Applying machine learning to international drug monitoring: classifying cannabis resin collected in Europe using cannabinoid concentrations

Tom P Freeman; Edward Beeching; Sam Craft; Marta Di Forti; Giampietro Frison; Christian Lindholst; Pieter E Oomen; David Potter; Sander Rigter; Kristine Rømer Thomsen; Luca Zamengo; Andrew Cunningham; Teodora Groshkova; Roumen Sedefov

doi:10.1007/s00406-024-01816-w

Applying machine learning to international drug monitoring: classifying cannabis resin collected in Europe using cannabinoid concentrations

Eur Arch Psychiatry Clin Neurosci. 2024 May 21. doi: 10.1007/s00406-024-01816-w. Online ahead of print.

Authors

Affiliations

¹ Addiction and Mental Health Group (AIM), Department of Psychology, University of Bath, Bath, UK. t.p.freeman@bath.ac.uk.
² Hugging Face, Paris, France.
³ Addiction and Mental Health Group (AIM), Department of Psychology, University of Bath, Bath, UK.
⁴ Department of Social, Genetic and Developmental Psychiatry, Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, UK.
⁵ Cannabis Clinic for Psychosis, South London and Maudsley Foundation Trust, London, UK.
⁶ Laboratory of Clinical and Forensic Toxicology, DMPO Department, AULSS 3, Venice, Italy.
⁷ Section for Toxicology and Drug Analysis, Department of Forensic Medicine, Aarhus University, Aarhus, Denmark.
⁸ Trimbos Institute, The Netherlands Institute of Mental Health and Addiction, Utrecht, Netherlands.
⁹ , Canterbury, UK.
¹⁰ Centre for Alcohol and Drug Research, Department of Psychology and Behavioural Sciences, Aarhus University, Aarhus, Denmark.
¹¹ European Monitoring Centre for Drugs and Drug Addiction (EMCDDA), Lisbon, Portugal.

PMID: 38771330
DOI: 10.1007/s00406-024-01816-w

Abstract

In Europe, concentrations of ∆⁹-tetrahydrocannabinol (THC) in cannabis resin (also known as hash) have risen markedly in the past decade, potentially increasing risks of mental health disorders. Current approaches to international drug monitoring cannot distinguish between different types of cannabis resin which may have contrasting health effects due to THC and cannabidiol (CBD) content. Here, we compared concentrations of THC and CBD in different types of cannabis resin collected in Europe (either Moroccan-type, or Dutch-type). We then tested the ability of machine learning algorithms to classify the type of cannabis resin (either Moroccan-type, or Dutch-type) using routinely collected monitoring data on THC and CBD. Finally, we applied the optimal algorithm to new samples collected in countries where the type of cannabis resin was unknown, the UK and Denmark. Results showed that overall, Dutch-type samples had higher THC (Hedges' g = 2.39) and lower CBD (Hedges' g = 0.81) than Moroccan-type samples. A Support Vector Machine algorithm achieved classification accuracy exceeding 95%, with little variation in this estimate, good interpretability, and plausibility. It made contrasting predictions about the type of cannabis resin collected in the UK (94% Moroccan-type; 6% Dutch-type) and Denmark (36% Moroccan-type; 64% Dutch-type). In conclusion, we provide proof-of-concept evidence for the potential of machine learning to inform international drug monitoring. Our findings should not be interpreted as objective confirmatory evidence but suggest that Dutch-type cannabis resin has higher THC concentrations than Moroccan-type cannabis resin, which may contribute to variation in drug markets and health outcomes for people who use cannabis in Europe.

Keywords: Cannabidiol; Delta-9-tetrahydrocannabinol; Drug policy; Psychiatric disorders.