Drug classification with IR spectra: Evaluating the effect of analyst decisions in the context of simpler and harder classification challenges

Résumé fourni par la source

We studied the effect of three specific analyst decisions when computationally classifying single component drug samples by their infrared (IR) spectra: (i) how spectra are encoded as vectors, (ii) whether dimensionality is reduced, and (iii) what classification algorithm is used. We evaluated these decisions in the context of two classification tasks representing different levels of challenge: (i) the easier distinct-class challenge, and (ii) the appreciably more difficult fuzzy-class challenge. All spectra used to evaluate these challenges were sourced from the most recent version of the SWGDRUG IR library (Version 3.1). Eight models (combinations of analyst decision variables) were created for each task, and 10000 trials were completed using each model. All eight models performed well for the distinct-class challenge, with one model configuration having a median accuracy of 100%. The most important driver for accuracy with the distinct-class challenge was choice of classification algorithm. The models did not perform as well with the fuzzy-class challenge, with median classification accuracies ranging between 50% and 75%. However, we found that all three choices affected classification accuracy, with a strong interaction between choice of encoding and whether to employ dimensionality reduction. We posit that developing robust methods for feature selection will be important when trying to classify more closely related drugs. We also suggest that deeper exploration of how we label drugs (i.e., define classification labels) is necessary to meaningfully evaluate classification procedures. All data and source code used to prepare this article are freely available at https://github.com/Brandon12G/GiotisNumericalTreatments.