Comparing numerical treatments for drug classification using IR spectra

In this article, we investigate the numerical relationships between illicit drugs based on IR spectra from a publicly available library. Classification models were trained to test how well single component drugs can be classified using their IR spectra alone. Two challenges were considered: classifying spectra representing various amphetamine, fentanyl, and cocaine compounds, and classifying spectra specifically of fentanyl analogs. All spectra were sourced from the most recent version of the SWGDRUG IR library (Version 3.1). Multiple classification models with varied parameters were created for each task. The models for challenge 1 performed very well, with several models having a median reported accuracy of 100% The models did not perform well with the appreciably more difficult challenge 2, with median classification accuracies ranging between 50% and 75%. Our results demonstrate the appropriateness of classifying drugs based on IR spectra at the major class level, while confirming the limitation of trying to use IR spectroscopy for analog level discrimination. Efforts to determine the “bounds” of drug classification using IR spectra is on-going in our lab. All source code used to generate the results presented in this article are available at https://github.com/Brandon12G/GiotisNumericalTreatments.