Using machine learning techniques for the classification of ultra-low concentrations of cannabis in biological fluids

In this work, the application of three different Machine Learning algorithms, random forest (RF), support vector machine (SVM), and artificial neural network (ANN), to accurately classify ultra-low concentrations of Δ9-tetrahydrocannabinol in biological fluids such as saliva was successfully demonstrated. In doing so, experimental data consisting of the voltammetry signals of 0, 2, and 5 ng/mL of Δ9-tetrahydrocannabinol (THC) in synthetic and human saliva was employed. The dataset consists of person-to-person saliva variation and experimental variabilities in the procedure and artifacts. The results showed that RF was the most robust ML technique to classify different concentration levels of THC in the presence of a variety of variabilities in the experimental data, with an average training accuracy of 96% and an average testing accuracy of 87%. On the other hand, SVM and ANN models were susceptible to incoherencies in the experimental data.