A Machine Learning Approach to Classifying Harm Reduction Street Substance Categories

<span style="font-weight: 400;">This study aimed to use harm reduction related variables to categorize street substances in order for greater ease in education, research and prevention of overdoses. The categorization process was achieved through statistical modeling and machine learning, and included variables relating to the chemical and pharmacological properties, contextual usage characteristics, prevalence, and co-occurrence of substances in our sample set. The statistical modeling approach included two feature selection methods accompanied by Latent Class Analysis (LCA) and Latent Profile Analysis (LPA). The machine learning approach experimented with dimension reduction techniques including Uniform Manifold Approximation and Projection (UMAP) to reduce dimensions, and K-Means clustering to achieve the distinct drug classes. Each aforementioned approach revealed some level of class separation, although interpretability of classes was a challenge due to low sample size. As the street drug supply is in constant flux, manual categorization is no longer an efficient or comprehensive method to understand the supply. Our approach to modeling drug categories using machine learning, sets the stage for an on-going, low maintenance method to model categories. Follow up studies can be done using more comprehensive chemical data for the creation of more accurate substance classes using minimal harm reduction information. </span>