Substance Use Prediction Using Artificial Intelligence Techniques

The study examines whether artificial intelligence can reliably predict the use of specific substances in Finland, including cannabis, ecstasy, amphetamine, cocaine and non-prescribed medicines. Using data from the 2022 Finnish National Drug Survey, the authors model substance use based on demographic characteristics, health status, prior drug experiences and social environment.

They evaluate a wide range of models, from conventional machine learning methods to more advanced deep learning (LSTM) architectures. After addressing the imbalance between users and non-users in the data, all models demonstrate very high performance, with accuracies typically around 98–99%. In several cases, traditional machine learning models perform slightly better than deep learning models.

To interpret the results, the authors apply SHAP analysis to identify the most influential predictors. For cannabis use, key factors include having friends who use drugs, receiving drug offers, use of snus, vaping and more permissive attitudes toward drugs and their risks. For other substances, such as ecstasy, important patterns involve co-occurring health problems, polydrug use, prescribed medications, place of residence and risk-taking attitudes.

The authors conclude that AI-based models show strong potential as tools for drug monitoring and targeted prevention efforts. However, they emphasise that the findings are based on self-reported data from a single national context and on resampling-intensive methods, and therefore should be interpreted with caution and validated in future research.