One of the most significant topics over recent years has been the identification of stress using physiological data, which has gained tremendous potential towards improving mental health interventions and quality of life. It deals with the effectiveness of numerous machine learning and deep learning techniques in identifying stress states. We apply preprocessing to the physiological data, study the classical machine learning model that includes Logistic Regression and Support Vector Machines, and present an architecture of a new neural network specially designed for this application. Our custom architecture utilizes attention mechanisms to extract detailed features from high-dimensional data. The results of the analysis show that traditional models such as Gaussian Naïve Bayes have high accuracy. However, our proposed neural network outperforms them in terms of precision and adaptability, especially with imbalanced datasets. A few key findings highlighted the relevance of robust preprocessing and architecture customization for extracting meaningful insights. The proposed approach is useful for early stress detection and yields significant benefits in real applications such as workplace wellness programs and clinical stress monitoring.