RT journal article
T1 Machine learning approaches over ion mobility spectra for the discrimination of ignitable liquids residues from interfering substrates
A1 Pérez Calle, José Luis
A1 Falatová, Barbara
A1 Aliaño González, María José
A1 Ferreiro González, Marta
A1 Palma Lovillo, Miguel
A2 Química Analítica
K1 Machine learning algorithms
K1 Fire debris
K1 Ignitable liquids
K1 Ion mobility spectrometry
K1 Petroleum-derived substrates
AB In arson fires, ignitable liquids (ILs) are frequently used to start combustion. For this reason, detecting IL residues (ILRs) at the fire scene is a key factor in fire investigation to determine whether a crime has been committed as well as to establish the modus operandi of the perpetrator. In the present study, the application of headspace-gas chromatography-ion mobility spectrometry (HS-GC-IMS) for the detection of ILRs in fire debris from complex matrices in combination with machine learning (ML) tools is proposed as an alternative to the traditional method, based on gas chromatography–mass spectrometry (GC-MS), described by the ASTM E1618 standard method. For this purpose, petroleum-derived substrates (vinyl, nylon, and polyester) and natural substrates (cotton, cork and linoleum) burned alone and with different ILs (gasoline, diesel, ethanol and charcoal starter with kerosene) were used. In addition, samples were taken at different times (0, 1, 6, 12, 24 and 48 h) after the fire was finished. The ion mobility sum spectrum (IMSS) of each sample was obtained and different ML algorithms were applied. The first derivative was performed at the IMSS, as well as a Savitzky-Golay filter. Hierarchical cluster analysis (HCA) revealed a clustering trend as a function of substrate and ILs used, where the studied sampling times did not affect the resulting clusters. The classification models for the detection of the presence of ILRs have high performance with an accuracy of 100% for support vector machines (SVM) and random forest model (RF), followed by linear discriminant analysis (LDA) with an accuracy of 86.67%. When discriminating the type of ILs used, the RF model obtained an accuracy of 100%, followed by the LDA with 97.22% and finally the SVM model with an accuracy of 93.06%. In addition, a simple web application has been developed where the trained models can be used, so any researcher can apply the method to detect ILRs in fire debris.
PB ELSEVIER
SN 2666-8319
YR 2022
FD 2022
LK http://hdl.handle.net/10498/27752
UL http://hdl.handle.net/10498/27752
LA eng
DS Repositorio Institucional de la Universidad de Cádiz
RD 13-nov-2026