Identifying Key Predictors of Smoking Cessation Success: Text-Based Feature Selection Using a Large Language Model

Area of Science:

• Tobacco control research
• Artificial intelligence in health research
• Behavioral science

Background:

• Smoking cessation is vital for reducing mortality and morbidity.
• Despite high quit attempt rates, success rates remain low (around 10%).
• Identifying predictors of success can enhance intervention efficacy.

Purpose of the Study:

• To identify key predictors of 12-month smoking abstinence.
• To evaluate the efficacy of AI (GPT-4.1) in variable selection for smoking cessation research.
• To derive actionable insights from top predictive variables.

Main Methods:

• Analysis of Population Assessment of Tobacco and Health (PATH) study data (Waves 5 & 6).
• Utilized OpenAI's GPT-4.1 for initial selection of 45 predictive variables from textual descriptions.
• Validated variable importance using eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP).

Main Results:

• XGBoost models showed nearly identical performance whether trained on all variables or GPT-4.1 selected variables (AUC: ~0.75).
• Top predictors included smoking frequency, time to first cigarette, social influences (peers, important others), and emotional dependence.
• Concerns about health harms and daily electronic nicotine product use also emerged as significant factors.

Conclusions:

• GPT-4.1 demonstrated efficient and effective variable selection for smoking cessation predictors.
• The identified top variables align with known risk factors, offering refined targets for interventions.
• AI integration in tobacco research can optimize resource allocation for targeted cessation strategies.