JMIR Publications

ABSTRACT

Background:

We aimed to address gaps in global understanding of cultural and social variations by employing a high-performance machine learning model to predict adolescent substance use across three national datasets.

Objective:

This study aims to develop a predictive model for adolescent substance use using multinational datasets and machine learning(ML).

Methods:

The study utilized the Korea Youth Risk Behavior Web-Based Survey(KYRBS) from South Korea(n=1,145,178) to train ML models. For external validation, we employed the Youth Risk Behavior Survey(YRBS) from the USA(n=1,690,108) and Norwegian nationwide Ungdata surveys(Ungdata) from Norway(n=793,879). After developing diverse tree-based models, we further evaluated feature importance.

Results:

The study utilized nationwide adolescent datasets for ML model development and validation, analyzing data from 1,145,178 KYRBS adolescents, 54,709 YRBS Asian subset participants, and 720,812 from Ungdata. The random forest model was the top performer on the KYRBS, achieving an AUROC of 80.8%(95% CI, 80.7-80.8) with sensitivity of 72.9%(72.8-73.0), specificity of 72.9%(72.9-73.0), accuracy of 72.9%(72.8-73.0), and balanced accuracy of 72.9%(72.8-73.0). The model's AUROC scores were 73.2% for YRBS and 75.7% for Ungdata in external validation. The top features for predicting substance use were smoking status, body mass index (BMI), and alcoholic consumption.

Conclusions:

With multinational datasets from South Korea, USA, and Norway, the findings of this study underscore the potential efficacy of ML models in predicting adolescent substance use with smoking status, body mass index, and alcoholic consumption identified as key predictors. The random forest model exhibited notable performance in this prediction. These findings could be a basis for future studies exploring more comprehensive factors influencing adolescent substance use or developing intervention strategies based on these predictors.