Semester
Summer
Date of Graduation
2024
Document Type
Thesis
Degree Type
MS
College
Eberly College of Arts and Sciences
Department
Forensic and Investigative Science
Committee Chair
Glen P. Jackson
Committee Co-Chair
Jacqueline Speir
Committee Member
Shikha Sharma
Abstract
Arson investigators require reliable and objective methods to interrogate fire debris for the presence of ignitable liquid residues. Although standard methods like ASTM E1618-19 exist for such procedures, these methods typically do not rely on quantitative measures but instead rely on the subjective judgment of fire debris analysts. To provide a more objective means to interpret fire debris evidence, we examined the ability of the Expert Algorithm for Substance Identification (EASI) to identify ignitable liquid classes from the total ion spectrum (TIS) of gas chromatography-mass spectrometry (GC-MS) data files. TISs have been promoted by others to help remove variability caused by the GC dimension. The TIS data are first preprocessed by selecting 20 abundant and characteristic ions using two different methods and normalizing using one of two different methods. This work focused on modeling and identifying gasoline samples from the other ignitable liquid classes.
The algorithm is based on general linear modeling and uses mixed stepwise selection, where the 15 most abundant ions in the TIS act, iteratively, as the dependent variable while the remaining 19 ions serve as covariates. For identification of gasoline, EASI models were trained only with known positive gasoline samples from the NCFS-UCF ignitable liquid reference collection (ILRC) that were weathered to varying extents and analyzed on an Agilent GC-MS. The models were validated and tested on hundreds of ground-truth samples from the ILRC collection and in-house data on two different instruments.
The accuracy of spectral predictions were evaluated using both Pearson product moment correlations (PPMCs) and the mean absolute residual (MAR). These measures were then used for binary classification—e.g., gasoline present or not present—and assessed using Receiver Operating Characteristic (ROC) curves. MARs produced the most accurate spectral predictions, as determined by evaluating the area under the ROC curves (AUROC). More specifically, the greatest AUROC of 0.937 was observed for the MAR using ions selected according to suggested extracted ion chromatograms (EICs) from ASTM E1618-19 and normalized by summing the TIS to one. When minimizing the true positive rate (TPR) at 0.8, the trade-off false positive rate (FPR) was 0.058, and when maximizing the FPR at 0.2, the TPR was 0.938. Most of the false positives for gasoline were caused by samples from the aromatic class, which is understandable given their chemical similarity. EASI was also compared to the Mahalanobis distance as a binary classifier using the training set of gasoline TIS data to define the true positive distribution. The AUROC of the Mahalanobis distance using EIC ions from ASTM E1618-19 and normalized by summing TISs to one was 0.571, which is just above the useless decision threshold AUROC value of 0.50. The poor classification for Mahalanobis distance indicates that EASI is a more effective means to model and classify ignitable liquids into their ASTM E1618-19 classifications when each class of ignitable liquid contains correlated fragment ion abundances caused by weathering.
Recommended Citation
McMillen, Hannah L., "Application of the expert algorithm for substance identification (EASI) to predict ASTM E1618-19 ignitable liquid classes from gas chromatography-mass spectrometry data of ignitable liquid residues" (2024). Graduate Theses, Dissertations, and Problem Reports. 12574.
https://researchrepository.wvu.edu/etd/12574
Embargo Reason
Publication Pending