Modeling Factors Associated with Smoking Dependence Severity Using Binary Classification-Based Machine Learning:...

Background/Objectives: The Fagerström Test for Nicotine Dependence (FTND) is a subjective, self-report-based instrument used to grade the level of nicotine dependence. Complete blood count (CBC) parameters reflect markers of smoking-related systemic inflammation. In the present study, we aimed to develop a leakage-free and interpretable machine learning approach for classifying FTND level using demographic data, CBC parameters, and duration of smoking exposure. Methods: In this retrospectively designed study, the records of 660 patients with complete data who had presented to the Smoking Cessation Outpatient Clinic of Malatya Training and Research Hospital were included. Because one of the six FTND items directly scores the number of cigarettes smoked per day, the number of packs and the features related to it constitute data leakage and were therefore excluded from the data pool. Smoking duration (years), which is not an FTND item, was retained as the sole predictor of smoking exposure. Feature engineering generated 49 candidate features. Six base learners (logistic regression, random forest, gradient boosting, SVM, LightGBM, XGBoost) and a stacking ensemble were evaluated with an 80-20% training/test split and 10-fold stratified cross-validation. To prevent optimism bias, feature selection and scaling were performed only within the training folds, and class imbalance was addressed through class weighting. The models were interpreted with SHAP and LIME, and the contribution of each feature family was quantified by ablation analysis. Results: In leakage-free 10-fold cross-validation (the primary performance estimate), discrimination was modest, with the highest AUC (0.652) obtained by both SVM (RBF) and logistic regression; on the single hold-out test set, the stacking ensemble reached an AUC of 0.749 (95% CI 0.667-0.831). The full model did not significantly outperform smoking duration alone (DeLong p = 0.22), and extensive feature engineering did not provide a measurable gain (ΔAUC = -0.015). In the SHAP analysis, smoking duration and the age-based interaction terms made the largest predictive contributions. Conclusions: Routine CBC parameters do not by themselves constitute an independent diagnostic tool for determining dependence severity; they carry only weak, complementary information and should not be regarded as a screening biomarker. This exploratory study classifies FTND-based nicotine dependence severity at the individual patient level by combining CBC parameters with machine learning and explainable artificial intelligence; the model is not ready for clinical use.