To include more information later!!
This project was submitted to NeurIPS in 2025. Code to be uploaded later.
Summary: Use of illicit or illegal substances in adolescents is a substantial public health concern, but reliably detecting substance use in non-anonymous clinical settings is challenging due to under-reporting. This study introduces a neural network approach using ordinal embeddings of health-related behaviors to predict substance use risk. We propose that adjacent health-related behaviors can predict substance use, and that a model with ordinal embeddings can successfully distinguish between adolescents using substances and non-users with enough success to be clinically feasible. We also demonstrate improved generalizability to unseen data when including ordinal embedding in the model compared to traditional one-hot encoding of data, at least within the context of predicting adolescent substance use.