Predictors of Smoking Cessation in the U.S.: A Machine Learning Analysis

Project Title: Predictors of Smoking Cessation in the U.S.: A Machine Learning Analysis Revealing the Socioeconomic Paradox in BRFSS Data (2018-2023)

Team Memmbers: Armin KHoojavi, Hamed Hesami, Mahdieh Ebrahimi

Project Type & Duration: Team Project (Academic Project), Summer 2025

Objective

This research aimed to provide a deep, multifaceted, and stable understanding of factors influencing successful smoking cessation in adults. Utilizing extensive BRFSS survey data from a six-year period (2018-2023), the goal was to derive data-driven evidence for designing personalized public health interventions.

Methodology

This longitudinal study leveraged a suite of machine learning models for smoking cessation prediction, including:

• Logistic Regression
• LDA
• Decision Tree
• LightGBM
• CatBoost
• XGBoost

Model performance was rigorously evaluated comparing a "full features" strategy against a "VIF-based feature reduction" approach. To enhance prediction accuracy and stability, comprehensive modeling techniques were implemented:

• Data Pooling
• Temporal Ensemble
• Stacking Ensemble (specifically utilized for its superior predictive power)

The influence and directionality of each factor were assessed using SHAP (SHapley Additive exPlanations) analysis, also examining their stability over time.

Key Contributions

• Developed and evaluated multiple advanced machine learning models, with the Stacking Ensemble model achieving superior overall performance:
  • AUC: 0.7881
  • Accuracy: 0.7446
  • F1-Score: 0.7330
• Identified and validated remarkably stable key predictors of smoking cessation across the six-year study period. These included:
  • Older age (consistently the strongest factor increasing successful cessation probability).
  • Being married.
  • Higher levels of income and education.
  • Higher BMI and increased weight.
• Contributed to robust data-driven insights for public health policy and tobacco control by emphasizing the necessity of personalized, multi-faceted interventions. Insights focused on specific age groups and an integrated approach to supporting mental health, physical health, and socioeconomic factors.

Technologies Used

• Programming Languages: Python
• Machine Learning Libraries: scikit-learn, LightGBM, CatBoost, XGBoost
• Interpretability: SHAP
• Data Manipulation & Analysis: Pandas, NumPy
• Data Visualization: Matplotlib, Seaborn