Machine Learning Classification of Smoking Behaviours—From Social Environment to the Prefrontal Cortex

Pablo Reinhardt, Norman Zacharias, Marinus Fislage, Justin Böhmer, Barbara Hollunder, Zala Reppmann, Anton Wiehe, Rebecca Rajwich, Nanne Dominick, Kerstin Ritter, Malek Bajbouj, Thomas F. Wienker, Jürgen Gallinat, Norbert Thürauf, Johannes Kornhuber, Falk Kiefer, Michael Wagner, Oliver Tuescher, Henrik Walter, Georg WintererAddiction 2025 · DOI

TL;DR

This study leverages machine learning to identify key predictors of smoking behavior and long-term cessation success by integrating diverse data sources, including electroencephalography (EEG), neuropsychological tests, and social environment factors. The findings highlight the crucial role of prefrontal cortex function and cognitive control, offering a path towards more personalized smoking cessation strategies.

What Problem Does This Paper Address?

Smoking is a complex behavior with significant interindividual variability in how people start, maintain, and quit. Traditional statistical methods often struggle to capture these nuances. This paper addresses the challenge of understanding and predicting these diverse smoking trajectories by applying machine learning to a rich, multimodal dataset. The goal is to move beyond broad classifications and identify specific markers that can distinguish between smokers and non-smokers, and even predict long-term success in quitting, paving the way for more tailored interventions.

Methods

The researchers employed a machine learning pipeline using scikit-learn, imbalanced-learn, and TensorFlow. This pipeline involved several steps: feature imputation, outlier removal using isolation forests, class balancing with SMOTETomek, and finally, training an XGBoost classifier. For feature importance, they utilized SHAP (Shapley Additive Explanations), a model-agnostic approach that quantifies the contribution of each feature to the model's predictions. To ensure robustness, they used a 10-fold repeated nested cross-validation strategy. For the longitudinal classification of quitters versus non-quitters, they also incorporated an autoencoder to engineer latent components from baseline EEG data.

Datasets

The study utilized data from the population-based Nicotine Cohort Study (NCS).

  • Baseline data (T1): Collected between 2007-2009 from 2,396 participants (1,116 smokers, 1,280 never smokers). This included a wide array of sociodemographic, environmental, clinical, neuropsychological, physiological, and electrophysiological (EEG) variables.
  • Follow-up data (T2): Collected approximately 10 years later (2019-2022) from a subsample of NCS participants. This online assessment included sociodemographic, environmental, clinical (mood, stress), and substance use data.

The final samples for classification were:

  • Smokers vs. Never Smokers (Baseline): N=707 smokers (including 222 heavy smokers) and N=864 never smokers.
  • Quitters vs. Non-Quitters (10-Year Follow-Up): N=60 successful quitters and N=81 non-quitters.

Key Insights and Ideas

  • Multimodal Data is Key: The study strongly suggests that integrating data from diverse domains (EEG, neuropsychology, social environment) is crucial for accurately predicting smoking behavior and cessation success. No single domain holds all the answers.
  • Prefrontal Cortex Dominance: Frontal lobe function, particularly P50 gamma band activity (related to sensory gating and inhibition) and cognitive control (as measured by the Stroop task), consistently emerged as highly influential predictors. This reinforces the prefrontal cortex's central role in addiction.
  • Social Environment Matters, Long-Term: Smoking behavior within one's social environment (personal and professional) was a significant predictor, not just for current smoking status but also for long-term cessation success. This highlights the enduring impact of social influences.
  • Cognitive Control as a Quitting Predictor: For predicting long-term quitting success, cognitive control measures, such as performance on the Stroop task and dietary self-regulation, were more important than baseline nicotine dependence levels. This suggests that the ability to exert self-control is a critical factor in sustained abstinence.
  • Machine Learning for Precision Psychiatry: The study demonstrates the power of machine learning to identify complex patterns in heterogeneous data, offering a promising avenue for developing personalized risk stratification and intervention strategies in mental health.

