Citation

Feeney W, Menking-Hoggatt K, Arroyo L, Curran J, Bell S, Trejos T. Forensic Chem. 2022; 27: e100389.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.forc.2021.100389

PMID

unavailable

Abstract

This work investigated the prevalence of organic and inorganic gunshot residue within two main subpopulations, 1) non-shooters, including groups with low- and high-risk of potentially containing GSR-like residues, and 2) individuals involved in a firing event (shooters, bystanders, and shooters performing post-shooting activities). The study analyzed over 400 samples via a liquid chromatography-mass spectrometry (LC-MS/MS) methodology with complexing agents. Exploratory statistical tools and machine learning algorithms (neural networks, NN) were used to evaluate the resulting mass spectral and quantitative data. This study observed lower occurrences of OGSR compounds in the non-shooter populations compared to IGSR analytes. The presence of GSR on authentic shooters versus other potential sources of false positives, such as bystanders and professions including police officers, agricultural workers, and mechanics, were further assessed by utilizing machine learning algorithms trained with the observed OGSR/IGSR traces. The probability of false negatives was also estimated with groups who performed regular activities after firing. Additionally, the low-risk background set allowed documentation of GSR occurrence in the general population. The probabilistic outputs of the neural network models were utilized to calculate likelihood ratios (LR) to measure the weight of the evidence. Using both the IGSR and OGSR profiles, the NN model's accuracy ranged from 90 to 99%, depending on the subpopulation complexity. The log-LR histograms and Tippet plots show the method can discriminate between each sub-population and low rates of misleading evidence, suggesting that the proposed approach can be effectively used for a probabilistic interpretation of GSR evidence.

Language: en

Keywords

Gunshot residue; LCMS; Machine learning; Probabilistic interpretation; Survey