<p>Nomophobia is emerging as a developing psychological condition characterized by fear and anxiety are caused from the lack of access to smartphones. Furthermore, the high dependence on smartphone use for interactions, academic production, and fun has worsened the problem, especially among teens and young adults. Still, despite being widespread, existing studies lack significant automated approaches that can be implemented for early nomophobia detection. This research paper attempts to bridge the gap between traditional methods and design an effective and potentially scalable framework for the early detection of nomophobia by adopting advanced machine learning techniques. Thus, an effective and domain-adapted approach using the TabTransformer model was proposed, which is a robust Deep Learning model developed to handle tabular data with categorical and numerical features in which a self-attention mechanism can capture the more complex inter-feature relationships. A newly designed dataset, surveyed from 2013 respondents with structured surveys, consisted of smartphone usage patterns, psychological indicators, and behavioral metrics like insomnia, late-night usage, and loneliness scores. The preprocessing steps applied include feature scaling, encoding, and class balancing through SVM_SMOTE to prepare the data for analysis. It indicates that the proposed TabTransformer model was superior and achieved a greater accuracy of 91% along with recall 88%, and had outperformed the other classical models, namely Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting (GB) and Ensemble classifier (EC). The proposed model achieved a 6.12% improvement in recall and a 7.31% improvement in F1-score compared to existing state-of-the-art methods like DT, RF, GB, SVM, EC, and TabNet. Although these improvements indicate better performance compared to baseline models, statistical significance testing and variability analysis were not explicitly conducted in this study. Therefore, the results should be interpreted as indicative rather than definitive. Importantly, the model showed high precision at 97% for absent nomophobia cases while achieving the highest F1-score at 92% for mild cases, thus ensuring balanced performance across all severity levels. The results indicate the possibility of using Deep Learning and Machine Learning with behavioral psychology to identify nomophobia early. Results contribute not only to research in mental health but also offer a practical framework for developing potential real-time monitoring applications. The experimental design and evaluation framework have been carefully structured to ensure reproducibility, robustness, and fair comparison across models.</p>