Growing stress levels and anxiety levels in modern-day fast-paced work environments require good tools for early detection and management. The interest in research into deep learning techniques for the recognition of emotions from EEG leads to the development of the present study. It is research for the formation of a strong model to classify people under various levels of stress based on real-time EEG signals from 30 participants at different times of the day when they were working. The proposed methodology involves the application of advanced deep learning models, such as EEGNet, VGG19, VGG16, and ResNet, integrated with Long Short-Term Memory (LSTM) networks to analyze the temporal aspects of the collected EEG data. The dataset contains 3,500 spatial and temporal data points that are distinctive of various states of stress and anxiety. For the result after training and validation of the models, it concludes that the highest accuracy in the range of 97.89% in the EEGNet + LSTM configuration, which has a better True Positive Rate and minimal False Positive Rate, therefore illustrating it is more efficient in stress detection. Coming next in the rank are the VGG19 + LSTM with an accuracy of 95.6% and VGG16 + LSTM with 92.3% accuracy, while a minimum of 89.9% accuracy was reported for the ResNet + LSTM. As the above results demonstrate, it is possible to use deep learning methodologies to achieve real-time emotion recognition with more serious implications for mental health monitoring systems. The above models are useful for applications that generally require immediate feedback and interventions based on stress management, thus improving working conditions.