Quantitative profile–profile relationship (QPPR) modelling: a novel machine learning approach to predict and associate chemical characteristics of unspent ammunition from gunshot residue (GSR)

Abstract

Evidence association in forensic cases involving gunshot residue (GSR) remains very challenging. Herein, a new in silico approach, called quantitative profile–profile relationship (QPPR) modelling, is reported. This is based on the application of modern machine learning techniques to predict the pre-discharge chemical profiles of selected ammunition components from those of the respective post-discharge GSR. The obtained profiles can then be compared with one another and/or with other measured profiles to make evidential links during forensic investigations. In particular, the approach was optimised and successfully tested for the prediction of GC-MS profiles of smokeless powders (SLPs) from organic GSR in spent cases, for nine ammunition types. Results showed a high degree of similarity between predicted and experimentally measured profiles, after adequate combination and evaluation of fourteen machine learning techniques (median correlation of 0.982). Areas under the curve (AUCs) of 0.976 and 0.824 were observed after receiver operating characteristic (ROC) analysis of the results obtained in the comparisons between predicted–predicted and predicted–measured profiles, respectively, in the specific case that the ammunition types of interest were excluded from the training dataset (i.e., extrapolation). Furthermore, AUCs of 0.962 and 0.894 were observed in interpolation mode. These values were close to those of the comparison of the measured SLP profiles between themselves (AUC = 0.998), demonstrating excellent potential to correctly associate evidence in a number of different forensic scenarios. This work represents the first time that a quantitative approach has successfully been applied to associate a GSR to a specific ammunition.