A machine learning approach to predict e-cigarette use and dependence among Ontario youth

Abstract

Introduction We developed separate random forest algorithms to predict e-cigarette (vaping) ever use and daily use among Ontario youth, and subsequently examined predictor importance and statistical interaction. Methods This cross-sectional study used a representative sample of Ontario elementary and high school students in 2019 (N = 6471). Vaping frequency over the last 12 months was used to define ever-vaping and daily vaping. We considered a large set of individual characteristics as potential correlates for ever-vaping (176 variables) and daily vaping (179 variables). Using cross-validation, we developed random forest algorithms and evaluated model performance based on the C-index, a measure to assess the discriminatory ability of a model, for both outcomes. Further, the top 10 correlates were identified by relative importance score calculation and their interaction with sociodemographic characteristics. Results There were 2064 (31.9%) ever-vapers, and 490 (7.6%) of the respondents were daily users. The random forest algorithms for both outcomes achieved high performance, with C-index over 0.90. The top 10 correlates of daily vaping included use of caffeine, cannabis and tobacco, source and type of e-cigarette and absence in last 20 school days. Those of ever-vaping included school size, use of alcohol, cannabis and tobacco; 9 of the top 10 ever-vaping correlates demonstrated interactions with ethnicity. Conclusion Machine learning is a promising methodology for identifying the risks of ever-vaping and daily vaping. Furthermore, it enables the identification of important correlates and the assessment of complex intersections, which may inform future longitudinal studies to customize public health policies for targeted population subgroups.

Are the relevant risk factors being adequately captured in empirical studies of smoking initiation? A machine learning analysis based on the Population Assessment of Tobacco and Health study

“…Thus, the rankings of predictors produced by the RF classifier may be biased due to variable correlations [28, 29]. Focusing on justifying only a few top RF-ranked variables as the most important predictors of the interesting outcomes like in previous studies [14, 15, 18] is unreliable. Therefore, we considered all variables relevant to the outcome variables.…”

“…In tobacco control, a growing number of research articles have focused on predicting smoking and vaping behaviors and identifying the most significant predictors of such behaviors through ML algorithms. For instance, researchers used ML methods such as RF classifiers and penalized logistic regression models to predict ENDS use status and identify the most important predictors of this behavior based on various surveys [13-15]. Coughlin et.…”

“…Most studies used data from clinical trials or local representative surveys [13, 15, 16, 18]. Therefore, the findings may not be generalizable to the whole US population.…”

Towards a Canadian evidence base to inform action to prevent and control vaping in Canada

“…Cynthia Callard, MM (1); Thierry Gagné, PhD (2); Jennifer L. O'Loughlin, PhD (3,4) Tweet this article Pound et al 7 describe a simulation study showing the relative impact of ENDS on population health across contrasting regulatory scenarios, including a complete ban and a prescription-only scenario (i.e. wherein vaping is available to smokers only).…”

“…Two papers, one by Ahmad et al 3 and one by Shi et al, 4 identify determinants of vaping initiation and daily use among Canadian youth. These papers indicate that key determinants of youth vaping in Canada likely include ease of access in addition to the constellation of vulnerabilities underpinning substance use in general, as evidenced by the close associations between vaping and other risk-taking behaviours such as cigarette smoking and use of alcohol, energy drinks and marijuana.…”