Transitioning from Univariate to Multivariate Predictive Models for Respirable Quartz via Infrared Spectroscopy

1722 

Professional Poster 

J Stancil¹

¹NIOSH, N/A

Jordan Stancil  
NIOSH

The quantification of quartz in respirable dust samples is possible using a variety of established analytical methods. These methods have been established by national institutions, such as the National Institute for Occupational Safety and Health (NIOSH), standardization bodies, such as the American Society for Testing and Materials (ASTM International), and the International Organization for Standardization (ISO). These methods are generally based on either X-ray diffraction or infrared spectroscopy analytical techniques. The analysis is executed either directly on the filter used for sampling or after extraction of the sample and redeposition on new media. Every method is calibrated, independently of the analytical technique or preparation of the sample, using samples of a known quantity of material characterized by high purity quartz. In addition, every method is calibrated using a quantification model based on univariate calibration. The response of the analytical technique is condensed into one variable, typically the height or area of the region of interest. The values for the selected variable for different standard reference or calibration samples are correlated with the known amount of quartz in the same samples. The resulting regression becomes the calibration model and the quantification model for future unknown samples. While this approach is well established and validated, it does not take advantage of the full potential of the data created by modern diffractomet

"The quantification of quartz in respirable dust samples is possible using a variety of established analytical methods. These methods have been established by national institutions, such as the National Institute for Occupational Safety and Health (NIOSH), standardization bodies, such as the American Society for Testing and Materials (ASTM International), and the International Organization for Standardization (ISO). These methods are generally based on either X-ray diffraction or infrared spectroscopy analytical techniques. The analysis is executed either directly on the filter used for sampling or after extraction of the sample and redeposition on new media. Every method is calibrated, independently of the analytical technique or preparation of the sample, using samples of a known quantity of material characterized by high purity quartz. In addition, every method is calibrated using a quantification model based on univariate calibration. The response of the analytical technique is condensed into one variable, typically the height or area of the region of interest. The values for the selected variable for different standard reference or calibration samples are correlated with the known amount of quartz in the same samples. The resulting regression becomes the calibration model and the quantification model for future unknown samples. While this approach is well established and validated, it does not take advantage of the full potential of the data created by modern diffractome

"Respirable aerosol samples were collected via aerosolization and filter deposition inside aerosol chambers. Conceptually, the samples were grouped as calibration and validation samples. All of the respirable aerosol samples were collected on 37-mm, 5.0-µm pore PVC filters with 10-mm nylon Dorr Oliver cyclones at a flow rate of 1.7 lpm. The calibration samples were generated by aerosolizing particles of Minusil5, Minusil10, and Minusil15 powder products. The alpha quartz content in the Minusil5 powder was found to be >95% by analyzing the sample via X-Ray powder diffraction full pattern analysis with an internal standard. The validation samples with only quartz were generated using the same Minusil5 powder. The validation samples with quartz and albite, a known analytical confounder for quartz for infrared spectroscopy, were generated by layering computationally the spectrum of a single sample of Minusil5 and the spectra from samples with varying masses of respirable albite obtained via separate aerosol sampling from crushed albite powder. This computational approach was chosen so that the effect of varying the amount of albite on the spectrum could be examined while keeping the mass of quartz constant. The validation samples with quartz only were prepared in a hexagonal calm air chamber. Sample collection is described in more detail in the manuscript. Samples were collected in six sets to investigate the precision of the analysis for the quantification of quartz by the different quantification models. The validation data with quartz and albite were prepared via aerosolization of albite in the same chamber used for the calibration samples and computational layering. Crushed albite powder was aerosolized in the chamber and respirable samples were collected with different target loadings. Then, one single respirable sample of quartz from aerosolization of Minusil5 with known loading was selected for computational layering. Each sample collected for calibration or validation was analyzed using two independent analytical techniques. Filter samples were pre- and post-weighed in a temperature- and humidity-controlled weighing room using a microbalance. Following gravimetric analysis, samples were analyzed direct-on-filter with a portable transmission infrared spectroscopy instrument; specific parameters can be found in the manuscript. A blank PVC filter (from the same lot as the set of sample filters) was used as the background spectrum. For validation data containing albite, albite spectra of various mass loadings were computationally combined with the single quartz spectrum utilizing the additive property of infrared absorption.
For both the calibration and validation samples, the results of the gravimetric analysis were used as reference values for the mass of quartz and albite in each sample. The infrared spectrum for each sample was used as an indirect measurement specific to predicting the quartz mass in each sample. Two different predictive mo

"This study compares the performance of univariate and multivariate predictive models for the quantification of respirable quartz direct-on-filter. Both models were trained on samples of aerosolized quartz and underwent validation with two sample sets, one of respirable quartz and one with respirable quartz and albite mixtures. The performance of the two models was comparable, and when challenged with a validation sample set of pure quartz, the RMSEPs were 8.2 μg and 9.4 μg, respectively, for the univariate and multivariate models. Using the quartz and albite mixtures as an example, we also explain the ability to use the multivariate summary statistics, Q residuals, and Hotelling T2 values as an objective way to assess analytical interferences. The direct advantage of the multivariate modeling technique over the univariate technique was the ability to objectively assess the abnormality of a spectrum, compared to the training set in a systematic way. Thus, we demonstrate the superiority of a multivariate calibration approach and lay the foundations for the first analytical method for quantification of respirable quartz that utilizes a multivariate predictive model. In addition, this superiority can be critical in the implementation of the model in an analytical method used by non-experts, such as in-field occupational hygienists.

Results including other summary statistics for the models & tables, plots, and figures displayed in the manuscript will be presented on the poster as well for a closer look at these specifics."

Exposure Assessment