forest fires. Identifying the source of PM can inform effective mitigation strategies and policies, but this is difficult to do using current techniques. Here we present a method for identifying PM source using excitation emission matrix (EEM) fluorescence spectroscopy and a machine learning algorithm. We collected combustion generated PM2.5 from wood burning, diesel exhaust, and cigarettes using filters. Filters were weighted to determine mass concentration followed by extraction into cyclohexane and analysis by EEM fluorescence spectroscopy. Spectra obtained from each source served as training data for a convolutional neural network (CNN) used for source identification in mixed samples. This method can predict the presence or absence of the three laboratory sources with an overall accuracy of 89% when the threshold for classifying a source as present is 1.1 µg/m3 in air over a 24-hour sampling time. The limit of detection for cigarette, diesel and wood are 0.7, 2.6, 0.9 µg/m3, respectively, in air assuming a 24-hour sampling time at an air sampling rate of 1.8 liters per minute. We applied the CNN algorithm developed using the laboratory training data to a small set of field samples and found the algorithm was effective in some cases but would require a training data set containing more samples to be more broadly applicable.