Substance use among university students is a growing health concern that is often overlooked until it escalates into a full-grown disorder. This study presents a multiclass machine learning model for predicting substance use risk levels based on DSM-5 diagnostic criteria and psychosocial factors such as trauma, academic stress and social networks. Data were collected through a survey answered by university students, the resulting dataset was used to train and compare multiple models. After performing feature selection, class balancing and hyperparameter tuning, the best performing and most accurate model, was a logistic-regression model that achieved a macro F1-score of 0.946. More notably however, the model showed improved sensitivity for mild-risk cases, which tend to go underdetected in binary classification schemes. These results support the integration of clinically based machine learning models, into educational institutions health protocols.