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

Background: The most effective way to reduce mortality and morbidity among current smokers is to quit smoking. Although about half of smokers attempted to quit, only one-tenth succeeded in 2022.

Objective: To identify key predictors of smoking cessation success to inform cessation interventions and increase quitting rates.

Methods: We analyzed data from waves 5 and 6 of the Population Assessment of Tobacco and Health (PATH) study (December 2018 to November 2021). Using OpenAI's GPT-4.1, we identified the top 45 variables from wave 5 that are highly predictive of 12-month smoking abstinence in wave 6, based on descriptions of survey variables. We then validated the predictive power of the GPT-4.1-selected variables by comparing the performance of eXtreme Gradient Boosting (XGBoost) trained on different sets of variables. Finally, we derived insights into the top 10 variables, ranked according to their SHapley Additive exPlanations values.

Results: The performance of XGBoost trained with all possible wave 5 variables and the 45 selected variables was almost identical (AUC:0.749 vs AUC:0.752). The top 10 variables included past 30-day smoking frequency, minutes from waking up to smoking first cigarette, important people's views on tobacco use, prevalence of tobacco use among close associates, daily electronic nicotine product use, emotional dependence, and health harm concerns.

Conclusion: This study demonstrates the ability of OpenAI's GPT-4.1 to identify the top 45 PATH wave 5 variables associated with 12-month smoking abstinence using only their descriptions. This approach could help researchers design more effective survey questionnaires and improve efficiency of data collection.