Predictors of smoking cessation outcomes identified by machine learning: A systematic review

“

TL;DR

This systematic review of 12 studies found that machine learning methods, particularly tree-based approaches, can identify predictors of smoking cessation such as delay discounting, self-efficacy, and smoking severity. The most robust predictors spanned seven categories including psychological, neurocognitive, and sociodemographic factors. However, the field is nascent and limited by small samples, reliance on self-report, and lack of neural mechanism inclusion, suggesting opportunities for future research to improve clinical outcomes.

Full summary

5,004 CHARS

**Background:** Smoking is the leading preventable cause of death in the US, contributing to ~500,000 deaths annually and over $300 billion in costs. Although effective treatments exist, relapse rates remain high (~70% within 6 months). Identifying predictors of cessation could enable patient-treatment matching. Machine learning (ML) offers a powerful approach to discover such predictors from complex datasets. This systematic review aimed to characterize ML methods used to identify predictors of smoking cessation outcomes and to identify knowledge gaps.

**Methods:** The review followed PRISMA guidelines and was registered in PROSPERO (CRD42022323796). Searches were conducted through December 9, 2022, in nine databases (MEDLINE, Science Citation Index, Social Science Citation Index, EMBASE, CINAHL Plus, APA PsycINFO, PubMed, Cochrane Central Register of Controlled Trials, IEEE Xplore). Search terms combined ML terms (e.g., "machine learning", "deep learning") with smoking/cessation terms. Inclusion criteria: studies using ML to evaluate cigarette smoking cessation outcomes (smoking status or number of cigarettes) in individuals who smoke; any experimental design (cross-sectional, longitudinal, clinical trials). Exclusion criteria: non-cigarette smoking, non-human studies, gray literature. Two independent reviewers screened titles/abstracts and full texts; conflicts resolved by a third. Data extraction and quality assessment (using the Mixed Method Appraisal Tool, MMAT) were performed by the team.

**Key Results:** The search yielded 4,306 citations; after deduplication (2,025 removed), 2,283 titles/abstracts were screened, 161 full texts sought, and 136 reviewed. 124 studies were excluded (61 for study design, 35 for outcomes, 3 for population, 25 other). Twelve studies met inclusion criteria. Total participants across studies: 40,208 (range 39 to 14,443). Studies were published 2006–2022; 6/12 from the US, others from the Netherlands (2), Canada (1), South Korea (1), China (1), New Zealand (1). Quality varied: MMAT scores ranged from 0% to 100% (one study 0%, one 20%, four 40%, four 60%, one 80%, one 100%). All 12 studies used supervised ML techniques (random forest, classification/regression trees, logistic regression). Seven reported sensitivity/specificity, five reported AUC, one reported positive/negative predictive value, three did not report relevant metrics. Smoking cessation outcomes were assessed by self-report (e.g., point prevalence abstinence, cigarettes per day, relapse) and/or biochemical validation (expired CO, salivary cotinine); only one study used both. Predictors were grouped into seven categories: biomarkers (2 studies; e.g., exhaled CO, neuroimaging), economic/environmental/sociodemographic (9 studies; e.g., gender, race, household income, cigarette availability), engagement (3 studies; e.g., attendance, response rate), neurocognitive (1 study; e.g., delay discounting, memory), physical health-related (7 studies; e.g., BMI, alcohol consumption, sleep quality), psychological (9 studies; e.g., motivation, self-efficacy, affect, perceived stress), and smoking severity/history (8 studies; e.g., Fagerström Test for Cigarette Dependence, age of initiation, quantity). The strongest predictors identified across studies included: delay discounting (correctly identified smoking status of 80% of participants post-treatment in one study), having daily smokers at home, attendance in cessation education, more positive vaping experiences, higher perceived odds of smoking today, increased confidence to avoid smoking, belief that medication is dangerous, Wisconsin Inventory of Smoking Dependence Motives scores (cognitive enhancement, primary dependence, taste/sensory), parent ethnicity, smoking intervals, boredom, male sex, consumption during intervention commencement, engagement, irritability, cigarette availability, exposure to smokers, smoking restrictions, recent alcohol consumption, resting state brain networks, motivation, and education level.

**Clinical Implications:** The review highlights that ML can identify a diverse set of predictors spanning multiple domains, with tree-based methods offering interpretable models. For example, delay discounting emerged as a strong predictor, consistent with the competing neurobehavioral decision systems theory, suggesting that interventions targeting delay discounting could improve cessation outcomes. However, the field is limited by small, homogeneous datasets, reliance on self-report, lack of external validation, and underuse of unsupervised or reinforcement learning. Only one study included neuroimaging biomarkers. The authors recommend future research incorporate larger datasets, biochemical verification, and neural-based decision-making constructs to enhance predictive power and clinical utility. Overall, while ML has not yet dramatically improved smoking cessation outcomes, it holds promise for personalized treatment matching if these gaps are addressed.