Key Results

  • High Accuracy for Smoking Status: The models achieved high performance in classifying smoking status at baseline:
  • Smokers vs. Never Smokers: AUROC = 0.85 ± 0.03
    • Heavy Smokers vs. Never Smokers: AUROC = 0.92 ± 0.03
  • Moderate Accuracy for Long-Term Quitting: The model for predicting long-term cessation success (quitters vs. non-quitters) showed moderate performance:
  • Quitters vs. Non-Quitters (10-year follow-up): AUROC = 0.68 ± 0.13
  • Top Predictors:
    • Smokers vs. Never Smokers: AUDIT score, smoking in private/professional environments, and frontal P50 gamma activity.
  • Heavy Smokers vs. Never Smokers: Frontal P50 gamma activity, workplace/social environment smoking, and resting-state frontal average power.
  • Quitters vs. Non-Quitters: Stroop task congruent condition, residential population size, cognitive control of eating behavior, and CPT reaction time hits.
  • Domain Performance: Sensitivity analyses showed that EEG and sociodemographic-environmental domains were strong contributors across classifications. For quitters vs. non-quitters, the sociodemographic-environmental domain outperformed EEG.

Limitations and Future Work

  • Representativeness and Attrition: The study acknowledges limitations due to potential biases in cohort representativeness and attrition over the 10-year follow-up period, which could affect the generalizability of long-term predictions.
  • Causality: The identified variables predict smoking behavior but do not establish a causal relationship.
  • Unmeasured Confounders: The extended follow-up period may have introduced unmeasured confounders influencing the results.

Future Work: * Longitudinal Integration: Prioritize longitudinal studies that integrate electrophysiology, neuroimaging, and advanced machine learning to capture the dynamic processes of nicotine dependence and cessation. * Clinical Translation: Focus on translating these predictive insights into clinically viable tools for personalized interventions. * Targeted Interventions: Develop strategies that target specific identified risk profiles, such as individuals with prefrontal vulnerabilities or those in smoking-conducive social environments. * Broader Application: Explore the utility of multimodal predictive tools in clinical decision-making for more effective, individualized prevention and cessation treatment.

Summary updated

Abstract

The pronounced heterogeneity in smoking trajectories-ranging from occasional or heavy use to successful quitting -highlights substantial interindividual variation within the smoking population. Machine learning is particularly well suited to capture these complex patterns that may be challenging for traditional inferential statistics to uncover. In this study, we applied machine learning to data from a population-based cohort to identify multimodal markers that distinguish smokers from never smokers at baseline and predict long-term cessation success at a 10-year follow-up. We employed 10 times repeated nested cross-validation (10 outer folds, 5 inner folds) to analyse baseline data (T1) from 707 smokers-including 222 heavy smokers (FTND ≥ 4)-and 864 never smokers for smoking status classification. At the 10-year follow-up (T2), we further classified 60 successful quitters (≥ 1 year abstinent) versus 81 non-quitters. Feature importance was assessed using averaged SHAP values derived from test set predictions. Classification models achieved the following performance, expressed by the area under the receiver operating characteristic curve (AUROC; mean ± SD): smokers versus never smokers, 0.85 ± 0.03; heavy smokers versus never smokers, 0.92 ± 0.03; and quitters versus non-quitters, 0.68 ± 0.13. SHAP analysis identified markers of frontal functioning, cognitive control and smoking behaviour within the social environment among the most influential predictors of both smoking status and cessation success. In conclusion, our machine learning analyses support a multifactorial model of smoking behaviour and cessation success, which may inform nuanced risk stratification to advance the development of personalized cessation strategies.
BibTeX
@inproceedings{reinhardt2025machine,
  author = {Pablo Reinhardt and Norman Zacharias and Marinus Fislage and Justin Böhmer and Barbara Hollunder and Zala Reppmann and Anton Wiehe and Rebecca Rajwich and Nanne Dominick and Kerstin Ritter and Malek Bajbouj and Thomas F. Wienker and Jürgen Gallinat and Norbert Thürauf and Johannes Kornhuber and Falk Kiefer and Michael Wagner and Oliver Tuescher and Henrik Walter and Georg Winterer},
  title = {{Machine Learning Classification of Smoking Behaviours—From Social Environment to the Prefrontal Cortex}},
  booktitle = {Addiction},
  year = {2025},
  doi = {10.1111/adb.70056},
}
Figure