Semi-Quantitative Estimation of MDMA Tablet Dosage and Cocaine/Ketamine Purity Using a Simple to Operate Field-Portable Device

A low-cost, field-portable device using Hybridized Spectral Fingerprinting (HSF) and deep learning can rapidly screen illicit drug samples for MDMA dosage and cocaine/ketamine purity. A convolutional neural network trained on 62 GC-EI-MS quantified MDMA tablets classified samples into 1–170 mg or 170–300 mg dosage brackets, achieving 98% accuracy on 195 external test samples. Another model trained on 1H NMR quantified cocaine and ketamine samples provided presumptive identification and semi-quantitative purity estimation with 96% accuracy on 47 external samples. The approach supports harm reduction and police screening by alerting on potentially harmful MDMA tablets and estimating cocaine and ketamine purity.

Study at a glance

Abstract

High-dose MDMA tablets and variable-purity cocaine and ketamine samples are commonly encountered by drug checking services and police forces in the United Kingdom and Europe. At the time of writing, there are few simpleto-operate field-portable technologies that provide semi-quantitative information on MDMA tablet dosage or cocaine and ketamine purity at the point of sampling. We previously reported on the development of a low-cost, fieldportable device that can be used to rapidly screen illicit drug samples through Hybridized Spectral Fingerprinting (HSF); the combined measurement of fluorescence emission and diffuse LED reflectance. Here, we describe the development and testing of deep learning models for both MDMA tablet dosage and cocaine/ketamine purity that can be used in on this device in the field. We used 62 GC-EI-MS quantified MDMA tablet samples to train a convolutional neural network model to classify tablet samples in either 1–170 mg or 170–300 mg MDMA HCl dosage brackets. External testing on 195 positive and negative MDMA samples yielded a 98% device accuracy. We used 1H NMR quantified cocaine and ketamine samples to train a second convolutional neural network model for presumptive identification and semi-quantitative purity estimation of cocaine and ketamine samples, yielding an accuracy of 96% on an external test set of 47 samples. These findings demonstrate that Hybridized Spectral Fingerprinting paired with deep learning can be reliably used to alert on potentially harmful MDMA tablets and estimate the purity of cocaine and ketamine samples, supporting harm reduction activities and police screening or intelligence gathering workflows.