Behavior Change Techniques Associated With Abstinence in a Chat-Based Smoking Cessation Intervention: Secondary Analysis of a Randomized Clinical Trial Using Machine Learning
Journal of Medical Internet Research ·
Background: Isolating the candidate intervention components that drive the success of digital behavioral interventions is crucial for understanding intervention mechanisms and optimizing future interventions. Objective: This study aimed to identify behavior change techniques (BCTs) delivered in a chat-based smoking cessation intervention and determine their independent associations with biochemically validated abstinence. Methods: This study is a secondary analysis of a pragmatic, cluster-randomized clinical trial conducted in Hong Kong. From June 18 to September 30, 2017, 624 adult daily smokers were recruited from 34 community sites and randomized to the intervention group. Participants received 3 months of personalized, real-time chat-based cessation support delivered by a live counselor via a mobile instant messaging app. The analytic sample included 230 participants (n=170, 73.9% male; mean age 43.0, SD 13.6 years) who engaged with the chat support. All chat dialogues during the real-time support were retrieved and systematically coded using an established BCT taxonomy for smoking cessation (BCTTsm). Four machine learning methods, least absolute shrinkage and selection operator regression, Extreme Gradient Boosting, the Boruta algorithm, and recursive feature elimination, were used to identify the BCTs that are most predictive of biochemically validated tobacco abstinence outcomes at 3 and 6 months. A prespecified supermajority rule retained only BCTs selected by at least 3 of the 4 models for subsequent analysis. The stability of the selected BCTs was assessed using 500 bootstrap resamples for each model. Poisson regression with robust variance was subsequently used to examine the independent associations between the selected BCTs and abstinence outcomes. Covariates included sex, age, educational level, nicotine dependence, intention to quit in 30 days, and quit attempts at baseline. Results: Systematic coding identified 35 distinct BCTs from the BCTTsm as having been delivered. Three BCTs were consistently selected by machine learning as the most predictive of both 3- and 6-month abstinence, with moderate to high stability (ranging from 61% to 100%): “provide rewards contingent on successfully stopping smoking” (BM4), “provide information on withdrawal symptoms” (RC6), and “provide information on consequences of smoking and smoking cessation” (BM1). In multivariable analyses, 6-month abstinence was significantly associated with BM4 (risk ratio [RR] 7.9, 95% CI 4.2-14.6), RC6 (RR 3.3, 95% CI 1.6-6.5), and BM1 (RR 2.7, 95% CI 1.4-5.5), with consistent results for 3-month abstinence. Conclusions: Three BCTs, contingent rewards (BM4), information on withdrawal symptoms (RC6), and information on consequences (BM1), were identified as candidate intervention components associated with success in a chat-based smoking cessation intervention. This study is the first to apply machine learning for BCT identification, moving beyond simple frequency counts to data-driven feature selection. These findings provide empirical evidence for the candidate components of effective digital behavioral intervention that may warrant further testing and offer guidance for future intervention optimization. Trial Registration: ClinicalTrials.gov NCT03182790; https://clinicaltrials.gov/study/NCT03182790