The Use of Machine Learning to Predict Offline Adolescent e-Cigarette Use: a Proof-of-Concept

Identification of adolescent e-cigarette use could inform prevention and intervention programming and reduce associated consequences. One way to predict those engaging in use is by examining social media profiles and metrics. Most studies examining substance use content on social media employ self-report or human coding that have methodological limitations. Thus, the current study developed a supervised machine learning algorithm to classify participants into e-cigarette use categories based on Instagram metrics. Participants (n = 67, M_age = 18.27; 64.2% female, 82.1% Hispanic/Latino[a/x], 91% White) in the study provided their Instagram data downloaded through the app. Instagram metrics (i.e., number of followers, number following, number of liked comments, number of liked posts, number of posts, and number of messages) were extracted and included as input features in the model. Adolescents reported their e-cigarette use on a self-report measure. A classification tree method was used to classify participants as engaging in e-cigarette use or not. Data was partitioned into a training and test set using stratified sampling. All analyses were performed in Python. Three input features (number of followers, number of liked posts, and number of messages) were selected through hyperparameter-optimized feature selection. The final model accurately detected e-cigarette use 71% of the time. Findings indicate that supervised learning can predict adolescent e-cigarette use with accuracy consistent with other clinical populations. This study establishes that universal aspects of social media may be harbingers for policy makers and tech companies to provide targeted support and messaging.