Evaluation and Classification of Fentanyl Related Compounds using EC-SERS and Machine Learning

• NCJ Number: 308300
• Journal: Journal of Forensic Sciences Volume: 68 Issue: 5 Dated: 22 May 2023 Pages: 1520-1526
• Author(s): Travon Cooman; Colby E. Ott; Luis E. Arroyo
• Date Published: May 2023
• Length: 7 pages
• Grant Number(s): 2019-DU-BX-0030
• Agencies: NIJ-Sponsored
• Publication Type: Research (Applied/Empirical)
• Award: Fast On-site Screening of Seized Drugs by Electrochemical and Spectroscopic Tools: Identification of Fentanyl and Novel Psychoactive Substances

Annotation

This paper reports on an evaluation of fentanyl-related compounds and other drugs of abuse using EC-SERS and to process the acquired data using machine learning-convolutional neural networks.

Abstract

Multiple analytical techniques for the screening of fentanyl-related compounds exist. High discriminatory methods such as GC–MS and LC–MS are expensive, time-consuming, and less amenable to onsite analysis. Raman spectroscopy provides a rapid, inexpensive alternative. Raman variants such as electrochemical surface-enhanced Raman scattering (EC-SERS) can provide signal enhancements with 1010 magnitudes, allowing for the detection of low-concentration analytes, otherwise undetected using conventional Raman. Library search algorithms embedded in instruments utilizing SERS may suffer from accuracy when multicomponent mixtures involving fentanyl derivatives are analyzed. The complexing of machine learning techniques to Raman spectra demonstrates an improvement in the discrimination of drugs even when present in multicomponent mixtures of various ratios. Additionally, these algorithms are capable of identifying spectral features difficult to detect by manual comparisons. Therefore, the goal of this study was to evaluate fentanyl-related compounds and other drugs of abuse using EC-SERS and to process the acquired data using machine learning—convolutional neural networks (CNN). The CNN was created using Keras v 2.4.0 with Tensorflow v 2.9.1 backend. In-house binary mixtures and authentic adjudicated case samples were used to evaluate the created machine-learning models. The overall accuracy of the model was 98.4 ± 0.1 percent after 10-fold cross-validation. The correct identification for the in-house binary mixtures was 92 percent, while the authentic case samples were 85 percent. The high accuracies achieved in this study demonstrate the advantage of using machine learning to process spectral data when screening seized drug materials comprised of multiple components. (Published Abstract Provided)