Using games to identify adolescents at risk for substance misuse—a proof-of-concept study

Introduction Early identification of adolescents at risk for substance use is critical for timely intervention. However, standard screening tools like CRAFFT (Car, Relax, Alone, Forget, Family and Friends, Trouble) and S2BI (Screening to Brief Intervention) face significant barriers, including adolescent disclosure reluctance, limited clinic privacy, and administrative challenges with paper-based forms. Digital games offer a promising alternative by generating behavioral data that may serve as digital biomarkers of substance use risk through engaging, low-burden gameplay. The objective of this proof-of-concept study was to explore the utility of game log data collected during gameplay to predict substance use. Methods We analyzed game log data from 160 adolescents aged 11–14 years who played an HIV prevention game targeting high-risk behaviors including substance use, drawn from a larger randomized controlled trial ( N = 333) conducted in schools. We extracted 240 gameplay-derived behavioral metrics reflecting executive function, decision-making, and inhibitory control—cognitive domains affected by substance use. Machine learning models (Support Vector Machine, Logistic Regression, Gradient Boosting, Neural Networks, Decision Tree, Random Forest) were trained to predict lifetime substance use and drug-refusal self-efficacy. A validation study with 36 newly recruited high school students tested associations with standardized measures of emotion regulation, impulsivity, and cognitive performance. Model performance was assessed using discrimination (e.g., AUC) and clinical utility (e.g., sensitivity) metrics. Results Predictive accuracy across all models was insufficient for clinical translation. For drug use prediction, AUC (SD) ranged from 0.458 (0.11) to 0.593 (0.07) (mean AUC <0.6 across all models), sensitivity ranged from 0 to 0.265 (0.18), and F-1 scores ranged from 0.398 (0.0) to 0.482 (0.08). For drug-refusal self-efficacy prediction, AUC ranged from 0.425 (0.10) to 0.592 (0.09), sensitivity ranged from 0.386 (0.14) to 0.785 (0.19), and F-1 scores ranged from 0.431 (0.06) to 0.598 (0.07). No model achieved the minimum threshold (AUC ≥0.7) suggested for clinical utility. Discussion Our study elucidates the challenges associated with extracting behavioral markers from naturalistic digital environments. We discuss the potential of game-based digital biomarkers as scalable, low-burden tools for screening and monitoring, and the limitations of our study that can inform future studies seeking to understand the feasibility of using in-game data as digital biomarkers of substance use risk in adolescents.