Application of knowledge-driven ensemble algorithms to improve preventative measures for underage tobacco consumption in the U.S.

Tobacco consumption by underage individuals remains a significant public health concern. Despite strict regulations and public awareness campaigns, many adolescents experiment with tobacco products, including cigarettes and smokeless tobacco. Previous research focuses heavily on adult datasets, broad global trends, and single tobacco delivery methods. Our model fills a knowledge gap by examining fine-scale predictors of American adolescent consumption aggregated across 12 tobacco delivery methods. Additionally, we analyzed surveys collected after the federal ruling to increase the tobacco purchasing age from 18 to 21. We explored the roles of environment, socioeconomic status, and identity on youth tobacco use with the National Youth Tobacco Survey, a comprehensive, diverse dataset from 341 schools across the United States. Our best predictive model, trained with one-hot-encoded data, achieved an F1 accuracy score of 95.78% with the Random Forest classifier, and a score of 92.84% when applied to data from another year. This accuracy is higher compared to other related studies. Our top ten predictors from this model ranked by permutation feature importance performed at 91.54% accuracy and 91.09% with our validation data. These predictors relate to media and advertising (‘Seen Warning Label on E-Cig—Did Not See,’ ‘Seen Tobacco Ads in Store—Sometimes,’ ‘Seen Tobacco Used by On-Screen Characters—Sometimes,’ ‘Seen Tobacco Ads in Store—Rarely,’ and ‘Seen Anti-Tobacco Ads (12 months)—Yes’), age (‘Age—17′), grades (‘Grades (12 months)—Ave B’), and perceptions of tobacco (‘Tobacco Consumption is Dangerous—Strongly Agree,’ ‘Tobacco Consumption is Dangerous—Disagree,’ and ‘Addictiveness: Cigars vs Cigs—Unsure’). Our work demonstrates that these categories should be monitored by health practitioners for early intervention in high-risk patients.