New perspective for the in-field analysis of cannabis samples using handheld near-infrared spectroscopy: A case study focusing on the determination of Δ9-tetrahydrocannabinol

The aim of the present study was to explore the feasibility of applying near-infrared (NIR) spectroscopy for the quantitative analysis of Δ9-tetrahydrocannabinol (THC) in cannabis products using handheld devices. A preliminary study was conducted on different physical forms (entire, ground and sieved) of cannabis inflorescences in order to evaluate the impact of sample homogeneity on THC content predictions. Since entire cannabis inflorescences represent the most common types of samples found in both the pharmaceutical and illicit markets, they have been considered priority analytical targets. Two handheld NIR spectrophotometers (a low-cost device and a mid-cost device) were used to perform the analyses and their predictive performance was compared. Six partial least square (PLS) models based on reference data obtained by UHPLC-UV were built. The importance of the technical features of the spectrophotometer for quantitative applications was highlighted. The mid-cost system outperformed the low-cost system in terms of predictive performance, especially when analyzing entire cannabis inflorescences. In contrast, for the more homogeneous forms, the results were comparable.

The mid-cost system was selected as the best-suited spectrophotometer for this application. The number of cannabis inflorescence samples was augmented with new real samples, and a chemometric model based on machine learning ensemble algorithms was developed to predict the concentration of THC in those samples. Good predictive performance was obtained with a root mean squared error of prediction of 1.75% (w/w). The Bland-Altman method was then used to compare the NIR predictions to the quantitative results obtained by UHPLC-UV and to evaluate the degree of accordance between the two analytical techniques. Each result fell within the established limits of agreement, demonstrating the feasibility of this chemometric model for analytical purposes.

Finally, resin samples were investigated by both NIR devices. Two PLS models were built by using a sample set of 45 samples. When the analytical performances were compared, the mid-cost spectrophotometer significantly outperformed the low-cost device for prediction accuracy and reproducibility.