Machine Learning Helps in Prediction of Tobacco Smoking in Adolescents

Abstract

Background: Considering the increasing prevalence of adolescent smoking in recent years, this study proposes a machine learning (ML) approach for distinguishing adolescents who are prone to start smoking and those who do not directly confess to smoking.

Methods: We used two repeated measures cross-sectional studies, including data from 7940 individuals as distinct training and test datasets. Utilizing the randomized least absolute shrinkage and selector operator (LASSO), the most influential factors were selected. We then investigated the performance of different ML approaches for the automatic classification of students into smoker/nonsmoker and low-risk/high-risk categories.

Results: Randomized LASSO feature selection prioritized 15 factors, including peer influence, risky behaviors, attitude and school policy toward smoking, family factors, depression, and sex as the most influential factors in smoking. Applying different ML approaches to the three study plans yielded an AUC of up to 0.92, sensitivity of up to 0.88, PPV of up to 0.72, specificity of up to 0.98, and NPV of up to 0.99.

Conclusions: The results showed the capability of our ML approach to distinguish between classes of smokers and nonsmokers. This model can be used as a brief screening tool for automated prediction of individuals susceptible to smoking for more precise preventive intervention plans focusing on adolescents.

Keywords: Adolescent; classification; machine learning; prediction; tobacco.

Figure 1

Description of the three output definition (OD)

Figure 2

Flowchart of the proposed method. The utilized classifier are Logistic Regression Classifier (LRC), Support Vector Machine (SVM), Random Forest Classifier (RFC), Gradient Boosting classifier (GBC), and AdaBoost Classifier (ADB). The performance evaluation metrics are: Area Under Curve (AUC), Sensitivity, Specificity, Positive Predictive Value (PPV) and Negative Predictive Value (NPV)

Figure 3

Average of the answers to the 15 selected variables, in training and test data sets. PI, Peer Influences; SP, School Policies toward smoking; A, Attitude toward smoking; F, Family factor; RB, Risky Behaviors; D, Depression; and S, Sex. Figures (I) and (II) are related to the study population of 2010 (Training samples) and, Figures (III) and (IV) are related to the study population of 2015 (Test samples)