Smoking remains a major global public health issue, contributing to millions of preventable deaths each year and placing a significant burden on healthcare systems, particularly in low- and middle-income countries. Despite widespread awareness of its harmful effects, tobacco use continues to rise in certain populations, including educated individuals. Traditional self-reported methods for identifying smokers often suffer from inaccuracy, underscoring the need for reliable, data-driven alternatives. This study explores a predictive modeling approach to classify individuals as smokers or non-smokers based on a range of demographic, behavioral, and psychological factors. A custom dataset was developed, comprising 223 instances and 17 features related to personal background and smoking-related influences. To address the complexity of feature interactions, we propose a hybrid deep learning architecture that integrates Graph Neural Networks (GNN) and a Feature Tokenizer-based Transformer. This model leverages both relational and contextual information to improve the identification of smoking patterns. The findings highlight the potential of advanced machine learning methods in supporting early intervention strategies and enhancing public health planning.